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#include <vm/vm_compressor.h>
30
31#if CONFIG_PHANTOM_CACHE
32#include <vm/vm_phantom_cache.h>
33#endif
34
35#include <vm/vm_map.h>
36#include <vm/vm_pageout.h>
37#include <vm/memory_object.h>
38#include <vm/vm_compressor_algorithms.h>
39#include <vm/vm_compressor_backing_store.h>
40#include <vm/vm_fault.h>
41#include <vm/vm_protos.h>
42#include <mach/mach_host.h> /* for host_info() */
43#if DEVELOPMENT || DEBUG
44#include <kern/hvg_hypercall.h>
45#endif
46#include <kern/ledger.h>
47#include <kern/policy_internal.h>
48#include <kern/thread_group.h>
49#include <san/kasan.h>
50#include <os/log.h>
51#include <pexpert/pexpert.h>
52#include <pexpert/device_tree.h>
53
54#if defined(__x86_64__)
55#include <i386/misc_protos.h>
56#endif
57#if defined(__arm64__)
58#include <arm/machine_routines.h>
59#endif
60
61#include <IOKit/IOHibernatePrivate.h>
62
63/*
64 * The segment buffer size is a tradeoff.
65 * A larger buffer leads to faster I/O throughput, better compression ratios
66 * (since fewer bytes are wasted at the end of the segment),
67 * and less overhead (both in time and space).
68 * However, a smaller buffer causes less swap when the system is overcommited
69 * b/c a higher percentage of the swapped-in segment is definitely accessed
70 * before it goes back out to storage.
71 *
72 * So on systems without swap, a larger segment is a clear win.
73 * On systems with swap, the choice is murkier. Empirically, we've
74 * found that a 64KB segment provides a better tradeoff both in terms of
75 * performance and swap writes than a 256KB segment on systems with fast SSDs
76 * and a HW compression block.
77 */
78#define C_SEG_BUFSIZE_ARM_SWAP (1024 * 64)
79#if XNU_TARGET_OS_OSX && defined(__arm64__)
80#define C_SEG_BUFSIZE_DEFAULT C_SEG_BUFSIZE_ARM_SWAP
81#else
82#define C_SEG_BUFSIZE_DEFAULT (1024 * 256)
83#endif /* TARGET_OS_OSX && defined(__arm64__) */
84uint32_t c_seg_bufsize;
85
86uint32_t c_seg_max_pages, c_seg_off_limit, c_seg_allocsize, c_seg_slot_var_array_min_len;
87
88extern boolean_t vm_darkwake_mode;
89extern zone_t vm_page_zone;
90
91#if DEVELOPMENT || DEBUG
92/* sysctl defined in bsd/dev/arm64/sysctl.c */
93int do_cseg_wedge_thread(void);
94int do_cseg_unwedge_thread(void);
95static event_t debug_cseg_wait_event = NULL;
96#endif /* DEVELOPMENT || DEBUG */
97
98#if CONFIG_FREEZE
99bool freezer_incore_cseg_acct = TRUE; /* Only count incore compressed memory for jetsams. */
100void task_disown_frozen_csegs(task_t owner_task);
101#endif /* CONFIG_FREEZE */
102
103#if POPCOUNT_THE_COMPRESSED_DATA
104boolean_t popcount_c_segs = TRUE;
105
106static inline uint32_t
107vmc_pop(uintptr_t ins, int sz)
108{
109 uint32_t rv = 0;
110
111 if (__probable(popcount_c_segs == FALSE)) {
112 return 0xDEAD707C;
113 }
114
115 while (sz >= 16) {
116 uint32_t rv1, rv2;
117 uint64_t *ins64 = (uint64_t *) ins;
118 uint64_t *ins642 = (uint64_t *) (ins + 8);
119 rv1 = __builtin_popcountll(*ins64);
120 rv2 = __builtin_popcountll(*ins642);
121 rv += rv1 + rv2;
122 sz -= 16;
123 ins += 16;
124 }
125
126 while (sz >= 4) {
127 uint32_t *ins32 = (uint32_t *) ins;
128 rv += __builtin_popcount(*ins32);
129 sz -= 4;
130 ins += 4;
131 }
132
133 while (sz > 0) {
134 char *ins8 = (char *)ins;
135 rv += __builtin_popcount(*ins8);
136 sz--;
137 ins++;
138 }
139 return rv;
140}
141#endif
142
143#if VALIDATE_C_SEGMENTS
144boolean_t validate_c_segs = TRUE;
145#endif
146/*
147 * vm_compressor_mode has a hierarchy of control to set its value.
148 * boot-args are checked first, then device-tree, and finally
149 * the default value that is defined below. See vm_fault_init() for
150 * the boot-arg & device-tree code.
151 */
152
153#if !XNU_TARGET_OS_OSX
154
155#if CONFIG_FREEZE
156int vm_compressor_mode = VM_PAGER_FREEZER_DEFAULT;
157struct freezer_context freezer_context_global;
158#else /* CONFIG_FREEZE */
159int vm_compressor_mode = VM_PAGER_NOT_CONFIGURED;
160#endif /* CONFIG_FREEZE */
161
162#else /* !XNU_TARGET_OS_OSX */
163int vm_compressor_mode = VM_PAGER_COMPRESSOR_WITH_SWAP;
164
165#endif /* !XNU_TARGET_OS_OSX */
166
167TUNABLE(uint32_t, vm_compression_limit, "vm_compression_limit", 0);
168int vm_compressor_is_active = 0;
169int vm_compressor_available = 0;
170
171extern uint64_t vm_swap_get_max_configured_space(void);
172extern void vm_pageout_io_throttle(void);
173bool vm_compressor_swapout_is_ripe(void);
174
175#if CHECKSUM_THE_DATA || CHECKSUM_THE_SWAP || CHECKSUM_THE_COMPRESSED_DATA
176extern unsigned int hash_string(char *cp, int len);
177static unsigned int vmc_hash(char *, int);
178boolean_t checksum_c_segs = TRUE;
179
180unsigned int
181vmc_hash(char *cp, int len)
182{
183 if (__probable(checksum_c_segs == FALSE)) {
