1/*
2 * Copyright (c) 2000-2020 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28/*
29 * @OSF_COPYRIGHT@
30 */
31/*
32 * Mach Operating System
33 * Copyright (c) 1991,1990,1989,1988,1987 Carnegie Mellon University
34 * All Rights Reserved.
35 *
36 * Permission to use, copy, modify and distribute this software and its
37 * documentation is hereby granted, provided that both the copyright
38 * notice and this permission notice appear in all copies of the
39 * software, derivative works or modified versions, and any portions
40 * thereof, and that both notices appear in supporting documentation.
41 *
42 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
43 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR
44 * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
45 *
46 * Carnegie Mellon requests users of this software to return to
47 *
48 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
49 * School of Computer Science
50 * Carnegie Mellon University
51 * Pittsburgh PA 15213-3890
52 *
53 * any improvements or extensions that they make and grant Carnegie Mellon
54 * the rights to redistribute these changes.
55 */
56/*
57 */
58/*
59 * File: vm/vm_object.c
60 * Author: Avadis Tevanian, Jr., Michael Wayne Young
61 *
62 * Virtual memory object module.
63 */
64
65#include <debug.h>
66
67#include <mach/mach_types.h>
68#include <mach/memory_object.h>
69#include <mach/vm_param.h>
70
71#include <mach/sdt.h>
72
73#include <ipc/ipc_types.h>
74#include <ipc/ipc_port.h>
75
76#include <kern/kern_types.h>
77#include <kern/assert.h>
78#include <kern/queue.h>
79#include <kern/kalloc.h>
80#include <kern/zalloc.h>
81#include <kern/host.h>
82#include <kern/host_statistics.h>
83#include <kern/processor.h>
84#include <kern/misc_protos.h>
85#include <kern/policy_internal.h>
86
87#include <sys/kdebug_triage.h>
88
89#include <vm/memory_object.h>
90#include <vm/vm_compressor_pager.h>
91#include <vm/vm_fault.h>
92#include <vm/vm_map.h>
93#include <vm/vm_object.h>
94#include <vm/vm_page.h>
95#include <vm/vm_pageout.h>
96#include <vm/vm_protos.h>
97#include <vm/vm_purgeable_internal.h>
98
99#include <vm/vm_compressor.h>
100
101#if CONFIG_PHANTOM_CACHE
102#include <vm/vm_phantom_cache.h>
103#endif
104
105#if VM_OBJECT_ACCESS_TRACKING
106uint64_t vm_object_access_tracking_reads = 0;
107uint64_t vm_object_access_tracking_writes = 0;
108#endif /* VM_OBJECT_ACCESS_TRACKING */
109
110boolean_t vm_object_collapse_compressor_allowed = TRUE;
111
112struct vm_counters vm_counters;
113
114#if DEVELOPMENT || DEBUG
115extern struct memory_object_pager_ops shared_region_pager_ops;
116extern unsigned int shared_region_pagers_resident_count;
117extern unsigned int shared_region_pagers_resident_peak;
118#endif /* DEVELOPMENT || DEBUG */
119
120#if VM_OBJECT_TRACKING
121btlog_t vm_object_tracking_btlog;
122
123void
124vm_object_tracking_init(void)
125{
126 int vm_object_tracking;
127
128 vm_object_tracking = 1;
129 PE_parse_boot_argn("vm_object_tracking", &vm_object_tracking,
130 sizeof(vm_object_tracking));
131
132 if (vm_object_tracking) {
133 vm_object_tracking_btlog = btlog_create(BTLOG_HASH,
134 VM_OBJECT_TRACKING_NUM_RECORDS);
135 assert(vm_object_tracking_btlog);
136 }
137}
138#endif /* VM_OBJECT_TRACKING */
139
140/*
141 * Virtual memory objects maintain the actual data
142 * associated with allocated virtual memory. A given
143 * page of memory exists within exactly one object.
144 *
145 * An object is only deallocated when all "references"
146 * are given up.
147 *
148 * Associated with each object is a list of all resident
149 * memory pages belonging to that object; this list is
150 * maintained by the "vm_page" module, but locked by the object's
151 * lock.
152 *
153 * Each object also records the memory object reference
154 * that is used by the kernel to request and write
155 * back data (the memory object, field "pager"), etc...
156 *
157 * Virtual memory objects are allocated to provide
158 * zero-filled memory (vm_allocate) or map a user-defined
159 * memory object into a virtual address space (vm_map).
160 *
161 * Virtual memory objects that refer to a user-defined
162 * memory object are called "permanent", because all changes
163 * made in virtual memory are reflected back to the
164 * memory manager, which may then store it permanently.
165 * Other virtual memory objects are called "temporary",
166 * meaning that changes need be written back only when
167 * necessary to reclaim pages, and that storage associated
168 * with the object can be discarded once it is no longer
169 * mapped.
170 *
171 * A permanent memory object may be mapped into more
172 * than one virtual address space. Moreover, two threads
173 * may attempt to make the first mapping of a memory
174 * object concurrently. Only one thread is allowed to
175 * complete this mapping; all others wait for the
176 * "pager_initialized" field is asserted, indicating
177 * that the first thread has initialized all of the
178 * necessary fields in the virtual memory object structure.
179 *
180 * The kernel relies on a *default memory manager* to
181 * provide backing storage for the zero-filled virtual
182 * memory objects. The pager memory objects associated
183 * with these temporary virtual memory objects are only
184 * requested from the default memory manager when it
185 * becomes necessary. Virtual memory objects
186 * that depend on the default memory manager are called
187 * "internal". The "pager_created" field is provided to
188 * indicate whether these ports have ever been allocated.
189 *
190 * The kernel may also create virtual memory objects to
191 * hold changed pages after a copy-on-write operation.
192 * In this case, the virtual memory object (and its
193 * backing storage -- its memory object) only contain
194 * those pages that have been changed. The "shadow"
195 * field refers to the virtual memory object that contains
196 * the remainder of the contents. The "shadow_offset"
197 * field indicates where in the "shadow" these contents begin.
198 * The "copy" field refers to a virtual memory object
199 * to which changed pages must be copied before changing
200 * this object, in order to implement another form
201 * of copy-on-write optimization.
202 *
203 * The virtual memory object structure also records
204 * the attributes associated with its memory object.
205 * The "pager_ready", "can_persist" and "copy_strategy"
206 * fields represent those attributes. The "cached_list"
207 * field is used in the implementation of the persistence
208 * attribute.
209 *
210 * ZZZ Continue this comment.
211 */
212
213/* Forward declarations for internal functions. */
214static kern_return_t vm_object_terminate(
215 vm_object_t object);
216
217static void vm_object_do_collapse(
218 vm_object_t object,
219 vm_object_t backing_object);
220
221static void vm_object_do_bypass(
222 vm_object_t object,
223 vm_object_t backing_object);
224
225static void vm_object_release_pager(
226 memory_object_t pager);
227
228SECURITY_READ_ONLY_LATE(zone_t) vm_object_zone; /* vm backing store zone */
229
230/*
231 * All wired-down kernel memory belongs to this memory object
232 * memory object (kernel_object) by default to avoid wasting data structures.
233 */
234static struct vm_object kernel_object_store VM_PAGE_PACKED_ALIGNED;
235const vm_object_t kernel_object_default = &kernel_object_store;
236
237static struct vm_object compressor_object_store VM_PAGE_PACKED_ALIGNED;
238const vm_object_t compressor_object = &compressor_object_store;
239
240/*
241 * This object holds all pages that have been retired due to errors like ECC.
242 * The system should never use the page or look at its contents. The offset
243 * in this object is the same as the page's physical address.
244 */
245static struct vm_object retired_pages_object_store VM_PAGE_PACKED_ALIGNED;
246const vm_object_t retired_pages_object = &retired_pages_object_store;
247
248static struct vm_object exclaves_object_store VM_PAGE_PACKED_ALIGNED;
249const vm_object_t exclaves_object = &exclaves_object_store;
250
251
252/*
253 * Virtual memory objects are initialized from
254 * a template (see vm_object_allocate).
255 *
256 * When adding a new field to the virtual memory
257 * object structure, be sure to add initialization
258 * (see _vm_object_allocate()).
259 */
260static const struct vm_object vm_object_template = {
261 .memq.prev = 0,
262 .memq.next = 0,
263 /*
264 * The lock will be initialized for each allocated object in
265 * _vm_object_allocate(), so we don't need to initialize it in
266 * the vm_object_template.
267 */
268 .vo_size = 0,
269 .memq_hint = VM_PAGE_NULL,
270 .ref_count = 1,
271 .resident_page_count = 0,
272 .wired_page_count = 0,
273 .reusable_page_count = 0,
274 .vo_copy = VM_OBJECT_NULL,
275 .vo_copy_version = 0,
276 .shadow = VM_OBJECT_NULL,
277 .vo_shadow_offset = (vm_object_offset_t) 0,
278 .pager = MEMORY_OBJECT_NULL,
279 .paging_offset = 0,
280 .pager_control = MEMORY_OBJECT_CONTROL_NULL,
281 .copy_strategy = MEMORY_OBJECT_COPY_SYMMETRIC,
282 .paging_in_progress = 0,
283 .vo_size_delta = 0,
284 .activity_in_progress = 0,
285
286 /* Begin bitfields */
287 .all_wanted = 0, /* all bits FALSE */
288 .pager_created = FALSE,
289 .pager_initialized = FALSE,
290 .pager_ready = FALSE,
291 .pager_trusted = FALSE,
292 .can_persist = FALSE,
293 .internal = TRUE,
294 .private = FALSE,
295 .pageout = FALSE,
296 .alive = TRUE,
297 .purgable = VM_PURGABLE_DENY,
298 .purgeable_when_ripe = FALSE,
299 .purgeable_only_by_kernel = FALSE,
300 .shadowed = FALSE,
301 .true_share = FALSE,
302 .terminating = FALSE,
303 .named = FALSE,
304 .shadow_severed = FALSE,
305 .phys_contiguous = FALSE,
306 .nophyscache = FALSE,
307 /* End bitfields */
308
309 .cached_list.prev = NULL,
310 .cached_list.next = NULL,
311
312 .last_alloc = (vm_object_offset_t) 0,
313 .sequential = (vm_object_offset_t) 0,
314 .pages_created = 0,
315 .pages_used = 0,
316 .scan_collisions = 0,
317#if CONFIG_PHANTOM_CACHE
318 .phantom_object_id = 0,
319#endif
320 .cow_hint = ~(vm_offset_t)0,
321
322 /* cache bitfields */
323 .wimg_bits = VM_WIMG_USE_DEFAULT,
324 .set_cache_attr = FALSE,
325 .object_is_shared_cache = FALSE,
326 .code_signed = FALSE,
327 .transposed = FALSE,
328 .mapping_in_progress = FALSE,
329 .phantom_isssd = FALSE,
330 .volatile_empty = FALSE,
331 .volatile_fault = FALSE,
332 .all_reusable = FALSE,
333 .blocked_access = FALSE,
334 .vo_ledger_tag = VM_LEDGER_TAG_NONE,
335 .vo_no_footprint = FALSE,
336#if CONFIG_IOSCHED || UPL_DEBUG
337 .uplq.prev = NULL,
338 .uplq.next = NULL,
339#endif /* UPL_DEBUG */
340#ifdef VM_PIP_DEBUG
341 .pip_holders = {0},
342#endif /* VM_PIP_DEBUG */
343
344 .objq.next = NULL,
345 .objq.prev = NULL,
346 .task_objq.next = NULL,
347 .task_objq.prev = NULL,
348
349 .purgeable_queue_type = PURGEABLE_Q_TYPE_MAX,
350 .purgeable_queue_group = 0,
351
352 .wire_tag = VM_KERN_MEMORY_NONE,
353#if !VM_TAG_ACTIVE_UPDATE
354 .wired_objq.next = NULL,
355 .wired_objq.prev = NULL,
356#endif /* ! VM_TAG_ACTIVE_UPDATE */
357
358 .io_tracking = FALSE,
359
360#if CONFIG_SECLUDED_MEMORY
361 .eligible_for_secluded = FALSE,
362 .can_grab_secluded = FALSE,
363#else /* CONFIG_SECLUDED_MEMORY */
364 .__object3_unused_bits = 0,
365#endif /* CONFIG_SECLUDED_MEMORY */
366
367 .for_realtime = false,
368 .no_pager_reason = VM_OBJECT_DESTROY_UNKNOWN_REASON,
369
370#if VM_OBJECT_ACCESS_TRACKING
371 .access_tracking = FALSE,
372 .access_tracking_reads = 0,
373 .access_tracking_writes = 0,
374#endif /* VM_OBJECT_ACCESS_TRACKING */
375
376#if DEBUG
377 .purgeable_owner_bt = {0},
378 .vo_purgeable_volatilizer = NULL,
379 .purgeable_volatilizer_bt = {0},
380#endif /* DEBUG */
381};
382
383LCK_GRP_DECLARE(vm_object_lck_grp, "vm_object");
384LCK_GRP_DECLARE(vm_object_cache_lck_grp, "vm_object_cache");
385LCK_ATTR_DECLARE(vm_object_lck_attr, 0, 0);
386LCK_ATTR_DECLARE(kernel_object_lck_attr, 0, LCK_ATTR_DEBUG);
387LCK_ATTR_DECLARE(compressor_object_lck_attr, 0, LCK_ATTR_DEBUG);
388
389unsigned int vm_page_purged_wired = 0;
390unsigned int vm_page_purged_busy = 0;
391unsigned int vm_page_purged_others = 0;
392
393static queue_head_t vm_object_cached_list;
394static uint32_t vm_object_cache_pages_freed = 0;
395static uint32_t vm_object_cache_pages_moved = 0;
396static uint32_t vm_object_cache_pages_skipped = 0;
397static uint32_t vm_object_cache_adds = 0;
398static uint32_t vm_object_cached_count = 0;
399static LCK_MTX_DECLARE_ATTR(vm_object_cached_lock_data,
400 &vm_object_cache_lck_grp, &vm_object_lck_attr);
401
402static uint32_t vm_object_page_grab_failed = 0;
403static uint32_t vm_object_page_grab_skipped = 0;
404static uint32_t vm_object_page_grab_returned = 0;
405static uint32_t vm_object_page_grab_pmapped = 0;
406static uint32_t vm_object_page_grab_reactivations = 0;
407
408#define vm_object_cache_lock_spin() \
409 lck_mtx_lock_spin(&vm_object_cached_lock_data)
410#define vm_object_cache_unlock() \
411 lck_mtx_unlock(&vm_object_cached_lock_data)
412
413static void vm_object_cache_remove_locked(vm_object_t);
414
415
416static void vm_object_reap(vm_object_t object);
417static void vm_object_reap_async(vm_object_t object);
418static void vm_object_reaper_thread(void);
419
420static LCK_MTX_DECLARE_ATTR(vm_object_reaper_lock_data,
421 &vm_object_lck_grp, &vm_object_lck_attr);
422
423static queue_head_t vm_object_reaper_queue; /* protected by vm_object_reaper_lock() */
424unsigned int vm_object_reap_count = 0;
425unsigned int vm_object_reap_count_async = 0;
426
427#define vm_object_reaper_lock() \
428 lck_mtx_lock(&vm_object_reaper_lock_data)
429#define vm_object_reaper_lock_spin() \
430 lck_mtx_lock_spin(&vm_object_reaper_lock_data)
431#define vm_object_reaper_unlock() \
432 lck_mtx_unlock(&vm_object_reaper_lock_data)
433
434#if CONFIG_IOSCHED
435/* I/O Re-prioritization request list */
436queue_head_t io_reprioritize_list = QUEUE_HEAD_INITIALIZER(io_reprioritize_list);
437
438LCK_SPIN_DECLARE_ATTR(io_reprioritize_list_lock,
439 &vm_object_lck_grp, &vm_object_lck_attr);
440
441#define IO_REPRIORITIZE_LIST_LOCK() \
442 lck_spin_lock_grp(&io_reprioritize_list_lock, &vm_object_lck_grp)
443#define IO_REPRIORITIZE_LIST_UNLOCK() \
444 lck_spin_unlock(&io_reprioritize_list_lock)
445
446ZONE_DEFINE_TYPE(io_reprioritize_req_zone, "io_reprioritize_req",
447 struct io_reprioritize_req, ZC_NONE);
448
449/* I/O Re-prioritization thread */
450int io_reprioritize_wakeup = 0;
451static void io_reprioritize_thread(void *param __unused, wait_result_t wr __unused);
452
453#define IO_REPRIO_THREAD_WAKEUP() thread_wakeup((event_t)&io_reprioritize_wakeup)
454#define IO_REPRIO_THREAD_CONTINUATION() \
455{ \
456 assert_wait(&io_reprioritize_wakeup, THREAD_UNINT); \
457 thread_block(io_reprioritize_thread); \
458}
459
460void vm_page_request_reprioritize(vm_object_t, uint64_t, uint32_t, int);
461void vm_page_handle_prio_inversion(vm_object_t, vm_page_t);
462void vm_decmp_upl_reprioritize(upl_t, int);
463#endif
464
465#if 0
466#undef KERNEL_DEBUG
467#define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT
468#endif
469
470
471void
472vm_object_set_size(
473 vm_object_t object,
474 vm_object_size_t outer_size,
475 vm_object_size_t inner_size)
476{
477 object->vo_size = vm_object_round_page(outer_size);
478#if KASAN
479 assert(object->vo_size - inner_size <= USHRT_MAX);
480 object->vo_size_delta = (unsigned short)(object->vo_size - inner_size);
481#else
482 (void)inner_size;
483#endif
484}
485
486
487/*
488 * vm_object_allocate:
489 *
490 * Returns a new object with the given size.
491 */
492
493__private_extern__ void
494_vm_object_allocate(
495 vm_object_size_t size,
496 vm_object_t object)
497{
498 *object = vm_object_template;
499 vm_page_queue_init(&object->memq);
500#if UPL_DEBUG || CONFIG_IOSCHED
501 queue_init(&object->uplq);
502#endif
503 vm_object_lock_init(object);
504 vm_object_set_size(object, outer_size: size, inner_size: size);
505
506#if VM_OBJECT_TRACKING_OP_CREATED
507 if (vm_object_tracking_btlog) {
508 btlog_record(vm_object_tracking_btlog, object,
509 VM_OBJECT_TRACKING_OP_CREATED,
510 btref_get(__builtin_frame_address(0), 0));
511 }
512#endif /* VM_OBJECT_TRACKING_OP_CREATED */
513}
514
515__private_extern__ vm_object_t
516vm_object_allocate(
517 vm_object_size_t size)
518{
519 vm_object_t object;
520
521 object = zalloc_flags(vm_object_zone, Z_WAITOK | Z_NOFAIL);
522 _vm_object_allocate(size, object);
523
524 return object;
525}
526
527TUNABLE(bool, workaround_41447923, "workaround_41447923", false);
528
529/*
530 * vm_object_bootstrap:
531 *
532 * Initialize the VM objects module.
533 */
534__startup_func
535void
536vm_object_bootstrap(void)
537{
538 vm_size_t vm_object_size;
539
540 assert(sizeof(mo_ipc_object_bits_t) == sizeof(ipc_object_bits_t));
541
542 vm_object_size = (sizeof(struct vm_object) + (VM_PAGE_PACKED_PTR_ALIGNMENT - 1)) &
543 ~(VM_PAGE_PACKED_PTR_ALIGNMENT - 1);
544
545 vm_object_zone = zone_create(name: "vm objects", size: vm_object_size,
546 flags: ZC_NOENCRYPT | ZC_ALIGNMENT_REQUIRED | ZC_VM | ZC_NOTBITAG);
547
548 queue_init(&vm_object_cached_list);
549
550 queue_init(&vm_object_reaper_queue);
551
552 /*
553 * Initialize the "kernel object"
554 */
555
556 /*
557 * Note that in the following size specifications, we need to add 1 because
558 * VM_MAX_KERNEL_ADDRESS (vm_last_addr) is a maximum address, not a size.
559 */
560 _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, object: kernel_object_default);
561 _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, object: compressor_object);
562 kernel_object_default->copy_strategy = MEMORY_OBJECT_COPY_NONE;
563 compressor_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
564 kernel_object_default->no_tag_update = TRUE;
565
566 /*
567 * The object to hold retired VM pages.
568 */
569 _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, object: retired_pages_object);
570 retired_pages_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
571
572 /**
573 * The object to hold pages owned by exclaves.
574 */
575 _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, object: exclaves_object);
576 exclaves_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
577}
578
579#if CONFIG_IOSCHED
580void
581vm_io_reprioritize_init(void)
582{
583 kern_return_t result;
584 thread_t thread = THREAD_NULL;
585
586 result = kernel_thread_start_priority(continuation: io_reprioritize_thread, NULL, priority: 95 /* MAXPRI_KERNEL */, new_thread: &thread);
587 if (result == KERN_SUCCESS) {
588 thread_set_thread_name(th: thread, name: "VM_io_reprioritize_thread");
589 thread_deallocate(thread);
590 } else {
591 panic("Could not create io_reprioritize_thread");
592 }
593}
594#endif
595
596void
597vm_object_reaper_init(void)
598{
599 kern_return_t kr;
600 thread_t thread;
601
602 kr = kernel_thread_start_priority(
603 continuation: (thread_continue_t) vm_object_reaper_thread,
604 NULL,
605 BASEPRI_VM,
606 new_thread: &thread);
607 if (kr != KERN_SUCCESS) {
608 panic("failed to launch vm_object_reaper_thread kr=0x%x", kr);
609 }
610 thread_set_thread_name(th: thread, name: "VM_object_reaper_thread");
611 thread_deallocate(thread);
612}
613
614
615/*
616 * vm_object_deallocate:
617 *
618 * Release a reference to the specified object,
619 * gained either through a vm_object_allocate
620 * or a vm_object_reference call. When all references
621 * are gone, storage associated with this object
622 * may be relinquished.
623 *
624 * No object may be locked.
625 */
626unsigned long vm_object_deallocate_shared_successes = 0;
627unsigned long vm_object_deallocate_shared_failures = 0;
628unsigned long vm_object_deallocate_shared_swap_failures = 0;
629
630__private_extern__ void
631vm_object_deallocate(
632 vm_object_t object)
633{
634 vm_object_t shadow = VM_OBJECT_NULL;
635
636// if(object)dbgLog(object, object->ref_count, object->can_persist, 3); /* (TEST/DEBUG) */
637// else dbgLog(object, 0, 0, 3); /* (TEST/DEBUG) */
638
639 if (object == VM_OBJECT_NULL) {
640 return;
641 }
642
643 if (is_kernel_object(object) || object == compressor_object || object == retired_pages_object) {
644 vm_object_lock_shared(object);
645
646 OSAddAtomic(-1, &object->ref_count);
647
648 if (object->ref_count == 0) {
649 if (is_kernel_object(object)) {
650 panic("vm_object_deallocate: losing a kernel_object");
651 } else if (object == retired_pages_object) {
652 panic("vm_object_deallocate: losing retired_pages_object");
653 } else {
654 panic("vm_object_deallocate: losing compressor_object");
655 }
656 }
657 vm_object_unlock(object);
658 return;
659 }
660
661 if (object->ref_count == 2 &&
662 object->named) {
663 /*
664 * This "named" object's reference count is about to
665 * drop from 2 to 1:
666 * we'll need to call memory_object_last_unmap().
667 */
668 } else if (object->ref_count == 2 &&
669 object->internal &&
670 object->shadow != VM_OBJECT_NULL) {
671 /*
672 * This internal object's reference count is about to
673 * drop from 2 to 1 and it has a shadow object:
674 * we'll want to try and collapse this object with its
675 * shadow.
676 */
677 } else if (object->ref_count >= 2) {
678 UInt32 original_ref_count;
679 volatile UInt32 *ref_count_p;
680 Boolean atomic_swap;
681
682 /*
683 * The object currently looks like it is not being
684 * kept alive solely by the reference we're about to release.
685 * Let's try and release our reference without taking
686 * all the locks we would need if we had to terminate the
687 * object (cache lock + exclusive object lock).
688 * Lock the object "shared" to make sure we don't race with
689 * anyone holding it "exclusive".
690 */
691 vm_object_lock_shared(object);
692 ref_count_p = (volatile UInt32 *) &object->ref_count;
693 original_ref_count = object->ref_count;
694 /*
695 * Test again as "ref_count" could have changed.
696 * "named" shouldn't change.
697 */
698 if (original_ref_count == 2 &&
699 object->named) {
700 /* need to take slow path for m_o_last_unmap() */
701 atomic_swap = FALSE;
702 } else if (original_ref_count == 2 &&
703 object->internal &&
704 object->shadow != VM_OBJECT_NULL) {
705 /* need to take slow path for vm_object_collapse() */
706 atomic_swap = FALSE;
707 } else if (original_ref_count < 2) {
708 /* need to take slow path for vm_object_terminate() */
709 atomic_swap = FALSE;
710 } else {
711 /* try an atomic update with the shared lock */
712 atomic_swap = OSCompareAndSwap(
713 original_ref_count,
714 original_ref_count - 1,
715 (UInt32 *) &object->ref_count);
716 if (atomic_swap == FALSE) {
717 vm_object_deallocate_shared_swap_failures++;
718 /* fall back to the slow path... */
719 }
720 }
721
722 vm_object_unlock(object);
723
724 if (atomic_swap) {
725 /*
726 * ref_count was updated atomically !
727 */
728 vm_object_deallocate_shared_successes++;
729 return;
730 }
731
732 /*
733 * Someone else updated the ref_count at the same
734 * time and we lost the race. Fall back to the usual
735 * slow but safe path...
736 */
737 vm_object_deallocate_shared_failures++;
738 }
739
740 while (object != VM_OBJECT_NULL) {
741 vm_object_lock(object);
742
743 assert(object->ref_count > 0);
744
745 /*
746 * If the object has a named reference, and only
747 * that reference would remain, inform the pager
748 * about the last "mapping" reference going away.
749 */
750 if ((object->ref_count == 2) && (object->named)) {
751 memory_object_t pager = object->pager;
752
753 /* Notify the Pager that there are no */
754 /* more mappers for this object */
755
756 if (pager != MEMORY_OBJECT_NULL) {
757 vm_object_mapping_wait(object, THREAD_UNINT);
758 vm_object_mapping_begin(object);
759 vm_object_unlock(object);
760
761 memory_object_last_unmap(memory_object: pager);
762
763 vm_object_lock(object);
764 vm_object_mapping_end(object);
765 }
766 assert(object->ref_count > 0);
767 }
768
769 /*
770 * Lose the reference. If other references
771 * remain, then we are done, unless we need
772 * to retry a cache trim.
773 * If it is the last reference, then keep it
774 * until any pending initialization is completed.
775 */
776
777 /* if the object is terminating, it cannot go into */
778 /* the cache and we obviously should not call */
779 /* terminate again. */
780
781 if ((object->ref_count > 1) || object->terminating) {
782 vm_object_lock_assert_exclusive(object);
783 object->ref_count--;
784
785 if (object->ref_count == 1 &&
786 object->shadow != VM_OBJECT_NULL) {
787 /*
788 * There's only one reference left on this
789 * VM object. We can't tell if it's a valid
790 * one (from a mapping for example) or if this
791 * object is just part of a possibly stale and
792 * useless shadow chain.
793 * We would like to try and collapse it into
794 * its parent, but we don't have any pointers
795 * back to this parent object.
796 * But we can try and collapse this object with
797 * its own shadows, in case these are useless
798 * too...
799 * We can't bypass this object though, since we
800 * don't know if this last reference on it is
801 * meaningful or not.
802 */
803 vm_object_collapse(object, offset: 0, FALSE);
804 }
805 vm_object_unlock(object);
806 return;
807 }
808
809 /*
810 * We have to wait for initialization
811 * before destroying or caching the object.
812 */
813
814 if (object->pager_created && !object->pager_initialized) {
815 assert(!object->can_persist);
816 vm_object_assert_wait(object,
817 VM_OBJECT_EVENT_INITIALIZED,
818 THREAD_UNINT);
819 vm_object_unlock(object);
820
821 thread_block(THREAD_CONTINUE_NULL);
822 continue;
823 }
824
825 /*
826 * Terminate this object. If it had a shadow,
827 * then deallocate it; otherwise, if we need
828 * to retry a cache trim, do so now; otherwise,
829 * we are done. "pageout" objects have a shadow,
830 * but maintain a "paging reference" rather than
831 * a normal reference.
832 */
833 shadow = object->pageout?VM_OBJECT_NULL:object->shadow;
834
835 if (vm_object_terminate(object) != KERN_SUCCESS) {
836 return;
837 }
838 if (shadow != VM_OBJECT_NULL) {
839 object = shadow;
840 continue;
841 }
842 return;
843 }
844}
845
846
847
848vm_page_t
849vm_object_page_grab(
850 vm_object_t object)
851{
852 vm_page_t p, next_p;
853 int p_limit = 0;
854 int p_skipped = 0;
855
856 vm_object_lock_assert_exclusive(object);
857
858 next_p = (vm_page_t)vm_page_queue_first(&object->memq);
859 p_limit = MIN(50, object->resident_page_count);
860
861 while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next_p) && --p_limit > 0) {
862 p = next_p;
863 next_p = (vm_page_t)vm_page_queue_next(&next_p->vmp_listq);
864
865 if (VM_PAGE_WIRED(p) || p->vmp_busy || p->vmp_cleaning || p->vmp_laundry || p->vmp_fictitious) {
866 goto move_page_in_obj;
867 }
868
869 if (p->vmp_pmapped || p->vmp_dirty || p->vmp_precious) {
870 vm_page_lockspin_queues();
871
872 if (p->vmp_pmapped) {
873 int refmod_state;
874
875 vm_object_page_grab_pmapped++;
876
877 if (p->vmp_reference == FALSE || p->vmp_dirty == FALSE) {
878 refmod_state = pmap_get_refmod(pn: VM_PAGE_GET_PHYS_PAGE(m: p));
879
880 if (refmod_state & VM_MEM_REFERENCED) {
881 p->vmp_reference = TRUE;
882 }
883 if (refmod_state & VM_MEM_MODIFIED) {
884 SET_PAGE_DIRTY(p, FALSE);
885 }
886 }
887 if (p->vmp_dirty == FALSE && p->vmp_precious == FALSE) {
888 vm_page_lockconvert_queues();
889 refmod_state = pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: p));
890
891 if (refmod_state & VM_MEM_REFERENCED) {
892 p->vmp_reference = TRUE;
893 }
894 if (refmod_state & VM_MEM_MODIFIED) {
895 SET_PAGE_DIRTY(p, FALSE);
896 }
897
898 if (p->vmp_dirty == FALSE) {
899 goto take_page;
900 }
901 }
902 }
903 if ((p->vmp_q_state != VM_PAGE_ON_ACTIVE_Q) && p->vmp_reference == TRUE) {
904 vm_page_activate(page: p);
905
906 counter_inc(&vm_statistics_reactivations);
907 vm_object_page_grab_reactivations++;
908 }
909 vm_page_unlock_queues();
910move_page_in_obj:
911 vm_page_queue_remove(&object->memq, p, vmp_listq);
912 vm_page_queue_enter(&object->memq, p, vmp_listq);
913
914 p_skipped++;
915 continue;
916 }
917 vm_page_lockspin_queues();
918take_page:
919 vm_page_free_prepare_queues(page: p);
920 vm_object_page_grab_returned++;
921 vm_object_page_grab_skipped += p_skipped;
922
923 vm_page_unlock_queues();
924
925 vm_page_free_prepare_object(page: p, TRUE);
926
927 return p;
928 }
929 vm_object_page_grab_skipped += p_skipped;
930 vm_object_page_grab_failed++;
931
932 return NULL;
933}
934
935
936
937#define EVICT_PREPARE_LIMIT 64
938#define EVICT_AGE 10
939
940static clock_sec_t vm_object_cache_aging_ts = 0;
941
942static void
943vm_object_cache_remove_locked(
944 vm_object_t object)
945{
946 assert(object->purgable == VM_PURGABLE_DENY);
947
948 queue_remove(&vm_object_cached_list, object, vm_object_t, cached_list);
949 object->cached_list.next = NULL;
950 object->cached_list.prev = NULL;
951
952 vm_object_cached_count--;
953}
954
955void
956vm_object_cache_remove(
957 vm_object_t object)
958{
959 vm_object_cache_lock_spin();
960
961 if (object->cached_list.next &&
962 object->cached_list.prev) {
963 vm_object_cache_remove_locked(object);
964 }
965
966 vm_object_cache_unlock();
967}
968
969void
970vm_object_cache_add(
971 vm_object_t object)
972{
973 clock_sec_t sec;
974 clock_nsec_t nsec;
975
976 assert(object->purgable == VM_PURGABLE_DENY);
977
978 if (object->resident_page_count == 0) {
979 return;
980 }
981 clock_get_system_nanotime(secs: &sec, nanosecs: &nsec);
982
983 vm_object_cache_lock_spin();
984
985 if (object->cached_list.next == NULL &&
986 object->cached_list.prev == NULL) {
987 queue_enter(&vm_object_cached_list, object, vm_object_t, cached_list);
988 object->vo_cache_ts = sec + EVICT_AGE;
989 object->vo_cache_pages_to_scan = object->resident_page_count;
990
991 vm_object_cached_count++;
992 vm_object_cache_adds++;
993 }
994 vm_object_cache_unlock();
995}
996
997int
998vm_object_cache_evict(
999 int num_to_evict,
1000 int max_objects_to_examine)
1001{
1002 vm_object_t object = VM_OBJECT_NULL;
1003 vm_object_t next_obj = VM_OBJECT_NULL;
1004 vm_page_t local_free_q = VM_PAGE_NULL;
1005 vm_page_t p;
1006 vm_page_t next_p;
1007 int object_cnt = 0;
1008 vm_page_t ep_array[EVICT_PREPARE_LIMIT];
1009 int ep_count;
1010 int ep_limit;
1011 int ep_index;
1012 int ep_freed = 0;
1013 int ep_moved = 0;
1014 uint32_t ep_skipped = 0;
1015 clock_sec_t sec;
1016 clock_nsec_t nsec;
1017
1018 KERNEL_DEBUG(0x13001ec | DBG_FUNC_START, 0, 0, 0, 0, 0);
1019 /*
1020 * do a couple of quick checks to see if it's
1021 * worthwhile grabbing the lock
1022 */
1023 if (queue_empty(&vm_object_cached_list)) {
1024 KERNEL_DEBUG(0x13001ec | DBG_FUNC_END, 0, 0, 0, 0, 0);
1025 return 0;
1026 }
1027 clock_get_system_nanotime(secs: &sec, nanosecs: &nsec);
1028
1029 /*
1030 * the object on the head of the queue has not
1031 * yet sufficiently aged
1032 */
1033 if (sec < vm_object_cache_aging_ts) {
1034 KERNEL_DEBUG(0x13001ec | DBG_FUNC_END, 0, 0, 0, 0, 0);
1035 return 0;
1036 }
1037 /*
1038 * don't need the queue lock to find
1039 * and lock an object on the cached list
1040 */
1041 vm_page_unlock_queues();
1042
1043 vm_object_cache_lock_spin();
1044
1045 for (;;) {
1046 next_obj = (vm_object_t)queue_first(&vm_object_cached_list);
1047
1048 while (!queue_end(&vm_object_cached_list, (queue_entry_t)next_obj) && object_cnt++ < max_objects_to_examine) {
1049 object = next_obj;
1050 next_obj = (vm_object_t)queue_next(&next_obj->cached_list);
1051
1052 assert(object->purgable == VM_PURGABLE_DENY);
1053
1054 if (sec < object->vo_cache_ts) {
1055 KERNEL_DEBUG(0x130020c, object, object->resident_page_count, object->vo_cache_ts, sec, 0);
1056
1057 vm_object_cache_aging_ts = object->vo_cache_ts;
1058 object = VM_OBJECT_NULL;
1059 break;
1060 }
1061 if (!vm_object_lock_try_scan(object)) {
1062 /*
1063 * just skip over this guy for now... if we find
1064 * an object to steal pages from, we'll revist in a bit...
1065 * hopefully, the lock will have cleared
1066 */
1067 KERNEL_DEBUG(0x13001f8, object, object->resident_page_count, 0, 0, 0);
1068
1069 object = VM_OBJECT_NULL;
1070 continue;
1071 }
1072 if (vm_page_queue_empty(&object->memq) || object->vo_cache_pages_to_scan == 0) {
1073 /*
1074 * this case really shouldn't happen, but it's not fatal
1075 * so deal with it... if we don't remove the object from
1076 * the list, we'll never move past it.
1077 */
1078 KERNEL_DEBUG(0x13001fc, object, object->resident_page_count, ep_freed, ep_moved, 0);
1079
1080 vm_object_cache_remove_locked(object);
1081 vm_object_unlock(object);
1082 object = VM_OBJECT_NULL;
1083 continue;
1084 }
1085 /*
1086 * we have a locked object with pages...
1087 * time to start harvesting
1088 */
1089 break;
1090 }
1091 vm_object_cache_unlock();
1092
1093 if (object == VM_OBJECT_NULL) {
1094 break;
1095 }
1096
1097 /*
1098 * object is locked at this point and
1099 * has resident pages
1100 */
1101 next_p = (vm_page_t)vm_page_queue_first(&object->memq);
1102
1103 /*
1104 * break the page scan into 2 pieces to minimize the time spent
1105 * behind the page queue lock...
1106 * the list of pages on these unused objects is likely to be cold
1107 * w/r to the cpu cache which increases the time to scan the list
1108 * tenfold... and we may have a 'run' of pages we can't utilize that
1109 * needs to be skipped over...
1110 */
1111 if ((ep_limit = num_to_evict - (ep_freed + ep_moved)) > EVICT_PREPARE_LIMIT) {
1112 ep_limit = EVICT_PREPARE_LIMIT;
1113 }
1114 ep_count = 0;
1115
1116 while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next_p) && object->vo_cache_pages_to_scan && ep_count < ep_limit) {
1117 p = next_p;
1118 next_p = (vm_page_t)vm_page_queue_next(&next_p->vmp_listq);
1119
1120 object->vo_cache_pages_to_scan--;
1121
1122 if (VM_PAGE_WIRED(p) || p->vmp_busy || p->vmp_cleaning || p->vmp_laundry) {
1123 vm_page_queue_remove(&object->memq, p, vmp_listq);
1124 vm_page_queue_enter(&object->memq, p, vmp_listq);
1125
1126 ep_skipped++;
1127 continue;
1128 }
1129 if (p->vmp_wpmapped || p->vmp_dirty || p->vmp_precious) {
1130 vm_page_queue_remove(&object->memq, p, vmp_listq);
1131 vm_page_queue_enter(&object->memq, p, vmp_listq);
1132
1133 pmap_clear_reference(pn: VM_PAGE_GET_PHYS_PAGE(m: p));
1134 }
1135 ep_array[ep_count++] = p;
1136 }
1137 KERNEL_DEBUG(0x13001f4 | DBG_FUNC_START, object, object->resident_page_count, ep_freed, ep_moved, 0);
1138
1139 vm_page_lockspin_queues();
1140
1141 for (ep_index = 0; ep_index < ep_count; ep_index++) {
1142 p = ep_array[ep_index];
1143
1144 if (p->vmp_wpmapped || p->vmp_dirty || p->vmp_precious) {
1145 p->vmp_reference = FALSE;
1146 p->vmp_no_cache = FALSE;
1147
1148 /*
1149 * we've already filtered out pages that are in the laundry
1150 * so if we get here, this page can't be on the pageout queue
1151 */
1152 vm_page_queues_remove(mem: p, FALSE);
1153 vm_page_enqueue_inactive(mem: p, TRUE);
1154
1155 ep_moved++;
1156 } else {
1157#if CONFIG_PHANTOM_CACHE
1158 vm_phantom_cache_add_ghost(p);
1159#endif
1160 vm_page_free_prepare_queues(page: p);
1161
1162 assert(p->vmp_pageq.next == 0 && p->vmp_pageq.prev == 0);
1163 /*
1164 * Add this page to our list of reclaimed pages,
1165 * to be freed later.
1166 */
1167 p->vmp_snext = local_free_q;
1168 local_free_q = p;
1169
1170 ep_freed++;
1171 }
1172 }
1173 vm_page_unlock_queues();
1174
1175 KERNEL_DEBUG(0x13001f4 | DBG_FUNC_END, object, object->resident_page_count, ep_freed, ep_moved, 0);
1176
1177 if (local_free_q) {
1178 vm_page_free_list(mem: local_free_q, TRUE);
1179 local_free_q = VM_PAGE_NULL;
1180 }
1181 if (object->vo_cache_pages_to_scan == 0) {
1182 KERNEL_DEBUG(0x1300208, object, object->resident_page_count, ep_freed, ep_moved, 0);
1183
1184 vm_object_cache_remove(object);
1185
1186 KERNEL_DEBUG(0x13001fc, object, object->resident_page_count, ep_freed, ep_moved, 0);
1187 }
1188 /*
1189 * done with this object
1190 */
1191 vm_object_unlock(object);
1192 object = VM_OBJECT_NULL;
1193
1194 /*
1195 * at this point, we are not holding any locks
1196 */
1197 if ((ep_freed + ep_moved) >= num_to_evict) {
1198 /*
1199 * we've reached our target for the
1200 * number of pages to evict
1201 */
1202 break;
1203 }
1204 vm_object_cache_lock_spin();
1205 }
1206 /*
1207 * put the page queues lock back to the caller's
1208 * idea of it
1209 */
1210 vm_page_lock_queues();
1211
1212 vm_object_cache_pages_freed += ep_freed;
1213 vm_object_cache_pages_moved += ep_moved;
1214 vm_object_cache_pages_skipped += ep_skipped;
1215
1216 KERNEL_DEBUG(0x13001ec | DBG_FUNC_END, ep_freed, 0, 0, 0, 0);
1217 return ep_freed;
1218}
1219
1220/*
1221 * Routine: vm_object_terminate
1222 * Purpose:
1223 * Free all resources associated with a vm_object.
1224 * In/out conditions:
1225 * Upon entry, the object must be locked,
1226 * and the object must have exactly one reference.
1227 *
1228 * The shadow object reference is left alone.
1229 *
1230 * The object must be unlocked if its found that pages
1231 * must be flushed to a backing object. If someone
1232 * manages to map the object while it is being flushed
1233 * the object is returned unlocked and unchanged. Otherwise,
1234 * upon exit, the cache will be unlocked, and the
1235 * object will cease to exist.
1236 */
1237static kern_return_t
1238vm_object_terminate(
1239 vm_object_t object)
1240{
1241 vm_object_t shadow_object;
1242
1243 vm_object_lock_assert_exclusive(object);
1244
1245 if (!object->pageout && (!object->internal && object->can_persist) &&
1246 (object->pager != NULL || object->shadow_severed)) {
1247 /*
1248 * Clear pager_trusted bit so that the pages get yanked
1249 * out of the object instead of cleaned in place. This
1250 * prevents a deadlock in XMM and makes more sense anyway.
1251 */
1252 VM_OBJECT_SET_PAGER_TRUSTED(object, FALSE);
1253
1254 vm_object_reap_pages(object, REAP_TERMINATE);
1255 }
1256 /*
1257 * Make sure the object isn't already being terminated
1258 */
1259 if (object->terminating) {
1260 vm_object_lock_assert_exclusive(object);
1261 object->ref_count--;
1262 assert(object->ref_count > 0);
1263 vm_object_unlock(object);
1264 return KERN_FAILURE;
1265 }
1266
1267 /*
1268 * Did somebody get a reference to the object while we were
1269 * cleaning it?
1270 */
1271 if (object->ref_count != 1) {
1272 vm_object_lock_assert_exclusive(object);
1273 object->ref_count--;
1274 assert(object->ref_count > 0);
1275 vm_object_unlock(object);
1276 return KERN_FAILURE;
1277 }
1278
1279 /*
1280 * Make sure no one can look us up now.
1281 */
1282
1283 VM_OBJECT_SET_TERMINATING(object, TRUE);
1284 VM_OBJECT_SET_ALIVE(object, FALSE);
1285
1286 if (!object->internal &&
1287 object->cached_list.next &&
1288 object->cached_list.prev) {
1289 vm_object_cache_remove(object);
1290 }
1291
1292 /*
1293 * Detach the object from its shadow if we are the shadow's
1294 * copy. The reference we hold on the shadow must be dropped
1295 * by our caller.
1296 */
1297 if (((shadow_object = object->shadow) != VM_OBJECT_NULL) &&
1298 !(object->pageout)) {
1299 vm_object_lock(shadow_object);
1300 if (shadow_object->vo_copy == object) {
1301 VM_OBJECT_COPY_SET(object: shadow_object, VM_OBJECT_NULL);
1302 }
1303 vm_object_unlock(shadow_object);
1304 }
1305
1306 if (object->paging_in_progress != 0 ||
1307 object->activity_in_progress != 0) {
1308 /*
1309 * There are still some paging_in_progress references
1310 * on this object, meaning that there are some paging
1311 * or other I/O operations in progress for this VM object.
1312 * Such operations take some paging_in_progress references
1313 * up front to ensure that the object doesn't go away, but
1314 * they may also need to acquire a reference on the VM object,
1315 * to map it in kernel space, for example. That means that
1316 * they may end up releasing the last reference on the VM
1317 * object, triggering its termination, while still holding
1318 * paging_in_progress references. Waiting for these
1319 * pending paging_in_progress references to go away here would
1320 * deadlock.
1321 *
1322 * To avoid deadlocking, we'll let the vm_object_reaper_thread
1323 * complete the VM object termination if it still holds
1324 * paging_in_progress references at this point.
1325 *
1326 * No new paging_in_progress should appear now that the
1327 * VM object is "terminating" and not "alive".
1328 */
1329 vm_object_reap_async(object);
1330 vm_object_unlock(object);
1331 /*
1332 * Return KERN_FAILURE to let the caller know that we
1333 * haven't completed the termination and it can't drop this
1334 * object's reference on its shadow object yet.
1335 * The reaper thread will take care of that once it has
1336 * completed this object's termination.
1337 */
1338 return KERN_FAILURE;
1339 }
1340 /*
1341 * complete the VM object termination
1342 */
1343 vm_object_reap(object);
1344 object = VM_OBJECT_NULL;
1345
1346 /*
1347 * the object lock was released by vm_object_reap()
1348 *
1349 * KERN_SUCCESS means that this object has been terminated
1350 * and no longer needs its shadow object but still holds a
1351 * reference on it.
1352 * The caller is responsible for dropping that reference.
1353 * We can't call vm_object_deallocate() here because that
1354 * would create a recursion.
1355 */
1356 return KERN_SUCCESS;
1357}
1358
1359
1360/*
1361 * vm_object_reap():
1362 *
1363 * Complete the termination of a VM object after it's been marked
1364 * as "terminating" and "!alive" by vm_object_terminate().
1365 *
1366 * The VM object must be locked by caller.
1367 * The lock will be released on return and the VM object is no longer valid.
1368 */
1369
1370void
1371vm_object_reap(
1372 vm_object_t object)
1373{
1374 memory_object_t pager;
1375
1376 vm_object_lock_assert_exclusive(object);
1377 assert(object->paging_in_progress == 0);
1378 assert(object->activity_in_progress == 0);
1379
1380 vm_object_reap_count++;
1381
1382 /*
1383 * Disown this purgeable object to cleanup its owner's purgeable
1384 * ledgers. We need to do this before disconnecting the object
1385 * from its pager, to properly account for compressed pages.
1386 */
1387 if (object->internal &&
1388 (object->purgable != VM_PURGABLE_DENY ||
1389 object->vo_ledger_tag)) {
1390 int ledger_flags;
1391 kern_return_t kr;
1392
1393 ledger_flags = 0;
1394 if (object->vo_no_footprint) {
1395 ledger_flags |= VM_LEDGER_FLAG_NO_FOOTPRINT;
1396 }
1397 assert(!object->alive);
1398 assert(object->terminating);
1399 kr = vm_object_ownership_change(object,
1400 new_ledger_tag: object->vo_ledger_tag, /* unchanged */
1401 NULL, /* no owner */
1402 new_ledger_flags: ledger_flags,
1403 FALSE); /* task_objq not locked */
1404 assert(kr == KERN_SUCCESS);
1405 assert(object->vo_owner == NULL);
1406 }
1407
1408#if DEVELOPMENT || DEBUG
1409 if (object->object_is_shared_cache &&
1410 object->pager != NULL &&
1411 object->pager->mo_pager_ops == &shared_region_pager_ops) {
1412 OSAddAtomic(-object->resident_page_count, &shared_region_pagers_resident_count);
1413 }
1414#endif /* DEVELOPMENT || DEBUG */
1415
1416 pager = object->pager;
1417 object->pager = MEMORY_OBJECT_NULL;
1418
1419 if (pager != MEMORY_OBJECT_NULL) {
1420 memory_object_control_disable(control: &object->pager_control);
1421 }
1422
1423 object->ref_count--;
1424 assert(object->ref_count == 0);
1425
1426 /*
1427 * remove from purgeable queue if it's on
1428 */
1429 if (object->internal) {
1430 assert(VM_OBJECT_OWNER(object) == TASK_NULL);
1431
1432 VM_OBJECT_UNWIRED(object);
1433
1434 if (object->purgable == VM_PURGABLE_DENY) {
1435 /* not purgeable: nothing to do */
1436 } else if (object->purgable == VM_PURGABLE_VOLATILE) {
1437 purgeable_q_t queue;
1438
1439 queue = vm_purgeable_object_remove(object);
1440 assert(queue);
1441
1442 if (object->purgeable_when_ripe) {
1443 /*
1444 * Must take page lock for this -
1445 * using it to protect token queue
1446 */
1447 vm_page_lock_queues();
1448 vm_purgeable_token_delete_first(queue);
1449
1450 assert(queue->debug_count_objects >= 0);
1451 vm_page_unlock_queues();
1452 }
1453
1454 /*
1455 * Update "vm_page_purgeable_count" in bulk and mark
1456 * object as VM_PURGABLE_EMPTY to avoid updating
1457 * "vm_page_purgeable_count" again in vm_page_remove()
1458 * when reaping the pages.
1459 */
1460 unsigned int delta;
1461 assert(object->resident_page_count >=
1462 object->wired_page_count);
1463 delta = (object->resident_page_count -
1464 object->wired_page_count);
1465 if (delta != 0) {
1466 assert(vm_page_purgeable_count >= delta);
1467 OSAddAtomic(-delta,
1468 (SInt32 *)&vm_page_purgeable_count);
1469 }
1470 if (object->wired_page_count != 0) {
1471 assert(vm_page_purgeable_wired_count >=
1472 object->wired_page_count);
1473 OSAddAtomic(-object->wired_page_count,
1474 (SInt32 *)&vm_page_purgeable_wired_count);
1475 }
1476 VM_OBJECT_SET_PURGABLE(object, VM_PURGABLE_EMPTY);
1477 } else if (object->purgable == VM_PURGABLE_NONVOLATILE ||
1478 object->purgable == VM_PURGABLE_EMPTY) {
1479 /* remove from nonvolatile queue */
1480 vm_purgeable_nonvolatile_dequeue(object);
1481 } else {
1482 panic("object %p in unexpected purgeable state 0x%x",
1483 object, object->purgable);
1484 }
1485 if (object->transposed &&
1486 object->cached_list.next != NULL &&
1487 object->cached_list.prev == NULL) {
1488 /*
1489 * object->cached_list.next "points" to the
1490 * object that was transposed with this object.
1491 */
1492 } else {
1493 assert(object->cached_list.next == NULL);
1494 }
1495 assert(object->cached_list.prev == NULL);
1496 }
1497
1498 if (object->pageout) {
1499 /*
1500 * free all remaining pages tabled on
1501 * this object
1502 * clean up it's shadow
1503 */
1504 assert(object->shadow != VM_OBJECT_NULL);
1505
1506 vm_pageout_object_terminate(object);
1507 } else if (object->resident_page_count) {
1508 /*
1509 * free all remaining pages tabled on
1510 * this object
1511 */
1512 vm_object_reap_pages(object, REAP_REAP);
1513 }
1514 assert(vm_page_queue_empty(&object->memq));
1515 assert(object->paging_in_progress == 0);
1516 assert(object->activity_in_progress == 0);
1517 assert(object->ref_count == 0);
1518
1519 /*
1520 * If the pager has not already been released by
1521 * vm_object_destroy, we need to terminate it and
1522 * release our reference to it here.
1523 */
1524 if (pager != MEMORY_OBJECT_NULL) {
1525 vm_object_unlock(object);
1526 vm_object_release_pager(pager);
1527 vm_object_lock(object);
1528 }
1529
1530 /* kick off anyone waiting on terminating */
1531 VM_OBJECT_SET_TERMINATING(object, FALSE);
1532 vm_object_paging_begin(object);
1533 vm_object_paging_end(object);
1534 vm_object_unlock(object);
1535
1536 object->shadow = VM_OBJECT_NULL;
1537
1538#if VM_OBJECT_TRACKING
1539 if (vm_object_tracking_btlog) {
1540 btlog_erase(vm_object_tracking_btlog, object);
1541 }
1542#endif /* VM_OBJECT_TRACKING */
1543
1544 vm_object_lock_destroy(object);
1545 /*
1546 * Free the space for the object.
1547 */
1548 zfree(vm_object_zone, object);
1549 object = VM_OBJECT_NULL;
1550}
1551
1552
1553unsigned int vm_max_batch = 256;
1554
1555#define V_O_R_MAX_BATCH 128
1556
1557#define BATCH_LIMIT(max) (vm_max_batch >= max ? max : vm_max_batch)
1558
1559
1560#define VM_OBJ_REAP_FREELIST(_local_free_q, do_disconnect) \
1561 MACRO_BEGIN \
1562 if (_local_free_q) { \
1563 if (do_disconnect) { \
1564 vm_page_t m; \
1565 for (m = _local_free_q; \
1566 m != VM_PAGE_NULL; \
1567 m = m->vmp_snext) { \
1568 if (m->vmp_pmapped) { \
1569 pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)); \
1570 } \
1571 } \
1572 } \
1573 vm_page_free_list(_local_free_q, TRUE); \
1574 _local_free_q = VM_PAGE_NULL; \
1575 } \
1576 MACRO_END
1577
1578
1579void
1580vm_object_reap_pages(
1581 vm_object_t object,
1582 int reap_type)
1583{
1584 vm_page_t p;
1585 vm_page_t next;
1586 vm_page_t local_free_q = VM_PAGE_NULL;
1587 int loop_count;
1588 boolean_t disconnect_on_release;
1589 pmap_flush_context pmap_flush_context_storage;
1590
1591 if (reap_type == REAP_DATA_FLUSH) {
1592 /*
1593 * We need to disconnect pages from all pmaps before
1594 * releasing them to the free list
1595 */
1596 disconnect_on_release = TRUE;
1597 } else {
1598 /*
1599 * Either the caller has already disconnected the pages
1600 * from all pmaps, or we disconnect them here as we add
1601 * them to out local list of pages to be released.
1602 * No need to re-disconnect them when we release the pages
1603 * to the free list.
1604 */
1605 disconnect_on_release = FALSE;
1606 }
1607
1608restart_after_sleep:
1609 if (vm_page_queue_empty(&object->memq)) {
1610 return;
1611 }
1612 loop_count = BATCH_LIMIT(V_O_R_MAX_BATCH);
1613
1614 if (reap_type == REAP_PURGEABLE) {
1615 pmap_flush_context_init(&pmap_flush_context_storage);
1616 }
1617
1618 vm_page_lock_queues();
1619
1620 next = (vm_page_t)vm_page_queue_first(&object->memq);
1621
1622 while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next)) {
1623 p = next;
1624 next = (vm_page_t)vm_page_queue_next(&next->vmp_listq);
1625
1626 if (--loop_count == 0) {
1627 vm_page_unlock_queues();
1628
1629 if (local_free_q) {
1630 if (reap_type == REAP_PURGEABLE) {
1631 pmap_flush(&pmap_flush_context_storage);
1632 pmap_flush_context_init(&pmap_flush_context_storage);
1633 }
1634 /*
1635 * Free the pages we reclaimed so far
1636 * and take a little break to avoid
1637 * hogging the page queue lock too long
1638 */
1639 VM_OBJ_REAP_FREELIST(local_free_q,
1640 disconnect_on_release);
1641 } else {
1642 mutex_pause(0);
1643 }
1644
1645 loop_count = BATCH_LIMIT(V_O_R_MAX_BATCH);
1646
1647 vm_page_lock_queues();
1648 }
1649 if (reap_type == REAP_DATA_FLUSH || reap_type == REAP_TERMINATE) {
1650 if (p->vmp_busy || p->vmp_cleaning) {
1651 vm_page_unlock_queues();
1652 /*
1653 * free the pages reclaimed so far
1654 */
1655 VM_OBJ_REAP_FREELIST(local_free_q,
1656 disconnect_on_release);
1657
1658 PAGE_SLEEP(object, p, THREAD_UNINT);
1659
1660 goto restart_after_sleep;
1661 }
1662 if (p->vmp_laundry) {
1663 vm_pageout_steal_laundry(page: p, TRUE);
1664 }
1665 }
1666 switch (reap_type) {
1667 case REAP_DATA_FLUSH:
1668 if (VM_PAGE_WIRED(p)) {
1669 /*
1670 * this is an odd case... perhaps we should
1671 * zero-fill this page since we're conceptually
1672 * tossing its data at this point, but leaving
1673 * it on the object to honor the 'wire' contract
1674 */
1675 continue;
1676 }
1677 break;
1678
1679 case REAP_PURGEABLE:
1680 if (VM_PAGE_WIRED(p)) {
1681 /*
1682 * can't purge a wired page
1683 */
1684 vm_page_purged_wired++;
1685 continue;
1686 }
1687 if (p->vmp_laundry && !p->vmp_busy && !p->vmp_cleaning) {
1688 vm_pageout_steal_laundry(page: p, TRUE);
1689 }
1690
1691 if (p->vmp_cleaning || p->vmp_laundry || p->vmp_absent) {
1692 /*
1693 * page is being acted upon,
1694 * so don't mess with it
1695 */
1696 vm_page_purged_others++;
1697 continue;
1698 }
1699 if (p->vmp_busy) {
1700 /*
1701 * We can't reclaim a busy page but we can
1702 * make it more likely to be paged (it's not wired) to make
1703 * sure that it gets considered by
1704 * vm_pageout_scan() later.
1705 */
1706 if (VM_PAGE_PAGEABLE(p)) {
1707 vm_page_deactivate(page: p);
1708 }
1709 vm_page_purged_busy++;
1710 continue;
1711 }
1712
1713 assert(!is_kernel_object(VM_PAGE_OBJECT(p)));
1714
1715 /*
1716 * we can discard this page...
1717 */
1718 if (p->vmp_pmapped == TRUE) {
1719 /*
1720 * unmap the page
1721 */
1722 pmap_disconnect_options(phys: VM_PAGE_GET_PHYS_PAGE(m: p), PMAP_OPTIONS_NOFLUSH | PMAP_OPTIONS_NOREFMOD, arg: (void *)&pmap_flush_context_storage);
1723 }
1724 vm_page_purged_count++;
1725
1726 break;
1727
1728 case REAP_TERMINATE:
1729 if (p->vmp_absent || p->vmp_private) {
1730 /*
1731 * For private pages, VM_PAGE_FREE just
1732 * leaves the page structure around for
1733 * its owner to clean up. For absent
1734 * pages, the structure is returned to
1735 * the appropriate pool.
1736 */
1737 break;
1738 }
1739 if (p->vmp_fictitious) {
1740 assert(VM_PAGE_GET_PHYS_PAGE(p) == vm_page_guard_addr);
1741 break;
1742 }
1743 if (!p->vmp_dirty && p->vmp_wpmapped) {
1744 p->vmp_dirty = pmap_is_modified(pn: VM_PAGE_GET_PHYS_PAGE(m: p));
1745 }
1746
1747 if ((p->vmp_dirty || p->vmp_precious) && !VMP_ERROR_GET(p) && object->alive) {
1748 assert(!object->internal);
1749
1750 p->vmp_free_when_done = TRUE;
1751
1752 if (!p->vmp_laundry) {
1753 vm_page_queues_remove(mem: p, TRUE);
1754 /*
1755 * flush page... page will be freed
1756 * upon completion of I/O
1757 */
1758 vm_pageout_cluster(m: p);
1759 }
1760 vm_page_unlock_queues();
1761 /*
1762 * free the pages reclaimed so far
1763 */
1764 VM_OBJ_REAP_FREELIST(local_free_q,
1765 disconnect_on_release);
1766
1767 vm_object_paging_wait(object, THREAD_UNINT);
1768
1769 goto restart_after_sleep;
1770 }
1771 break;
1772
1773 case REAP_REAP:
1774 break;
1775 }
1776 vm_page_free_prepare_queues(page: p);
1777 assert(p->vmp_pageq.next == 0 && p->vmp_pageq.prev == 0);
1778 /*
1779 * Add this page to our list of reclaimed pages,
1780 * to be freed later.
1781 */
1782 p->vmp_snext = local_free_q;
1783 local_free_q = p;
1784 }
1785 vm_page_unlock_queues();
1786
1787 /*
1788 * Free the remaining reclaimed pages
1789 */
1790 if (reap_type == REAP_PURGEABLE) {
1791 pmap_flush(&pmap_flush_context_storage);
1792 }
1793
1794 VM_OBJ_REAP_FREELIST(local_free_q,
1795 disconnect_on_release);
1796}
1797
1798
1799void
1800vm_object_reap_async(
1801 vm_object_t object)
1802{
1803 vm_object_lock_assert_exclusive(object);
1804
1805 vm_object_reaper_lock_spin();
1806
1807 vm_object_reap_count_async++;
1808
1809 /* enqueue the VM object... */
1810 queue_enter(&vm_object_reaper_queue, object,
1811 vm_object_t, cached_list);
1812
1813 vm_object_reaper_unlock();
1814
1815 /* ... and wake up the reaper thread */
1816 thread_wakeup((event_t) &vm_object_reaper_queue);
1817}
1818
1819
1820void
1821vm_object_reaper_thread(void)
1822{
1823 vm_object_t object, shadow_object;
1824
1825 vm_object_reaper_lock_spin();
1826
1827 while (!queue_empty(&vm_object_reaper_queue)) {
1828 queue_remove_first(&vm_object_reaper_queue,
1829 object,
1830 vm_object_t,
1831 cached_list);
1832
1833 vm_object_reaper_unlock();
1834 vm_object_lock(object);
1835
1836 assert(object->terminating);
1837 assert(!object->alive);
1838
1839 /*
1840 * The pageout daemon might be playing with our pages.
1841 * Now that the object is dead, it won't touch any more
1842 * pages, but some pages might already be on their way out.
1843 * Hence, we wait until the active paging activities have
1844 * ceased before we break the association with the pager
1845 * itself.
1846 */
1847 while (object->paging_in_progress != 0 ||
1848 object->activity_in_progress != 0) {
1849 vm_object_wait(object,
1850 VM_OBJECT_EVENT_PAGING_IN_PROGRESS,
1851 THREAD_UNINT);
1852 vm_object_lock(object);
1853 }
1854
1855 shadow_object =
1856 object->pageout ? VM_OBJECT_NULL : object->shadow;
1857
1858 vm_object_reap(object);
1859 /* cache is unlocked and object is no longer valid */
1860 object = VM_OBJECT_NULL;
1861
1862 if (shadow_object != VM_OBJECT_NULL) {
1863 /*
1864 * Drop the reference "object" was holding on
1865 * its shadow object.
1866 */
1867 vm_object_deallocate(object: shadow_object);
1868 shadow_object = VM_OBJECT_NULL;
1869 }
1870 vm_object_reaper_lock_spin();
1871 }
1872
1873 /* wait for more work... */
1874 assert_wait(event: (event_t) &vm_object_reaper_queue, THREAD_UNINT);
1875
1876 vm_object_reaper_unlock();
1877
1878 thread_block(continuation: (thread_continue_t) vm_object_reaper_thread);
1879 /*NOTREACHED*/
1880}
1881
1882/*
1883 * Routine: vm_object_release_pager
1884 * Purpose: Terminate the pager and, upon completion,
1885 * release our last reference to it.
1886 */
1887static void
1888vm_object_release_pager(
1889 memory_object_t pager)
1890{
1891 /*
1892 * Terminate the pager.
1893 */
1894
1895 (void) memory_object_terminate(memory_object: pager);
1896
1897 /*
1898 * Release reference to pager.
1899 */
1900 memory_object_deallocate(object: pager);
1901}
1902
1903/*
1904 * Routine: vm_object_destroy
1905 * Purpose:
1906 * Shut down a VM object, despite the
1907 * presence of address map (or other) references
1908 * to the vm_object.
1909 */
1910#if MACH_ASSERT
1911extern uint32_t system_inshutdown;
1912int fbdp_no_panic = 1;
1913#endif /* MACH_ASSERT */
1914kern_return_t
1915vm_object_destroy(
1916 vm_object_t object,
1917 vm_object_destroy_reason_t reason)
1918{
1919 memory_object_t old_pager;
1920
1921 if (object == VM_OBJECT_NULL) {
1922 return KERN_SUCCESS;
1923 }
1924
1925 /*
1926 * Remove the pager association immediately.
1927 *
1928 * This will prevent the memory manager from further
1929 * meddling. [If it wanted to flush data or make
1930 * other changes, it should have done so before performing
1931 * the destroy call.]
1932 */
1933
1934 vm_object_lock(object);
1935
1936#if FBDP_DEBUG_OBJECT_NO_PAGER
1937 static bool fbdp_no_panic_retrieved = false;
1938 if (!fbdp_no_panic_retrieved) {
1939 PE_parse_boot_argn("fbdp_no_panic4", &fbdp_no_panic, sizeof(fbdp_no_panic));
1940 fbdp_no_panic_retrieved = true;
1941 }
1942
1943 bool forced_unmount = false;
1944 if (object->named &&
1945 object->ref_count > 2 &&
1946 object->pager != NULL &&
1947 vnode_pager_get_forced_unmount(object->pager, &forced_unmount) == KERN_SUCCESS &&
1948 forced_unmount == false) {
1949 if (!fbdp_no_panic) {
1950 panic("FBDP rdar://99829401 object %p refs %d pager %p (no forced unmount)\n", object, object->ref_count, object->pager);
1951 }
1952 DTRACE_VM3(vm_object_destroy_no_forced_unmount,
1953 vm_object_t, object,
1954 int, object->ref_count,
1955 memory_object_t, object->pager);
1956 }
1957
1958 if (object->fbdp_tracked) {
1959 if (object->ref_count > 2 && !system_inshutdown) {
1960 if (!fbdp_no_panic) {
1961 panic("FBDP/4 rdar://99829401 object %p refs %d pager %p (tracked)\n", object, object->ref_count, object->pager);
1962 }
1963 }
1964 VM_OBJECT_SET_FBDP_TRACKED(object, false);
1965 }
1966#endif /* FBDP_DEBUG_OBJECT_NO_PAGER */
1967
1968 if (reason != VM_OBJECT_DESTROY_UNKNOWN_REASON) {
1969 VM_OBJECT_SET_NO_PAGER_REASON(object, value: reason);
1970 }
1971
1972 VM_OBJECT_SET_CAN_PERSIST(object, FALSE);
1973 VM_OBJECT_SET_NAMED(object, FALSE);
1974#if 00
1975 VM_OBJECT_SET_ALIVE(object, FALSE);
1976#endif /* 00 */
1977
1978#if DEVELOPMENT || DEBUG
1979 if (object->object_is_shared_cache &&
1980 object->pager != NULL &&
1981 object->pager->mo_pager_ops == &shared_region_pager_ops) {
1982 OSAddAtomic(-object->resident_page_count, &shared_region_pagers_resident_count);
1983 }
1984#endif /* DEVELOPMENT || DEBUG */
1985
1986 old_pager = object->pager;
1987 object->pager = MEMORY_OBJECT_NULL;
1988 if (old_pager != MEMORY_OBJECT_NULL) {
1989 memory_object_control_disable(control: &object->pager_control);
1990 }
1991
1992 /*
1993 * Wait for the existing paging activity (that got
1994 * through before we nulled out the pager) to subside.
1995 */
1996
1997 vm_object_paging_wait(object, THREAD_UNINT);
1998 vm_object_unlock(object);
1999
2000 /*
2001 * Terminate the object now.
2002 */
2003 if (old_pager != MEMORY_OBJECT_NULL) {
2004 vm_object_release_pager(pager: old_pager);
2005
2006 /*
2007 * JMM - Release the caller's reference. This assumes the
2008 * caller had a reference to release, which is a big (but
2009 * currently valid) assumption if this is driven from the
2010 * vnode pager (it is holding a named reference when making
2011 * this call)..
2012 */
2013 vm_object_deallocate(object);
2014 }
2015 return KERN_SUCCESS;
2016}
2017
2018/*
2019 * The "chunk" macros are used by routines below when looking for pages to deactivate. These
2020 * exist because of the need to handle shadow chains. When deactivating pages, we only
2021 * want to deactive the ones at the top most level in the object chain. In order to do
2022 * this efficiently, the specified address range is divided up into "chunks" and we use
2023 * a bit map to keep track of which pages have already been processed as we descend down
2024 * the shadow chain. These chunk macros hide the details of the bit map implementation
2025 * as much as we can.
2026 *
2027 * For convenience, we use a 64-bit data type as the bit map, and therefore a chunk is
2028 * set to 64 pages. The bit map is indexed from the low-order end, so that the lowest
2029 * order bit represents page 0 in the current range and highest order bit represents
2030 * page 63.
2031 *
2032 * For further convenience, we also use negative logic for the page state in the bit map.
2033 * The bit is set to 1 to indicate it has not yet been seen, and to 0 to indicate it has
2034 * been processed. This way we can simply test the 64-bit long word to see if it's zero
2035 * to easily tell if the whole range has been processed. Therefore, the bit map starts
2036 * out with all the bits set. The macros below hide all these details from the caller.
2037 */
2038
2039#define PAGES_IN_A_CHUNK 64 /* The number of pages in the chunk must */
2040 /* be the same as the number of bits in */
2041 /* the chunk_state_t type. We use 64 */
2042 /* just for convenience. */
2043
2044#define CHUNK_SIZE (PAGES_IN_A_CHUNK * PAGE_SIZE_64) /* Size of a chunk in bytes */
2045
2046typedef uint64_t chunk_state_t;
2047
2048/*
2049 * The bit map uses negative logic, so we start out with all 64 bits set to indicate
2050 * that no pages have been processed yet. Also, if len is less than the full CHUNK_SIZE,
2051 * then we mark pages beyond the len as having been "processed" so that we don't waste time
2052 * looking at pages in that range. This can save us from unnecessarily chasing down the
2053 * shadow chain.
2054 */
2055
2056#define CHUNK_INIT(c, len) \
2057 MACRO_BEGIN \
2058 uint64_t p; \
2059 \
2060 (c) = 0xffffffffffffffffLL; \
2061 \
2062 for (p = (len) / PAGE_SIZE_64; p < PAGES_IN_A_CHUNK; p++) \
2063 MARK_PAGE_HANDLED(c, p); \
2064 MACRO_END
2065
2066
2067/*
2068 * Return true if all pages in the chunk have not yet been processed.
2069 */
2070
2071#define CHUNK_NOT_COMPLETE(c) ((c) != 0)
2072
2073/*
2074 * Return true if the page at offset 'p' in the bit map has already been handled
2075 * while processing a higher level object in the shadow chain.
2076 */
2077
2078#define PAGE_ALREADY_HANDLED(c, p) (((c) & (1ULL << (p))) == 0)
2079
2080/*
2081 * Mark the page at offset 'p' in the bit map as having been processed.
2082 */
2083
2084#define MARK_PAGE_HANDLED(c, p) \
2085MACRO_BEGIN \
2086 (c) = (c) & ~(1ULL << (p)); \
2087MACRO_END
2088
2089
2090/*
2091 * Return true if the page at the given offset has been paged out. Object is
2092 * locked upon entry and returned locked.
2093 */
2094
2095static boolean_t
2096page_is_paged_out(
2097 vm_object_t object,
2098 vm_object_offset_t offset)
2099{
2100 if (object->internal &&
2101 object->alive &&
2102 !object->terminating &&
2103 object->pager_ready) {
2104 if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset)
2105 == VM_EXTERNAL_STATE_EXISTS) {
2106 return TRUE;
2107 }
2108 }
2109 return FALSE;
2110}
2111
2112
2113
2114/*
2115 * madvise_free_debug
2116 *
2117 * To help debug madvise(MADV_FREE*) mis-usage, this triggers a
2118 * zero-fill as soon as a page is affected by a madvise(MADV_FREE*), to
2119 * simulate the loss of the page's contents as if the page had been
2120 * reclaimed and then re-faulted.
2121 */
2122#if DEVELOPMENT || DEBUG
2123int madvise_free_debug = 0;
2124int madvise_free_debug_sometimes = 1;
2125#else /* DEBUG */
2126int madvise_free_debug = 0;
2127int madvise_free_debug_sometimes = 0;
2128#endif /* DEBUG */
2129int madvise_free_counter = 0;
2130
2131__options_decl(deactivate_flags_t, uint32_t, {
2132 DEACTIVATE_KILL = 0x1,
2133 DEACTIVATE_REUSABLE = 0x2,
2134 DEACTIVATE_ALL_REUSABLE = 0x4,
2135 DEACTIVATE_CLEAR_REFMOD = 0x8,
2136 DEACTIVATE_REUSABLE_NO_WRITE = 0x10
2137});
2138
2139/*
2140 * Deactivate the pages in the specified object and range. If kill_page is set, also discard any
2141 * page modified state from the pmap. Update the chunk_state as we go along. The caller must specify
2142 * a size that is less than or equal to the CHUNK_SIZE.
2143 */
2144
2145static void
2146deactivate_pages_in_object(
2147 vm_object_t object,
2148 vm_object_offset_t offset,
2149 vm_object_size_t size,
2150 deactivate_flags_t flags,
2151 chunk_state_t *chunk_state,
2152 pmap_flush_context *pfc,
2153 struct pmap *pmap,
2154 vm_map_offset_t pmap_offset)
2155{
2156 vm_page_t m;
2157 int p;
2158 struct vm_page_delayed_work dw_array;
2159 struct vm_page_delayed_work *dwp, *dwp_start;
2160 bool dwp_finish_ctx = TRUE;
2161 int dw_count;
2162 int dw_limit;
2163 unsigned int reusable = 0;
2164
2165 /*
2166 * Examine each page in the chunk. The variable 'p' is the page number relative to the start of the
2167 * chunk. Since this routine is called once for each level in the shadow chain, the chunk_state may
2168 * have pages marked as having been processed already. We stop the loop early if we find we've handled
2169 * all the pages in the chunk.
2170 */
2171
2172 dwp_start = dwp = NULL;
2173 dw_count = 0;
2174 dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT);
2175 dwp_start = vm_page_delayed_work_get_ctx();
2176 if (dwp_start == NULL) {
2177 dwp_start = &dw_array;
2178 dw_limit = 1;
2179 dwp_finish_ctx = FALSE;
2180 }
2181
2182 dwp = dwp_start;
2183
2184 for (p = 0; size && CHUNK_NOT_COMPLETE(*chunk_state); p++, size -= PAGE_SIZE_64, offset += PAGE_SIZE_64, pmap_offset += PAGE_SIZE_64) {
2185 /*
2186 * If this offset has already been found and handled in a higher level object, then don't
2187 * do anything with it in the current shadow object.
2188 */
2189
2190 if (PAGE_ALREADY_HANDLED(*chunk_state, p)) {
2191 continue;
2192 }
2193
2194 /*
2195 * See if the page at this offset is around. First check to see if the page is resident,
2196 * then if not, check the existence map or with the pager.
2197 */
2198
2199 if ((m = vm_page_lookup(object, offset)) != VM_PAGE_NULL) {
2200 /*
2201 * We found a page we were looking for. Mark it as "handled" now in the chunk_state
2202 * so that we won't bother looking for a page at this offset again if there are more
2203 * shadow objects. Then deactivate the page.
2204 */
2205
2206 MARK_PAGE_HANDLED(*chunk_state, p);
2207
2208 if ((!VM_PAGE_WIRED(m)) && (!m->vmp_private) && (!m->vmp_gobbled) && (!m->vmp_busy) &&
2209 (!m->vmp_laundry) && (!m->vmp_cleaning) && !(m->vmp_free_when_done)) {
2210 int clear_refmod_mask;
2211 int pmap_options;
2212 dwp->dw_mask = 0;
2213
2214 pmap_options = 0;
2215 clear_refmod_mask = VM_MEM_REFERENCED;
2216 dwp->dw_mask |= DW_clear_reference;
2217
2218 if ((flags & DEACTIVATE_KILL) && (object->internal)) {
2219 if (!(flags & DEACTIVATE_REUSABLE_NO_WRITE) &&
2220 (madvise_free_debug ||
2221 (madvise_free_debug_sometimes &&
2222 madvise_free_counter++ & 0x1))) {
2223 /*
2224 * zero-fill the page (or every
2225 * other page) now to simulate
2226 * it being reclaimed and
2227 * re-faulted.
2228 */
2229#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
2230 if (!m->vmp_unmodified_ro) {
2231#else /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
2232 if (true) {
2233#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
2234 pmap_zero_page(pn: VM_PAGE_GET_PHYS_PAGE(m));
2235 }
2236 }
2237 m->vmp_precious = FALSE;
2238 m->vmp_dirty = FALSE;
2239
2240 clear_refmod_mask |= VM_MEM_MODIFIED;
2241 if (m->vmp_q_state == VM_PAGE_ON_THROTTLED_Q) {
2242 /*
2243 * This page is now clean and
2244 * reclaimable. Move it out
2245 * of the throttled queue, so
2246 * that vm_pageout_scan() can
2247 * find it.
2248 */
2249 dwp->dw_mask |= DW_move_page;
2250 }
2251
2252#if 0
2253#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
2254 /*
2255 * COMMENT BLOCK ON WHY THIS SHOULDN'T BE DONE.
2256 *
2257 * Since we are about to do a VM_COMPRESSOR_PAGER_STATE_CLR
2258 * below for this page, which drops any existing compressor
2259 * storage of this page (eg side-effect of a CoW operation or
2260 * a collapse operation), it is tempting to think that we should
2261 * treat this page as if it was just decompressed (during which
2262 * we also drop existing compressor storage) and so start its life
2263 * out with vmp_unmodified_ro set to FALSE.
2264 *
2265 * However, we can't do that here because we could swing around
2266 * and re-access this page in a read-only fault.
2267 * Clearing this bit means we'll try to zero it up above
2268 * and fail.
2269 *
2270 * Note that clearing the bit is unnecessary regardless because
2271 * dirty state has been cleared. During the next soft fault, the
2272 * right state will be restored and things will progress just fine.
2273 */
2274 if (m->vmp_unmodified_ro == true) {
2275 /* Need object and pageq locks for bit manipulation*/
2276 m->vmp_unmodified_ro = false;
2277 os_atomic_dec(&compressor_ro_uncompressed);
2278 }
2279#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
2280#endif /* 0 */
2281 VM_COMPRESSOR_PAGER_STATE_CLR(object, offset);
2282
2283 if ((flags & DEACTIVATE_REUSABLE) && !m->vmp_reusable) {
2284 assert(!(flags & DEACTIVATE_ALL_REUSABLE));
2285 assert(!object->all_reusable);
2286 m->vmp_reusable = TRUE;
2287 object->reusable_page_count++;
2288 assert(object->resident_page_count >= object->reusable_page_count);
2289 reusable++;
2290 /*
2291 * Tell pmap this page is now
2292 * "reusable" (to update pmap
2293 * stats for all mappings).
2294 */
2295 pmap_options |= PMAP_OPTIONS_SET_REUSABLE;
2296 }
2297 }
2298 if (flags & DEACTIVATE_CLEAR_REFMOD) {
2299 /*
2300 * The caller didn't clear the refmod bits in advance.
2301 * Clear them for this page now.
2302 */
2303 pmap_options |= PMAP_OPTIONS_NOFLUSH;
2304 pmap_clear_refmod_options(pn: VM_PAGE_GET_PHYS_PAGE(m),
2305 mask: clear_refmod_mask,
2306 options: pmap_options,
2307 (void *)pfc);
2308 }
2309
2310 if ((m->vmp_q_state != VM_PAGE_ON_THROTTLED_Q) &&
2311 !(flags & (DEACTIVATE_REUSABLE | DEACTIVATE_ALL_REUSABLE))) {
2312 dwp->dw_mask |= DW_move_page;
2313 }
2314
2315 if (dwp->dw_mask) {
2316 VM_PAGE_ADD_DELAYED_WORK(dwp, m,
2317 dw_count);
2318 }
2319
2320 if (dw_count >= dw_limit) {
2321 if (reusable) {
2322 OSAddAtomic(reusable,
2323 &vm_page_stats_reusable.reusable_count);
2324 vm_page_stats_reusable.reusable += reusable;
2325 reusable = 0;
2326 }
2327 vm_page_do_delayed_work(object, VM_KERN_MEMORY_NONE, dwp: dwp_start, dw_count);
2328
2329 dwp = dwp_start;
2330 dw_count = 0;
2331 }
2332 }
2333 } else {
2334 /*
2335 * The page at this offset isn't memory resident, check to see if it's
2336 * been paged out. If so, mark it as handled so we don't bother looking
2337 * for it in the shadow chain.
2338 */
2339
2340 if (page_is_paged_out(object, offset)) {
2341 MARK_PAGE_HANDLED(*chunk_state, p);
2342
2343 /*
2344 * If we're killing a non-resident page, then clear the page in the existence
2345 * map so we don't bother paging it back in if it's touched again in the future.
2346 */
2347
2348 if ((flags & DEACTIVATE_KILL) && (object->internal)) {
2349 VM_COMPRESSOR_PAGER_STATE_CLR(object, offset);
2350
2351 if (pmap != PMAP_NULL) {
2352 /*
2353 * Tell pmap that this page
2354 * is no longer mapped, to
2355 * adjust the footprint ledger
2356 * because this page is no
2357 * longer compressed.
2358 */
2359 pmap_remove_options(
2360 map: pmap,
2361 s: pmap_offset,
2362 e: (pmap_offset +
2363 PAGE_SIZE),
2364 PMAP_OPTIONS_REMOVE);
2365 }
2366 }
2367 }
2368 }
2369 }
2370
2371 if (reusable) {
2372 OSAddAtomic(reusable, &vm_page_stats_reusable.reusable_count);
2373 vm_page_stats_reusable.reusable += reusable;
2374 reusable = 0;
2375 }
2376
2377 if (dw_count) {
2378 vm_page_do_delayed_work(object, VM_KERN_MEMORY_NONE, dwp: dwp_start, dw_count);
2379 dwp = dwp_start;
2380 dw_count = 0;
2381 }
2382
2383 if (dwp_start && dwp_finish_ctx) {
2384 vm_page_delayed_work_finish_ctx(dwp: dwp_start);
2385 dwp_start = dwp = NULL;
2386 }
2387}
2388
2389
2390/*
2391 * Deactive a "chunk" of the given range of the object starting at offset. A "chunk"
2392 * will always be less than or equal to the given size. The total range is divided up
2393 * into chunks for efficiency and performance related to the locks and handling the shadow
2394 * chain. This routine returns how much of the given "size" it actually processed. It's
2395 * up to the caler to loop and keep calling this routine until the entire range they want
2396 * to process has been done.
2397 * Iff clear_refmod is true, pmap_clear_refmod_options is called for each physical page in this range.
2398 */
2399
2400static vm_object_size_t
2401deactivate_a_chunk(
2402 vm_object_t orig_object,
2403 vm_object_offset_t offset,
2404 vm_object_size_t size,
2405 deactivate_flags_t flags,
2406 pmap_flush_context *pfc,
2407 struct pmap *pmap,
2408 vm_map_offset_t pmap_offset)
2409{
2410 vm_object_t object;
2411 vm_object_t tmp_object;
2412 vm_object_size_t length;
2413 chunk_state_t chunk_state;
2414
2415
2416 /*
2417 * Get set to do a chunk. We'll do up to CHUNK_SIZE, but no more than the
2418 * remaining size the caller asked for.
2419 */
2420
2421 length = MIN(size, CHUNK_SIZE);
2422
2423 /*
2424 * The chunk_state keeps track of which pages we've already processed if there's
2425 * a shadow chain on this object. At this point, we haven't done anything with this
2426 * range of pages yet, so initialize the state to indicate no pages processed yet.
2427 */
2428
2429 CHUNK_INIT(chunk_state, length);
2430 object = orig_object;
2431
2432 /*
2433 * Start at the top level object and iterate around the loop once for each object
2434 * in the shadow chain. We stop processing early if we've already found all the pages
2435 * in the range. Otherwise we stop when we run out of shadow objects.
2436 */
2437
2438 while (object && CHUNK_NOT_COMPLETE(chunk_state)) {
2439 vm_object_paging_begin(object);
2440
2441 deactivate_pages_in_object(object, offset, size: length, flags, chunk_state: &chunk_state, pfc, pmap, pmap_offset);
2442
2443 vm_object_paging_end(object);
2444
2445 /*
2446 * We've finished with this object, see if there's a shadow object. If
2447 * there is, update the offset and lock the new object. We also turn off
2448 * kill_page at this point since we only kill pages in the top most object.
2449 */
2450
2451 tmp_object = object->shadow;
2452
2453 if (tmp_object) {
2454 assert(!(flags & DEACTIVATE_KILL) || (flags & DEACTIVATE_CLEAR_REFMOD));
2455 flags &= ~(DEACTIVATE_KILL | DEACTIVATE_REUSABLE | DEACTIVATE_ALL_REUSABLE);
2456 offset += object->vo_shadow_offset;
2457 vm_object_lock(tmp_object);
2458 }
2459
2460 if (object != orig_object) {
2461 vm_object_unlock(object);
2462 }
2463
2464 object = tmp_object;
2465 }
2466
2467 if (object && object != orig_object) {
2468 vm_object_unlock(object);
2469 }
2470
2471 return length;
2472}
2473
2474
2475
2476/*
2477 * Move any resident pages in the specified range to the inactive queue. If kill_page is set,
2478 * we also clear the modified status of the page and "forget" any changes that have been made
2479 * to the page.
2480 */
2481
2482__private_extern__ void
2483vm_object_deactivate_pages(
2484 vm_object_t object,
2485 vm_object_offset_t offset,
2486 vm_object_size_t size,
2487 boolean_t kill_page,
2488 boolean_t reusable_page,
2489 boolean_t reusable_no_write,
2490 struct pmap *pmap,
2491 vm_map_offset_t pmap_offset)
2492{
2493 vm_object_size_t length;
2494 boolean_t all_reusable;
2495 pmap_flush_context pmap_flush_context_storage;
2496 unsigned int pmap_clear_refmod_mask = VM_MEM_REFERENCED;
2497 unsigned int pmap_clear_refmod_options = 0;
2498 deactivate_flags_t flags = DEACTIVATE_CLEAR_REFMOD;
2499 bool refmod_cleared = false;
2500 if (kill_page) {
2501 flags |= DEACTIVATE_KILL;
2502 }
2503 if (reusable_page) {
2504 flags |= DEACTIVATE_REUSABLE;
2505 }
2506 if (reusable_no_write) {
2507 flags |= DEACTIVATE_REUSABLE_NO_WRITE;
2508 }
2509
2510 /*
2511 * We break the range up into chunks and do one chunk at a time. This is for
2512 * efficiency and performance while handling the shadow chains and the locks.
2513 * The deactivate_a_chunk() function returns how much of the range it processed.
2514 * We keep calling this routine until the given size is exhausted.
2515 */
2516
2517
2518 all_reusable = FALSE;
2519#if 11
2520 /*
2521 * For the sake of accurate "reusable" pmap stats, we need
2522 * to tell pmap about each page that is no longer "reusable",
2523 * so we can't do the "all_reusable" optimization.
2524 *
2525 * If we do go with the all_reusable optimization, we can't
2526 * return if size is 0 since we could have "all_reusable == TRUE"
2527 * In this case, we save the overhead of doing the pmap_flush_context
2528 * work.
2529 */
2530 if (size == 0) {
2531 return;
2532 }
2533#else
2534 if (reusable_page &&
2535 object->internal &&
2536 object->vo_size != 0 &&
2537 object->vo_size == size &&
2538 object->reusable_page_count == 0) {
2539 all_reusable = TRUE;
2540 reusable_page = FALSE;
2541 flags |= DEACTIVATE_ALL_REUSABLE;
2542 }
2543#endif
2544
2545 if ((reusable_page || all_reusable) && object->all_reusable) {
2546 /* This means MADV_FREE_REUSABLE has been called twice, which
2547 * is probably illegal. */
2548 return;
2549 }
2550
2551
2552 pmap_flush_context_init(&pmap_flush_context_storage);
2553
2554 /*
2555 * If we're deactivating multiple pages, try to perform one bulk pmap operation.
2556 * We can't do this if we're killing pages and there's a shadow chain as
2557 * we don't yet know which pages are in the top object (pages in shadow copies aren't
2558 * safe to kill).
2559 * And we can only do this on hardware that supports it.
2560 */
2561 if (size > PAGE_SIZE && (!kill_page || !object->shadow)) {
2562 if (kill_page && object->internal) {
2563 pmap_clear_refmod_mask |= VM_MEM_MODIFIED;
2564 }
2565 if (reusable_page) {
2566 pmap_clear_refmod_options |= PMAP_OPTIONS_SET_REUSABLE;
2567 }
2568
2569 refmod_cleared = pmap_clear_refmod_range_options(pmap, start: pmap_offset, end: pmap_offset + size, mask: pmap_clear_refmod_mask, options: pmap_clear_refmod_options);
2570 if (refmod_cleared) {
2571 // We were able to clear all the refmod bits. So deactivate_a_chunk doesn't need to do it.
2572 flags &= ~DEACTIVATE_CLEAR_REFMOD;
2573 }
2574 }
2575
2576 while (size) {
2577 length = deactivate_a_chunk(orig_object: object, offset, size, flags,
2578 pfc: &pmap_flush_context_storage, pmap, pmap_offset);
2579
2580 size -= length;
2581 offset += length;
2582 pmap_offset += length;
2583 }
2584 pmap_flush(&pmap_flush_context_storage);
2585
2586 if (all_reusable) {
2587 if (!object->all_reusable) {
2588 unsigned int reusable;
2589
2590 object->all_reusable = TRUE;
2591 assert(object->reusable_page_count == 0);
2592 /* update global stats */
2593 reusable = object->resident_page_count;
2594 OSAddAtomic(reusable,
2595 &vm_page_stats_reusable.reusable_count);
2596 vm_page_stats_reusable.reusable += reusable;
2597 vm_page_stats_reusable.all_reusable_calls++;
2598 }
2599 } else if (reusable_page) {
2600 vm_page_stats_reusable.partial_reusable_calls++;
2601 }
2602}
2603
2604void
2605vm_object_reuse_pages(
2606 vm_object_t object,
2607 vm_object_offset_t start_offset,
2608 vm_object_offset_t end_offset,
2609 boolean_t allow_partial_reuse)
2610{
2611 vm_object_offset_t cur_offset;
2612 vm_page_t m;
2613 unsigned int reused, reusable;
2614
2615#define VM_OBJECT_REUSE_PAGE(object, m, reused) \
2616 MACRO_BEGIN \
2617 if ((m) != VM_PAGE_NULL && \
2618 (m)->vmp_reusable) { \
2619 assert((object)->reusable_page_count <= \
2620 (object)->resident_page_count); \
2621 assert((object)->reusable_page_count > 0); \
2622 (object)->reusable_page_count--; \
2623 (m)->vmp_reusable = FALSE; \
2624 (reused)++; \
2625 /* \
2626 * Tell pmap that this page is no longer \
2627 * "reusable", to update the "reusable" stats \
2628 * for all the pmaps that have mapped this \
2629 * page. \
2630 */ \
2631 pmap_clear_refmod_options(VM_PAGE_GET_PHYS_PAGE((m)), \
2632 0, /* refmod */ \
2633 (PMAP_OPTIONS_CLEAR_REUSABLE \
2634 | PMAP_OPTIONS_NOFLUSH), \
2635 NULL); \
2636 } \
2637 MACRO_END
2638
2639 reused = 0;
2640 reusable = 0;
2641
2642 vm_object_lock_assert_exclusive(object);
2643
2644 if (object->all_reusable) {
2645 panic("object %p all_reusable: can't update pmap stats",
2646 object);
2647 assert(object->reusable_page_count == 0);
2648 object->all_reusable = FALSE;
2649 if (end_offset - start_offset == object->vo_size ||
2650 !allow_partial_reuse) {
2651 vm_page_stats_reusable.all_reuse_calls++;
2652 reused = object->resident_page_count;
2653 } else {
2654 vm_page_stats_reusable.partial_reuse_calls++;
2655 vm_page_queue_iterate(&object->memq, m, vmp_listq) {
2656 if (m->vmp_offset < start_offset ||
2657 m->vmp_offset >= end_offset) {
2658 m->vmp_reusable = TRUE;
2659 object->reusable_page_count++;
2660 assert(object->resident_page_count >= object->reusable_page_count);
2661 continue;
2662 } else {
2663 assert(!m->vmp_reusable);
2664 reused++;
2665 }
2666 }
2667 }
2668 } else if (object->resident_page_count >
2669 ((end_offset - start_offset) >> PAGE_SHIFT)) {
2670 vm_page_stats_reusable.partial_reuse_calls++;
2671 for (cur_offset = start_offset;
2672 cur_offset < end_offset;
2673 cur_offset += PAGE_SIZE_64) {
2674 if (object->reusable_page_count == 0) {
2675 break;
2676 }
2677 m = vm_page_lookup(object, offset: cur_offset);
2678 VM_OBJECT_REUSE_PAGE(object, m, reused);
2679 }
2680 } else {
2681 vm_page_stats_reusable.partial_reuse_calls++;
2682 vm_page_queue_iterate(&object->memq, m, vmp_listq) {
2683 if (object->reusable_page_count == 0) {
2684 break;
2685 }
2686 if (m->vmp_offset < start_offset ||
2687 m->vmp_offset >= end_offset) {
2688 continue;
2689 }
2690 VM_OBJECT_REUSE_PAGE(object, m, reused);
2691 }
2692 }
2693
2694 /* update global stats */
2695 OSAddAtomic(reusable - reused, &vm_page_stats_reusable.reusable_count);
2696 vm_page_stats_reusable.reused += reused;
2697 vm_page_stats_reusable.reusable += reusable;
2698}
2699
2700/*
2701 * This function determines if the zero operation can be run on the
2702 * object. The checks on the entry have already been performed by
2703 * vm_map_zero_entry_preflight.
2704 */
2705static kern_return_t
2706vm_object_zero_preflight(
2707 vm_object_t object,
2708 vm_object_offset_t start,
2709 vm_object_offset_t end)
2710{
2711 /*
2712 * Zeroing is further restricted to anonymous memory.
2713 */
2714 if (!object->internal) {
2715 return KERN_PROTECTION_FAILURE;
2716 }
2717
2718 /*
2719 * Zeroing for copy on write isn't yet supported
2720 */
2721 if (object->shadow != NULL ||
2722 object->vo_copy != NULL) {
2723 return KERN_NO_ACCESS;
2724 }
2725
2726 /*
2727 * Ensure the that bounds makes sense wrt the object
2728 */
2729 if (end - start > object->vo_size) {
2730 return KERN_INVALID_ADDRESS;
2731 }
2732
2733 return KERN_SUCCESS;
2734}
2735
2736/*
2737 * This function looks up a page and waits if it is busy or being cleaned.
2738 * It returns false when the page found is busy and it needs to wait. Caller
2739 * of this function should restart the operation for the cur_offset when
2740 * this function returns false.
2741 */
2742static bool
2743vm_object_lookup_page_wait_busy(
2744 vm_object_t object,
2745 vm_object_offset_t cur_offset,
2746 vm_page_t *page)
2747{
2748 vm_page_t m;
2749
2750 m = vm_page_lookup(object, offset: cur_offset);
2751 if ((m == VM_PAGE_NULL) ||
2752 (!m->vmp_busy && !m->vmp_cleaning)) {
2753 *page = m;
2754 return true;
2755 }
2756
2757 PAGE_SLEEP(object, m, THREAD_UNINT);
2758 return false;
2759}
2760
2761static void
2762vm_object_zero_page(vm_page_t m)
2763{
2764 if (m != VM_PAGE_NULL) {
2765 ppnum_t phy_page_num = VM_PAGE_GET_PHYS_PAGE(m);
2766
2767 /*
2768 * Skip fictitious guard pages
2769 */
2770 if (m->vmp_fictitious) {
2771 assert(phy_page_num == vm_page_guard_addr);
2772 return;
2773 }
2774 pmap_zero_page(pn: phy_page_num);
2775 }
2776}
2777
2778/*
2779 * This function iterates the range of pages specified in the object and
2780 * discards the ones that are compressed and zeroes the ones that are wired.
2781 * This function may drop the object lock while waiting for a page that is
2782 * busy and will restart the operation for the specific offset.
2783 */
2784kern_return_t
2785vm_object_zero(
2786 vm_object_t object,
2787 vm_object_offset_t cur_offset,
2788 vm_object_offset_t end_offset)
2789{
2790 kern_return_t ret;
2791
2792 ret = vm_object_zero_preflight(object, start: cur_offset, end: end_offset);
2793 if (ret != KERN_SUCCESS) {
2794 return ret;
2795 }
2796
2797 while (cur_offset < end_offset) {
2798 vm_page_t m;
2799
2800 /*
2801 * If the compressor has the page then just discard it instead
2802 * of faulting it in and zeroing it else zero the page if it exists. If
2803 * we dropped the object lock during the lookup retry the lookup for the
2804 * cur_offset.
2805 */
2806 if (page_is_paged_out(object, offset: cur_offset)) {
2807 VM_COMPRESSOR_PAGER_STATE_CLR(object, cur_offset);
2808 } else if (vm_object_lookup_page_wait_busy(object, cur_offset, page: &m)) {
2809 vm_object_zero_page(m);
2810 } else {
2811 /*
2812 * If we dropped the lock then relookup the cur_offset in the object
2813 */
2814 ret = vm_object_zero_preflight(object, start: cur_offset, end: end_offset);
2815 if (ret != KERN_SUCCESS) {
2816 return ret;
2817 }
2818 continue;
2819 }
2820 cur_offset += PAGE_SIZE_64;
2821 /*
2822 * TODO: May need a vm_object_lock_yield_shared in this loop if it takes
2823 * too long, as holding the object lock for too long can stall pageout
2824 * scan (or other users of the object)
2825 */
2826 }
2827
2828 return KERN_SUCCESS;
2829}
2830
2831/*
2832 * Routine: vm_object_pmap_protect
2833 *
2834 * Purpose:
2835 * Reduces the permission for all physical
2836 * pages in the specified object range.
2837 *
2838 * If removing write permission only, it is
2839 * sufficient to protect only the pages in
2840 * the top-level object; only those pages may
2841 * have write permission.
2842 *
2843 * If removing all access, we must follow the
2844 * shadow chain from the top-level object to
2845 * remove access to all pages in shadowed objects.
2846 *
2847 * The object must *not* be locked. The object must
2848 * be internal.
2849 *
2850 * If pmap is not NULL, this routine assumes that
2851 * the only mappings for the pages are in that
2852 * pmap.
2853 */
2854
2855__private_extern__ void
2856vm_object_pmap_protect(
2857 vm_object_t object,
2858 vm_object_offset_t offset,
2859 vm_object_size_t size,
2860 pmap_t pmap,
2861 vm_map_size_t pmap_page_size,
2862 vm_map_offset_t pmap_start,
2863 vm_prot_t prot)
2864{
2865 vm_object_pmap_protect_options(object, offset, size, pmap,
2866 pmap_page_size,
2867 pmap_start, prot, options: 0);
2868}
2869
2870__private_extern__ void
2871vm_object_pmap_protect_options(
2872 vm_object_t object,
2873 vm_object_offset_t offset,
2874 vm_object_size_t size,
2875 pmap_t pmap,
2876 vm_map_size_t pmap_page_size,
2877 vm_map_offset_t pmap_start,
2878 vm_prot_t prot,
2879 int options)
2880{
2881 pmap_flush_context pmap_flush_context_storage;
2882 boolean_t delayed_pmap_flush = FALSE;
2883 vm_object_offset_t offset_in_object;
2884 vm_object_size_t size_in_object;
2885
2886 if (object == VM_OBJECT_NULL) {
2887 return;
2888 }
2889 if (pmap_page_size > PAGE_SIZE) {
2890 /* for 16K map on 4K device... */
2891 pmap_page_size = PAGE_SIZE;
2892 }
2893 /*
2894 * If we decide to work on the object itself, extend the range to
2895 * cover a full number of native pages.
2896 */
2897 size_in_object = vm_object_round_page(offset + size) - vm_object_trunc_page(offset);
2898 offset_in_object = vm_object_trunc_page(offset);
2899 /*
2900 * If we decide to work on the pmap, use the exact range specified,
2901 * so no rounding/truncating offset and size. They should already
2902 * be aligned to pmap_page_size.
2903 */
2904 assertf(!(offset & (pmap_page_size - 1)) && !(size & (pmap_page_size - 1)),
2905 "offset 0x%llx size 0x%llx pmap_page_size 0x%llx",
2906 offset, size, (uint64_t)pmap_page_size);
2907
2908 vm_object_lock(object);
2909
2910 if (object->phys_contiguous) {
2911 if (pmap != NULL) {
2912 vm_object_unlock(object);
2913 pmap_protect_options(map: pmap,
2914 s: pmap_start,
2915 e: pmap_start + size,
2916 prot,
2917 options: options & ~PMAP_OPTIONS_NOFLUSH,
2918 NULL);
2919 } else {
2920 vm_object_offset_t phys_start, phys_end, phys_addr;
2921
2922 phys_start = object->vo_shadow_offset + offset_in_object;
2923 phys_end = phys_start + size_in_object;
2924 assert(phys_start <= phys_end);
2925 assert(phys_end <= object->vo_shadow_offset + object->vo_size);
2926 vm_object_unlock(object);
2927
2928 pmap_flush_context_init(&pmap_flush_context_storage);
2929 delayed_pmap_flush = FALSE;
2930
2931 for (phys_addr = phys_start;
2932 phys_addr < phys_end;
2933 phys_addr += PAGE_SIZE_64) {
2934 pmap_page_protect_options(
2935 phys: (ppnum_t) (phys_addr >> PAGE_SHIFT),
2936 prot,
2937 options: options | PMAP_OPTIONS_NOFLUSH,
2938 arg: (void *)&pmap_flush_context_storage);
2939 delayed_pmap_flush = TRUE;
2940 }
2941 if (delayed_pmap_flush == TRUE) {
2942 pmap_flush(&pmap_flush_context_storage);
2943 }
2944 }
2945 return;
2946 }
2947
2948 assert(object->internal);
2949
2950 while (TRUE) {
2951 if (ptoa_64(object->resident_page_count) > size_in_object / 2 && pmap != PMAP_NULL) {
2952 vm_object_unlock(object);
2953 if (pmap_page_size < PAGE_SIZE) {
2954 DEBUG4K_PMAP("pmap %p start 0x%llx end 0x%llx prot 0x%x: pmap_protect()\n", pmap, (uint64_t)pmap_start, pmap_start + size, prot);
2955 }
2956 pmap_protect_options(map: pmap, s: pmap_start, e: pmap_start + size, prot,
2957 options: options & ~PMAP_OPTIONS_NOFLUSH, NULL);
2958 return;
2959 }
2960
2961 if (pmap_page_size < PAGE_SIZE) {
2962 DEBUG4K_PMAP("pmap %p start 0x%llx end 0x%llx prot 0x%x: offset 0x%llx size 0x%llx object %p offset 0x%llx size 0x%llx\n", pmap, (uint64_t)pmap_start, pmap_start + size, prot, offset, size, object, offset_in_object, size_in_object);
2963 }
2964
2965 pmap_flush_context_init(&pmap_flush_context_storage);
2966 delayed_pmap_flush = FALSE;
2967
2968 /*
2969 * if we are doing large ranges with respect to resident
2970 * page count then we should interate over pages otherwise
2971 * inverse page look-up will be faster
2972 */
2973 if (ptoa_64(object->resident_page_count / 4) < size_in_object) {
2974 vm_page_t p;
2975 vm_object_offset_t end;
2976
2977 end = offset_in_object + size_in_object;
2978
2979 vm_page_queue_iterate(&object->memq, p, vmp_listq) {
2980 if (!p->vmp_fictitious && (offset_in_object <= p->vmp_offset) && (p->vmp_offset < end)) {
2981 vm_map_offset_t start;
2982
2983 /*
2984 * XXX FBDP 4K: intentionally using "offset" here instead
2985 * of "offset_in_object", since "start" is a pmap address.
2986 */
2987 start = pmap_start + p->vmp_offset - offset;
2988
2989 if (pmap != PMAP_NULL) {
2990 vm_map_offset_t curr;
2991 for (curr = start;
2992 curr < start + PAGE_SIZE_64;
2993 curr += pmap_page_size) {
2994 if (curr < pmap_start) {
2995 continue;
2996 }
2997 if (curr >= pmap_start + size) {
2998 break;
2999 }
3000 pmap_protect_options(
3001 map: pmap,
3002 s: curr,
3003 e: curr + pmap_page_size,
3004 prot,
3005 options: options | PMAP_OPTIONS_NOFLUSH,
3006 arg: &pmap_flush_context_storage);
3007 }
3008 } else {
3009 pmap_page_protect_options(
3010 phys: VM_PAGE_GET_PHYS_PAGE(m: p),
3011 prot,
3012 options: options | PMAP_OPTIONS_NOFLUSH,
3013 arg: &pmap_flush_context_storage);
3014 }
3015 delayed_pmap_flush = TRUE;
3016 }
3017 }
3018 } else {
3019 vm_page_t p;
3020 vm_object_offset_t end;
3021 vm_object_offset_t target_off;
3022
3023 end = offset_in_object + size_in_object;
3024
3025 for (target_off = offset_in_object;
3026 target_off < end; target_off += PAGE_SIZE) {
3027 p = vm_page_lookup(object, offset: target_off);
3028
3029 if (p != VM_PAGE_NULL) {
3030 vm_object_offset_t start;
3031
3032 /*
3033 * XXX FBDP 4K: intentionally using "offset" here instead
3034 * of "offset_in_object", since "start" is a pmap address.
3035 */
3036 start = pmap_start + (p->vmp_offset - offset);
3037
3038 if (pmap != PMAP_NULL) {
3039 vm_map_offset_t curr;
3040 for (curr = start;
3041 curr < start + PAGE_SIZE;
3042 curr += pmap_page_size) {
3043 if (curr < pmap_start) {
3044 continue;
3045 }
3046 if (curr >= pmap_start + size) {
3047 break;
3048 }
3049 pmap_protect_options(
3050 map: pmap,
3051 s: curr,
3052 e: curr + pmap_page_size,
3053 prot,
3054 options: options | PMAP_OPTIONS_NOFLUSH,
3055 arg: &pmap_flush_context_storage);
3056 }
3057 } else {
3058 pmap_page_protect_options(
3059 phys: VM_PAGE_GET_PHYS_PAGE(m: p),
3060 prot,
3061 options: options | PMAP_OPTIONS_NOFLUSH,
3062 arg: &pmap_flush_context_storage);
3063 }
3064 delayed_pmap_flush = TRUE;
3065 }
3066 }
3067 }
3068 if (delayed_pmap_flush == TRUE) {
3069 pmap_flush(&pmap_flush_context_storage);
3070 }
3071
3072 if (prot == VM_PROT_NONE) {
3073 /*
3074 * Must follow shadow chain to remove access
3075 * to pages in shadowed objects.
3076 */
3077 vm_object_t next_object;
3078
3079 next_object = object->shadow;
3080 if (next_object != VM_OBJECT_NULL) {
3081 offset_in_object += object->vo_shadow_offset;
3082 offset += object->vo_shadow_offset;
3083 vm_object_lock(next_object);
3084 vm_object_unlock(object);
3085 object = next_object;
3086 } else {
3087 /*
3088 * End of chain - we are done.
3089 */
3090 break;
3091 }
3092 } else {
3093 /*
3094 * Pages in shadowed objects may never have
3095 * write permission - we may stop here.
3096 */
3097 break;
3098 }
3099 }
3100
3101 vm_object_unlock(object);
3102}
3103
3104uint32_t vm_page_busy_absent_skipped = 0;
3105
3106/*
3107 * Routine: vm_object_copy_slowly
3108 *
3109 * Description:
3110 * Copy the specified range of the source
3111 * virtual memory object without using
3112 * protection-based optimizations (such
3113 * as copy-on-write). The pages in the
3114 * region are actually copied.
3115 *
3116 * In/out conditions:
3117 * The caller must hold a reference and a lock
3118 * for the source virtual memory object. The source
3119 * object will be returned *unlocked*.
3120 *
3121 * Results:
3122 * If the copy is completed successfully, KERN_SUCCESS is
3123 * returned. If the caller asserted the interruptible
3124 * argument, and an interruption occurred while waiting
3125 * for a user-generated event, MACH_SEND_INTERRUPTED is
3126 * returned. Other values may be returned to indicate
3127 * hard errors during the copy operation.
3128 *
3129 * A new virtual memory object is returned in a
3130 * parameter (_result_object). The contents of this
3131 * new object, starting at a zero offset, are a copy
3132 * of the source memory region. In the event of
3133 * an error, this parameter will contain the value
3134 * VM_OBJECT_NULL.
3135 */
3136__private_extern__ kern_return_t
3137vm_object_copy_slowly(
3138 vm_object_t src_object,
3139 vm_object_offset_t src_offset,
3140 vm_object_size_t size,
3141 boolean_t interruptible,
3142 vm_object_t *_result_object) /* OUT */
3143{
3144 vm_object_t new_object;
3145 vm_object_offset_t new_offset;
3146
3147 struct vm_object_fault_info fault_info = {};
3148
3149 if (size == 0) {
3150 vm_object_unlock(src_object);
3151 *_result_object = VM_OBJECT_NULL;
3152 return KERN_INVALID_ARGUMENT;
3153 }
3154
3155 /*
3156 * Prevent destruction of the source object while we copy.
3157 */
3158
3159 vm_object_reference_locked(src_object);
3160 vm_object_unlock(src_object);
3161
3162 /*
3163 * Create a new object to hold the copied pages.
3164 * A few notes:
3165 * We fill the new object starting at offset 0,
3166 * regardless of the input offset.
3167 * We don't bother to lock the new object within
3168 * this routine, since we have the only reference.
3169 */
3170
3171 size = vm_object_round_page(src_offset + size) - vm_object_trunc_page(src_offset);
3172 src_offset = vm_object_trunc_page(src_offset);
3173 new_object = vm_object_allocate(size);
3174 new_offset = 0;
3175
3176 assert(size == trunc_page_64(size)); /* Will the loop terminate? */
3177
3178 fault_info.interruptible = interruptible;
3179 fault_info.behavior = VM_BEHAVIOR_SEQUENTIAL;
3180 fault_info.lo_offset = src_offset;
3181 fault_info.hi_offset = src_offset + size;
3182 fault_info.stealth = TRUE;
3183
3184 for (;
3185 size != 0;
3186 src_offset += PAGE_SIZE_64,
3187 new_offset += PAGE_SIZE_64, size -= PAGE_SIZE_64
3188 ) {
3189 vm_page_t new_page;
3190 vm_fault_return_t result;
3191
3192 vm_object_lock(new_object);
3193
3194 while ((new_page = vm_page_alloc(object: new_object, offset: new_offset))
3195 == VM_PAGE_NULL) {
3196 vm_object_unlock(new_object);
3197
3198 if (!vm_page_wait(interruptible)) {
3199 vm_object_deallocate(object: new_object);
3200 vm_object_deallocate(object: src_object);
3201 *_result_object = VM_OBJECT_NULL;
3202 return MACH_SEND_INTERRUPTED;
3203 }
3204 vm_object_lock(new_object);
3205 }
3206 vm_object_unlock(new_object);
3207
3208 do {
3209 vm_prot_t prot = VM_PROT_READ;
3210 vm_page_t _result_page;
3211 vm_page_t top_page;
3212 vm_page_t result_page;
3213 kern_return_t error_code;
3214 vm_object_t result_page_object;
3215
3216
3217 vm_object_lock(src_object);
3218
3219 if (src_object->internal &&
3220 src_object->shadow == VM_OBJECT_NULL &&
3221 (src_object->pager == NULL ||
3222 (VM_COMPRESSOR_PAGER_STATE_GET(src_object,
3223 src_offset) ==
3224 VM_EXTERNAL_STATE_ABSENT))) {
3225 boolean_t can_skip_page;
3226
3227 _result_page = vm_page_lookup(object: src_object,
3228 offset: src_offset);
3229 if (_result_page == VM_PAGE_NULL) {
3230 /*
3231 * This page is neither resident nor
3232 * compressed and there's no shadow
3233 * object below "src_object", so this
3234 * page is really missing.
3235 * There's no need to zero-fill it just
3236 * to copy it: let's leave it missing
3237 * in "new_object" and get zero-filled
3238 * on demand.
3239 */
3240 can_skip_page = TRUE;
3241 } else if (workaround_41447923 &&
3242 src_object->pager == NULL &&
3243 _result_page != VM_PAGE_NULL &&
3244 _result_page->vmp_busy &&
3245 _result_page->vmp_absent &&
3246 src_object->purgable == VM_PURGABLE_DENY &&
3247 !src_object->blocked_access) {
3248 /*
3249 * This page is "busy" and "absent"
3250 * but not because we're waiting for
3251 * it to be decompressed. It must
3252 * be because it's a "no zero fill"
3253 * page that is currently not
3254 * accessible until it gets overwritten
3255 * by a device driver.
3256 * Since its initial state would have
3257 * been "zero-filled", let's leave the
3258 * copy page missing and get zero-filled
3259 * on demand.
3260 */
3261 assert(src_object->internal);
3262 assert(src_object->shadow == NULL);
3263 assert(src_object->pager == NULL);
3264 can_skip_page = TRUE;
3265 vm_page_busy_absent_skipped++;
3266 } else {
3267 can_skip_page = FALSE;
3268 }
3269 if (can_skip_page) {
3270 vm_object_unlock(src_object);
3271 /* free the unused "new_page"... */
3272 vm_object_lock(new_object);
3273 VM_PAGE_FREE(new_page);
3274 new_page = VM_PAGE_NULL;
3275 vm_object_unlock(new_object);
3276 /* ...and go to next page in "src_object" */
3277 result = VM_FAULT_SUCCESS;
3278 break;
3279 }
3280 }
3281
3282 vm_object_paging_begin(src_object);
3283
3284 /* cap size at maximum UPL size */
3285 upl_size_t cluster_size;
3286 if (os_convert_overflow(size, &cluster_size)) {
3287 cluster_size = 0 - (upl_size_t)PAGE_SIZE;
3288 }
3289 fault_info.cluster_size = cluster_size;
3290
3291 _result_page = VM_PAGE_NULL;
3292 result = vm_fault_page(first_object: src_object, first_offset: src_offset,
3293 VM_PROT_READ, FALSE,
3294 FALSE, /* page not looked up */
3295 protection: &prot, result_page: &_result_page, top_page: &top_page,
3296 type_of_fault: (int *)0,
3297 error_code: &error_code, FALSE, fault_info: &fault_info);
3298
3299 switch (result) {
3300 case VM_FAULT_SUCCESS:
3301 result_page = _result_page;
3302 result_page_object = VM_PAGE_OBJECT(result_page);
3303
3304 /*
3305 * Copy the page to the new object.
3306 *
3307 * POLICY DECISION:
3308 * If result_page is clean,
3309 * we could steal it instead
3310 * of copying.
3311 */
3312
3313 vm_page_copy(src_page: result_page, dest_page: new_page);
3314 vm_object_unlock(result_page_object);
3315
3316 /*
3317 * Let go of both pages (make them
3318 * not busy, perform wakeup, activate).
3319 */
3320 vm_object_lock(new_object);
3321 SET_PAGE_DIRTY(new_page, FALSE);
3322 PAGE_WAKEUP_DONE(new_page);
3323 vm_object_unlock(new_object);
3324
3325 vm_object_lock(result_page_object);
3326 PAGE_WAKEUP_DONE(result_page);
3327
3328 vm_page_lockspin_queues();
3329 if ((result_page->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q) ||
3330 (result_page->vmp_q_state == VM_PAGE_NOT_ON_Q)) {
3331 vm_page_activate(page: result_page);
3332 }
3333 vm_page_activate(page: new_page);
3334 vm_page_unlock_queues();
3335
3336 /*
3337 * Release paging references and
3338 * top-level placeholder page, if any.
3339 */
3340
3341 vm_fault_cleanup(object: result_page_object,
3342 top_page);
3343
3344 break;
3345
3346 case VM_FAULT_RETRY:
3347 break;
3348
3349 case VM_FAULT_MEMORY_SHORTAGE:
3350 if (vm_page_wait(interruptible)) {
3351 break;
3352 }
3353 ktriage_record(thread_id: thread_tid(thread: current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_VM, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_VM_FAULT_OBJCOPYSLOWLY_MEMORY_SHORTAGE), arg: 0 /* arg */);
3354 OS_FALLTHROUGH;
3355
3356 case VM_FAULT_INTERRUPTED:
3357 vm_object_lock(new_object);
3358 VM_PAGE_FREE(new_page);
3359 vm_object_unlock(new_object);
3360
3361 vm_object_deallocate(object: new_object);
3362 vm_object_deallocate(object: src_object);
3363 *_result_object = VM_OBJECT_NULL;
3364 return MACH_SEND_INTERRUPTED;
3365
3366 case VM_FAULT_SUCCESS_NO_VM_PAGE:
3367 /* success but no VM page: fail */
3368 vm_object_paging_end(src_object);
3369 vm_object_unlock(src_object);
3370 OS_FALLTHROUGH;
3371 case VM_FAULT_MEMORY_ERROR:
3372 /*
3373 * A policy choice:
3374 * (a) ignore pages that we can't
3375 * copy
3376 * (b) return the null object if
3377 * any page fails [chosen]
3378 */
3379
3380 vm_object_lock(new_object);
3381 VM_PAGE_FREE(new_page);
3382 vm_object_unlock(new_object);
3383
3384 vm_object_deallocate(object: new_object);
3385 vm_object_deallocate(object: src_object);
3386 *_result_object = VM_OBJECT_NULL;
3387 return error_code ? error_code:
3388 KERN_MEMORY_ERROR;
3389
3390 default:
3391 panic("vm_object_copy_slowly: unexpected error"
3392 " 0x%x from vm_fault_page()\n", result);
3393 }
3394 } while (result != VM_FAULT_SUCCESS);
3395 }
3396
3397 /*
3398 * Lose the extra reference, and return our object.
3399 */
3400 vm_object_deallocate(object: src_object);
3401 *_result_object = new_object;
3402 return KERN_SUCCESS;
3403}
3404
3405/*
3406 * Routine: vm_object_copy_quickly
3407 *
3408 * Purpose:
3409 * Copy the specified range of the source virtual
3410 * memory object, if it can be done without waiting
3411 * for user-generated events.
3412 *
3413 * Results:
3414 * If the copy is successful, the copy is returned in
3415 * the arguments; otherwise, the arguments are not
3416 * affected.
3417 *
3418 * In/out conditions:
3419 * The object should be unlocked on entry and exit.
3420 */
3421
3422/*ARGSUSED*/
3423__private_extern__ boolean_t
3424vm_object_copy_quickly(
3425 vm_object_t object, /* IN */
3426 __unused vm_object_offset_t offset, /* IN */
3427 __unused vm_object_size_t size, /* IN */
3428 boolean_t *_src_needs_copy, /* OUT */
3429 boolean_t *_dst_needs_copy) /* OUT */
3430{
3431 memory_object_copy_strategy_t copy_strategy;
3432
3433 if (object == VM_OBJECT_NULL) {
3434 *_src_needs_copy = FALSE;
3435 *_dst_needs_copy = FALSE;
3436 return TRUE;
3437 }
3438
3439 vm_object_lock(object);
3440
3441 copy_strategy = object->copy_strategy;
3442
3443 switch (copy_strategy) {
3444 case MEMORY_OBJECT_COPY_SYMMETRIC:
3445
3446 /*
3447 * Symmetric copy strategy.
3448 * Make another reference to the object.
3449 * Leave object/offset unchanged.
3450 */
3451
3452 vm_object_reference_locked(object);
3453 VM_OBJECT_SET_SHADOWED(object, TRUE);
3454 vm_object_unlock(object);
3455
3456 /*
3457 * Both source and destination must make
3458 * shadows, and the source must be made
3459 * read-only if not already.
3460 */
3461
3462 *_src_needs_copy = TRUE;
3463 *_dst_needs_copy = TRUE;
3464
3465 break;
3466
3467 case MEMORY_OBJECT_COPY_DELAY:
3468 vm_object_unlock(object);
3469 return FALSE;
3470
3471 default:
3472 vm_object_unlock(object);
3473 return FALSE;
3474 }
3475 return TRUE;
3476}
3477
3478static uint32_t copy_delayed_lock_collisions;
3479static uint32_t copy_delayed_max_collisions;
3480static uint32_t copy_delayed_lock_contention;
3481static uint32_t copy_delayed_protect_iterate;
3482
3483/*
3484 * Routine: vm_object_copy_delayed [internal]
3485 *
3486 * Description:
3487 * Copy the specified virtual memory object, using
3488 * the asymmetric copy-on-write algorithm.
3489 *
3490 * In/out conditions:
3491 * The src_object must be locked on entry. It will be unlocked
3492 * on exit - so the caller must also hold a reference to it.
3493 *
3494 * This routine will not block waiting for user-generated
3495 * events. It is not interruptible.
3496 */
3497__private_extern__ vm_object_t
3498vm_object_copy_delayed(
3499 vm_object_t src_object,
3500 vm_object_offset_t src_offset,
3501 vm_object_size_t size,
3502 boolean_t src_object_shared)
3503{
3504 vm_object_t new_copy = VM_OBJECT_NULL;
3505 vm_object_t old_copy;
3506 vm_page_t p;
3507 vm_object_size_t copy_size = src_offset + size;
3508 pmap_flush_context pmap_flush_context_storage;
3509 boolean_t delayed_pmap_flush = FALSE;
3510
3511
3512 uint32_t collisions = 0;
3513 /*
3514 * The user-level memory manager wants to see all of the changes
3515 * to this object, but it has promised not to make any changes on
3516 * its own.
3517 *
3518 * Perform an asymmetric copy-on-write, as follows:
3519 * Create a new object, called a "copy object" to hold
3520 * pages modified by the new mapping (i.e., the copy,
3521 * not the original mapping).
3522 * Record the original object as the backing object for
3523 * the copy object. If the original mapping does not
3524 * change a page, it may be used read-only by the copy.
3525 * Record the copy object in the original object.
3526 * When the original mapping causes a page to be modified,
3527 * it must be copied to a new page that is "pushed" to
3528 * the copy object.
3529 * Mark the new mapping (the copy object) copy-on-write.
3530 * This makes the copy object itself read-only, allowing
3531 * it to be reused if the original mapping makes no
3532 * changes, and simplifying the synchronization required
3533 * in the "push" operation described above.
3534 *
3535 * The copy-on-write is said to be assymetric because the original
3536 * object is *not* marked copy-on-write. A copied page is pushed
3537 * to the copy object, regardless which party attempted to modify
3538 * the page.
3539 *
3540 * Repeated asymmetric copy operations may be done. If the
3541 * original object has not been changed since the last copy, its
3542 * copy object can be reused. Otherwise, a new copy object can be
3543 * inserted between the original object and its previous copy
3544 * object. Since any copy object is read-only, this cannot affect
3545 * affect the contents of the previous copy object.
3546 *
3547 * Note that a copy object is higher in the object tree than the
3548 * original object; therefore, use of the copy object recorded in
3549 * the original object must be done carefully, to avoid deadlock.
3550 */
3551
3552 copy_size = vm_object_round_page(copy_size);
3553Retry:
3554
3555 /*
3556 * Wait for paging in progress.
3557 */
3558 if (!src_object->true_share &&
3559 (src_object->paging_in_progress != 0 ||
3560 src_object->activity_in_progress != 0)) {
3561 if (src_object_shared == TRUE) {
3562 vm_object_unlock(src_object);
3563 vm_object_lock(src_object);
3564 src_object_shared = FALSE;
3565 goto Retry;
3566 }
3567 vm_object_paging_wait(src_object, THREAD_UNINT);
3568 }
3569 /*
3570 * See whether we can reuse the result of a previous
3571 * copy operation.
3572 */
3573
3574 old_copy = src_object->vo_copy;
3575 if (old_copy != VM_OBJECT_NULL) {
3576 int lock_granted;
3577
3578 /*
3579 * Try to get the locks (out of order)
3580 */
3581 if (src_object_shared == TRUE) {
3582 lock_granted = vm_object_lock_try_shared(old_copy);
3583 } else {
3584 lock_granted = vm_object_lock_try(old_copy);
3585 }
3586
3587 if (!lock_granted) {
3588 vm_object_unlock(src_object);
3589
3590 if (collisions++ == 0) {
3591 copy_delayed_lock_contention++;
3592 }
3593 mutex_pause(collisions);
3594
3595 /* Heisenberg Rules */
3596 copy_delayed_lock_collisions++;
3597
3598 if (collisions > copy_delayed_max_collisions) {
3599 copy_delayed_max_collisions = collisions;
3600 }
3601
3602 if (src_object_shared == TRUE) {
3603 vm_object_lock_shared(src_object);
3604 } else {
3605 vm_object_lock(src_object);
3606 }
3607
3608 goto Retry;
3609 }
3610
3611 /*
3612 * Determine whether the old copy object has
3613 * been modified.
3614 */
3615
3616 if (old_copy->resident_page_count == 0 &&
3617 !old_copy->pager_created) {
3618 /*
3619 * It has not been modified.
3620 *
3621 * Return another reference to
3622 * the existing copy-object if
3623 * we can safely grow it (if
3624 * needed).
3625 */
3626
3627 if (old_copy->vo_size < copy_size) {
3628 if (src_object_shared == TRUE) {
3629 vm_object_unlock(old_copy);
3630 vm_object_unlock(src_object);
3631
3632 vm_object_lock(src_object);
3633 src_object_shared = FALSE;
3634 goto Retry;
3635 }
3636 /*
3637 * We can't perform a delayed copy if any of the
3638 * pages in the extended range are wired (because
3639 * we can't safely take write permission away from
3640 * wired pages). If the pages aren't wired, then
3641 * go ahead and protect them.
3642 */
3643 copy_delayed_protect_iterate++;
3644
3645 pmap_flush_context_init(&pmap_flush_context_storage);
3646 delayed_pmap_flush = FALSE;
3647
3648 vm_page_queue_iterate(&src_object->memq, p, vmp_listq) {
3649 if (!p->vmp_fictitious &&
3650 p->vmp_offset >= old_copy->vo_size &&
3651 p->vmp_offset < copy_size) {
3652 if (VM_PAGE_WIRED(p)) {
3653 vm_object_unlock(old_copy);
3654 vm_object_unlock(src_object);
3655
3656 if (new_copy != VM_OBJECT_NULL) {
3657 vm_object_unlock(new_copy);
3658 vm_object_deallocate(object: new_copy);
3659 }
3660 if (delayed_pmap_flush == TRUE) {
3661 pmap_flush(&pmap_flush_context_storage);
3662 }
3663
3664 return VM_OBJECT_NULL;
3665 } else {
3666 pmap_page_protect_options(phys: VM_PAGE_GET_PHYS_PAGE(m: p),
3667 prot: (p->vmp_xpmapped ? (VM_PROT_READ | VM_PROT_EXECUTE) : VM_PROT_READ),
3668 PMAP_OPTIONS_NOFLUSH, arg: (void *)&pmap_flush_context_storage);
3669 delayed_pmap_flush = TRUE;
3670 }
3671 }
3672 }
3673 if (delayed_pmap_flush == TRUE) {
3674 pmap_flush(&pmap_flush_context_storage);
3675 }
3676
3677 assertf(page_aligned(copy_size),
3678 "object %p size 0x%llx",
3679 old_copy, (uint64_t)copy_size);
3680 old_copy->vo_size = copy_size;
3681
3682 /*
3683 * src_object's "vo_copy" object now covers
3684 * a larger portion of src_object.
3685 * Increment src_object's "vo_copy_version"
3686 * to make any racing vm_fault() on
3687 * "src_object" re-check if it needs to honor
3688 * any new copy-on-write obligation.
3689 */
3690 src_object->vo_copy_version++;
3691 }
3692 if (src_object_shared == TRUE) {
3693 vm_object_reference_shared(old_copy);
3694 } else {
3695 vm_object_reference_locked(old_copy);
3696 }
3697 vm_object_unlock(old_copy);
3698 vm_object_unlock(src_object);
3699
3700 if (new_copy != VM_OBJECT_NULL) {
3701 vm_object_unlock(new_copy);
3702 vm_object_deallocate(object: new_copy);
3703 }
3704 return old_copy;
3705 }
3706
3707
3708
3709 /*
3710 * Adjust the size argument so that the newly-created
3711 * copy object will be large enough to back either the
3712 * old copy object or the new mapping.
3713 */
3714 if (old_copy->vo_size > copy_size) {
3715 copy_size = old_copy->vo_size;
3716 }
3717
3718 if (new_copy == VM_OBJECT_NULL) {
3719 vm_object_unlock(old_copy);
3720 vm_object_unlock(src_object);
3721 new_copy = vm_object_allocate(size: copy_size);
3722 vm_object_lock(src_object);
3723 vm_object_lock(new_copy);
3724
3725 src_object_shared = FALSE;
3726 goto Retry;
3727 }
3728 assertf(page_aligned(copy_size),
3729 "object %p size 0x%llx",
3730 new_copy, (uint64_t)copy_size);
3731 new_copy->vo_size = copy_size;
3732
3733 /*
3734 * The copy-object is always made large enough to
3735 * completely shadow the original object, since
3736 * it may have several users who want to shadow
3737 * the original object at different points.
3738 */
3739
3740 assert((old_copy->shadow == src_object) &&
3741 (old_copy->vo_shadow_offset == (vm_object_offset_t) 0));
3742 } else if (new_copy == VM_OBJECT_NULL) {
3743 vm_object_unlock(src_object);
3744 new_copy = vm_object_allocate(size: copy_size);
3745 vm_object_lock(src_object);
3746 vm_object_lock(new_copy);
3747
3748 src_object_shared = FALSE;
3749 goto Retry;
3750 }
3751
3752 /*
3753 * We now have the src object locked, and the new copy object
3754 * allocated and locked (and potentially the old copy locked).
3755 * Before we go any further, make sure we can still perform
3756 * a delayed copy, as the situation may have changed.
3757 *
3758 * Specifically, we can't perform a delayed copy if any of the
3759 * pages in the range are wired (because we can't safely take
3760 * write permission away from wired pages). If the pages aren't
3761 * wired, then go ahead and protect them.
3762 */
3763 copy_delayed_protect_iterate++;
3764
3765 pmap_flush_context_init(&pmap_flush_context_storage);
3766 delayed_pmap_flush = FALSE;
3767
3768 vm_page_queue_iterate(&src_object->memq, p, vmp_listq) {
3769 if (!p->vmp_fictitious && p->vmp_offset < copy_size) {
3770 if (VM_PAGE_WIRED(p)) {
3771 if (old_copy) {
3772 vm_object_unlock(old_copy);
3773 }
3774 vm_object_unlock(src_object);
3775 vm_object_unlock(new_copy);
3776 vm_object_deallocate(object: new_copy);
3777
3778 if (delayed_pmap_flush == TRUE) {
3779 pmap_flush(&pmap_flush_context_storage);
3780 }
3781
3782 return VM_OBJECT_NULL;
3783 } else {
3784 pmap_page_protect_options(phys: VM_PAGE_GET_PHYS_PAGE(m: p),
3785 prot: (p->vmp_xpmapped ? (VM_PROT_READ | VM_PROT_EXECUTE) : VM_PROT_READ),
3786 PMAP_OPTIONS_NOFLUSH, arg: (void *)&pmap_flush_context_storage);
3787 delayed_pmap_flush = TRUE;
3788 }
3789 }
3790 }
3791 if (delayed_pmap_flush == TRUE) {
3792 pmap_flush(&pmap_flush_context_storage);
3793 }
3794
3795 if (old_copy != VM_OBJECT_NULL) {
3796 /*
3797 * Make the old copy-object shadow the new one.
3798 * It will receive no more pages from the original
3799 * object.
3800 */
3801
3802 /* remove ref. from old_copy */
3803 vm_object_lock_assert_exclusive(src_object);
3804 src_object->ref_count--;
3805 assert(src_object->ref_count > 0);
3806 vm_object_lock_assert_exclusive(old_copy);
3807 old_copy->shadow = new_copy;
3808 vm_object_lock_assert_exclusive(new_copy);
3809 assert(new_copy->ref_count > 0);
3810 new_copy->ref_count++; /* for old_copy->shadow ref. */
3811
3812 vm_object_unlock(old_copy); /* done with old_copy */
3813 }
3814
3815 /*
3816 * Point the new copy at the existing object.
3817 */
3818 vm_object_lock_assert_exclusive(new_copy);
3819 new_copy->shadow = src_object;
3820 new_copy->vo_shadow_offset = 0;
3821 VM_OBJECT_SET_SHADOWED(object: new_copy, TRUE); /* caller must set needs_copy */
3822
3823 vm_object_lock_assert_exclusive(src_object);
3824 vm_object_reference_locked(src_object);
3825 VM_OBJECT_COPY_SET(object: src_object, copy: new_copy);
3826 vm_object_unlock(src_object);
3827 vm_object_unlock(new_copy);
3828
3829 return new_copy;
3830}
3831
3832/*
3833 * Routine: vm_object_copy_strategically
3834 *
3835 * Purpose:
3836 * Perform a copy according to the source object's
3837 * declared strategy. This operation may block,
3838 * and may be interrupted.
3839 */
3840__private_extern__ kern_return_t
3841vm_object_copy_strategically(
3842 vm_object_t src_object,
3843 vm_object_offset_t src_offset,
3844 vm_object_size_t size,
3845 bool forking,
3846 vm_object_t *dst_object, /* OUT */
3847 vm_object_offset_t *dst_offset, /* OUT */
3848 boolean_t *dst_needs_copy) /* OUT */
3849{
3850 boolean_t result;
3851 boolean_t interruptible = THREAD_ABORTSAFE; /* XXX */
3852 boolean_t object_lock_shared = FALSE;
3853 memory_object_copy_strategy_t copy_strategy;
3854
3855 assert(src_object != VM_OBJECT_NULL);
3856
3857 copy_strategy = src_object->copy_strategy;
3858
3859 if (copy_strategy == MEMORY_OBJECT_COPY_DELAY) {
3860 vm_object_lock_shared(src_object);
3861 object_lock_shared = TRUE;
3862 } else {
3863 vm_object_lock(src_object);
3864 }
3865
3866 /*
3867 * The copy strategy is only valid if the memory manager
3868 * is "ready". Internal objects are always ready.
3869 */
3870
3871 while (!src_object->internal && !src_object->pager_ready) {
3872 wait_result_t wait_result;
3873
3874 if (object_lock_shared == TRUE) {
3875 vm_object_unlock(src_object);
3876 vm_object_lock(src_object);
3877 object_lock_shared = FALSE;
3878 continue;
3879 }
3880 wait_result = vm_object_sleep( object: src_object,
3881 VM_OBJECT_EVENT_PAGER_READY,
3882 interruptible);
3883 if (wait_result != THREAD_AWAKENED) {
3884 vm_object_unlock(src_object);
3885 *dst_object = VM_OBJECT_NULL;
3886 *dst_offset = 0;
3887 *dst_needs_copy = FALSE;
3888 return MACH_SEND_INTERRUPTED;
3889 }
3890 }
3891
3892 /*
3893 * Use the appropriate copy strategy.
3894 */
3895
3896 if (copy_strategy == MEMORY_OBJECT_COPY_DELAY_FORK) {
3897 if (forking) {
3898 copy_strategy = MEMORY_OBJECT_COPY_DELAY;
3899 } else {
3900 copy_strategy = MEMORY_OBJECT_COPY_NONE;
3901 if (object_lock_shared) {
3902 vm_object_unlock(src_object);
3903 vm_object_lock(src_object);
3904 object_lock_shared = FALSE;
3905 }
3906 }
3907 }
3908
3909 switch (copy_strategy) {
3910 case MEMORY_OBJECT_COPY_DELAY:
3911 *dst_object = vm_object_copy_delayed(src_object,
3912 src_offset, size, src_object_shared: object_lock_shared);
3913 if (*dst_object != VM_OBJECT_NULL) {
3914 *dst_offset = src_offset;
3915 *dst_needs_copy = TRUE;
3916 result = KERN_SUCCESS;
3917 break;
3918 }
3919 vm_object_lock(src_object);
3920 OS_FALLTHROUGH; /* fall thru when delayed copy not allowed */
3921
3922 case MEMORY_OBJECT_COPY_NONE:
3923 result = vm_object_copy_slowly(src_object, src_offset, size,
3924 interruptible, result_object: dst_object);
3925 if (result == KERN_SUCCESS) {
3926 *dst_offset = src_offset - vm_object_trunc_page(src_offset);
3927 *dst_needs_copy = FALSE;
3928 }
3929 break;
3930
3931 case MEMORY_OBJECT_COPY_SYMMETRIC:
3932 vm_object_unlock(src_object);
3933 result = KERN_MEMORY_RESTART_COPY;
3934 break;
3935
3936 default:
3937 panic("copy_strategically: bad strategy %d for object %p",
3938 copy_strategy, src_object);
3939 result = KERN_INVALID_ARGUMENT;
3940 }
3941 return result;
3942}
3943
3944/*
3945 * vm_object_shadow:
3946 *
3947 * Create a new object which is backed by the
3948 * specified existing object range. The source
3949 * object reference is deallocated.
3950 *
3951 * The new object and offset into that object
3952 * are returned in the source parameters.
3953 */
3954boolean_t vm_object_shadow_check = TRUE;
3955uint64_t vm_object_shadow_forced = 0;
3956uint64_t vm_object_shadow_skipped = 0;
3957
3958__private_extern__ boolean_t
3959vm_object_shadow(
3960 vm_object_t *object, /* IN/OUT */
3961 vm_object_offset_t *offset, /* IN/OUT */
3962 vm_object_size_t length,
3963 boolean_t always_shadow)
3964{
3965 vm_object_t source;
3966 vm_object_t result;
3967
3968 source = *object;
3969 assert(source != VM_OBJECT_NULL);
3970 if (source == VM_OBJECT_NULL) {
3971 return FALSE;
3972 }
3973
3974 assert(source->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC);
3975
3976 /*
3977 * Determine if we really need a shadow.
3978 *
3979 * If the source object is larger than what we are trying
3980 * to create, then force the shadow creation even if the
3981 * ref count is 1. This will allow us to [potentially]
3982 * collapse the underlying object away in the future
3983 * (freeing up the extra data it might contain and that
3984 * we don't need).
3985 */
3986
3987 assert(source->copy_strategy != MEMORY_OBJECT_COPY_NONE); /* Purgeable objects shouldn't have shadow objects. */
3988
3989 /*
3990 * The following optimization does not work in the context of submaps
3991 * (the shared region, in particular).
3992 * This object might have only 1 reference (in the submap) but that
3993 * submap can itself be mapped multiple times, so the object is
3994 * actually indirectly referenced more than once...
3995 * The caller can specify to "always_shadow" to bypass the optimization.
3996 */
3997 if (vm_object_shadow_check &&
3998 source->vo_size == length &&
3999 source->ref_count == 1) {
4000 if (always_shadow) {
4001 vm_object_shadow_forced++;
4002 } else {
4003 /*
4004 * Lock the object and check again.
4005 * We also check to see if there's
4006 * a shadow or copy object involved.
4007 * We can't do that earlier because
4008 * without the object locked, there
4009 * could be a collapse and the chain
4010 * gets modified leaving us with an
4011 * invalid pointer.
4012 */
4013 vm_object_lock(source);
4014 if (source->vo_size == length &&
4015 source->ref_count == 1 &&
4016 (source->shadow == VM_OBJECT_NULL ||
4017 source->shadow->vo_copy == VM_OBJECT_NULL)) {
4018 VM_OBJECT_SET_SHADOWED(object: source, FALSE);
4019 vm_object_unlock(source);
4020 vm_object_shadow_skipped++;
4021 return FALSE;
4022 }
4023 /* things changed while we were locking "source"... */
4024 vm_object_unlock(source);
4025 }
4026 }
4027
4028 /*
4029 * *offset is the map entry's offset into the VM object and
4030 * is aligned to the map's page size.
4031 * VM objects need to be aligned to the system's page size.
4032 * Record the necessary adjustment and re-align the offset so
4033 * that result->vo_shadow_offset is properly page-aligned.
4034 */
4035 vm_object_offset_t offset_adjustment;
4036 offset_adjustment = *offset - vm_object_trunc_page(*offset);
4037 length = vm_object_round_page(length + offset_adjustment);
4038 *offset = vm_object_trunc_page(*offset);
4039
4040 /*
4041 * Allocate a new object with the given length
4042 */
4043
4044 if ((result = vm_object_allocate(size: length)) == VM_OBJECT_NULL) {
4045 panic("vm_object_shadow: no object for shadowing");
4046 }
4047
4048 /*
4049 * The new object shadows the source object, adding
4050 * a reference to it. Our caller changes his reference
4051 * to point to the new object, removing a reference to
4052 * the source object. Net result: no change of reference
4053 * count.
4054 */
4055 result->shadow = source;
4056
4057 /*
4058 * Store the offset into the source object,
4059 * and fix up the offset into the new object.
4060 */
4061
4062 result->vo_shadow_offset = *offset;
4063 assertf(page_aligned(result->vo_shadow_offset),
4064 "result %p shadow offset 0x%llx",
4065 result, result->vo_shadow_offset);
4066
4067 /*
4068 * Return the new things
4069 */
4070
4071 *offset = 0;
4072 if (offset_adjustment) {
4073 /*
4074 * Make the map entry point to the equivalent offset
4075 * in the new object.
4076 */
4077 DEBUG4K_COPY("adjusting offset @ %p from 0x%llx to 0x%llx for object %p length: 0x%llx\n", offset, *offset, *offset + offset_adjustment, result, length);
4078 *offset += offset_adjustment;
4079 }
4080 *object = result;
4081 return TRUE;
4082}
4083
4084/*
4085 * The relationship between vm_object structures and
4086 * the memory_object requires careful synchronization.
4087 *
4088 * All associations are created by memory_object_create_named
4089 * for external pagers and vm_object_compressor_pager_create for internal
4090 * objects as follows:
4091 *
4092 * pager: the memory_object itself, supplied by
4093 * the user requesting a mapping (or the kernel,
4094 * when initializing internal objects); the
4095 * kernel simulates holding send rights by keeping
4096 * a port reference;
4097 *
4098 * pager_request:
4099 * the memory object control port,
4100 * created by the kernel; the kernel holds
4101 * receive (and ownership) rights to this
4102 * port, but no other references.
4103 *
4104 * When initialization is complete, the "initialized" field
4105 * is asserted. Other mappings using a particular memory object,
4106 * and any references to the vm_object gained through the
4107 * port association must wait for this initialization to occur.
4108 *
4109 * In order to allow the memory manager to set attributes before
4110 * requests (notably virtual copy operations, but also data or
4111 * unlock requests) are made, a "ready" attribute is made available.
4112 * Only the memory manager may affect the value of this attribute.
4113 * Its value does not affect critical kernel functions, such as
4114 * internal object initialization or destruction. [Furthermore,
4115 * memory objects created by the kernel are assumed to be ready
4116 * immediately; the default memory manager need not explicitly
4117 * set the "ready" attribute.]
4118 *
4119 * [Both the "initialized" and "ready" attribute wait conditions
4120 * use the "pager" field as the wait event.]
4121 *
4122 * The port associations can be broken down by any of the
4123 * following routines:
4124 * vm_object_terminate:
4125 * No references to the vm_object remain, and
4126 * the object cannot (or will not) be cached.
4127 * This is the normal case, and is done even
4128 * though one of the other cases has already been
4129 * done.
4130 * memory_object_destroy:
4131 * The memory manager has requested that the
4132 * kernel relinquish references to the memory
4133 * object. [The memory manager may not want to
4134 * destroy the memory object, but may wish to
4135 * refuse or tear down existing memory mappings.]
4136 *
4137 * Each routine that breaks an association must break all of
4138 * them at once. At some later time, that routine must clear
4139 * the pager field and release the memory object references.
4140 * [Furthermore, each routine must cope with the simultaneous
4141 * or previous operations of the others.]
4142 *
4143 * Because the pager field may be cleared spontaneously, it
4144 * cannot be used to determine whether a memory object has
4145 * ever been associated with a particular vm_object. [This
4146 * knowledge is important to the shadow object mechanism.]
4147 * For this reason, an additional "created" attribute is
4148 * provided.
4149 *
4150 * During various paging operations, the pager reference found in the
4151 * vm_object must be valid. To prevent this from being released,
4152 * (other than being removed, i.e., made null), routines may use
4153 * the vm_object_paging_begin/end routines [actually, macros].
4154 * The implementation uses the "paging_in_progress" and "wanted" fields.
4155 * [Operations that alter the validity of the pager values include the
4156 * termination routines and vm_object_collapse.]
4157 */
4158
4159
4160/*
4161 * Routine: vm_object_memory_object_associate
4162 * Purpose:
4163 * Associate a VM object to the given pager.
4164 * If a VM object is not provided, create one.
4165 * Initialize the pager.
4166 */
4167vm_object_t
4168vm_object_memory_object_associate(
4169 memory_object_t pager,
4170 vm_object_t object,
4171 vm_object_size_t size,
4172 boolean_t named)
4173{
4174 memory_object_control_t control;
4175
4176 assert(pager != MEMORY_OBJECT_NULL);
4177
4178 if (object != VM_OBJECT_NULL) {
4179 vm_object_lock(object);
4180 assert(object->internal);
4181 assert(object->pager_created);
4182 assert(!object->pager_initialized);
4183 assert(!object->pager_ready);
4184 assert(object->pager_trusted);
4185 } else {
4186 object = vm_object_allocate(size);
4187 assert(object != VM_OBJECT_NULL);
4188 vm_object_lock(object);
4189 VM_OBJECT_SET_INTERNAL(object, FALSE);
4190 VM_OBJECT_SET_PAGER_TRUSTED(object, FALSE);
4191 /* copy strategy invalid until set by memory manager */
4192 object->copy_strategy = MEMORY_OBJECT_COPY_INVALID;
4193 }
4194
4195 /*
4196 * Allocate request port.
4197 */
4198
4199 control = memory_object_control_allocate(object);
4200 assert(control != MEMORY_OBJECT_CONTROL_NULL);
4201
4202 assert(!object->pager_ready);
4203 assert(!object->pager_initialized);
4204 assert(object->pager == NULL);
4205 assert(object->pager_control == NULL);
4206
4207 /*
4208 * Copy the reference we were given.
4209 */
4210
4211 memory_object_reference(object: pager);
4212 VM_OBJECT_SET_PAGER_CREATED(object, TRUE);
4213 object->pager = pager;
4214 object->pager_control = control;
4215 VM_OBJECT_SET_PAGER_READY(object, FALSE);
4216
4217 vm_object_unlock(object);
4218
4219 /*
4220 * Let the pager know we're using it.
4221 */
4222
4223 (void) memory_object_init(memory_object: pager,
4224 memory_control: object->pager_control,
4225 PAGE_SIZE);
4226
4227 vm_object_lock(object);
4228 if (named) {
4229 VM_OBJECT_SET_NAMED(object, TRUE);
4230 }
4231 if (object->internal) {
4232 VM_OBJECT_SET_PAGER_READY(object, TRUE);
4233 vm_object_wakeup(object, VM_OBJECT_EVENT_PAGER_READY);
4234 }
4235
4236 VM_OBJECT_SET_PAGER_INITIALIZED(object, TRUE);
4237 vm_object_wakeup(object, VM_OBJECT_EVENT_INITIALIZED);
4238
4239 vm_object_unlock(object);
4240
4241 return object;
4242}
4243
4244/*
4245 * Routine: vm_object_compressor_pager_create
4246 * Purpose:
4247 * Create a memory object for an internal object.
4248 * In/out conditions:
4249 * The object is locked on entry and exit;
4250 * it may be unlocked within this call.
4251 * Limitations:
4252 * Only one thread may be performing a
4253 * vm_object_compressor_pager_create on an object at
4254 * a time. Presumably, only the pageout
4255 * daemon will be using this routine.
4256 */
4257
4258void
4259vm_object_compressor_pager_create(
4260 vm_object_t object)
4261{
4262 memory_object_t pager;
4263 vm_object_t pager_object = VM_OBJECT_NULL;
4264
4265 assert(!is_kernel_object(object));
4266
4267 /*
4268 * Prevent collapse or termination by holding a paging reference
4269 */
4270
4271 vm_object_paging_begin(object);
4272 if (object->pager_created) {
4273 /*
4274 * Someone else got to it first...
4275 * wait for them to finish initializing the ports
4276 */
4277 while (!object->pager_initialized) {
4278 vm_object_sleep(object,
4279 VM_OBJECT_EVENT_INITIALIZED,
4280 THREAD_UNINT);
4281 }
4282 vm_object_paging_end(object);
4283 return;
4284 }
4285
4286 if ((uint32_t) (object->vo_size / PAGE_SIZE) !=
4287 (object->vo_size / PAGE_SIZE)) {
4288#if DEVELOPMENT || DEBUG
4289 printf("vm_object_compressor_pager_create(%p): "
4290 "object size 0x%llx >= 0x%llx\n",
4291 object,
4292 (uint64_t) object->vo_size,
4293 0x0FFFFFFFFULL * PAGE_SIZE);
4294#endif /* DEVELOPMENT || DEBUG */
4295 vm_object_paging_end(object);
4296 return;
4297 }
4298
4299 /*
4300 * Indicate that a memory object has been assigned
4301 * before dropping the lock, to prevent a race.
4302 */
4303
4304 VM_OBJECT_SET_PAGER_CREATED(object, TRUE);
4305 VM_OBJECT_SET_PAGER_TRUSTED(object, TRUE);
4306 object->paging_offset = 0;
4307
4308 vm_object_unlock(object);
4309
4310 /*
4311 * Create the [internal] pager, and associate it with this object.
4312 *
4313 * We make the association here so that vm_object_enter()
4314 * can look up the object to complete initializing it. No
4315 * user will ever map this object.
4316 */
4317 {
4318 /* create our new memory object */
4319 assert((uint32_t) (object->vo_size / PAGE_SIZE) ==
4320 (object->vo_size / PAGE_SIZE));
4321 (void) compressor_memory_object_create(
4322 (memory_object_size_t) object->vo_size,
4323 &pager);
4324 if (pager == NULL) {
4325 panic("vm_object_compressor_pager_create(): "
4326 "no pager for object %p size 0x%llx\n",
4327 object, (uint64_t) object->vo_size);
4328 }
4329 }
4330
4331 /*
4332 * A reference was returned by
4333 * memory_object_create(), and it is
4334 * copied by vm_object_memory_object_associate().
4335 */
4336
4337 pager_object = vm_object_memory_object_associate(pager,
4338 object,
4339 size: object->vo_size,
4340 FALSE);
4341 if (pager_object != object) {
4342 panic("vm_object_compressor_pager_create: mismatch (pager: %p, pager_object: %p, orig_object: %p, orig_object size: 0x%llx)", pager, pager_object, object, (uint64_t) object->vo_size);
4343 }
4344
4345 /*
4346 * Drop the reference we were passed.
4347 */
4348 memory_object_deallocate(object: pager);
4349
4350 vm_object_lock(object);
4351
4352 /*
4353 * Release the paging reference
4354 */
4355 vm_object_paging_end(object);
4356}
4357
4358/*
4359 * Global variables for vm_object_collapse():
4360 *
4361 * Counts for normal collapses and bypasses.
4362 * Debugging variables, to watch or disable collapse.
4363 */
4364static long object_collapses = 0;
4365static long object_bypasses = 0;
4366
4367static boolean_t vm_object_collapse_allowed = TRUE;
4368static boolean_t vm_object_bypass_allowed = TRUE;
4369
4370void vm_object_do_collapse_compressor(vm_object_t object,
4371 vm_object_t backing_object);
4372void
4373vm_object_do_collapse_compressor(
4374 vm_object_t object,
4375 vm_object_t backing_object)
4376{
4377 vm_object_offset_t new_offset, backing_offset;
4378 vm_object_size_t size;
4379
4380 vm_counters.do_collapse_compressor++;
4381
4382 vm_object_lock_assert_exclusive(object);
4383 vm_object_lock_assert_exclusive(backing_object);
4384
4385 size = object->vo_size;
4386
4387 /*
4388 * Move all compressed pages from backing_object
4389 * to the parent.
4390 */
4391
4392 for (backing_offset = object->vo_shadow_offset;
4393 backing_offset < object->vo_shadow_offset + object->vo_size;
4394 backing_offset += PAGE_SIZE) {
4395 memory_object_offset_t backing_pager_offset;
4396
4397 /* find the next compressed page at or after this offset */
4398 backing_pager_offset = (backing_offset +
4399 backing_object->paging_offset);
4400 backing_pager_offset = vm_compressor_pager_next_compressed(
4401 mem_obj: backing_object->pager,
4402 offset: backing_pager_offset);
4403 if (backing_pager_offset == (memory_object_offset_t) -1) {
4404 /* no more compressed pages */
4405 break;
4406 }
4407 backing_offset = (backing_pager_offset -
4408 backing_object->paging_offset);
4409
4410 new_offset = backing_offset - object->vo_shadow_offset;
4411
4412 if (new_offset >= object->vo_size) {
4413 /* we're out of the scope of "object": done */
4414 break;
4415 }
4416
4417 if ((vm_page_lookup(object, offset: new_offset) != VM_PAGE_NULL) ||
4418 (vm_compressor_pager_state_get(mem_obj: object->pager,
4419 offset: (new_offset +
4420 object->paging_offset)) ==
4421 VM_EXTERNAL_STATE_EXISTS)) {
4422 /*
4423 * This page already exists in object, resident or
4424 * compressed.
4425 * We don't need this compressed page in backing_object
4426 * and it will be reclaimed when we release
4427 * backing_object.
4428 */
4429 continue;
4430 }
4431
4432 /*
4433 * backing_object has this page in the VM compressor and
4434 * we need to transfer it to object.
4435 */
4436 vm_counters.do_collapse_compressor_pages++;
4437 vm_compressor_pager_transfer(
4438 /* destination: */
4439 dst_mem_obj: object->pager,
4440 dst_offset: (new_offset + object->paging_offset),
4441 /* source: */
4442 src_mem_obj: backing_object->pager,
4443 src_offset: (backing_offset + backing_object->paging_offset));
4444 }
4445}
4446
4447/*
4448 * Routine: vm_object_do_collapse
4449 * Purpose:
4450 * Collapse an object with the object backing it.
4451 * Pages in the backing object are moved into the
4452 * parent, and the backing object is deallocated.
4453 * Conditions:
4454 * Both objects and the cache are locked; the page
4455 * queues are unlocked.
4456 *
4457 */
4458static void
4459vm_object_do_collapse(
4460 vm_object_t object,
4461 vm_object_t backing_object)
4462{
4463 vm_page_t p, pp;
4464 vm_object_offset_t new_offset, backing_offset;
4465 vm_object_size_t size;
4466
4467 vm_object_lock_assert_exclusive(object);
4468 vm_object_lock_assert_exclusive(backing_object);
4469
4470 assert(object->purgable == VM_PURGABLE_DENY);
4471 assert(backing_object->purgable == VM_PURGABLE_DENY);
4472
4473 backing_offset = object->vo_shadow_offset;
4474 size = object->vo_size;
4475
4476 /*
4477 * Move all in-memory pages from backing_object
4478 * to the parent. Pages that have been paged out
4479 * will be overwritten by any of the parent's
4480 * pages that shadow them.
4481 */
4482
4483 while (!vm_page_queue_empty(&backing_object->memq)) {
4484 p = (vm_page_t) vm_page_queue_first(&backing_object->memq);
4485
4486 new_offset = (p->vmp_offset - backing_offset);
4487
4488 assert(!p->vmp_busy || p->vmp_absent);
4489
4490 /*
4491 * If the parent has a page here, or if
4492 * this page falls outside the parent,
4493 * dispose of it.
4494 *
4495 * Otherwise, move it as planned.
4496 */
4497
4498 if (p->vmp_offset < backing_offset || new_offset >= size) {
4499 VM_PAGE_FREE(p);
4500 } else {
4501 pp = vm_page_lookup(object, offset: new_offset);
4502 if (pp == VM_PAGE_NULL) {
4503 if (VM_COMPRESSOR_PAGER_STATE_GET(object,
4504 new_offset)
4505 == VM_EXTERNAL_STATE_EXISTS) {
4506 /*
4507 * Parent object has this page
4508 * in the VM compressor.
4509 * Throw away the backing
4510 * object's page.
4511 */
4512 VM_PAGE_FREE(p);
4513 } else {
4514 /*
4515 * Parent now has no page.
4516 * Move the backing object's page
4517 * up.
4518 */
4519 vm_page_rename(page: p, new_object: object, new_offset);
4520 }
4521 } else {
4522 assert(!pp->vmp_absent);
4523
4524 /*
4525 * Parent object has a real page.
4526 * Throw away the backing object's
4527 * page.
4528 */
4529 VM_PAGE_FREE(p);
4530 }
4531 }
4532 }
4533
4534 if (vm_object_collapse_compressor_allowed &&
4535 object->pager != MEMORY_OBJECT_NULL &&
4536 backing_object->pager != MEMORY_OBJECT_NULL) {
4537 /* move compressed pages from backing_object to object */
4538 vm_object_do_collapse_compressor(object, backing_object);
4539 } else if (backing_object->pager != MEMORY_OBJECT_NULL) {
4540 assert((!object->pager_created &&
4541 (object->pager == MEMORY_OBJECT_NULL)) ||
4542 (!backing_object->pager_created &&
4543 (backing_object->pager == MEMORY_OBJECT_NULL)));
4544 /*
4545 * Move the pager from backing_object to object.
4546 *
4547 * XXX We're only using part of the paging space
4548 * for keeps now... we ought to discard the
4549 * unused portion.
4550 */
4551
4552 assert(!object->paging_in_progress);
4553 assert(!object->activity_in_progress);
4554 assert(!object->pager_created);
4555 assert(object->pager == NULL);
4556 object->pager = backing_object->pager;
4557
4558 VM_OBJECT_SET_PAGER_CREATED(object, value: backing_object->pager_created);
4559 object->pager_control = backing_object->pager_control;
4560 VM_OBJECT_SET_PAGER_READY(object, value: backing_object->pager_ready);
4561 VM_OBJECT_SET_PAGER_INITIALIZED(object, value: backing_object->pager_initialized);
4562 object->paging_offset =
4563 backing_object->paging_offset + backing_offset;
4564 if (object->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
4565 memory_object_control_collapse(control: &object->pager_control,
4566 object);
4567 }
4568 /* the backing_object has lost its pager: reset all fields */
4569 VM_OBJECT_SET_PAGER_CREATED(object: backing_object, FALSE);
4570 backing_object->pager_control = NULL;
4571 VM_OBJECT_SET_PAGER_READY(object: backing_object, FALSE);
4572 backing_object->paging_offset = 0;
4573 backing_object->pager = NULL;
4574 }
4575 /*
4576 * Object now shadows whatever backing_object did.
4577 * Note that the reference to backing_object->shadow
4578 * moves from within backing_object to within object.
4579 */
4580
4581 assert(!object->phys_contiguous);
4582 assert(!backing_object->phys_contiguous);
4583 object->shadow = backing_object->shadow;
4584 if (object->shadow) {
4585 assertf(page_aligned(object->vo_shadow_offset),
4586 "object %p shadow_offset 0x%llx",
4587 object, object->vo_shadow_offset);
4588 assertf(page_aligned(backing_object->vo_shadow_offset),
4589 "backing_object %p shadow_offset 0x%llx",
4590 backing_object, backing_object->vo_shadow_offset);
4591 object->vo_shadow_offset += backing_object->vo_shadow_offset;
4592 /* "backing_object" gave its shadow to "object" */
4593 backing_object->shadow = VM_OBJECT_NULL;
4594 backing_object->vo_shadow_offset = 0;
4595 } else {
4596 /* no shadow, therefore no shadow offset... */
4597 object->vo_shadow_offset = 0;
4598 }
4599 assert((object->shadow == VM_OBJECT_NULL) ||
4600 (object->shadow->vo_copy != backing_object));
4601
4602 /*
4603 * Discard backing_object.
4604 *
4605 * Since the backing object has no pages, no
4606 * pager left, and no object references within it,
4607 * all that is necessary is to dispose of it.
4608 */
4609 object_collapses++;
4610
4611 assert(backing_object->ref_count == 1);
4612 assert(backing_object->resident_page_count == 0);
4613 assert(backing_object->paging_in_progress == 0);
4614 assert(backing_object->activity_in_progress == 0);
4615 assert(backing_object->shadow == VM_OBJECT_NULL);
4616 assert(backing_object->vo_shadow_offset == 0);
4617
4618 if (backing_object->pager != MEMORY_OBJECT_NULL) {
4619 /* ... unless it has a pager; need to terminate pager too */
4620 vm_counters.do_collapse_terminate++;
4621 if (vm_object_terminate(object: backing_object) != KERN_SUCCESS) {
4622 vm_counters.do_collapse_terminate_failure++;
4623 }
4624 return;
4625 }
4626
4627 assert(backing_object->pager == NULL);
4628
4629 VM_OBJECT_SET_ALIVE(object: backing_object, FALSE);
4630 vm_object_unlock(backing_object);
4631
4632#if VM_OBJECT_TRACKING
4633 if (vm_object_tracking_btlog) {
4634 btlog_erase(vm_object_tracking_btlog, backing_object);
4635 }
4636#endif /* VM_OBJECT_TRACKING */
4637
4638 vm_object_lock_destroy(backing_object);
4639
4640 zfree(vm_object_zone, backing_object);
4641}
4642
4643static void
4644vm_object_do_bypass(
4645 vm_object_t object,
4646 vm_object_t backing_object)
4647{
4648 /*
4649 * Make the parent shadow the next object
4650 * in the chain.
4651 */
4652
4653 vm_object_lock_assert_exclusive(object);
4654 vm_object_lock_assert_exclusive(backing_object);
4655
4656 vm_object_reference(backing_object->shadow);
4657
4658 assert(!object->phys_contiguous);
4659 assert(!backing_object->phys_contiguous);
4660 object->shadow = backing_object->shadow;
4661 if (object->shadow) {
4662 assertf(page_aligned(object->vo_shadow_offset),
4663 "object %p shadow_offset 0x%llx",
4664 object, object->vo_shadow_offset);
4665 assertf(page_aligned(backing_object->vo_shadow_offset),
4666 "backing_object %p shadow_offset 0x%llx",
4667 backing_object, backing_object->vo_shadow_offset);
4668 object->vo_shadow_offset += backing_object->vo_shadow_offset;
4669 } else {
4670 /* no shadow, therefore no shadow offset... */
4671 object->vo_shadow_offset = 0;
4672 }
4673
4674 /*
4675 * Backing object might have had a copy pointer
4676 * to us. If it did, clear it.
4677 */
4678 if (backing_object->vo_copy == object) {
4679 VM_OBJECT_COPY_SET(object: backing_object, VM_OBJECT_NULL);
4680 }
4681
4682 /*
4683 * Drop the reference count on backing_object.
4684 #if TASK_SWAPPER
4685 * Since its ref_count was at least 2, it
4686 * will not vanish; so we don't need to call
4687 * vm_object_deallocate.
4688 * [with a caveat for "named" objects]
4689 *
4690 * The res_count on the backing object is
4691 * conditionally decremented. It's possible
4692 * (via vm_pageout_scan) to get here with
4693 * a "swapped" object, which has a 0 res_count,
4694 * in which case, the backing object res_count
4695 * is already down by one.
4696 #else
4697 * Don't call vm_object_deallocate unless
4698 * ref_count drops to zero.
4699 *
4700 * The ref_count can drop to zero here if the
4701 * backing object could be bypassed but not
4702 * collapsed, such as when the backing object
4703 * is temporary and cachable.
4704 #endif
4705 */
4706 if (backing_object->ref_count > 2 ||
4707 (!backing_object->named && backing_object->ref_count > 1)) {
4708 vm_object_lock_assert_exclusive(backing_object);
4709 backing_object->ref_count--;
4710 vm_object_unlock(backing_object);
4711 } else {
4712 /*
4713 * Drop locks so that we can deallocate
4714 * the backing object.
4715 */
4716
4717 /*
4718 * vm_object_collapse (the caller of this function) is
4719 * now called from contexts that may not guarantee that a
4720 * valid reference is held on the object... w/o a valid
4721 * reference, it is unsafe and unwise (you will definitely
4722 * regret it) to unlock the object and then retake the lock
4723 * since the object may be terminated and recycled in between.
4724 * The "activity_in_progress" reference will keep the object
4725 * 'stable'.
4726 */
4727 vm_object_activity_begin(object);
4728 vm_object_unlock(object);
4729
4730 vm_object_unlock(backing_object);
4731 vm_object_deallocate(object: backing_object);
4732
4733 /*
4734 * Relock object. We don't have to reverify
4735 * its state since vm_object_collapse will
4736 * do that for us as it starts at the
4737 * top of its loop.
4738 */
4739
4740 vm_object_lock(object);
4741 vm_object_activity_end(object);
4742 }
4743
4744 object_bypasses++;
4745}
4746
4747
4748/*
4749 * vm_object_collapse:
4750 *
4751 * Perform an object collapse or an object bypass if appropriate.
4752 * The real work of collapsing and bypassing is performed in
4753 * the routines vm_object_do_collapse and vm_object_do_bypass.
4754 *
4755 * Requires that the object be locked and the page queues be unlocked.
4756 *
4757 */
4758static unsigned long vm_object_collapse_calls = 0;
4759static unsigned long vm_object_collapse_objects = 0;
4760static unsigned long vm_object_collapse_do_collapse = 0;
4761static unsigned long vm_object_collapse_do_bypass = 0;
4762
4763__private_extern__ void
4764vm_object_collapse(
4765 vm_object_t object,
4766 vm_object_offset_t hint_offset,
4767 boolean_t can_bypass)
4768{
4769 vm_object_t backing_object;
4770 vm_object_size_t object_vcount, object_rcount;
4771 vm_object_t original_object;
4772 int object_lock_type;
4773 int backing_object_lock_type;
4774
4775 vm_object_collapse_calls++;
4776
4777 assertf(page_aligned(hint_offset), "hint_offset 0x%llx", hint_offset);
4778
4779 if (!vm_object_collapse_allowed &&
4780 !(can_bypass && vm_object_bypass_allowed)) {
4781 return;
4782 }
4783
4784 if (object == VM_OBJECT_NULL) {
4785 return;
4786 }
4787
4788 original_object = object;
4789
4790 /*
4791 * The top object was locked "exclusive" by the caller.
4792 * In the first pass, to determine if we can collapse the shadow chain,
4793 * take a "shared" lock on the shadow objects. If we can collapse,
4794 * we'll have to go down the chain again with exclusive locks.
4795 */
4796 object_lock_type = OBJECT_LOCK_EXCLUSIVE;
4797 backing_object_lock_type = OBJECT_LOCK_SHARED;
4798
4799retry:
4800 object = original_object;
4801 vm_object_lock_assert_exclusive(object);
4802
4803 while (TRUE) {
4804 vm_object_collapse_objects++;
4805 /*
4806 * Verify that the conditions are right for either
4807 * collapse or bypass:
4808 */
4809
4810 /*
4811 * There is a backing object, and
4812 */
4813
4814 backing_object = object->shadow;
4815 if (backing_object == VM_OBJECT_NULL) {
4816 if (object != original_object) {
4817 vm_object_unlock(object);
4818 }
4819 return;
4820 }
4821 if (backing_object_lock_type == OBJECT_LOCK_SHARED) {
4822 vm_object_lock_shared(backing_object);
4823 } else {
4824 vm_object_lock(backing_object);
4825 }
4826
4827 /*
4828 * No pages in the object are currently
4829 * being paged out, and
4830 */
4831 if (object->paging_in_progress != 0 ||
4832 object->activity_in_progress != 0) {
4833 /* try and collapse the rest of the shadow chain */
4834 if (object != original_object) {
4835 vm_object_unlock(object);
4836 }
4837 object = backing_object;
4838 object_lock_type = backing_object_lock_type;
4839 continue;
4840 }
4841
4842 /*
4843 * ...
4844 * The backing object is not read_only,
4845 * and no pages in the backing object are
4846 * currently being paged out.
4847 * The backing object is internal.
4848 *
4849 */
4850
4851 if (!backing_object->internal ||
4852 backing_object->paging_in_progress != 0 ||
4853 backing_object->activity_in_progress != 0) {
4854 /* try and collapse the rest of the shadow chain */
4855 if (object != original_object) {
4856 vm_object_unlock(object);
4857 }
4858 object = backing_object;
4859 object_lock_type = backing_object_lock_type;
4860 continue;
4861 }
4862
4863 /*
4864 * Purgeable objects are not supposed to engage in
4865 * copy-on-write activities, so should not have
4866 * any shadow objects or be a shadow object to another
4867 * object.
4868 * Collapsing a purgeable object would require some
4869 * updates to the purgeable compressed ledgers.
4870 */
4871 if (object->purgable != VM_PURGABLE_DENY ||
4872 backing_object->purgable != VM_PURGABLE_DENY) {
4873 panic("vm_object_collapse() attempting to collapse "
4874 "purgeable object: %p(%d) %p(%d)\n",
4875 object, object->purgable,
4876 backing_object, backing_object->purgable);
4877 /* try and collapse the rest of the shadow chain */
4878 if (object != original_object) {
4879 vm_object_unlock(object);
4880 }
4881 object = backing_object;
4882 object_lock_type = backing_object_lock_type;
4883 continue;
4884 }
4885
4886 /*
4887 * The backing object can't be a copy-object:
4888 * the shadow_offset for the copy-object must stay
4889 * as 0. Furthermore (for the 'we have all the
4890 * pages' case), if we bypass backing_object and
4891 * just shadow the next object in the chain, old
4892 * pages from that object would then have to be copied
4893 * BOTH into the (former) backing_object and into the
4894 * parent object.
4895 */
4896 if (backing_object->shadow != VM_OBJECT_NULL &&
4897 backing_object->shadow->vo_copy == backing_object) {
4898 /* try and collapse the rest of the shadow chain */
4899 if (object != original_object) {
4900 vm_object_unlock(object);
4901 }
4902 object = backing_object;
4903 object_lock_type = backing_object_lock_type;
4904 continue;
4905 }
4906
4907 /*
4908 * We can now try to either collapse the backing
4909 * object (if the parent is the only reference to
4910 * it) or (perhaps) remove the parent's reference
4911 * to it.
4912 *
4913 * If there is exactly one reference to the backing
4914 * object, we may be able to collapse it into the
4915 * parent.
4916 *
4917 * As long as one of the objects is still not known
4918 * to the pager, we can collapse them.
4919 */
4920 if (backing_object->ref_count == 1 &&
4921 (vm_object_collapse_compressor_allowed ||
4922 !object->pager_created
4923 || (!backing_object->pager_created)
4924 ) && vm_object_collapse_allowed) {
4925 /*
4926 * We need the exclusive lock on the VM objects.
4927 */
4928 if (backing_object_lock_type != OBJECT_LOCK_EXCLUSIVE) {
4929 /*
4930 * We have an object and its shadow locked
4931 * "shared". We can't just upgrade the locks
4932 * to "exclusive", as some other thread might
4933 * also have these objects locked "shared" and
4934 * attempt to upgrade one or the other to
4935 * "exclusive". The upgrades would block
4936 * forever waiting for the other "shared" locks
4937 * to get released.
4938 * So we have to release the locks and go
4939 * down the shadow chain again (since it could
4940 * have changed) with "exclusive" locking.
4941 */
4942 vm_object_unlock(backing_object);
4943 if (object != original_object) {
4944 vm_object_unlock(object);
4945 }
4946 object_lock_type = OBJECT_LOCK_EXCLUSIVE;
4947 backing_object_lock_type = OBJECT_LOCK_EXCLUSIVE;
4948 goto retry;
4949 }
4950
4951 /*
4952 * Collapse the object with its backing
4953 * object, and try again with the object's
4954 * new backing object.
4955 */
4956
4957 vm_object_do_collapse(object, backing_object);
4958 vm_object_collapse_do_collapse++;
4959 continue;
4960 }
4961
4962 /*
4963 * Collapsing the backing object was not possible
4964 * or permitted, so let's try bypassing it.
4965 */
4966
4967 if (!(can_bypass && vm_object_bypass_allowed)) {
4968 /* try and collapse the rest of the shadow chain */
4969 if (object != original_object) {
4970 vm_object_unlock(object);
4971 }
4972 object = backing_object;
4973 object_lock_type = backing_object_lock_type;
4974 continue;
4975 }
4976
4977
4978 /*
4979 * If the object doesn't have all its pages present,
4980 * we have to make sure no pages in the backing object
4981 * "show through" before bypassing it.
4982 */
4983 object_vcount = object->vo_size >> PAGE_SHIFT;
4984 object_rcount = (vm_object_size_t)object->resident_page_count;
4985
4986 if (object_rcount != object_vcount) {
4987 vm_object_offset_t offset;
4988 vm_object_offset_t backing_offset;
4989 vm_object_size_t backing_rcount, backing_vcount;
4990
4991 /*
4992 * If the backing object has a pager but no pagemap,
4993 * then we cannot bypass it, because we don't know
4994 * what pages it has.
4995 */
4996 if (backing_object->pager_created) {
4997 /* try and collapse the rest of the shadow chain */
4998 if (object != original_object) {
4999 vm_object_unlock(object);
5000 }
5001 object = backing_object;
5002 object_lock_type = backing_object_lock_type;
5003 continue;
5004 }
5005
5006 /*
5007 * If the object has a pager but no pagemap,
5008 * then we cannot bypass it, because we don't know
5009 * what pages it has.
5010 */
5011 if (object->pager_created) {
5012 /* try and collapse the rest of the shadow chain */
5013 if (object != original_object) {
5014 vm_object_unlock(object);
5015 }
5016 object = backing_object;
5017 object_lock_type = backing_object_lock_type;
5018 continue;
5019 }
5020
5021 backing_offset = object->vo_shadow_offset;
5022 backing_vcount = backing_object->vo_size >> PAGE_SHIFT;
5023 backing_rcount = (vm_object_size_t)backing_object->resident_page_count;
5024 assert(backing_vcount >= object_vcount);
5025
5026 if (backing_rcount > (backing_vcount - object_vcount) &&
5027 backing_rcount - (backing_vcount - object_vcount) > object_rcount) {
5028 /*
5029 * we have enough pages in the backing object to guarantee that
5030 * at least 1 of them must be 'uncovered' by a resident page
5031 * in the object we're evaluating, so move on and
5032 * try to collapse the rest of the shadow chain
5033 */
5034 if (object != original_object) {
5035 vm_object_unlock(object);
5036 }
5037 object = backing_object;
5038 object_lock_type = backing_object_lock_type;
5039 continue;
5040 }
5041
5042 /*
5043 * If all of the pages in the backing object are
5044 * shadowed by the parent object, the parent
5045 * object no longer has to shadow the backing
5046 * object; it can shadow the next one in the
5047 * chain.
5048 *
5049 * If the backing object has existence info,
5050 * we must check examine its existence info
5051 * as well.
5052 *
5053 */
5054
5055#define EXISTS_IN_OBJECT(obj, off, rc) \
5056 ((VM_COMPRESSOR_PAGER_STATE_GET((obj), (off)) \
5057 == VM_EXTERNAL_STATE_EXISTS) || \
5058 ((rc) && vm_page_lookup((obj), (off)) != VM_PAGE_NULL && (rc)--))
5059
5060 /*
5061 * Check the hint location first
5062 * (since it is often the quickest way out of here).
5063 */
5064 if (object->cow_hint != ~(vm_offset_t)0) {
5065 hint_offset = (vm_object_offset_t)object->cow_hint;
5066 } else {
5067 hint_offset = (hint_offset > 8 * PAGE_SIZE_64) ?
5068 (hint_offset - 8 * PAGE_SIZE_64) : 0;
5069 }
5070
5071 if (EXISTS_IN_OBJECT(backing_object, hint_offset +
5072 backing_offset, backing_rcount) &&
5073 !EXISTS_IN_OBJECT(object, hint_offset, object_rcount)) {
5074 /* dependency right at the hint */
5075 object->cow_hint = (vm_offset_t) hint_offset; /* atomic */
5076 /* try and collapse the rest of the shadow chain */
5077 if (object != original_object) {
5078 vm_object_unlock(object);
5079 }
5080 object = backing_object;
5081 object_lock_type = backing_object_lock_type;
5082 continue;
5083 }
5084
5085 /*
5086 * If the object's window onto the backing_object
5087 * is large compared to the number of resident
5088 * pages in the backing object, it makes sense to
5089 * walk the backing_object's resident pages first.
5090 *
5091 * NOTE: Pages may be in both the existence map and/or
5092 * resident, so if we don't find a dependency while
5093 * walking the backing object's resident page list
5094 * directly, and there is an existence map, we'll have
5095 * to run the offset based 2nd pass. Because we may
5096 * have to run both passes, we need to be careful
5097 * not to decrement 'rcount' in the 1st pass
5098 */
5099 if (backing_rcount && backing_rcount < (object_vcount / 8)) {
5100 vm_object_size_t rc = object_rcount;
5101 vm_page_t p;
5102
5103 backing_rcount = backing_object->resident_page_count;
5104 p = (vm_page_t)vm_page_queue_first(&backing_object->memq);
5105 do {
5106 offset = (p->vmp_offset - backing_offset);
5107
5108 if (offset < object->vo_size &&
5109 offset != hint_offset &&
5110 !EXISTS_IN_OBJECT(object, offset, rc)) {
5111 /* found a dependency */
5112 object->cow_hint = (vm_offset_t) offset; /* atomic */
5113
5114 break;
5115 }
5116 p = (vm_page_t) vm_page_queue_next(&p->vmp_listq);
5117 } while (--backing_rcount);
5118 if (backing_rcount != 0) {
5119 /* try and collapse the rest of the shadow chain */
5120 if (object != original_object) {
5121 vm_object_unlock(object);
5122 }
5123 object = backing_object;
5124 object_lock_type = backing_object_lock_type;
5125 continue;
5126 }
5127 }
5128
5129 /*
5130 * Walk through the offsets looking for pages in the
5131 * backing object that show through to the object.
5132 */
5133 if (backing_rcount) {
5134 offset = hint_offset;
5135
5136 while ((offset =
5137 (offset + PAGE_SIZE_64 < object->vo_size) ?
5138 (offset + PAGE_SIZE_64) : 0) != hint_offset) {
5139 if (EXISTS_IN_OBJECT(backing_object, offset +
5140 backing_offset, backing_rcount) &&
5141 !EXISTS_IN_OBJECT(object, offset, object_rcount)) {
5142 /* found a dependency */
5143 object->cow_hint = (vm_offset_t) offset; /* atomic */
5144 break;
5145 }
5146 }
5147 if (offset != hint_offset) {
5148 /* try and collapse the rest of the shadow chain */
5149 if (object != original_object) {
5150 vm_object_unlock(object);
5151 }
5152 object = backing_object;
5153 object_lock_type = backing_object_lock_type;
5154 continue;
5155 }
5156 }
5157 }
5158
5159 /*
5160 * We need "exclusive" locks on the 2 VM objects.
5161 */
5162 if (backing_object_lock_type != OBJECT_LOCK_EXCLUSIVE) {
5163 vm_object_unlock(backing_object);
5164 if (object != original_object) {
5165 vm_object_unlock(object);
5166 }
5167 object_lock_type = OBJECT_LOCK_EXCLUSIVE;
5168 backing_object_lock_type = OBJECT_LOCK_EXCLUSIVE;
5169 goto retry;
5170 }
5171
5172 /* reset the offset hint for any objects deeper in the chain */
5173 object->cow_hint = (vm_offset_t)0;
5174
5175 /*
5176 * All interesting pages in the backing object
5177 * already live in the parent or its pager.
5178 * Thus we can bypass the backing object.
5179 */
5180
5181 vm_object_do_bypass(object, backing_object);
5182 vm_object_collapse_do_bypass++;
5183
5184 /*
5185 * Try again with this object's new backing object.
5186 */
5187
5188 continue;
5189 }
5190
5191 /* NOT REACHED */
5192 /*
5193 * if (object != original_object) {
5194 * vm_object_unlock(object);
5195 * }
5196 */
5197}
5198
5199/*
5200 * Routine: vm_object_page_remove: [internal]
5201 * Purpose:
5202 * Removes all physical pages in the specified
5203 * object range from the object's list of pages.
5204 *
5205 * In/out conditions:
5206 * The object must be locked.
5207 * The object must not have paging_in_progress, usually
5208 * guaranteed by not having a pager.
5209 */
5210unsigned int vm_object_page_remove_lookup = 0;
5211unsigned int vm_object_page_remove_iterate = 0;
5212
5213__private_extern__ void
5214vm_object_page_remove(
5215 vm_object_t object,
5216 vm_object_offset_t start,
5217 vm_object_offset_t end)
5218{
5219 vm_page_t p, next;
5220
5221 /*
5222 * One and two page removals are most popular.
5223 * The factor of 16 here is somewhat arbitrary.
5224 * It balances vm_object_lookup vs iteration.
5225 */
5226
5227 if (atop_64(end - start) < (unsigned)object->resident_page_count / 16) {
5228 vm_object_page_remove_lookup++;
5229
5230 for (; start < end; start += PAGE_SIZE_64) {
5231 p = vm_page_lookup(object, offset: start);
5232 if (p != VM_PAGE_NULL) {
5233 assert(!p->vmp_cleaning && !p->vmp_laundry);
5234 if (!p->vmp_fictitious && p->vmp_pmapped) {
5235 pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: p));
5236 }
5237 VM_PAGE_FREE(p);
5238 }
5239 }
5240 } else {
5241 vm_object_page_remove_iterate++;
5242
5243 p = (vm_page_t) vm_page_queue_first(&object->memq);
5244 while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t) p)) {
5245 next = (vm_page_t) vm_page_queue_next(&p->vmp_listq);
5246 if ((start <= p->vmp_offset) && (p->vmp_offset < end)) {
5247 assert(!p->vmp_cleaning && !p->vmp_laundry);
5248 if (!p->vmp_fictitious && p->vmp_pmapped) {
5249 pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: p));
5250 }
5251 VM_PAGE_FREE(p);
5252 }
5253 p = next;
5254 }
5255 }
5256}
5257
5258
5259/*
5260 * Routine: vm_object_coalesce
5261 * Function: Coalesces two objects backing up adjoining
5262 * regions of memory into a single object.
5263 *
5264 * returns TRUE if objects were combined.
5265 *
5266 * NOTE: Only works at the moment if the second object is NULL -
5267 * if it's not, which object do we lock first?
5268 *
5269 * Parameters:
5270 * prev_object First object to coalesce
5271 * prev_offset Offset into prev_object
5272 * next_object Second object into coalesce
5273 * next_offset Offset into next_object
5274 *
5275 * prev_size Size of reference to prev_object
5276 * next_size Size of reference to next_object
5277 *
5278 * Conditions:
5279 * The object(s) must *not* be locked. The map must be locked
5280 * to preserve the reference to the object(s).
5281 */
5282static int vm_object_coalesce_count = 0;
5283
5284__private_extern__ boolean_t
5285vm_object_coalesce(
5286 vm_object_t prev_object,
5287 vm_object_t next_object,
5288 vm_object_offset_t prev_offset,
5289 __unused vm_object_offset_t next_offset,
5290 vm_object_size_t prev_size,
5291 vm_object_size_t next_size)
5292{
5293 vm_object_size_t newsize;
5294
5295#ifdef lint
5296 next_offset++;
5297#endif /* lint */
5298
5299 if (next_object != VM_OBJECT_NULL) {
5300 return FALSE;
5301 }
5302
5303 if (prev_object == VM_OBJECT_NULL) {
5304 return TRUE;
5305 }
5306
5307 vm_object_lock(prev_object);
5308
5309 /*
5310 * Try to collapse the object first
5311 */
5312 vm_object_collapse(object: prev_object, hint_offset: prev_offset, TRUE);
5313
5314 /*
5315 * Can't coalesce if pages not mapped to
5316 * prev_entry may be in use any way:
5317 * . more than one reference
5318 * . paged out
5319 * . shadows another object
5320 * . has a copy elsewhere
5321 * . is purgeable
5322 * . paging references (pages might be in page-list)
5323 */
5324
5325 if ((prev_object->ref_count > 1) ||
5326 prev_object->pager_created ||
5327 prev_object->phys_contiguous ||
5328 (prev_object->shadow != VM_OBJECT_NULL) ||
5329 (prev_object->vo_copy != VM_OBJECT_NULL) ||
5330 (prev_object->true_share != FALSE) ||
5331 (prev_object->purgable != VM_PURGABLE_DENY) ||
5332 (prev_object->paging_in_progress != 0) ||
5333 (prev_object->activity_in_progress != 0)) {
5334 vm_object_unlock(prev_object);
5335 return FALSE;
5336 }
5337 /* newsize = prev_offset + prev_size + next_size; */
5338 if (__improbable(os_add3_overflow(prev_offset, prev_size, next_size,
5339 &newsize))) {
5340 vm_object_unlock(prev_object);
5341 return FALSE;
5342 }
5343
5344 vm_object_coalesce_count++;
5345
5346 /*
5347 * Remove any pages that may still be in the object from
5348 * a previous deallocation.
5349 */
5350 vm_object_page_remove(object: prev_object,
5351 start: prev_offset + prev_size,
5352 end: prev_offset + prev_size + next_size);
5353
5354 /*
5355 * Extend the object if necessary.
5356 */
5357 if (newsize > prev_object->vo_size) {
5358 assertf(page_aligned(newsize),
5359 "object %p size 0x%llx",
5360 prev_object, (uint64_t)newsize);
5361 prev_object->vo_size = newsize;
5362 }
5363
5364 vm_object_unlock(prev_object);
5365 return TRUE;
5366}
5367
5368kern_return_t
5369vm_object_populate_with_private(
5370 vm_object_t object,
5371 vm_object_offset_t offset,
5372 ppnum_t phys_page,
5373 vm_size_t size)
5374{
5375 ppnum_t base_page;
5376 vm_object_offset_t base_offset;
5377
5378
5379 if (!object->private) {
5380 return KERN_FAILURE;
5381 }
5382
5383 base_page = phys_page;
5384
5385 vm_object_lock(object);
5386
5387 if (!object->phys_contiguous) {
5388 vm_page_t m;
5389
5390 if ((base_offset = trunc_page_64(offset)) != offset) {
5391 vm_object_unlock(object);
5392 return KERN_FAILURE;
5393 }
5394 base_offset += object->paging_offset;
5395
5396 while (size) {
5397 m = vm_page_lookup(object, offset: base_offset);
5398
5399 if (m != VM_PAGE_NULL) {
5400 if (m->vmp_fictitious) {
5401 if (VM_PAGE_GET_PHYS_PAGE(m) != vm_page_guard_addr) {
5402 vm_page_lockspin_queues();
5403 m->vmp_private = TRUE;
5404 vm_page_unlock_queues();
5405
5406 m->vmp_fictitious = FALSE;
5407 VM_PAGE_SET_PHYS_PAGE(m, base_page);
5408 }
5409 } else if (VM_PAGE_GET_PHYS_PAGE(m) != base_page) {
5410 if (!m->vmp_private) {
5411 /*
5412 * we'd leak a real page... that can't be right
5413 */
5414 panic("vm_object_populate_with_private - %p not private", m);
5415 }
5416 if (m->vmp_pmapped) {
5417 /*
5418 * pmap call to clear old mapping
5419 */
5420 pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m));
5421 }
5422 VM_PAGE_SET_PHYS_PAGE(m, base_page);
5423 }
5424 } else {
5425 m = vm_page_grab_fictitious(TRUE);
5426
5427 /*
5428 * private normally requires lock_queues but since we
5429 * are initializing the page, its not necessary here
5430 */
5431 m->vmp_private = TRUE;
5432 m->vmp_fictitious = FALSE;
5433 VM_PAGE_SET_PHYS_PAGE(m, base_page);
5434 m->vmp_unusual = TRUE;
5435 m->vmp_busy = FALSE;
5436
5437 vm_page_insert(page: m, object, offset: base_offset);
5438 }
5439 base_page++; /* Go to the next physical page */
5440 base_offset += PAGE_SIZE;
5441 size -= PAGE_SIZE;
5442 }
5443 } else {
5444 /* NOTE: we should check the original settings here */
5445 /* if we have a size > zero a pmap call should be made */
5446 /* to disable the range */
5447
5448 /* pmap_? */
5449
5450 /* shadows on contiguous memory are not allowed */
5451 /* we therefore can use the offset field */
5452 object->vo_shadow_offset = (vm_object_offset_t)phys_page << PAGE_SHIFT;
5453 assertf(page_aligned(size),
5454 "object %p size 0x%llx",
5455 object, (uint64_t)size);
5456 object->vo_size = size;
5457 }
5458 vm_object_unlock(object);
5459
5460 return KERN_SUCCESS;
5461}
5462
5463
5464kern_return_t
5465memory_object_create_named(
5466 memory_object_t pager,
5467 memory_object_offset_t size,
5468 memory_object_control_t *control)
5469{
5470 vm_object_t object;
5471
5472 *control = MEMORY_OBJECT_CONTROL_NULL;
5473 if (pager == MEMORY_OBJECT_NULL) {
5474 return KERN_INVALID_ARGUMENT;
5475 }
5476
5477 object = vm_object_memory_object_associate(pager,
5478 VM_OBJECT_NULL,
5479 size,
5480 TRUE);
5481 if (object == VM_OBJECT_NULL) {
5482 return KERN_INVALID_OBJECT;
5483 }
5484
5485 /* wait for object (if any) to be ready */
5486 if (object != VM_OBJECT_NULL) {
5487 vm_object_lock(object);
5488 VM_OBJECT_SET_NAMED(object, TRUE);
5489 while (!object->pager_ready) {
5490 vm_object_sleep(object,
5491 VM_OBJECT_EVENT_PAGER_READY,
5492 THREAD_UNINT);
5493 }
5494 *control = object->pager_control;
5495 vm_object_unlock(object);
5496 }
5497 return KERN_SUCCESS;
5498}
5499
5500
5501__private_extern__ kern_return_t
5502vm_object_lock_request(
5503 vm_object_t object,
5504 vm_object_offset_t offset,
5505 vm_object_size_t size,
5506 memory_object_return_t should_return,
5507 int flags,
5508 vm_prot_t prot)
5509{
5510 __unused boolean_t should_flush;
5511
5512 should_flush = flags & MEMORY_OBJECT_DATA_FLUSH;
5513
5514 /*
5515 * Check for bogus arguments.
5516 */
5517 if (object == VM_OBJECT_NULL) {
5518 return KERN_INVALID_ARGUMENT;
5519 }
5520
5521 if ((prot & ~VM_PROT_ALL) != 0 && prot != VM_PROT_NO_CHANGE) {
5522 return KERN_INVALID_ARGUMENT;
5523 }
5524
5525 /*
5526 * XXX TODO4K
5527 * extend range for conservative operations (copy-on-write, sync, ...)
5528 * truncate range for destructive operations (purge, ...)
5529 */
5530 size = vm_object_round_page(offset + size) - vm_object_trunc_page(offset);
5531 offset = vm_object_trunc_page(offset);
5532
5533 /*
5534 * Lock the object, and acquire a paging reference to
5535 * prevent the memory_object reference from being released.
5536 */
5537 vm_object_lock(object);
5538 vm_object_paging_begin(object);
5539
5540 (void)vm_object_update(object,
5541 offset, size, NULL, NULL, should_return, flags, prot);
5542
5543 vm_object_paging_end(object);
5544 vm_object_unlock(object);
5545
5546 return KERN_SUCCESS;
5547}
5548
5549/*
5550 * Empty a purgeable object by grabbing the physical pages assigned to it and
5551 * putting them on the free queue without writing them to backing store, etc.
5552 * When the pages are next touched they will be demand zero-fill pages. We
5553 * skip pages which are busy, being paged in/out, wired, etc. We do _not_
5554 * skip referenced/dirty pages, pages on the active queue, etc. We're more
5555 * than happy to grab these since this is a purgeable object. We mark the
5556 * object as "empty" after reaping its pages.
5557 *
5558 * On entry the object must be locked and it must be
5559 * purgeable with no delayed copies pending.
5560 */
5561uint64_t
5562vm_object_purge(vm_object_t object, int flags)
5563{
5564 unsigned int object_page_count = 0, pgcount = 0;
5565 uint64_t total_purged_pgcount = 0;
5566 boolean_t skipped_object = FALSE;
5567
5568 vm_object_lock_assert_exclusive(object);
5569
5570 if (object->purgable == VM_PURGABLE_DENY) {
5571 return 0;
5572 }
5573
5574 assert(object->vo_copy == VM_OBJECT_NULL);
5575 assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
5576
5577 /*
5578 * We need to set the object's state to VM_PURGABLE_EMPTY *before*
5579 * reaping its pages. We update vm_page_purgeable_count in bulk
5580 * and we don't want vm_page_remove() to update it again for each
5581 * page we reap later.
5582 *
5583 * For the purgeable ledgers, pages from VOLATILE and EMPTY objects
5584 * are all accounted for in the "volatile" ledgers, so this does not
5585 * make any difference.
5586 * If we transitioned directly from NONVOLATILE to EMPTY,
5587 * vm_page_purgeable_count must have been updated when the object
5588 * was dequeued from its volatile queue and the purgeable ledgers
5589 * must have also been updated accordingly at that time (in
5590 * vm_object_purgable_control()).
5591 */
5592 if (object->purgable == VM_PURGABLE_VOLATILE) {
5593 unsigned int delta;
5594 assert(object->resident_page_count >=
5595 object->wired_page_count);
5596 delta = (object->resident_page_count -
5597 object->wired_page_count);
5598 if (delta != 0) {
5599 assert(vm_page_purgeable_count >=
5600 delta);
5601 OSAddAtomic(-delta,
5602 (SInt32 *)&vm_page_purgeable_count);
5603 }
5604 if (object->wired_page_count != 0) {
5605 assert(vm_page_purgeable_wired_count >=
5606 object->wired_page_count);
5607 OSAddAtomic(-object->wired_page_count,
5608 (SInt32 *)&vm_page_purgeable_wired_count);
5609 }
5610 VM_OBJECT_SET_PURGABLE(object, VM_PURGABLE_EMPTY);
5611 }
5612 assert(object->purgable == VM_PURGABLE_EMPTY);
5613
5614 object_page_count = object->resident_page_count;
5615
5616 vm_object_reap_pages(object, REAP_PURGEABLE);
5617
5618 if (object->resident_page_count >= object_page_count) {
5619 total_purged_pgcount = 0;
5620 } else {
5621 total_purged_pgcount = object_page_count - object->resident_page_count;
5622 }
5623
5624 if (object->pager != NULL) {
5625 assert(VM_CONFIG_COMPRESSOR_IS_PRESENT);
5626
5627 if (object->activity_in_progress == 0 &&
5628 object->paging_in_progress == 0) {
5629 /*
5630 * Also reap any memory coming from this object
5631 * in the VM compressor.
5632 *
5633 * There are no operations in progress on the VM object
5634 * and no operation can start while we're holding the
5635 * VM object lock, so it's safe to reap the compressed
5636 * pages and update the page counts.
5637 */
5638 pgcount = vm_compressor_pager_get_count(mem_obj: object->pager);
5639 if (pgcount) {
5640 pgcount = vm_compressor_pager_reap_pages(mem_obj: object->pager, flags);
5641 vm_compressor_pager_count(mem_obj: object->pager,
5642 compressed_count_delta: -pgcount,
5643 FALSE, /* shared */
5644 object);
5645 vm_object_owner_compressed_update(object,
5646 delta: -pgcount);
5647 }
5648 if (!(flags & C_DONT_BLOCK)) {
5649 assert(vm_compressor_pager_get_count(object->pager)
5650 == 0);
5651 }
5652 } else {
5653 /*
5654 * There's some kind of paging activity in progress
5655 * for this object, which could result in a page
5656 * being compressed or decompressed, possibly while
5657 * the VM object is not locked, so it could race
5658 * with us.
5659 *
5660 * We can't really synchronize this without possibly
5661 * causing a deadlock when the compressor needs to
5662 * allocate or free memory while compressing or
5663 * decompressing a page from a purgeable object
5664 * mapped in the kernel_map...
5665 *
5666 * So let's not attempt to purge the compressor
5667 * pager if there's any kind of operation in
5668 * progress on the VM object.
5669 */
5670 skipped_object = TRUE;
5671 }
5672 }
5673
5674 vm_object_lock_assert_exclusive(object);
5675
5676 total_purged_pgcount += pgcount;
5677
5678 KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, (MACHDBG_CODE(DBG_MACH_VM, OBJECT_PURGE_ONE)),
5679 VM_KERNEL_UNSLIDE_OR_PERM(object), /* purged object */
5680 object_page_count,
5681 total_purged_pgcount,
5682 skipped_object,
5683 0);
5684
5685 return total_purged_pgcount;
5686}
5687
5688
5689/*
5690 * vm_object_purgeable_control() allows the caller to control and investigate the
5691 * state of a purgeable object. A purgeable object is created via a call to
5692 * vm_allocate() with VM_FLAGS_PURGABLE specified. A purgeable object will
5693 * never be coalesced with any other object -- even other purgeable objects --
5694 * and will thus always remain a distinct object. A purgeable object has
5695 * special semantics when its reference count is exactly 1. If its reference
5696 * count is greater than 1, then a purgeable object will behave like a normal
5697 * object and attempts to use this interface will result in an error return
5698 * of KERN_INVALID_ARGUMENT.
5699 *
5700 * A purgeable object may be put into a "volatile" state which will make the
5701 * object's pages elligable for being reclaimed without paging to backing
5702 * store if the system runs low on memory. If the pages in a volatile
5703 * purgeable object are reclaimed, the purgeable object is said to have been
5704 * "emptied." When a purgeable object is emptied the system will reclaim as
5705 * many pages from the object as it can in a convenient manner (pages already
5706 * en route to backing store or busy for other reasons are left as is). When
5707 * a purgeable object is made volatile, its pages will generally be reclaimed
5708 * before other pages in the application's working set. This semantic is
5709 * generally used by applications which can recreate the data in the object
5710 * faster than it can be paged in. One such example might be media assets
5711 * which can be reread from a much faster RAID volume.
5712 *
5713 * A purgeable object may be designated as "non-volatile" which means it will
5714 * behave like all other objects in the system with pages being written to and
5715 * read from backing store as needed to satisfy system memory needs. If the
5716 * object was emptied before the object was made non-volatile, that fact will
5717 * be returned as the old state of the purgeable object (see
5718 * VM_PURGABLE_SET_STATE below). In this case, any pages of the object which
5719 * were reclaimed as part of emptying the object will be refaulted in as
5720 * zero-fill on demand. It is up to the application to note that an object
5721 * was emptied and recreate the objects contents if necessary. When a
5722 * purgeable object is made non-volatile, its pages will generally not be paged
5723 * out to backing store in the immediate future. A purgeable object may also
5724 * be manually emptied.
5725 *
5726 * Finally, the current state (non-volatile, volatile, volatile & empty) of a
5727 * volatile purgeable object may be queried at any time. This information may
5728 * be used as a control input to let the application know when the system is
5729 * experiencing memory pressure and is reclaiming memory.
5730 *
5731 * The specified address may be any address within the purgeable object. If
5732 * the specified address does not represent any object in the target task's
5733 * virtual address space, then KERN_INVALID_ADDRESS will be returned. If the
5734 * object containing the specified address is not a purgeable object, then
5735 * KERN_INVALID_ARGUMENT will be returned. Otherwise, KERN_SUCCESS will be
5736 * returned.
5737 *
5738 * The control parameter may be any one of VM_PURGABLE_SET_STATE or
5739 * VM_PURGABLE_GET_STATE. For VM_PURGABLE_SET_STATE, the in/out parameter
5740 * state is used to set the new state of the purgeable object and return its
5741 * old state. For VM_PURGABLE_GET_STATE, the current state of the purgeable
5742 * object is returned in the parameter state.
5743 *
5744 * The in/out parameter state may be one of VM_PURGABLE_NONVOLATILE,
5745 * VM_PURGABLE_VOLATILE or VM_PURGABLE_EMPTY. These, respectively, represent
5746 * the non-volatile, volatile and volatile/empty states described above.
5747 * Setting the state of a purgeable object to VM_PURGABLE_EMPTY will
5748 * immediately reclaim as many pages in the object as can be conveniently
5749 * collected (some may have already been written to backing store or be
5750 * otherwise busy).
5751 *
5752 * The process of making a purgeable object non-volatile and determining its
5753 * previous state is atomic. Thus, if a purgeable object is made
5754 * VM_PURGABLE_NONVOLATILE and the old state is returned as
5755 * VM_PURGABLE_VOLATILE, then the purgeable object's previous contents are
5756 * completely intact and will remain so until the object is made volatile
5757 * again. If the old state is returned as VM_PURGABLE_EMPTY then the object
5758 * was reclaimed while it was in a volatile state and its previous contents
5759 * have been lost.
5760 */
5761/*
5762 * The object must be locked.
5763 */
5764kern_return_t
5765vm_object_purgable_control(
5766 vm_object_t object,
5767 vm_purgable_t control,
5768 int *state)
5769{
5770 int old_state;
5771 int new_state;
5772
5773 if (object == VM_OBJECT_NULL) {
5774 /*
5775 * Object must already be present or it can't be purgeable.
5776 */
5777 return KERN_INVALID_ARGUMENT;
5778 }
5779
5780 vm_object_lock_assert_exclusive(object);
5781
5782 /*
5783 * Get current state of the purgeable object.
5784 */
5785 old_state = object->purgable;
5786 if (old_state == VM_PURGABLE_DENY) {
5787 return KERN_INVALID_ARGUMENT;
5788 }
5789
5790 /* purgeable cant have delayed copies - now or in the future */
5791 assert(object->vo_copy == VM_OBJECT_NULL);
5792 assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
5793
5794 /*
5795 * Execute the desired operation.
5796 */
5797 if (control == VM_PURGABLE_GET_STATE) {
5798 *state = old_state;
5799 return KERN_SUCCESS;
5800 }
5801
5802 if (control == VM_PURGABLE_SET_STATE &&
5803 object->purgeable_only_by_kernel) {
5804 return KERN_PROTECTION_FAILURE;
5805 }
5806
5807 if (control != VM_PURGABLE_SET_STATE &&
5808 control != VM_PURGABLE_SET_STATE_FROM_KERNEL) {
5809 return KERN_INVALID_ARGUMENT;
5810 }
5811
5812 if ((*state) & VM_PURGABLE_DEBUG_EMPTY) {
5813 object->volatile_empty = TRUE;
5814 }
5815 if ((*state) & VM_PURGABLE_DEBUG_FAULT) {
5816 object->volatile_fault = TRUE;
5817 }
5818
5819 new_state = *state & VM_PURGABLE_STATE_MASK;
5820 if (new_state == VM_PURGABLE_VOLATILE) {
5821 if (old_state == VM_PURGABLE_EMPTY) {
5822 /* what's been emptied must stay empty */
5823 new_state = VM_PURGABLE_EMPTY;
5824 }
5825 if (object->volatile_empty) {
5826 /* debugging mode: go straight to empty */
5827 new_state = VM_PURGABLE_EMPTY;
5828 }
5829 }
5830
5831 switch (new_state) {
5832 case VM_PURGABLE_DENY:
5833 /*
5834 * Attempting to convert purgeable memory to non-purgeable:
5835 * not allowed.
5836 */
5837 return KERN_INVALID_ARGUMENT;
5838 case VM_PURGABLE_NONVOLATILE:
5839 VM_OBJECT_SET_PURGABLE(object, value: new_state);
5840
5841 if (old_state == VM_PURGABLE_VOLATILE) {
5842 unsigned int delta;
5843
5844 assert(object->resident_page_count >=
5845 object->wired_page_count);
5846 delta = (object->resident_page_count -
5847 object->wired_page_count);
5848
5849 assert(vm_page_purgeable_count >= delta);
5850
5851 if (delta != 0) {
5852 OSAddAtomic(-delta,
5853 (SInt32 *)&vm_page_purgeable_count);
5854 }
5855 if (object->wired_page_count != 0) {
5856 assert(vm_page_purgeable_wired_count >=
5857 object->wired_page_count);
5858 OSAddAtomic(-object->wired_page_count,
5859 (SInt32 *)&vm_page_purgeable_wired_count);
5860 }
5861
5862 vm_page_lock_queues();
5863
5864 /* object should be on a queue */
5865 assert(object->objq.next != NULL &&
5866 object->objq.prev != NULL);
5867 purgeable_q_t queue;
5868
5869 /*
5870 * Move object from its volatile queue to the
5871 * non-volatile queue...
5872 */
5873 queue = vm_purgeable_object_remove(object);
5874 assert(queue);
5875
5876 if (object->purgeable_when_ripe) {
5877 vm_purgeable_token_delete_last(queue);
5878 }
5879 assert(queue->debug_count_objects >= 0);
5880
5881 vm_page_unlock_queues();
5882 }
5883 if (old_state == VM_PURGABLE_VOLATILE ||
5884 old_state == VM_PURGABLE_EMPTY) {
5885 /*
5886 * Transfer the object's pages from the volatile to
5887 * non-volatile ledgers.
5888 */
5889 vm_purgeable_accounting(object, VM_PURGABLE_VOLATILE);
5890 }
5891
5892 break;
5893
5894 case VM_PURGABLE_VOLATILE:
5895 if (object->volatile_fault) {
5896 vm_page_t p;
5897 int refmod;
5898
5899 vm_page_queue_iterate(&object->memq, p, vmp_listq) {
5900 if (p->vmp_busy ||
5901 VM_PAGE_WIRED(p) ||
5902 p->vmp_fictitious) {
5903 continue;
5904 }
5905 refmod = pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: p));
5906 if ((refmod & VM_MEM_MODIFIED) &&
5907 !p->vmp_dirty) {
5908 SET_PAGE_DIRTY(p, FALSE);
5909 }
5910 }
5911 }
5912
5913 assert(old_state != VM_PURGABLE_EMPTY);
5914
5915 purgeable_q_t queue;
5916
5917 /* find the correct queue */
5918 if ((*state & VM_PURGABLE_ORDERING_MASK) == VM_PURGABLE_ORDERING_OBSOLETE) {
5919 queue = &purgeable_queues[PURGEABLE_Q_TYPE_OBSOLETE];
5920 } else {
5921 if ((*state & VM_PURGABLE_BEHAVIOR_MASK) == VM_PURGABLE_BEHAVIOR_FIFO) {
5922 queue = &purgeable_queues[PURGEABLE_Q_TYPE_FIFO];
5923 } else {
5924 queue = &purgeable_queues[PURGEABLE_Q_TYPE_LIFO];
5925 }
5926 }
5927
5928 if (old_state == VM_PURGABLE_NONVOLATILE ||
5929 old_state == VM_PURGABLE_EMPTY) {
5930 unsigned int delta;
5931
5932 if ((*state & VM_PURGABLE_NO_AGING_MASK) ==
5933 VM_PURGABLE_NO_AGING) {
5934 VM_OBJECT_SET_PURGEABLE_WHEN_RIPE(object, FALSE);
5935 } else {
5936 VM_OBJECT_SET_PURGEABLE_WHEN_RIPE(object, TRUE);
5937 }
5938
5939 if (object->purgeable_when_ripe) {
5940 kern_return_t result;
5941
5942 /* try to add token... this can fail */
5943 vm_page_lock_queues();
5944
5945 result = vm_purgeable_token_add(queue);
5946 if (result != KERN_SUCCESS) {
5947 vm_page_unlock_queues();
5948 return result;
5949 }
5950 vm_page_unlock_queues();
5951 }
5952
5953 assert(object->resident_page_count >=
5954 object->wired_page_count);
5955 delta = (object->resident_page_count -
5956 object->wired_page_count);
5957
5958 if (delta != 0) {
5959 OSAddAtomic(delta,
5960 &vm_page_purgeable_count);
5961 }
5962 if (object->wired_page_count != 0) {
5963 OSAddAtomic(object->wired_page_count,
5964 &vm_page_purgeable_wired_count);
5965 }
5966
5967 VM_OBJECT_SET_PURGABLE(object, value: new_state);
5968
5969 /* object should be on "non-volatile" queue */
5970 assert(object->objq.next != NULL);
5971 assert(object->objq.prev != NULL);
5972 } else if (old_state == VM_PURGABLE_VOLATILE) {
5973 purgeable_q_t old_queue;
5974 boolean_t purgeable_when_ripe;
5975
5976 /*
5977 * if reassigning priorities / purgeable groups, we don't change the
5978 * token queue. So moving priorities will not make pages stay around longer.
5979 * Reasoning is that the algorithm gives most priority to the most important
5980 * object. If a new token is added, the most important object' priority is boosted.
5981 * This biases the system already for purgeable queues that move a lot.
5982 * It doesn't seem more biasing is neccessary in this case, where no new object is added.
5983 */
5984 assert(object->objq.next != NULL && object->objq.prev != NULL); /* object should be on a queue */
5985
5986 old_queue = vm_purgeable_object_remove(object);
5987 assert(old_queue);
5988
5989 if ((*state & VM_PURGABLE_NO_AGING_MASK) ==
5990 VM_PURGABLE_NO_AGING) {
5991 purgeable_when_ripe = FALSE;
5992 } else {
5993 purgeable_when_ripe = TRUE;
5994 }
5995
5996 if (old_queue != queue ||
5997 (purgeable_when_ripe !=
5998 object->purgeable_when_ripe)) {
5999 kern_return_t result;
6000
6001 /* Changing queue. Have to move token. */
6002 vm_page_lock_queues();
6003 if (object->purgeable_when_ripe) {
6004 vm_purgeable_token_delete_last(queue: old_queue);
6005 }
6006 VM_OBJECT_SET_PURGEABLE_WHEN_RIPE(object, value: purgeable_when_ripe);
6007 if (object->purgeable_when_ripe) {
6008 result = vm_purgeable_token_add(queue);
6009 assert(result == KERN_SUCCESS); /* this should never fail since we just freed a token */
6010 }
6011 vm_page_unlock_queues();
6012 }
6013 }
6014 ;
6015 vm_purgeable_object_add(object, queue, group: (*state & VM_VOLATILE_GROUP_MASK) >> VM_VOLATILE_GROUP_SHIFT );
6016 if (old_state == VM_PURGABLE_NONVOLATILE) {
6017 vm_purgeable_accounting(object,
6018 VM_PURGABLE_NONVOLATILE);
6019 }
6020
6021 assert(queue->debug_count_objects >= 0);
6022
6023 break;
6024
6025
6026 case VM_PURGABLE_EMPTY:
6027 if (object->volatile_fault) {
6028 vm_page_t p;
6029 int refmod;
6030
6031 vm_page_queue_iterate(&object->memq, p, vmp_listq) {
6032 if (p->vmp_busy ||
6033 VM_PAGE_WIRED(p) ||
6034 p->vmp_fictitious) {
6035 continue;
6036 }
6037 refmod = pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: p));
6038 if ((refmod & VM_MEM_MODIFIED) &&
6039 !p->vmp_dirty) {
6040 SET_PAGE_DIRTY(p, FALSE);
6041 }
6042 }
6043 }
6044
6045 if (old_state == VM_PURGABLE_VOLATILE) {
6046 purgeable_q_t old_queue;
6047
6048 /* object should be on a queue */
6049 assert(object->objq.next != NULL &&
6050 object->objq.prev != NULL);
6051
6052 old_queue = vm_purgeable_object_remove(object);
6053 assert(old_queue);
6054 if (object->purgeable_when_ripe) {
6055 vm_page_lock_queues();
6056 vm_purgeable_token_delete_first(queue: old_queue);
6057 vm_page_unlock_queues();
6058 }
6059 }
6060
6061 if (old_state == VM_PURGABLE_NONVOLATILE) {
6062 /*
6063 * This object's pages were previously accounted as
6064 * "non-volatile" and now need to be accounted as
6065 * "volatile".
6066 */
6067 vm_purgeable_accounting(object,
6068 VM_PURGABLE_NONVOLATILE);
6069 /*
6070 * Set to VM_PURGABLE_EMPTY because the pages are no
6071 * longer accounted in the "non-volatile" ledger
6072 * and are also not accounted for in
6073 * "vm_page_purgeable_count".
6074 */
6075 VM_OBJECT_SET_PURGABLE(object, VM_PURGABLE_EMPTY);
6076 }
6077
6078 (void) vm_object_purge(object, flags: 0);
6079 assert(object->purgable == VM_PURGABLE_EMPTY);
6080
6081 break;
6082 }
6083
6084 *state = old_state;
6085
6086 vm_object_lock_assert_exclusive(object);
6087
6088 return KERN_SUCCESS;
6089}
6090
6091kern_return_t
6092vm_object_get_page_counts(
6093 vm_object_t object,
6094 vm_object_offset_t offset,
6095 vm_object_size_t size,
6096 unsigned int *resident_page_count,
6097 unsigned int *dirty_page_count)
6098{
6099 kern_return_t kr = KERN_SUCCESS;
6100 boolean_t count_dirty_pages = FALSE;
6101 vm_page_t p = VM_PAGE_NULL;
6102 unsigned int local_resident_count = 0;
6103 unsigned int local_dirty_count = 0;
6104 vm_object_offset_t cur_offset = 0;
6105 vm_object_offset_t end_offset = 0;
6106
6107 if (object == VM_OBJECT_NULL) {
6108 return KERN_INVALID_ARGUMENT;
6109 }
6110
6111
6112 cur_offset = offset;
6113
6114 end_offset = offset + size;
6115
6116 vm_object_lock_assert_exclusive(object);
6117
6118 if (dirty_page_count != NULL) {
6119 count_dirty_pages = TRUE;
6120 }
6121
6122 if (resident_page_count != NULL && count_dirty_pages == FALSE) {
6123 /*
6124 * Fast path when:
6125 * - we only want the resident page count, and,
6126 * - the entire object is exactly covered by the request.
6127 */
6128 if (offset == 0 && (object->vo_size == size)) {
6129 *resident_page_count = object->resident_page_count;
6130 goto out;
6131 }
6132 }
6133
6134 if (object->resident_page_count <= (size >> PAGE_SHIFT)) {
6135 vm_page_queue_iterate(&object->memq, p, vmp_listq) {
6136 if (p->vmp_offset >= cur_offset && p->vmp_offset < end_offset) {
6137 local_resident_count++;
6138
6139 if (count_dirty_pages) {
6140 if (p->vmp_dirty || (p->vmp_wpmapped && pmap_is_modified(pn: VM_PAGE_GET_PHYS_PAGE(m: p)))) {
6141 local_dirty_count++;
6142 }
6143 }
6144 }
6145 }
6146 } else {
6147 for (cur_offset = offset; cur_offset < end_offset; cur_offset += PAGE_SIZE_64) {
6148 p = vm_page_lookup(object, offset: cur_offset);
6149
6150 if (p != VM_PAGE_NULL) {
6151 local_resident_count++;
6152
6153 if (count_dirty_pages) {
6154 if (p->vmp_dirty || (p->vmp_wpmapped && pmap_is_modified(pn: VM_PAGE_GET_PHYS_PAGE(m: p)))) {
6155 local_dirty_count++;
6156 }
6157 }
6158 }
6159 }
6160 }
6161
6162 if (resident_page_count != NULL) {
6163 *resident_page_count = local_resident_count;
6164 }
6165
6166 if (dirty_page_count != NULL) {
6167 *dirty_page_count = local_dirty_count;
6168 }
6169
6170out:
6171 return kr;
6172}
6173
6174
6175/*
6176 * vm_object_reference:
6177 *
6178 * Gets another reference to the given object.
6179 */
6180#ifdef vm_object_reference
6181#undef vm_object_reference
6182#endif
6183__private_extern__ void
6184vm_object_reference(
6185 vm_object_t object)
6186{
6187 if (object == VM_OBJECT_NULL) {
6188 return;
6189 }
6190
6191 vm_object_lock(object);
6192 assert(object->ref_count > 0);
6193 vm_object_reference_locked(object);
6194 vm_object_unlock(object);
6195}
6196
6197/*
6198 * vm_object_transpose
6199 *
6200 * This routine takes two VM objects of the same size and exchanges
6201 * their backing store.
6202 * The objects should be "quiesced" via a UPL operation with UPL_SET_IO_WIRE
6203 * and UPL_BLOCK_ACCESS if they are referenced anywhere.
6204 *
6205 * The VM objects must not be locked by caller.
6206 */
6207unsigned int vm_object_transpose_count = 0;
6208kern_return_t
6209vm_object_transpose(
6210 vm_object_t object1,
6211 vm_object_t object2,
6212 vm_object_size_t transpose_size)
6213{
6214 vm_object_t tmp_object;
6215 kern_return_t retval;
6216 boolean_t object1_locked, object2_locked;
6217 vm_page_t page;
6218 vm_object_offset_t page_offset;
6219
6220 tmp_object = VM_OBJECT_NULL;
6221 object1_locked = FALSE; object2_locked = FALSE;
6222
6223 if (object1 == object2 ||
6224 object1 == VM_OBJECT_NULL ||
6225 object2 == VM_OBJECT_NULL) {
6226 /*
6227 * If the 2 VM objects are the same, there's
6228 * no point in exchanging their backing store.
6229 */
6230 retval = KERN_INVALID_VALUE;
6231 goto done;
6232 }
6233
6234 /*
6235 * Since we need to lock both objects at the same time,
6236 * make sure we always lock them in the same order to
6237 * avoid deadlocks.
6238 */
6239 if (object1 > object2) {
6240 tmp_object = object1;
6241 object1 = object2;
6242 object2 = tmp_object;
6243 }
6244
6245 /*
6246 * Allocate a temporary VM object to hold object1's contents
6247 * while we copy object2 to object1.
6248 */
6249 tmp_object = vm_object_allocate(size: transpose_size);
6250 vm_object_lock(tmp_object);
6251 VM_OBJECT_SET_CAN_PERSIST(object: tmp_object, FALSE);
6252
6253
6254 /*
6255 * Grab control of the 1st VM object.
6256 */
6257 vm_object_lock(object1);
6258 object1_locked = TRUE;
6259 if (!object1->alive || object1->terminating ||
6260 object1->vo_copy || object1->shadow || object1->shadowed ||
6261 object1->purgable != VM_PURGABLE_DENY) {
6262 /*
6263 * We don't deal with copy or shadow objects (yet).
6264 */
6265 retval = KERN_INVALID_VALUE;
6266 goto done;
6267 }
6268 /*
6269 * We're about to mess with the object's backing store and
6270 * taking a "paging_in_progress" reference wouldn't be enough
6271 * to prevent any paging activity on this object, so the caller should
6272 * have "quiesced" the objects beforehand, via a UPL operation with
6273 * UPL_SET_IO_WIRE (to make sure all the pages are there and wired)
6274 * and UPL_BLOCK_ACCESS (to mark the pages "busy").
6275 *
6276 * Wait for any paging operation to complete (but only paging, not
6277 * other kind of activities not linked to the pager). After we're
6278 * statisfied that there's no more paging in progress, we keep the
6279 * object locked, to guarantee that no one tries to access its pager.
6280 */
6281 vm_object_paging_only_wait(object1, THREAD_UNINT);
6282
6283 /*
6284 * Same as above for the 2nd object...
6285 */
6286 vm_object_lock(object2);
6287 object2_locked = TRUE;
6288 if (!object2->alive || object2->terminating ||
6289 object2->vo_copy || object2->shadow || object2->shadowed ||
6290 object2->purgable != VM_PURGABLE_DENY) {
6291 retval = KERN_INVALID_VALUE;
6292 goto done;
6293 }
6294 vm_object_paging_only_wait(object2, THREAD_UNINT);
6295
6296
6297 if (object1->vo_size != object2->vo_size ||
6298 object1->vo_size != transpose_size) {
6299 /*
6300 * If the 2 objects don't have the same size, we can't
6301 * exchange their backing stores or one would overflow.
6302 * If their size doesn't match the caller's
6303 * "transpose_size", we can't do it either because the
6304 * transpose operation will affect the entire span of
6305 * the objects.
6306 */
6307 retval = KERN_INVALID_VALUE;
6308 goto done;
6309 }
6310
6311
6312 /*
6313 * Transpose the lists of resident pages.
6314 * This also updates the resident_page_count and the memq_hint.
6315 */
6316 if (object1->phys_contiguous || vm_page_queue_empty(&object1->memq)) {
6317 /*
6318 * No pages in object1, just transfer pages
6319 * from object2 to object1. No need to go through
6320 * an intermediate object.
6321 */
6322 while (!vm_page_queue_empty(&object2->memq)) {
6323 page = (vm_page_t) vm_page_queue_first(&object2->memq);
6324 vm_page_rename(page, new_object: object1, new_offset: page->vmp_offset);
6325 }
6326 assert(vm_page_queue_empty(&object2->memq));
6327 } else if (object2->phys_contiguous || vm_page_queue_empty(&object2->memq)) {
6328 /*
6329 * No pages in object2, just transfer pages
6330 * from object1 to object2. No need to go through
6331 * an intermediate object.
6332 */
6333 while (!vm_page_queue_empty(&object1->memq)) {
6334 page = (vm_page_t) vm_page_queue_first(&object1->memq);
6335 vm_page_rename(page, new_object: object2, new_offset: page->vmp_offset);
6336 }
6337 assert(vm_page_queue_empty(&object1->memq));
6338 } else {
6339 /* transfer object1's pages to tmp_object */
6340 while (!vm_page_queue_empty(&object1->memq)) {
6341 page = (vm_page_t) vm_page_queue_first(&object1->memq);
6342 page_offset = page->vmp_offset;
6343 vm_page_remove(page, TRUE);
6344 page->vmp_offset = page_offset;
6345 vm_page_queue_enter(&tmp_object->memq, page, vmp_listq);
6346 }
6347 assert(vm_page_queue_empty(&object1->memq));
6348 /* transfer object2's pages to object1 */
6349 while (!vm_page_queue_empty(&object2->memq)) {
6350 page = (vm_page_t) vm_page_queue_first(&object2->memq);
6351 vm_page_rename(page, new_object: object1, new_offset: page->vmp_offset);
6352 }
6353 assert(vm_page_queue_empty(&object2->memq));
6354 /* transfer tmp_object's pages to object2 */
6355 while (!vm_page_queue_empty(&tmp_object->memq)) {
6356 page = (vm_page_t) vm_page_queue_first(&tmp_object->memq);
6357 vm_page_queue_remove(&tmp_object->memq, page, vmp_listq);
6358 vm_page_insert(page, object: object2, offset: page->vmp_offset);
6359 }
6360 assert(vm_page_queue_empty(&tmp_object->memq));
6361 }
6362
6363#define __TRANSPOSE_FIELD(field) \
6364MACRO_BEGIN \
6365 tmp_object->field = object1->field; \
6366 object1->field = object2->field; \
6367 object2->field = tmp_object->field; \
6368MACRO_END
6369
6370 /* "Lock" refers to the object not its contents */
6371 /* "size" should be identical */
6372 assert(object1->vo_size == object2->vo_size);
6373 /* "memq_hint" was updated above when transposing pages */
6374 /* "ref_count" refers to the object not its contents */
6375 assert(object1->ref_count >= 1);
6376 assert(object2->ref_count >= 1);
6377 /* "resident_page_count" was updated above when transposing pages */
6378 /* "wired_page_count" was updated above when transposing pages */
6379#if !VM_TAG_ACTIVE_UPDATE
6380 /* "wired_objq" was dealt with along with "wired_page_count" */
6381#endif /* ! VM_TAG_ACTIVE_UPDATE */
6382 /* "reusable_page_count" was updated above when transposing pages */
6383 /* there should be no "copy" */
6384 assert(!object1->vo_copy);
6385 assert(!object2->vo_copy);
6386 /* there should be no "shadow" */
6387 assert(!object1->shadow);
6388 assert(!object2->shadow);
6389 __TRANSPOSE_FIELD(vo_shadow_offset); /* used by phys_contiguous objects */
6390 __TRANSPOSE_FIELD(pager);
6391 __TRANSPOSE_FIELD(paging_offset);
6392 __TRANSPOSE_FIELD(pager_control);
6393 /* update the memory_objects' pointers back to the VM objects */
6394 if (object1->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
6395 memory_object_control_collapse(control: &object1->pager_control,
6396 object: object1);
6397 }
6398 if (object2->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
6399 memory_object_control_collapse(control: &object2->pager_control,
6400 object: object2);
6401 }
6402 __TRANSPOSE_FIELD(copy_strategy);
6403 /* "paging_in_progress" refers to the object not its contents */
6404 assert(!object1->paging_in_progress);
6405 assert(!object2->paging_in_progress);
6406 assert(object1->activity_in_progress);
6407 assert(object2->activity_in_progress);
6408 /* "all_wanted" refers to the object not its contents */
6409 __TRANSPOSE_FIELD(pager_created);
6410 __TRANSPOSE_FIELD(pager_initialized);
6411 __TRANSPOSE_FIELD(pager_ready);
6412 __TRANSPOSE_FIELD(pager_trusted);
6413 __TRANSPOSE_FIELD(can_persist);
6414 __TRANSPOSE_FIELD(internal);
6415 __TRANSPOSE_FIELD(private);
6416 __TRANSPOSE_FIELD(pageout);
6417 /* "alive" should be set */
6418 assert(object1->alive);
6419 assert(object2->alive);
6420 /* "purgeable" should be non-purgeable */
6421 assert(object1->purgable == VM_PURGABLE_DENY);
6422 assert(object2->purgable == VM_PURGABLE_DENY);
6423 /* "shadowed" refers to the the object not its contents */
6424 __TRANSPOSE_FIELD(purgeable_when_ripe);
6425 __TRANSPOSE_FIELD(true_share);
6426 /* "terminating" should not be set */
6427 assert(!object1->terminating);
6428 assert(!object2->terminating);
6429 /* transfer "named" reference if needed */
6430 if (object1->named && !object2->named) {
6431 assert(object1->ref_count >= 2);
6432 assert(object2->ref_count >= 1);
6433 object1->ref_count--;
6434 object2->ref_count++;
6435 } else if (!object1->named && object2->named) {
6436 assert(object1->ref_count >= 1);
6437 assert(object2->ref_count >= 2);
6438 object1->ref_count++;
6439 object2->ref_count--;
6440 }
6441 __TRANSPOSE_FIELD(named);
6442 /* "shadow_severed" refers to the object not its contents */
6443 __TRANSPOSE_FIELD(phys_contiguous);
6444 __TRANSPOSE_FIELD(nophyscache);
6445 __TRANSPOSE_FIELD(no_pager_reason);
6446 /* "cached_list.next" points to transposed object */
6447 object1->cached_list.next = (queue_entry_t) object2;
6448 object2->cached_list.next = (queue_entry_t) object1;
6449 /* "cached_list.prev" should be NULL */
6450 assert(object1->cached_list.prev == NULL);
6451 assert(object2->cached_list.prev == NULL);
6452 __TRANSPOSE_FIELD(last_alloc);
6453 __TRANSPOSE_FIELD(sequential);
6454 __TRANSPOSE_FIELD(pages_created);
6455 __TRANSPOSE_FIELD(pages_used);
6456 __TRANSPOSE_FIELD(scan_collisions);
6457 __TRANSPOSE_FIELD(cow_hint);
6458 __TRANSPOSE_FIELD(wimg_bits);
6459 __TRANSPOSE_FIELD(set_cache_attr);
6460 __TRANSPOSE_FIELD(code_signed);
6461 object1->transposed = TRUE;
6462 object2->transposed = TRUE;
6463 __TRANSPOSE_FIELD(mapping_in_progress);
6464 __TRANSPOSE_FIELD(volatile_empty);
6465 __TRANSPOSE_FIELD(volatile_fault);
6466 __TRANSPOSE_FIELD(all_reusable);
6467 assert(object1->blocked_access);
6468 assert(object2->blocked_access);
6469 __TRANSPOSE_FIELD(set_cache_attr);
6470 assert(!object1->object_is_shared_cache);
6471 assert(!object2->object_is_shared_cache);
6472 /* ignore purgeable_queue_type and purgeable_queue_group */
6473 assert(!object1->io_tracking);
6474 assert(!object2->io_tracking);
6475#if VM_OBJECT_ACCESS_TRACKING
6476 assert(!object1->access_tracking);
6477 assert(!object2->access_tracking);
6478#endif /* VM_OBJECT_ACCESS_TRACKING */
6479 __TRANSPOSE_FIELD(no_tag_update);
6480#if CONFIG_SECLUDED_MEMORY
6481 assert(!object1->eligible_for_secluded);
6482 assert(!object2->eligible_for_secluded);
6483 assert(!object1->can_grab_secluded);
6484 assert(!object2->can_grab_secluded);
6485#else /* CONFIG_SECLUDED_MEMORY */
6486 assert(object1->__object3_unused_bits == 0);
6487 assert(object2->__object3_unused_bits == 0);
6488#endif /* CONFIG_SECLUDED_MEMORY */
6489#if UPL_DEBUG
6490 /* "uplq" refers to the object not its contents (see upl_transpose()) */
6491#endif
6492 assert((object1->purgable == VM_PURGABLE_DENY) || (object1->objq.next == NULL));
6493 assert((object1->purgable == VM_PURGABLE_DENY) || (object1->objq.prev == NULL));
6494 assert((object2->purgable == VM_PURGABLE_DENY) || (object2->objq.next == NULL));
6495 assert((object2->purgable == VM_PURGABLE_DENY) || (object2->objq.prev == NULL));
6496
6497#undef __TRANSPOSE_FIELD
6498
6499 retval = KERN_SUCCESS;
6500
6501done:
6502 /*
6503 * Cleanup.
6504 */
6505 if (tmp_object != VM_OBJECT_NULL) {
6506 vm_object_unlock(tmp_object);
6507 /*
6508 * Re-initialize the temporary object to avoid
6509 * deallocating a real pager.
6510 */
6511 _vm_object_allocate(size: transpose_size, object: tmp_object);
6512 vm_object_deallocate(object: tmp_object);
6513 tmp_object = VM_OBJECT_NULL;
6514 }
6515
6516 if (object1_locked) {
6517 vm_object_unlock(object1);
6518 object1_locked = FALSE;
6519 }
6520 if (object2_locked) {
6521 vm_object_unlock(object2);
6522 object2_locked = FALSE;
6523 }
6524
6525 vm_object_transpose_count++;
6526
6527 return retval;
6528}
6529
6530
6531/*
6532 * vm_object_cluster_size
6533 *
6534 * Determine how big a cluster we should issue an I/O for...
6535 *
6536 * Inputs: *start == offset of page needed
6537 * *length == maximum cluster pager can handle
6538 * Outputs: *start == beginning offset of cluster
6539 * *length == length of cluster to try
6540 *
6541 * The original *start will be encompassed by the cluster
6542 *
6543 */
6544extern int speculative_reads_disabled;
6545
6546/*
6547 * Try to always keep these values an even multiple of PAGE_SIZE. We use these values
6548 * to derive min_ph_bytes and max_ph_bytes (IMP: bytes not # of pages) and expect those values to
6549 * always be page-aligned. The derivation could involve operations (e.g. division)
6550 * that could give us non-page-size aligned values if we start out with values that
6551 * are odd multiples of PAGE_SIZE.
6552 */
6553#if !XNU_TARGET_OS_OSX
6554unsigned int preheat_max_bytes = (1024 * 512);
6555#else /* !XNU_TARGET_OS_OSX */
6556unsigned int preheat_max_bytes = MAX_UPL_TRANSFER_BYTES;
6557#endif /* !XNU_TARGET_OS_OSX */
6558unsigned int preheat_min_bytes = (1024 * 32);
6559
6560
6561__private_extern__ void
6562vm_object_cluster_size(vm_object_t object, vm_object_offset_t *start,
6563 vm_size_t *length, vm_object_fault_info_t fault_info, uint32_t *io_streaming)
6564{
6565 vm_size_t pre_heat_size;
6566 vm_size_t tail_size;
6567 vm_size_t head_size;
6568 vm_size_t max_length;
6569 vm_size_t cluster_size;
6570 vm_object_offset_t object_size;
6571 vm_object_offset_t orig_start;
6572 vm_object_offset_t target_start;
6573 vm_object_offset_t offset;
6574 vm_behavior_t behavior;
6575 boolean_t look_behind = TRUE;
6576 boolean_t look_ahead = TRUE;
6577 boolean_t isSSD = FALSE;
6578 uint32_t throttle_limit;
6579 int sequential_run;
6580 int sequential_behavior = VM_BEHAVIOR_SEQUENTIAL;
6581 vm_size_t max_ph_size;
6582 vm_size_t min_ph_size;
6583
6584 assert( !(*length & PAGE_MASK));
6585 assert( !(*start & PAGE_MASK_64));
6586
6587 /*
6588 * remember maxiumum length of run requested
6589 */
6590 max_length = *length;
6591 /*
6592 * we'll always return a cluster size of at least
6593 * 1 page, since the original fault must always
6594 * be processed
6595 */
6596 *length = PAGE_SIZE;
6597 *io_streaming = 0;
6598
6599 if (speculative_reads_disabled || fault_info == NULL) {
6600 /*
6601 * no cluster... just fault the page in
6602 */
6603 return;
6604 }
6605 orig_start = *start;
6606 target_start = orig_start;
6607 cluster_size = round_page(x: fault_info->cluster_size);
6608 behavior = fault_info->behavior;
6609
6610 vm_object_lock(object);
6611
6612 if (object->pager == MEMORY_OBJECT_NULL) {
6613 goto out; /* pager is gone for this object, nothing more to do */
6614 }
6615 vnode_pager_get_isSSD(object->pager, &isSSD);
6616
6617 min_ph_size = round_page(x: preheat_min_bytes);
6618 max_ph_size = round_page(x: preheat_max_bytes);
6619
6620#if XNU_TARGET_OS_OSX
6621 if (isSSD) {
6622 min_ph_size /= 2;
6623 max_ph_size /= 8;
6624
6625 if (min_ph_size & PAGE_MASK_64) {
6626 min_ph_size = trunc_page(min_ph_size);
6627 }
6628
6629 if (max_ph_size & PAGE_MASK_64) {
6630 max_ph_size = trunc_page(max_ph_size);
6631 }
6632 }
6633#endif /* XNU_TARGET_OS_OSX */
6634
6635 if (min_ph_size < PAGE_SIZE) {
6636 min_ph_size = PAGE_SIZE;
6637 }
6638
6639 if (max_ph_size < PAGE_SIZE) {
6640 max_ph_size = PAGE_SIZE;
6641 } else if (max_ph_size > MAX_UPL_TRANSFER_BYTES) {
6642 max_ph_size = MAX_UPL_TRANSFER_BYTES;
6643 }
6644
6645 if (max_length > max_ph_size) {
6646 max_length = max_ph_size;
6647 }
6648
6649 if (max_length <= PAGE_SIZE) {
6650 goto out;
6651 }
6652
6653 if (object->internal) {
6654 object_size = object->vo_size;
6655 } else {
6656 vnode_pager_get_object_size(object->pager, &object_size);
6657 }
6658
6659 object_size = round_page_64(x: object_size);
6660
6661 if (orig_start >= object_size) {
6662 /*
6663 * fault occurred beyond the EOF...
6664 * we need to punt w/o changing the
6665 * starting offset
6666 */
6667 goto out;
6668 }
6669 if (object->pages_used > object->pages_created) {
6670 /*
6671 * must have wrapped our 32 bit counters
6672 * so reset
6673 */
6674 object->pages_used = object->pages_created = 0;
6675 }
6676 if ((sequential_run = object->sequential)) {
6677 if (sequential_run < 0) {
6678 sequential_behavior = VM_BEHAVIOR_RSEQNTL;
6679 sequential_run = 0 - sequential_run;
6680 } else {
6681 sequential_behavior = VM_BEHAVIOR_SEQUENTIAL;
6682 }
6683 }
6684 switch (behavior) {
6685 default:
6686 behavior = VM_BEHAVIOR_DEFAULT;
6687 OS_FALLTHROUGH;
6688
6689 case VM_BEHAVIOR_DEFAULT:
6690 if (object->internal && fault_info->user_tag == VM_MEMORY_STACK) {
6691 goto out;
6692 }
6693
6694 if (sequential_run >= (3 * PAGE_SIZE)) {
6695 pre_heat_size = sequential_run + PAGE_SIZE;
6696
6697 if (sequential_behavior == VM_BEHAVIOR_SEQUENTIAL) {
6698 look_behind = FALSE;
6699 } else {
6700 look_ahead = FALSE;
6701 }
6702
6703 *io_streaming = 1;
6704 } else {
6705 if (object->pages_created < (20 * (min_ph_size >> PAGE_SHIFT))) {
6706 /*
6707 * prime the pump
6708 */
6709 pre_heat_size = min_ph_size;
6710 } else {
6711 /*
6712 * Linear growth in PH size: The maximum size is max_length...
6713 * this cacluation will result in a size that is neither a
6714 * power of 2 nor a multiple of PAGE_SIZE... so round
6715 * it up to the nearest PAGE_SIZE boundary
6716 */
6717 pre_heat_size = (max_length * (uint64_t)object->pages_used) / object->pages_created;
6718
6719 if (pre_heat_size < min_ph_size) {
6720 pre_heat_size = min_ph_size;
6721 } else {
6722 pre_heat_size = round_page(x: pre_heat_size);
6723 }
6724 }
6725 }
6726 break;
6727
6728 case VM_BEHAVIOR_RANDOM:
6729 if ((pre_heat_size = cluster_size) <= PAGE_SIZE) {
6730 goto out;
6731 }
6732 break;
6733
6734 case VM_BEHAVIOR_SEQUENTIAL:
6735 if ((pre_heat_size = cluster_size) == 0) {
6736 pre_heat_size = sequential_run + PAGE_SIZE;
6737 }
6738 look_behind = FALSE;
6739 *io_streaming = 1;
6740
6741 break;
6742
6743 case VM_BEHAVIOR_RSEQNTL:
6744 if ((pre_heat_size = cluster_size) == 0) {
6745 pre_heat_size = sequential_run + PAGE_SIZE;
6746 }
6747 look_ahead = FALSE;
6748 *io_streaming = 1;
6749
6750 break;
6751 }
6752 throttle_limit = (uint32_t) max_length;
6753 assert(throttle_limit == max_length);
6754
6755 if (vnode_pager_get_throttle_io_limit(object->pager, &throttle_limit) == KERN_SUCCESS) {
6756 if (max_length > throttle_limit) {
6757 max_length = throttle_limit;
6758 }
6759 }
6760 if (pre_heat_size > max_length) {
6761 pre_heat_size = max_length;
6762 }
6763
6764 if (behavior == VM_BEHAVIOR_DEFAULT && (pre_heat_size > min_ph_size)) {
6765 unsigned int consider_free = vm_page_free_count + vm_page_cleaned_count;
6766
6767 if (consider_free < vm_page_throttle_limit) {
6768 pre_heat_size = trunc_page(pre_heat_size / 16);
6769 } else if (consider_free < vm_page_free_target) {
6770 pre_heat_size = trunc_page(pre_heat_size / 4);
6771 }
6772
6773 if (pre_heat_size < min_ph_size) {
6774 pre_heat_size = min_ph_size;
6775 }
6776 }
6777 if (look_ahead == TRUE) {
6778 if (look_behind == TRUE) {
6779 /*
6780 * if we get here its due to a random access...
6781 * so we want to center the original fault address
6782 * within the cluster we will issue... make sure
6783 * to calculate 'head_size' as a multiple of PAGE_SIZE...
6784 * 'pre_heat_size' is a multiple of PAGE_SIZE but not
6785 * necessarily an even number of pages so we need to truncate
6786 * the result to a PAGE_SIZE boundary
6787 */
6788 head_size = trunc_page(pre_heat_size / 2);
6789
6790 if (target_start > head_size) {
6791 target_start -= head_size;
6792 } else {
6793 target_start = 0;
6794 }
6795
6796 /*
6797 * 'target_start' at this point represents the beginning offset
6798 * of the cluster we are considering... 'orig_start' will be in
6799 * the center of this cluster if we didn't have to clip the start
6800 * due to running into the start of the file
6801 */
6802 }
6803 if ((target_start + pre_heat_size) > object_size) {
6804 pre_heat_size = (vm_size_t)(round_page_64(x: object_size - target_start));
6805 }
6806 /*
6807 * at this point caclulate the number of pages beyond the original fault
6808 * address that we want to consider... this is guaranteed not to extend beyond
6809 * the current EOF...
6810 */
6811 assert((vm_size_t)(orig_start - target_start) == (orig_start - target_start));
6812 tail_size = pre_heat_size - (vm_size_t)(orig_start - target_start) - PAGE_SIZE;
6813 } else {
6814 if (pre_heat_size > target_start) {
6815 /*
6816 * since pre_heat_size is always smaller then 2^32,
6817 * if it is larger then target_start (a 64 bit value)
6818 * it is safe to clip target_start to 32 bits
6819 */
6820 pre_heat_size = (vm_size_t) target_start;
6821 }
6822 tail_size = 0;
6823 }
6824 assert( !(target_start & PAGE_MASK_64));
6825 assert( !(pre_heat_size & PAGE_MASK_64));
6826
6827 if (pre_heat_size <= PAGE_SIZE) {
6828 goto out;
6829 }
6830
6831 if (look_behind == TRUE) {
6832 /*
6833 * take a look at the pages before the original
6834 * faulting offset... recalculate this in case
6835 * we had to clip 'pre_heat_size' above to keep
6836 * from running past the EOF.
6837 */
6838 head_size = pre_heat_size - tail_size - PAGE_SIZE;
6839
6840 for (offset = orig_start - PAGE_SIZE_64; head_size; offset -= PAGE_SIZE_64, head_size -= PAGE_SIZE) {
6841 /*
6842 * don't poke below the lowest offset
6843 */
6844 if (offset < fault_info->lo_offset) {
6845 break;
6846 }
6847 /*
6848 * for external objects or internal objects w/o a pager,
6849 * VM_COMPRESSOR_PAGER_STATE_GET will return VM_EXTERNAL_STATE_UNKNOWN
6850 */
6851 if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset) == VM_EXTERNAL_STATE_ABSENT) {
6852 break;
6853 }
6854 if (vm_page_lookup(object, offset) != VM_PAGE_NULL) {
6855 /*
6856 * don't bridge resident pages
6857 */
6858 break;
6859 }
6860 *start = offset;
6861 *length += PAGE_SIZE;
6862 }
6863 }
6864 if (look_ahead == TRUE) {
6865 for (offset = orig_start + PAGE_SIZE_64; tail_size; offset += PAGE_SIZE_64, tail_size -= PAGE_SIZE) {
6866 /*
6867 * don't poke above the highest offset
6868 */
6869 if (offset >= fault_info->hi_offset) {
6870 break;
6871 }
6872 assert(offset < object_size);
6873
6874 /*
6875 * for external objects or internal objects w/o a pager,
6876 * VM_COMPRESSOR_PAGER_STATE_GET will return VM_EXTERNAL_STATE_UNKNOWN
6877 */
6878 if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset) == VM_EXTERNAL_STATE_ABSENT) {
6879 break;
6880 }
6881 if (vm_page_lookup(object, offset) != VM_PAGE_NULL) {
6882 /*
6883 * don't bridge resident pages
6884 */
6885 break;
6886 }
6887 *length += PAGE_SIZE;
6888 }
6889 }
6890out:
6891 if (*length > max_length) {
6892 *length = max_length;
6893 }
6894
6895 vm_object_unlock(object);
6896
6897 DTRACE_VM1(clustersize, vm_size_t, *length);
6898}
6899
6900
6901/*
6902 * Allow manipulation of individual page state. This is actually part of
6903 * the UPL regimen but takes place on the VM object rather than on a UPL
6904 */
6905
6906kern_return_t
6907vm_object_page_op(
6908 vm_object_t object,
6909 vm_object_offset_t offset,
6910 int ops,
6911 ppnum_t *phys_entry,
6912 int *flags)
6913{
6914 vm_page_t dst_page;
6915
6916 vm_object_lock(object);
6917
6918 if (ops & UPL_POP_PHYSICAL) {
6919 if (object->phys_contiguous) {
6920 if (phys_entry) {
6921 *phys_entry = (ppnum_t)
6922 (object->vo_shadow_offset >> PAGE_SHIFT);
6923 }
6924 vm_object_unlock(object);
6925 return KERN_SUCCESS;
6926 } else {
6927 vm_object_unlock(object);
6928 return KERN_INVALID_OBJECT;
6929 }
6930 }
6931 if (object->phys_contiguous) {
6932 vm_object_unlock(object);
6933 return KERN_INVALID_OBJECT;
6934 }
6935
6936 while (TRUE) {
6937 if ((dst_page = vm_page_lookup(object, offset)) == VM_PAGE_NULL) {
6938 vm_object_unlock(object);
6939 return KERN_FAILURE;
6940 }
6941
6942 /* Sync up on getting the busy bit */
6943 if ((dst_page->vmp_busy || dst_page->vmp_cleaning) &&
6944 (((ops & UPL_POP_SET) &&
6945 (ops & UPL_POP_BUSY)) || (ops & UPL_POP_DUMP))) {
6946 /* someone else is playing with the page, we will */
6947 /* have to wait */
6948 PAGE_SLEEP(object, dst_page, THREAD_UNINT);
6949 continue;
6950 }
6951
6952 if (ops & UPL_POP_DUMP) {
6953 if (dst_page->vmp_pmapped == TRUE) {
6954 pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: dst_page));
6955 }
6956
6957 VM_PAGE_FREE(dst_page);
6958 break;
6959 }
6960
6961 if (flags) {
6962 *flags = 0;
6963
6964 /* Get the condition of flags before requested ops */
6965 /* are undertaken */
6966
6967 if (dst_page->vmp_dirty) {
6968 *flags |= UPL_POP_DIRTY;
6969 }
6970 if (dst_page->vmp_free_when_done) {
6971 *flags |= UPL_POP_PAGEOUT;
6972 }
6973 if (dst_page->vmp_precious) {
6974 *flags |= UPL_POP_PRECIOUS;
6975 }
6976 if (dst_page->vmp_absent) {
6977 *flags |= UPL_POP_ABSENT;
6978 }
6979 if (dst_page->vmp_busy) {
6980 *flags |= UPL_POP_BUSY;
6981 }
6982 }
6983
6984 /* The caller should have made a call either contingent with */
6985 /* or prior to this call to set UPL_POP_BUSY */
6986 if (ops & UPL_POP_SET) {
6987 /* The protection granted with this assert will */
6988 /* not be complete. If the caller violates the */
6989 /* convention and attempts to change page state */
6990 /* without first setting busy we may not see it */
6991 /* because the page may already be busy. However */
6992 /* if such violations occur we will assert sooner */
6993 /* or later. */
6994 assert(dst_page->vmp_busy || (ops & UPL_POP_BUSY));
6995 if (ops & UPL_POP_DIRTY) {
6996 SET_PAGE_DIRTY(dst_page, FALSE);
6997 }
6998 if (ops & UPL_POP_PAGEOUT) {
6999 dst_page->vmp_free_when_done = TRUE;
7000 }
7001 if (ops & UPL_POP_PRECIOUS) {
7002 dst_page->vmp_precious = TRUE;
7003 }
7004 if (ops & UPL_POP_ABSENT) {
7005 dst_page->vmp_absent = TRUE;
7006 }
7007 if (ops & UPL_POP_BUSY) {
7008 dst_page->vmp_busy = TRUE;
7009 }
7010 }
7011
7012 if (ops & UPL_POP_CLR) {
7013 assert(dst_page->vmp_busy);
7014 if (ops & UPL_POP_DIRTY) {
7015 dst_page->vmp_dirty = FALSE;
7016 }
7017 if (ops & UPL_POP_PAGEOUT) {
7018 dst_page->vmp_free_when_done = FALSE;
7019 }
7020 if (ops & UPL_POP_PRECIOUS) {
7021 dst_page->vmp_precious = FALSE;
7022 }
7023 if (ops & UPL_POP_ABSENT) {
7024 dst_page->vmp_absent = FALSE;
7025 }
7026 if (ops & UPL_POP_BUSY) {
7027 dst_page->vmp_busy = FALSE;
7028 PAGE_WAKEUP(dst_page);
7029 }
7030 }
7031 if (phys_entry) {
7032 /*
7033 * The physical page number will remain valid
7034 * only if the page is kept busy.
7035 */
7036 assert(dst_page->vmp_busy);
7037 *phys_entry = VM_PAGE_GET_PHYS_PAGE(m: dst_page);
7038 }
7039
7040 break;
7041 }
7042
7043 vm_object_unlock(object);
7044 return KERN_SUCCESS;
7045}
7046
7047/*
7048 * vm_object_range_op offers performance enhancement over
7049 * vm_object_page_op for page_op functions which do not require page
7050 * level state to be returned from the call. Page_op was created to provide
7051 * a low-cost alternative to page manipulation via UPLs when only a single
7052 * page was involved. The range_op call establishes the ability in the _op
7053 * family of functions to work on multiple pages where the lack of page level
7054 * state handling allows the caller to avoid the overhead of the upl structures.
7055 */
7056
7057kern_return_t
7058vm_object_range_op(
7059 vm_object_t object,
7060 vm_object_offset_t offset_beg,
7061 vm_object_offset_t offset_end,
7062 int ops,
7063 uint32_t *range)
7064{
7065 vm_object_offset_t offset;
7066 vm_page_t dst_page;
7067
7068 if (offset_end - offset_beg > (uint32_t) -1) {
7069 /* range is too big and would overflow "*range" */
7070 return KERN_INVALID_ARGUMENT;
7071 }
7072 if (object->resident_page_count == 0) {
7073 if (range) {
7074 if (ops & UPL_ROP_PRESENT) {
7075 *range = 0;
7076 } else {
7077 *range = (uint32_t) (offset_end - offset_beg);
7078 assert(*range == (offset_end - offset_beg));
7079 }
7080 }
7081 return KERN_SUCCESS;
7082 }
7083 vm_object_lock(object);
7084
7085 if (object->phys_contiguous) {
7086 vm_object_unlock(object);
7087 return KERN_INVALID_OBJECT;
7088 }
7089
7090 offset = offset_beg & ~PAGE_MASK_64;
7091
7092 while (offset < offset_end) {
7093 dst_page = vm_page_lookup(object, offset);
7094 if (dst_page != VM_PAGE_NULL) {
7095 if (ops & UPL_ROP_DUMP) {
7096 if (dst_page->vmp_busy || dst_page->vmp_cleaning) {
7097 /*
7098 * someone else is playing with the
7099 * page, we will have to wait
7100 */
7101 PAGE_SLEEP(object, dst_page, THREAD_UNINT);
7102 /*
7103 * need to relook the page up since it's
7104 * state may have changed while we slept
7105 * it might even belong to a different object
7106 * at this point
7107 */
7108 continue;
7109 }
7110 if (dst_page->vmp_laundry) {
7111 vm_pageout_steal_laundry(page: dst_page, FALSE);
7112 }
7113
7114 if (dst_page->vmp_pmapped == TRUE) {
7115 pmap_disconnect(phys: VM_PAGE_GET_PHYS_PAGE(m: dst_page));
7116 }
7117
7118 VM_PAGE_FREE(dst_page);
7119 } else if ((ops & UPL_ROP_ABSENT)
7120 && (!dst_page->vmp_absent || dst_page->vmp_busy)) {
7121 break;
7122 }
7123 } else if (ops & UPL_ROP_PRESENT) {
7124 break;
7125 }
7126
7127 offset += PAGE_SIZE;
7128 }
7129 vm_object_unlock(object);
7130
7131 if (range) {
7132 if (offset > offset_end) {
7133 offset = offset_end;
7134 }
7135 if (offset > offset_beg) {
7136 *range = (uint32_t) (offset - offset_beg);
7137 assert(*range == (offset - offset_beg));
7138 } else {
7139 *range = 0;
7140 }
7141 }
7142 return KERN_SUCCESS;
7143}
7144
7145/*
7146 * Used to point a pager directly to a range of memory (when the pager may be associated
7147 * with a non-device vnode). Takes a virtual address, an offset, and a size. We currently
7148 * expect that the virtual address will denote the start of a range that is physically contiguous.
7149 */
7150kern_return_t
7151pager_map_to_phys_contiguous(
7152 memory_object_control_t object,
7153 memory_object_offset_t offset,
7154 addr64_t base_vaddr,
7155 vm_size_t size)
7156{
7157 ppnum_t page_num;
7158 boolean_t clobbered_private;
7159 kern_return_t retval;
7160 vm_object_t pager_object;
7161
7162 page_num = pmap_find_phys(map: kernel_pmap, va: base_vaddr);
7163
7164 if (!page_num) {
7165 retval = KERN_FAILURE;
7166 goto out;
7167 }
7168
7169 pager_object = memory_object_control_to_vm_object(control: object);
7170
7171 if (!pager_object) {
7172 retval = KERN_FAILURE;
7173 goto out;
7174 }
7175
7176 clobbered_private = pager_object->private;
7177 if (pager_object->private != TRUE) {
7178 vm_object_lock(pager_object);
7179 VM_OBJECT_SET_PRIVATE(object: pager_object, TRUE);
7180 vm_object_unlock(pager_object);
7181 }
7182 retval = vm_object_populate_with_private(object: pager_object, offset, phys_page: page_num, size);
7183
7184 if (retval != KERN_SUCCESS) {
7185 if (pager_object->private != clobbered_private) {
7186 vm_object_lock(pager_object);
7187 VM_OBJECT_SET_PRIVATE(object: pager_object, value: clobbered_private);
7188 vm_object_unlock(pager_object);
7189 }
7190 }
7191
7192out:
7193 return retval;
7194}
7195
7196uint32_t scan_object_collision = 0;
7197
7198void
7199vm_object_lock(vm_object_t object)
7200{
7201 if (object == vm_pageout_scan_wants_object) {
7202 scan_object_collision++;
7203 mutex_pause(2);
7204 }
7205 DTRACE_VM(vm_object_lock_w);
7206 lck_rw_lock_exclusive(lck: &object->Lock);
7207}
7208
7209boolean_t
7210vm_object_lock_avoid(vm_object_t object)
7211{
7212 if (object == vm_pageout_scan_wants_object) {
7213 scan_object_collision++;
7214 return TRUE;
7215 }
7216 return FALSE;
7217}
7218
7219boolean_t
7220_vm_object_lock_try(vm_object_t object)
7221{
7222 boolean_t retval;
7223
7224 retval = lck_rw_try_lock_exclusive(lck: &object->Lock);
7225#if DEVELOPMENT || DEBUG
7226 if (retval == TRUE) {
7227 DTRACE_VM(vm_object_lock_w);
7228 }
7229#endif
7230 return retval;
7231}
7232
7233boolean_t
7234vm_object_lock_try(vm_object_t object)
7235{
7236 /*
7237 * Called from hibernate path so check before blocking.
7238 */
7239 if (vm_object_lock_avoid(object) && ml_get_interrupts_enabled() && get_preemption_level() == 0) {
7240 mutex_pause(2);
7241 }
7242 return _vm_object_lock_try(object);
7243}
7244
7245/*
7246 * Lock the object exclusive.
7247 *
7248 * Returns true iff the thread had to spin or block before
7249 * acquiring the lock.
7250 */
7251bool
7252vm_object_lock_check_contended(vm_object_t object)
7253{
7254 if (object == vm_pageout_scan_wants_object) {
7255 scan_object_collision++;
7256 mutex_pause(2);
7257 }
7258 DTRACE_VM(vm_object_lock_w);
7259 return lck_rw_lock_exclusive_check_contended(lck: &object->Lock);
7260}
7261
7262void
7263vm_object_lock_shared(vm_object_t object)
7264{
7265 if (vm_object_lock_avoid(object)) {
7266 mutex_pause(2);
7267 }
7268 DTRACE_VM(vm_object_lock_r);
7269 lck_rw_lock_shared(lck: &object->Lock);
7270}
7271
7272boolean_t
7273vm_object_lock_yield_shared(vm_object_t object)
7274{
7275 boolean_t retval = FALSE, force_yield = FALSE;
7276
7277 vm_object_lock_assert_shared(object);
7278
7279 force_yield = vm_object_lock_avoid(object);
7280
7281 retval = lck_rw_lock_yield_shared(lck: &object->Lock, force_yield);
7282 if (retval) {
7283 DTRACE_VM(vm_object_lock_yield);
7284 }
7285
7286 return retval;
7287}
7288
7289boolean_t
7290vm_object_lock_try_shared(vm_object_t object)
7291{
7292 boolean_t retval;
7293
7294 if (vm_object_lock_avoid(object)) {
7295 mutex_pause(2);
7296 }
7297 retval = lck_rw_try_lock_shared(lck: &object->Lock);
7298 if (retval) {
7299 DTRACE_VM(vm_object_lock_r);
7300 }
7301 return retval;
7302}
7303
7304boolean_t
7305vm_object_lock_upgrade(vm_object_t object)
7306{
7307 boolean_t retval;
7308
7309 retval = lck_rw_lock_shared_to_exclusive(lck: &object->Lock);
7310#if DEVELOPMENT || DEBUG
7311 if (retval == TRUE) {
7312 DTRACE_VM(vm_object_lock_w);
7313 }
7314#endif
7315 return retval;
7316}
7317
7318void
7319vm_object_unlock(vm_object_t object)
7320{
7321#if DEVELOPMENT || DEBUG
7322 DTRACE_VM(vm_object_unlock);
7323#endif
7324 lck_rw_done(lck: &object->Lock);
7325}
7326
7327
7328unsigned int vm_object_change_wimg_mode_count = 0;
7329
7330/*
7331 * The object must be locked
7332 */
7333void
7334vm_object_change_wimg_mode(vm_object_t object, unsigned int wimg_mode)
7335{
7336 vm_page_t p;
7337
7338 vm_object_lock_assert_exclusive(object);
7339
7340 vm_object_paging_only_wait(object, THREAD_UNINT);
7341
7342 vm_page_queue_iterate(&object->memq, p, vmp_listq) {
7343 if (!p->vmp_fictitious) {
7344 pmap_set_cache_attributes(VM_PAGE_GET_PHYS_PAGE(m: p), wimg_mode);
7345 }
7346 }
7347 if (wimg_mode == VM_WIMG_USE_DEFAULT) {
7348 object->set_cache_attr = FALSE;
7349 } else {
7350 object->set_cache_attr = TRUE;
7351 }
7352
7353 object->wimg_bits = wimg_mode;
7354
7355 vm_object_change_wimg_mode_count++;
7356}
7357
7358#if CONFIG_FREEZE
7359
7360extern struct freezer_context freezer_context_global;
7361
7362/*
7363 * This routine does the "relocation" of previously
7364 * compressed pages belonging to this object that are
7365 * residing in a number of compressed segments into
7366 * a set of compressed segments dedicated to hold
7367 * compressed pages belonging to this object.
7368 */
7369
7370extern AbsoluteTime c_freezer_last_yield_ts;
7371
7372#define MAX_FREE_BATCH 32
7373#define FREEZER_DUTY_CYCLE_ON_MS 5
7374#define FREEZER_DUTY_CYCLE_OFF_MS 5
7375
7376static int c_freezer_should_yield(void);
7377
7378
7379static int
7380c_freezer_should_yield()
7381{
7382 AbsoluteTime cur_time;
7383 uint64_t nsecs;
7384
7385 assert(c_freezer_last_yield_ts);
7386 clock_get_uptime(&cur_time);
7387
7388 SUB_ABSOLUTETIME(&cur_time, &c_freezer_last_yield_ts);
7389 absolutetime_to_nanoseconds(cur_time, &nsecs);
7390
7391 if (nsecs > 1000 * 1000 * FREEZER_DUTY_CYCLE_ON_MS) {
7392 return 1;
7393 }
7394 return 0;
7395}
7396
7397
7398void
7399vm_object_compressed_freezer_done()
7400{
7401 vm_compressor_finished_filling( &(freezer_context_global.freezer_ctx_chead));
7402}
7403
7404
7405uint32_t
7406vm_object_compressed_freezer_pageout(
7407 vm_object_t object, uint32_t dirty_budget)
7408{
7409 vm_page_t p;
7410 vm_page_t local_freeq = NULL;
7411 int local_freed = 0;
7412 kern_return_t retval = KERN_SUCCESS;
7413 int obj_resident_page_count_snapshot = 0;
7414 uint32_t paged_out_count = 0;
7415
7416 assert(object != VM_OBJECT_NULL);
7417 assert(object->internal);
7418
7419 vm_object_lock(object);
7420
7421 if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) {
7422 if (!object->pager_initialized) {
7423 vm_object_collapse(object, (vm_object_offset_t) 0, TRUE);
7424
7425 if (!object->pager_initialized) {
7426 vm_object_compressor_pager_create(object);
7427 }
7428 }
7429
7430 if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) {
7431 vm_object_unlock(object);
7432 return paged_out_count;
7433 }
7434 }
7435
7436 /*
7437 * We could be freezing a shared internal object that might
7438 * be part of some other thread's current VM operations.
7439 * We skip it if there's a paging-in-progress or activity-in-progress
7440 * because we could be here a long time with the map lock held.
7441 *
7442 * Note: We are holding the map locked while we wait.
7443 * This is fine in the freezer path because the task
7444 * is suspended and so this latency is acceptable.
7445 */
7446 if (object->paging_in_progress || object->activity_in_progress) {
7447 vm_object_unlock(object);
7448 return paged_out_count;
7449 }
7450
7451 if (VM_CONFIG_FREEZER_SWAP_IS_ACTIVE) {
7452 vm_object_offset_t curr_offset = 0;
7453
7454 /*
7455 * Go through the object and make sure that any
7456 * previously compressed pages are relocated into
7457 * a compressed segment associated with our "freezer_chead".
7458 */
7459 while (curr_offset < object->vo_size) {
7460 curr_offset = vm_compressor_pager_next_compressed(object->pager, curr_offset);
7461
7462 if (curr_offset == (vm_object_offset_t) -1) {
7463 break;
7464 }
7465
7466 retval = vm_compressor_pager_relocate(object->pager, curr_offset, &(freezer_context_global.freezer_ctx_chead));
7467
7468 if (retval != KERN_SUCCESS) {
7469 break;
7470 }
7471
7472 curr_offset += PAGE_SIZE_64;
7473 }
7474 }
7475
7476 /*
7477 * We can't hold the object lock while heading down into the compressed pager
7478 * layer because we might need the kernel map lock down there to allocate new
7479 * compressor data structures. And if this same object is mapped in the kernel
7480 * and there's a fault on it, then that thread will want the object lock while
7481 * holding the kernel map lock.
7482 *
7483 * Since we are going to drop/grab the object lock repeatedly, we must make sure
7484 * we won't be stuck in an infinite loop if the same page(s) keep getting
7485 * decompressed. So we grab a snapshot of the number of pages in the object and
7486 * we won't process any more than that number of pages.
7487 */
7488
7489 obj_resident_page_count_snapshot = object->resident_page_count;
7490
7491 vm_object_activity_begin(object);
7492
7493 while ((obj_resident_page_count_snapshot--) && !vm_page_queue_empty(&object->memq) && paged_out_count < dirty_budget) {
7494 p = (vm_page_t)vm_page_queue_first(&object->memq);
7495
7496 KERNEL_DEBUG(0xe0430004 | DBG_FUNC_START, object, local_freed, 0, 0, 0);
7497
7498 vm_page_lockspin_queues();
7499
7500 if (p->vmp_cleaning || p->vmp_fictitious || p->vmp_busy || p->vmp_absent || p->vmp_unusual || VMP_ERROR_GET(p) || VM_PAGE_WIRED(p)) {
7501 vm_page_unlock_queues();
7502
7503 KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 1, 0, 0);
7504
7505 vm_page_queue_remove(&object->memq, p, vmp_listq);
7506 vm_page_queue_enter(&object->memq, p, vmp_listq);
7507
7508 continue;
7509 }
7510
7511 if (p->vmp_pmapped == TRUE) {
7512 int refmod_state, pmap_flags;
7513
7514 if (p->vmp_dirty || p->vmp_precious) {
7515 pmap_flags = PMAP_OPTIONS_COMPRESSOR;
7516 } else {
7517 pmap_flags = PMAP_OPTIONS_COMPRESSOR_IFF_MODIFIED;
7518 }
7519
7520 vm_page_lockconvert_queues();
7521 refmod_state = pmap_disconnect_options(VM_PAGE_GET_PHYS_PAGE(p), pmap_flags, NULL);
7522 if (refmod_state & VM_MEM_MODIFIED) {
7523 SET_PAGE_DIRTY(p, FALSE);
7524 }
7525 }
7526
7527 if (p->vmp_dirty == FALSE && p->vmp_precious == FALSE) {
7528 /*
7529 * Clean and non-precious page.
7530 */
7531 vm_page_unlock_queues();
7532 VM_PAGE_FREE(p);
7533
7534 KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 2, 0, 0);
7535 continue;
7536 }
7537
7538 if (p->vmp_laundry) {
7539 vm_pageout_steal_laundry(p, TRUE);
7540 }
7541
7542 vm_page_queues_remove(p, TRUE);
7543
7544 vm_page_unlock_queues();
7545
7546
7547 /*
7548 * In case the compressor fails to compress this page, we need it at
7549 * the back of the object memq so that we don't keep trying to process it.
7550 * Make the move here while we have the object lock held.
7551 */
7552
7553 vm_page_queue_remove(&object->memq, p, vmp_listq);
7554 vm_page_queue_enter(&object->memq, p, vmp_listq);
7555
7556 /*
7557 * Grab an activity_in_progress here for vm_pageout_compress_page() to consume.
7558 *
7559 * Mark the page busy so no one messes with it while we have the object lock dropped.
7560 */
7561 p->vmp_busy = TRUE;
7562
7563 vm_object_activity_begin(object);
7564
7565 vm_object_unlock(object);
7566
7567 if (vm_pageout_compress_page(&(freezer_context_global.freezer_ctx_chead),
7568 (freezer_context_global.freezer_ctx_compressor_scratch_buf),
7569 p) == KERN_SUCCESS) {
7570 /*
7571 * page has already been un-tabled from the object via 'vm_page_remove'
7572 */
7573 p->vmp_snext = local_freeq;
7574 local_freeq = p;
7575 local_freed++;
7576 paged_out_count++;
7577
7578 if (local_freed >= MAX_FREE_BATCH) {
7579 OSAddAtomic64(local_freed, &vm_pageout_vminfo.vm_pageout_compressions);
7580
7581 vm_page_free_list(local_freeq, TRUE);
7582
7583 local_freeq = NULL;
7584 local_freed = 0;
7585 }
7586 freezer_context_global.freezer_ctx_uncompressed_pages++;
7587 }
7588 KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 0, 0, 0);
7589
7590 if (local_freed == 0 && c_freezer_should_yield()) {
7591 thread_yield_internal(FREEZER_DUTY_CYCLE_OFF_MS);
7592 clock_get_uptime(&c_freezer_last_yield_ts);
7593 }
7594
7595 vm_object_lock(object);
7596 }
7597
7598 if (local_freeq) {
7599 OSAddAtomic64(local_freed, &vm_pageout_vminfo.vm_pageout_compressions);
7600
7601 vm_page_free_list(local_freeq, TRUE);
7602
7603 local_freeq = NULL;
7604 local_freed = 0;
7605 }
7606
7607 vm_object_activity_end(object);
7608
7609 vm_object_unlock(object);
7610
7611 if (c_freezer_should_yield()) {
7612 thread_yield_internal(FREEZER_DUTY_CYCLE_OFF_MS);
7613 clock_get_uptime(&c_freezer_last_yield_ts);
7614 }
7615 return paged_out_count;
7616}
7617
7618#endif /* CONFIG_FREEZE */
7619
7620
7621void
7622vm_object_pageout(
7623 vm_object_t object)
7624{
7625 vm_page_t p, next;
7626 struct vm_pageout_queue *iq;
7627
7628 if (!VM_CONFIG_COMPRESSOR_IS_PRESENT) {
7629 return;
7630 }
7631
7632 iq = &vm_pageout_queue_internal;
7633
7634 assert(object != VM_OBJECT_NULL );
7635
7636 vm_object_lock(object);
7637
7638 if (!object->internal ||
7639 object->terminating ||
7640 !object->alive) {
7641 vm_object_unlock(object);
7642 return;
7643 }
7644
7645 if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) {
7646 if (!object->pager_initialized) {
7647 vm_object_collapse(object, hint_offset: (vm_object_offset_t) 0, TRUE);
7648
7649 if (!object->pager_initialized) {
7650 vm_object_compressor_pager_create(object);
7651 }
7652 }
7653
7654 if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) {
7655 vm_object_unlock(object);
7656 return;
7657 }
7658 }
7659
7660ReScan:
7661 next = (vm_page_t)vm_page_queue_first(&object->memq);
7662
7663 while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next)) {
7664 p = next;
7665 next = (vm_page_t)vm_page_queue_next(&next->vmp_listq);
7666
7667 assert(p->vmp_q_state != VM_PAGE_ON_FREE_Q);
7668
7669 if ((p->vmp_q_state == VM_PAGE_ON_THROTTLED_Q) ||
7670 p->vmp_cleaning ||
7671 p->vmp_laundry ||
7672 p->vmp_busy ||
7673 p->vmp_absent ||
7674 VMP_ERROR_GET(p) ||
7675 p->vmp_fictitious ||
7676 VM_PAGE_WIRED(p)) {
7677 /*
7678 * Page is already being cleaned or can't be cleaned.
7679 */
7680 continue;
7681 }
7682 if (vm_compressor_low_on_space()) {
7683 break;
7684 }
7685
7686 /* Throw to the pageout queue */
7687
7688 vm_page_lockspin_queues();
7689
7690 if (VM_PAGE_Q_THROTTLED(iq)) {
7691 iq->pgo_draining = TRUE;
7692
7693 assert_wait(event: (event_t) (&iq->pgo_laundry + 1),
7694 THREAD_INTERRUPTIBLE);
7695 vm_page_unlock_queues();
7696 vm_object_unlock(object);
7697
7698 thread_block(THREAD_CONTINUE_NULL);
7699
7700 vm_object_lock(object);
7701 goto ReScan;
7702 }
7703
7704 assert(!p->vmp_fictitious);
7705 assert(!p->vmp_busy);
7706 assert(!p->vmp_absent);
7707 assert(!p->vmp_unusual);
7708 assert(!VMP_ERROR_GET(p)); /* XXX there's a window here where we could have an ECC error! */
7709 assert(!VM_PAGE_WIRED(p));
7710 assert(!p->vmp_cleaning);
7711
7712 if (p->vmp_pmapped == TRUE) {
7713 int refmod_state;
7714 int pmap_options;
7715
7716 /*
7717 * Tell pmap the page should be accounted
7718 * for as "compressed" if it's been modified.
7719 */
7720 pmap_options =
7721 PMAP_OPTIONS_COMPRESSOR_IFF_MODIFIED;
7722 if (p->vmp_dirty || p->vmp_precious) {
7723 /*
7724 * We already know it's been modified,
7725 * so tell pmap to account for it
7726 * as "compressed".
7727 */
7728 pmap_options = PMAP_OPTIONS_COMPRESSOR;
7729 }
7730 vm_page_lockconvert_queues();
7731 refmod_state = pmap_disconnect_options(phys: VM_PAGE_GET_PHYS_PAGE(m: p),
7732 options: pmap_options,
7733 NULL);
7734 if (refmod_state & VM_MEM_MODIFIED) {
7735 SET_PAGE_DIRTY(p, FALSE);
7736 }
7737 }
7738
7739 if (!p->vmp_dirty && !p->vmp_precious) {
7740 vm_page_unlock_queues();
7741 VM_PAGE_FREE(p);
7742 continue;
7743 }
7744 vm_page_queues_remove(mem: p, TRUE);
7745
7746 vm_pageout_cluster(m: p);
7747
7748 vm_page_unlock_queues();
7749 }
7750 vm_object_unlock(object);
7751}
7752
7753
7754#if CONFIG_IOSCHED
7755void
7756vm_page_request_reprioritize(vm_object_t o, uint64_t blkno, uint32_t len, int prio)
7757{
7758 io_reprioritize_req_t req;
7759 struct vnode *devvp = NULL;
7760
7761 if (vnode_pager_get_object_devvp(o->pager, (uintptr_t *)&devvp) != KERN_SUCCESS) {
7762 return;
7763 }
7764
7765 /*
7766 * Create the request for I/O reprioritization.
7767 * We use the noblock variant of zalloc because we're holding the object
7768 * lock here and we could cause a deadlock in low memory conditions.
7769 */
7770 req = (io_reprioritize_req_t)zalloc_noblock(zone: io_reprioritize_req_zone);
7771 if (req == NULL) {
7772 return;
7773 }
7774 req->blkno = blkno;
7775 req->len = len;
7776 req->priority = prio;
7777 req->devvp = devvp;
7778
7779 /* Insert request into the reprioritization list */
7780 IO_REPRIORITIZE_LIST_LOCK();
7781 queue_enter(&io_reprioritize_list, req, io_reprioritize_req_t, io_reprioritize_list);
7782 IO_REPRIORITIZE_LIST_UNLOCK();
7783
7784 /* Wakeup reprioritize thread */
7785 IO_REPRIO_THREAD_WAKEUP();
7786
7787 return;
7788}
7789
7790void
7791vm_decmp_upl_reprioritize(upl_t upl, int prio)
7792{
7793 int offset;
7794 vm_object_t object;
7795 io_reprioritize_req_t req;
7796 struct vnode *devvp = NULL;
7797 uint64_t blkno;
7798 uint32_t len;
7799 upl_t io_upl;
7800 uint64_t *io_upl_reprio_info;
7801 int io_upl_size;
7802
7803 if ((upl->flags & UPL_TRACKED_BY_OBJECT) == 0 || (upl->flags & UPL_EXPEDITE_SUPPORTED) == 0) {
7804 return;
7805 }
7806
7807 /*
7808 * We dont want to perform any allocations with the upl lock held since that might
7809 * result in a deadlock. If the system is low on memory, the pageout thread would
7810 * try to pageout stuff and might wait on this lock. If we are waiting for the memory to
7811 * be freed up by the pageout thread, it would be a deadlock.
7812 */
7813
7814
7815 /* First step is just to get the size of the upl to find out how big the reprio info is */
7816 if (!upl_try_lock(upl)) {
7817 return;
7818 }
7819
7820 if (upl->decmp_io_upl == NULL) {
7821 /* The real I/O upl was destroyed by the time we came in here. Nothing to do. */
7822 upl_unlock(upl);
7823 return;
7824 }
7825
7826 io_upl = upl->decmp_io_upl;
7827 assert((io_upl->flags & UPL_DECMP_REAL_IO) != 0);
7828 assertf(page_aligned(io_upl->u_offset) && page_aligned(io_upl->u_size),
7829 "upl %p offset 0x%llx size 0x%x\n",
7830 io_upl, io_upl->u_offset, io_upl->u_size);
7831 io_upl_size = io_upl->u_size;
7832 upl_unlock(upl);
7833
7834 /* Now perform the allocation */
7835 io_upl_reprio_info = kalloc_data(sizeof(uint64_t) * atop(io_upl_size), Z_WAITOK);
7836 if (io_upl_reprio_info == NULL) {
7837 return;
7838 }
7839
7840 /* Now again take the lock, recheck the state and grab out the required info */
7841 if (!upl_try_lock(upl)) {
7842 goto out;
7843 }
7844
7845 if (upl->decmp_io_upl == NULL || upl->decmp_io_upl != io_upl) {
7846 /* The real I/O upl was destroyed by the time we came in here. Nothing to do. */
7847 upl_unlock(upl);
7848 goto out;
7849 }
7850 memcpy(dst: io_upl_reprio_info, src: io_upl->upl_reprio_info,
7851 n: sizeof(uint64_t) * atop(io_upl_size));
7852
7853 /* Get the VM object for this UPL */
7854 if (io_upl->flags & UPL_SHADOWED) {
7855 object = io_upl->map_object->shadow;
7856 } else {
7857 object = io_upl->map_object;
7858 }
7859
7860 /* Get the dev vnode ptr for this object */
7861 if (!object || !object->pager ||
7862 vnode_pager_get_object_devvp(object->pager, (uintptr_t *)&devvp) != KERN_SUCCESS) {
7863 upl_unlock(upl);
7864 goto out;
7865 }
7866
7867 upl_unlock(upl);
7868
7869 /* Now we have all the information needed to do the expedite */
7870
7871 offset = 0;
7872 while (offset < io_upl_size) {
7873 blkno = io_upl_reprio_info[atop(offset)] & UPL_REPRIO_INFO_MASK;
7874 len = (io_upl_reprio_info[atop(offset)] >> UPL_REPRIO_INFO_SHIFT) & UPL_REPRIO_INFO_MASK;
7875
7876 /*
7877 * This implementation may cause some spurious expedites due to the
7878 * fact that we dont cleanup the blkno & len from the upl_reprio_info
7879 * even after the I/O is complete.
7880 */
7881
7882 if (blkno != 0 && len != 0) {
7883 /* Create the request for I/O reprioritization */
7884 req = zalloc_flags(io_reprioritize_req_zone,
7885 Z_WAITOK | Z_NOFAIL);
7886 req->blkno = blkno;
7887 req->len = len;
7888 req->priority = prio;
7889 req->devvp = devvp;
7890
7891 /* Insert request into the reprioritization list */
7892 IO_REPRIORITIZE_LIST_LOCK();
7893 queue_enter(&io_reprioritize_list, req, io_reprioritize_req_t, io_reprioritize_list);
7894 IO_REPRIORITIZE_LIST_UNLOCK();
7895
7896 offset += len;
7897 } else {
7898 offset += PAGE_SIZE;
7899 }
7900 }
7901
7902 /* Wakeup reprioritize thread */
7903 IO_REPRIO_THREAD_WAKEUP();
7904
7905out:
7906 kfree_data(io_upl_reprio_info, sizeof(uint64_t) * atop(io_upl_size));
7907}
7908
7909void
7910vm_page_handle_prio_inversion(vm_object_t o, vm_page_t m)
7911{
7912 upl_t upl;
7913 upl_page_info_t *pl;
7914 unsigned int i, num_pages;
7915 int cur_tier;
7916
7917 cur_tier = proc_get_effective_thread_policy(thread: current_thread(), TASK_POLICY_IO);
7918
7919 /*
7920 * Scan through all UPLs associated with the object to find the
7921 * UPL containing the contended page.
7922 */
7923 queue_iterate(&o->uplq, upl, upl_t, uplq) {
7924 if (((upl->flags & UPL_EXPEDITE_SUPPORTED) == 0) || upl->upl_priority <= cur_tier) {
7925 continue;
7926 }
7927 pl = UPL_GET_INTERNAL_PAGE_LIST(upl);
7928 assertf(page_aligned(upl->u_offset) && page_aligned(upl->u_size),
7929 "upl %p offset 0x%llx size 0x%x\n",
7930 upl, upl->u_offset, upl->u_size);
7931 num_pages = (upl->u_size / PAGE_SIZE);
7932
7933 /*
7934 * For each page in the UPL page list, see if it matches the contended
7935 * page and was issued as a low prio I/O.
7936 */
7937 for (i = 0; i < num_pages; i++) {
7938 if (UPL_PAGE_PRESENT(pl, i) && VM_PAGE_GET_PHYS_PAGE(m) == pl[i].phys_addr) {
7939 if ((upl->flags & UPL_DECMP_REQ) && upl->decmp_io_upl) {
7940 KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_EXPEDITE)) | DBG_FUNC_NONE, VM_KERNEL_UNSLIDE_OR_PERM(upl->upl_creator), VM_KERNEL_UNSLIDE_OR_PERM(m),
7941 VM_KERNEL_UNSLIDE_OR_PERM(upl), upl->upl_priority, 0);
7942 vm_decmp_upl_reprioritize(upl, prio: cur_tier);
7943 break;
7944 }
7945 KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_EXPEDITE)) | DBG_FUNC_NONE, VM_KERNEL_UNSLIDE_OR_PERM(upl->upl_creator), VM_KERNEL_UNSLIDE_OR_PERM(m),
7946 upl->upl_reprio_info[i], upl->upl_priority, 0);
7947 if (UPL_REPRIO_INFO_BLKNO(upl, i) != 0 && UPL_REPRIO_INFO_LEN(upl, i) != 0) {
7948 vm_page_request_reprioritize(o, UPL_REPRIO_INFO_BLKNO(upl, i), UPL_REPRIO_INFO_LEN(upl, i), prio: cur_tier);
7949 }
7950 break;
7951 }
7952 }
7953 /* Check if we found any hits */
7954 if (i != num_pages) {
7955 break;
7956 }
7957 }
7958
7959 return;
7960}
7961
7962wait_result_t
7963vm_page_sleep(vm_object_t o, vm_page_t m, int interruptible)
7964{
7965 wait_result_t ret;
7966
7967 KERNEL_DEBUG((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_SLEEP)) | DBG_FUNC_START, o, m, 0, 0, 0);
7968
7969 if (o->io_tracking && ((m->vmp_busy == TRUE) || (m->vmp_cleaning == TRUE) || VM_PAGE_WIRED(m))) {
7970 /*
7971 * Indicates page is busy due to an I/O. Issue a reprioritize request if necessary.
7972 */
7973 vm_page_handle_prio_inversion(o, m);
7974 }
7975 m->vmp_wanted = TRUE;
7976 ret = thread_sleep_vm_object(object: o, event: m, interruptible);
7977 KERNEL_DEBUG((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_SLEEP)) | DBG_FUNC_END, o, m, 0, 0, 0);
7978 return ret;
7979}
7980
7981static void
7982io_reprioritize_thread(void *param __unused, wait_result_t wr __unused)
7983{
7984 io_reprioritize_req_t req = NULL;
7985
7986 while (1) {
7987 IO_REPRIORITIZE_LIST_LOCK();
7988 if (queue_empty(&io_reprioritize_list)) {
7989 IO_REPRIORITIZE_LIST_UNLOCK();
7990 break;
7991 }
7992
7993 queue_remove_first(&io_reprioritize_list, req, io_reprioritize_req_t, io_reprioritize_list);
7994 IO_REPRIORITIZE_LIST_UNLOCK();
7995
7996 vnode_pager_issue_reprioritize_io(devvp: req->devvp, blkno: req->blkno, len: req->len, priority: req->priority);
7997 zfree(io_reprioritize_req_zone, req);
7998 }
7999
8000 IO_REPRIO_THREAD_CONTINUATION();
8001}
8002#endif
8003
8004#if VM_OBJECT_ACCESS_TRACKING
8005void
8006vm_object_access_tracking(
8007 vm_object_t object,
8008 int *access_tracking_p,
8009 uint32_t *access_tracking_reads_p,
8010 uint32_t *access_tracking_writes_p)
8011{
8012 int access_tracking;
8013
8014 access_tracking = !!*access_tracking_p;
8015
8016 vm_object_lock(object);
8017 *access_tracking_p = object->access_tracking;
8018 if (access_tracking_reads_p) {
8019 *access_tracking_reads_p = object->access_tracking_reads;
8020 }
8021 if (access_tracking_writes_p) {
8022 *access_tracking_writes_p = object->access_tracking_writes;
8023 }
8024 object->access_tracking = access_tracking;
8025 object->access_tracking_reads = 0;
8026 object->access_tracking_writes = 0;
8027 vm_object_unlock(object);
8028
8029 if (access_tracking) {
8030 vm_object_pmap_protect_options(object,
8031 0,
8032 object->vo_size,
8033 PMAP_NULL,
8034 PAGE_SIZE,
8035 0,
8036 VM_PROT_NONE,
8037 0);
8038 }
8039}
8040#endif /* VM_OBJECT_ACCESS_TRACKING */
8041
8042void
8043vm_object_ledger_tag_ledgers(
8044 vm_object_t object,
8045 int *ledger_idx_volatile,
8046 int *ledger_idx_nonvolatile,
8047 int *ledger_idx_volatile_compressed,
8048 int *ledger_idx_nonvolatile_compressed,
8049 boolean_t *do_footprint)
8050{
8051 assert(object->shadow == VM_OBJECT_NULL);
8052
8053 *do_footprint = !object->vo_no_footprint;
8054
8055 switch (object->vo_ledger_tag) {
8056 case VM_LEDGER_TAG_NONE:
8057 /*
8058 * Regular purgeable memory:
8059 * counts in footprint only when nonvolatile.
8060 */
8061 *do_footprint = TRUE;
8062 assert(object->purgable != VM_PURGABLE_DENY);
8063 *ledger_idx_volatile = task_ledgers.purgeable_volatile;
8064 *ledger_idx_nonvolatile = task_ledgers.purgeable_nonvolatile;
8065 *ledger_idx_volatile_compressed = task_ledgers.purgeable_volatile_compressed;
8066 *ledger_idx_nonvolatile_compressed = task_ledgers.purgeable_nonvolatile_compressed;
8067 break;
8068 case VM_LEDGER_TAG_DEFAULT:
8069 /*
8070 * "default" tagged memory:
8071 * counts in footprint only when nonvolatile and not marked
8072 * as "no_footprint".
8073 */
8074 *ledger_idx_volatile = task_ledgers.tagged_nofootprint;
8075 *ledger_idx_volatile_compressed = task_ledgers.tagged_nofootprint_compressed;
8076 if (*do_footprint) {
8077 *ledger_idx_nonvolatile = task_ledgers.tagged_footprint;
8078 *ledger_idx_nonvolatile_compressed = task_ledgers.tagged_footprint_compressed;
8079 } else {
8080 *ledger_idx_nonvolatile = task_ledgers.tagged_nofootprint;
8081 *ledger_idx_nonvolatile_compressed = task_ledgers.tagged_nofootprint_compressed;
8082 }
8083 break;
8084 case VM_LEDGER_TAG_NETWORK:
8085 /*
8086 * "network" tagged memory:
8087 * never counts in footprint.
8088 */
8089 *do_footprint = FALSE;
8090 *ledger_idx_volatile = task_ledgers.network_volatile;
8091 *ledger_idx_volatile_compressed = task_ledgers.network_volatile_compressed;
8092 *ledger_idx_nonvolatile = task_ledgers.network_nonvolatile;
8093 *ledger_idx_nonvolatile_compressed = task_ledgers.network_nonvolatile_compressed;
8094 break;
8095 case VM_LEDGER_TAG_MEDIA:
8096 /*
8097 * "media" tagged memory:
8098 * counts in footprint only when nonvolatile and not marked
8099 * as "no footprint".
8100 */
8101 *ledger_idx_volatile = task_ledgers.media_nofootprint;
8102 *ledger_idx_volatile_compressed = task_ledgers.media_nofootprint_compressed;
8103 if (*do_footprint) {
8104 *ledger_idx_nonvolatile = task_ledgers.media_footprint;
8105 *ledger_idx_nonvolatile_compressed = task_ledgers.media_footprint_compressed;
8106 } else {
8107 *ledger_idx_nonvolatile = task_ledgers.media_nofootprint;
8108 *ledger_idx_nonvolatile_compressed = task_ledgers.media_nofootprint_compressed;
8109 }
8110 break;
8111 case VM_LEDGER_TAG_GRAPHICS:
8112 /*
8113 * "graphics" tagged memory:
8114 * counts in footprint only when nonvolatile and not marked
8115 * as "no footprint".
8116 */
8117 *ledger_idx_volatile = task_ledgers.graphics_nofootprint;
8118 *ledger_idx_volatile_compressed = task_ledgers.graphics_nofootprint_compressed;
8119 if (*do_footprint) {
8120 *ledger_idx_nonvolatile = task_ledgers.graphics_footprint;
8121 *ledger_idx_nonvolatile_compressed = task_ledgers.graphics_footprint_compressed;
8122 } else {
8123 *ledger_idx_nonvolatile = task_ledgers.graphics_nofootprint;
8124 *ledger_idx_nonvolatile_compressed = task_ledgers.graphics_nofootprint_compressed;
8125 }
8126 break;
8127 case VM_LEDGER_TAG_NEURAL:
8128 /*
8129 * "neural" tagged memory:
8130 * counts in footprint only when nonvolatile and not marked
8131 * as "no footprint".
8132 */
8133 *ledger_idx_volatile = task_ledgers.neural_nofootprint;
8134 *ledger_idx_volatile_compressed = task_ledgers.neural_nofootprint_compressed;
8135 if (*do_footprint) {
8136 *ledger_idx_nonvolatile = task_ledgers.neural_footprint;
8137 *ledger_idx_nonvolatile_compressed = task_ledgers.neural_footprint_compressed;
8138 } else {
8139 *ledger_idx_nonvolatile = task_ledgers.neural_nofootprint;
8140 *ledger_idx_nonvolatile_compressed = task_ledgers.neural_nofootprint_compressed;
8141 }
8142 break;
8143 default:
8144 panic("%s: object %p has unsupported ledger_tag %d",
8145 __FUNCTION__, object, object->vo_ledger_tag);
8146 }
8147}
8148
8149kern_return_t
8150vm_object_ownership_change(
8151 vm_object_t object,
8152 int new_ledger_tag,
8153 task_t new_owner,
8154 int new_ledger_flags,
8155 boolean_t old_task_objq_locked)
8156{
8157 int old_ledger_tag;
8158 task_t old_owner;
8159 int resident_count, wired_count;
8160 unsigned int compressed_count;
8161 int ledger_idx_volatile;
8162 int ledger_idx_nonvolatile;
8163 int ledger_idx_volatile_compressed;
8164 int ledger_idx_nonvolatile_compressed;
8165 int ledger_idx;
8166 int ledger_idx_compressed;
8167 boolean_t do_footprint, old_no_footprint, new_no_footprint;
8168 boolean_t new_task_objq_locked;
8169
8170 vm_object_lock_assert_exclusive(object);
8171
8172 if (!object->internal) {
8173 return KERN_INVALID_ARGUMENT;
8174 }
8175 if (new_owner == VM_OBJECT_OWNER_UNCHANGED) {
8176 /* leave owner unchanged */
8177 new_owner = VM_OBJECT_OWNER(object);
8178 }
8179 if (new_ledger_tag == VM_LEDGER_TAG_UNCHANGED) {
8180 /* leave ledger_tag unchanged */
8181 new_ledger_tag = object->vo_ledger_tag;
8182 }
8183 if (new_ledger_tag == VM_LEDGER_TAG_NONE &&
8184 object->purgable == VM_PURGABLE_DENY) {
8185 /* non-purgeable memory must have a valid non-zero ledger tag */
8186 return KERN_INVALID_ARGUMENT;
8187 }
8188 if (new_ledger_tag < 0 ||
8189 new_ledger_tag > VM_LEDGER_TAG_MAX) {
8190 return KERN_INVALID_ARGUMENT;
8191 }
8192 if (new_ledger_flags & ~VM_LEDGER_FLAGS) {
8193 return KERN_INVALID_ARGUMENT;
8194 }
8195 if (object->vo_ledger_tag == VM_LEDGER_TAG_NONE &&
8196 object->purgable == VM_PURGABLE_DENY) {
8197 /*
8198 * This VM object is neither ledger-tagged nor purgeable.
8199 * We can convert it to "ledger tag" ownership iff it
8200 * has not been used at all yet (no resident pages and
8201 * no pager) and it's going to be assigned to a valid task.
8202 */
8203 if (object->resident_page_count != 0 ||
8204 object->pager != NULL ||
8205 object->pager_created ||
8206 object->ref_count != 1 ||
8207 object->vo_owner != TASK_NULL ||
8208 object->copy_strategy != MEMORY_OBJECT_COPY_NONE ||
8209 new_owner == TASK_NULL) {
8210 return KERN_FAILURE;
8211 }
8212 }
8213
8214 if (new_ledger_flags & VM_LEDGER_FLAG_NO_FOOTPRINT) {
8215 new_no_footprint = TRUE;
8216 } else {
8217 new_no_footprint = FALSE;
8218 }
8219#if __arm64__
8220 if (!new_no_footprint &&
8221 object->purgable != VM_PURGABLE_DENY &&
8222 new_owner != TASK_NULL &&
8223 new_owner != VM_OBJECT_OWNER_DISOWNED &&
8224 new_owner->task_legacy_footprint) {
8225 /*
8226 * This task has been granted "legacy footprint" and should
8227 * not be charged for its IOKit purgeable memory. Since we
8228 * might now change the accounting of such memory to the
8229 * "graphics" ledger, for example, give it the "no footprint"
8230 * option.
8231 */
8232 new_no_footprint = TRUE;
8233 }
8234#endif /* __arm64__ */
8235 assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
8236 assert(object->shadow == VM_OBJECT_NULL);
8237 assert(object->vo_copy == VM_OBJECT_NULL);
8238
8239 old_ledger_tag = object->vo_ledger_tag;
8240 old_no_footprint = object->vo_no_footprint;
8241 old_owner = VM_OBJECT_OWNER(object);
8242
8243 if (__improbable(vm_debug_events)) {
8244 DTRACE_VM8(object_ownership_change,
8245 vm_object_t, object,
8246 task_t, old_owner,
8247 int, old_ledger_tag,
8248 int, old_no_footprint,
8249 task_t, new_owner,
8250 int, new_ledger_tag,
8251 int, new_no_footprint,
8252 int, VM_OBJECT_ID(object));
8253 }
8254
8255 assert(object->internal);
8256 resident_count = object->resident_page_count - object->wired_page_count;
8257 wired_count = object->wired_page_count;
8258 compressed_count = vm_compressor_pager_get_count(mem_obj: object->pager);
8259
8260 /*
8261 * Deal with the old owner and/or ledger tag, if needed.
8262 */
8263 if (old_owner != TASK_NULL &&
8264 ((old_owner != new_owner) /* new owner ... */
8265 || /* ... or ... */
8266 (old_no_footprint != new_no_footprint) /* new "no_footprint" */
8267 || /* ... or ... */
8268 old_ledger_tag != new_ledger_tag)) { /* ... new ledger */
8269 /*
8270 * Take this object off of the old owner's ledgers.
8271 */
8272 vm_object_ledger_tag_ledgers(object,
8273 ledger_idx_volatile: &ledger_idx_volatile,
8274 ledger_idx_nonvolatile: &ledger_idx_nonvolatile,
8275 ledger_idx_volatile_compressed: &ledger_idx_volatile_compressed,
8276 ledger_idx_nonvolatile_compressed: &ledger_idx_nonvolatile_compressed,
8277 do_footprint: &do_footprint);
8278 if (object->purgable == VM_PURGABLE_VOLATILE ||
8279 object->purgable == VM_PURGABLE_EMPTY) {
8280 ledger_idx = ledger_idx_volatile;
8281 ledger_idx_compressed = ledger_idx_volatile_compressed;
8282 } else {
8283 ledger_idx = ledger_idx_nonvolatile;
8284 ledger_idx_compressed = ledger_idx_nonvolatile_compressed;
8285 }
8286 if (resident_count) {
8287 /*
8288 * Adjust the appropriate old owners's ledgers by the
8289 * number of resident pages.
8290 */
8291 ledger_debit(ledger: old_owner->ledger,
8292 entry: ledger_idx,
8293 ptoa_64(resident_count));
8294 /* adjust old owner's footprint */
8295 if (do_footprint &&
8296 object->purgable != VM_PURGABLE_VOLATILE &&
8297 object->purgable != VM_PURGABLE_EMPTY) {
8298 ledger_debit(ledger: old_owner->ledger,
8299 entry: task_ledgers.phys_footprint,
8300 ptoa_64(resident_count));
8301 }
8302 }
8303 if (wired_count) {
8304 /* wired pages are always nonvolatile */
8305 ledger_debit(ledger: old_owner->ledger,
8306 entry: ledger_idx_nonvolatile,
8307 ptoa_64(wired_count));
8308 if (do_footprint) {
8309 ledger_debit(ledger: old_owner->ledger,
8310 entry: task_ledgers.phys_footprint,
8311 ptoa_64(wired_count));
8312 }
8313 }
8314 if (compressed_count) {
8315 /*
8316 * Adjust the appropriate old owner's ledgers
8317 * by the number of compressed pages.
8318 */
8319 ledger_debit(ledger: old_owner->ledger,
8320 entry: ledger_idx_compressed,
8321 ptoa_64(compressed_count));
8322 if (do_footprint &&
8323 object->purgable != VM_PURGABLE_VOLATILE &&
8324 object->purgable != VM_PURGABLE_EMPTY) {
8325 ledger_debit(ledger: old_owner->ledger,
8326 entry: task_ledgers.phys_footprint,
8327 ptoa_64(compressed_count));
8328 }
8329 }
8330 if (old_owner != new_owner) {
8331 /* remove object from old_owner's list of owned objects */
8332 DTRACE_VM2(object_owner_remove,
8333 vm_object_t, object,
8334 task_t, old_owner);
8335 if (!old_task_objq_locked) {
8336 task_objq_lock(old_owner);
8337 }
8338 old_owner->task_owned_objects--;
8339 queue_remove(&old_owner->task_objq, object,
8340 vm_object_t, task_objq);
8341 switch (object->purgable) {
8342 case VM_PURGABLE_NONVOLATILE:
8343 case VM_PURGABLE_EMPTY:
8344 vm_purgeable_nonvolatile_owner_update(owner: old_owner,
8345 delta: -1);
8346 break;
8347 case VM_PURGABLE_VOLATILE:
8348 vm_purgeable_volatile_owner_update(owner: old_owner,
8349 delta: -1);
8350 break;
8351 default:
8352 break;
8353 }
8354 if (!old_task_objq_locked) {
8355 task_objq_unlock(old_owner);
8356 }
8357 }
8358 }
8359
8360 /*
8361 * Switch to new ledger tag and/or owner.
8362 */
8363
8364 new_task_objq_locked = FALSE;
8365 if (new_owner != old_owner &&
8366 new_owner != TASK_NULL &&
8367 new_owner != VM_OBJECT_OWNER_DISOWNED) {
8368 /*
8369 * If the new owner is not accepting new objects ("disowning"),
8370 * the object becomes "disowned" and will be added to
8371 * the kernel's task_objq.
8372 *
8373 * Check first without locking, to avoid blocking while the
8374 * task is disowning its objects.
8375 */
8376 if (new_owner->task_objects_disowning) {
8377 new_owner = VM_OBJECT_OWNER_DISOWNED;
8378 } else {
8379 task_objq_lock(new_owner);
8380 /* check again now that we have the lock */
8381 if (new_owner->task_objects_disowning) {
8382 new_owner = VM_OBJECT_OWNER_DISOWNED;
8383 task_objq_unlock(new_owner);
8384 } else {
8385 new_task_objq_locked = TRUE;
8386 }
8387 }
8388 }
8389
8390 object->vo_ledger_tag = new_ledger_tag;
8391 object->vo_owner = new_owner;
8392 object->vo_no_footprint = new_no_footprint;
8393
8394 if (new_owner == VM_OBJECT_OWNER_DISOWNED) {
8395 /*
8396 * Disowned objects are added to the kernel's task_objq but
8397 * are marked as owned by "VM_OBJECT_OWNER_DISOWNED" to
8398 * differentiate them from objects intentionally owned by
8399 * the kernel.
8400 */
8401 assert(old_owner != kernel_task);
8402 new_owner = kernel_task;
8403 assert(!new_task_objq_locked);
8404 task_objq_lock(new_owner);
8405 new_task_objq_locked = TRUE;
8406 }
8407
8408 /*
8409 * Deal with the new owner and/or ledger tag, if needed.
8410 */
8411 if (new_owner != TASK_NULL &&
8412 ((new_owner != old_owner) /* new owner ... */
8413 || /* ... or ... */
8414 (new_no_footprint != old_no_footprint) /* ... new "no_footprint" */
8415 || /* ... or ... */
8416 new_ledger_tag != old_ledger_tag)) { /* ... new ledger */
8417 /*
8418 * Add this object to the new owner's ledgers.
8419 */
8420 vm_object_ledger_tag_ledgers(object,
8421 ledger_idx_volatile: &ledger_idx_volatile,
8422 ledger_idx_nonvolatile: &ledger_idx_nonvolatile,
8423 ledger_idx_volatile_compressed: &ledger_idx_volatile_compressed,
8424 ledger_idx_nonvolatile_compressed: &ledger_idx_nonvolatile_compressed,
8425 do_footprint: &do_footprint);
8426 if (object->purgable == VM_PURGABLE_VOLATILE ||
8427 object->purgable == VM_PURGABLE_EMPTY) {
8428 ledger_idx = ledger_idx_volatile;
8429 ledger_idx_compressed = ledger_idx_volatile_compressed;
8430 } else {
8431 ledger_idx = ledger_idx_nonvolatile;
8432 ledger_idx_compressed = ledger_idx_nonvolatile_compressed;
8433 }
8434 if (resident_count) {
8435 /*
8436 * Adjust the appropriate new owners's ledgers by the
8437 * number of resident pages.
8438 */
8439 ledger_credit(ledger: new_owner->ledger,
8440 entry: ledger_idx,
8441 ptoa_64(resident_count));
8442 /* adjust new owner's footprint */
8443 if (do_footprint &&
8444 object->purgable != VM_PURGABLE_VOLATILE &&
8445 object->purgable != VM_PURGABLE_EMPTY) {
8446 ledger_credit(ledger: new_owner->ledger,
8447 entry: task_ledgers.phys_footprint,
8448 ptoa_64(resident_count));
8449 }
8450 }
8451 if (wired_count) {
8452 /* wired pages are always nonvolatile */
8453 ledger_credit(ledger: new_owner->ledger,
8454 entry: ledger_idx_nonvolatile,
8455 ptoa_64(wired_count));
8456 if (do_footprint) {
8457 ledger_credit(ledger: new_owner->ledger,
8458 entry: task_ledgers.phys_footprint,
8459 ptoa_64(wired_count));
8460 }
8461 }
8462 if (compressed_count) {
8463 /*
8464 * Adjust the new owner's ledgers by the number of
8465 * compressed pages.
8466 */
8467 ledger_credit(ledger: new_owner->ledger,
8468 entry: ledger_idx_compressed,
8469 ptoa_64(compressed_count));
8470 if (do_footprint &&
8471 object->purgable != VM_PURGABLE_VOLATILE &&
8472 object->purgable != VM_PURGABLE_EMPTY) {
8473 ledger_credit(ledger: new_owner->ledger,
8474 entry: task_ledgers.phys_footprint,
8475 ptoa_64(compressed_count));
8476 }
8477 }
8478 if (new_owner != old_owner) {
8479 /* add object to new_owner's list of owned objects */
8480 DTRACE_VM2(object_owner_add,
8481 vm_object_t, object,
8482 task_t, new_owner);
8483 assert(new_task_objq_locked);
8484 new_owner->task_owned_objects++;
8485 queue_enter(&new_owner->task_objq, object,
8486 vm_object_t, task_objq);
8487 switch (object->purgable) {
8488 case VM_PURGABLE_NONVOLATILE:
8489 case VM_PURGABLE_EMPTY:
8490 vm_purgeable_nonvolatile_owner_update(owner: new_owner,
8491 delta: +1);
8492 break;
8493 case VM_PURGABLE_VOLATILE:
8494 vm_purgeable_volatile_owner_update(owner: new_owner,
8495 delta: +1);
8496 break;
8497 default:
8498 break;
8499 }
8500 }
8501 }
8502
8503 if (new_task_objq_locked) {
8504 task_objq_unlock(new_owner);
8505 }
8506
8507 return KERN_SUCCESS;
8508}
8509
8510void
8511vm_owned_objects_disown(
8512 task_t task)
8513{
8514 vm_object_t next_object;
8515 vm_object_t object;
8516 int collisions;
8517 kern_return_t kr;
8518
8519 if (task == NULL) {
8520 return;
8521 }
8522
8523 collisions = 0;
8524
8525again:
8526 if (task->task_objects_disowned) {
8527 /* task has already disowned its owned objects */
8528 assert(task->task_volatile_objects == 0);
8529 assert(task->task_nonvolatile_objects == 0);
8530 assert(task->task_owned_objects == 0);
8531 return;
8532 }
8533
8534 task_objq_lock(task);
8535
8536 task->task_objects_disowning = TRUE;
8537
8538 for (object = (vm_object_t) queue_first(&task->task_objq);
8539 !queue_end(&task->task_objq, (queue_entry_t) object);
8540 object = next_object) {
8541 if (task->task_nonvolatile_objects == 0 &&
8542 task->task_volatile_objects == 0 &&
8543 task->task_owned_objects == 0) {
8544 /* no more objects owned by "task" */
8545 break;
8546 }
8547
8548 next_object = (vm_object_t) queue_next(&object->task_objq);
8549
8550#if DEBUG
8551 assert(object->vo_purgeable_volatilizer == NULL);
8552#endif /* DEBUG */
8553 assert(object->vo_owner == task);
8554 if (!vm_object_lock_try(object)) {
8555 task_objq_unlock(task);
8556 mutex_pause(collisions++);
8557 goto again;
8558 }
8559 /* transfer ownership to the kernel */
8560 assert(VM_OBJECT_OWNER(object) != kernel_task);
8561 kr = vm_object_ownership_change(
8562 object,
8563 new_ledger_tag: object->vo_ledger_tag, /* unchanged */
8564 VM_OBJECT_OWNER_DISOWNED, /* new owner */
8565 new_ledger_flags: 0, /* new_ledger_flags */
8566 TRUE); /* old_owner->task_objq locked */
8567 assert(kr == KERN_SUCCESS);
8568 assert(object->vo_owner == VM_OBJECT_OWNER_DISOWNED);
8569 vm_object_unlock(object);
8570 }
8571
8572 if (__improbable(task->task_owned_objects != 0)) {
8573 panic("%s(%p): volatile=%d nonvolatile=%d owned=%d q=%p q_first=%p q_last=%p",
8574 __FUNCTION__,
8575 task,
8576 task->task_volatile_objects,
8577 task->task_nonvolatile_objects,
8578 task->task_owned_objects,
8579 &task->task_objq,
8580 queue_first(&task->task_objq),
8581 queue_last(&task->task_objq));
8582 }
8583
8584 /* there shouldn't be any objects owned by task now */
8585 assert(task->task_volatile_objects == 0);
8586 assert(task->task_nonvolatile_objects == 0);
8587 assert(task->task_owned_objects == 0);
8588 assert(task->task_objects_disowning);
8589
8590 /* and we don't need to try and disown again */
8591 task->task_objects_disowned = TRUE;
8592
8593 task_objq_unlock(task);
8594}
8595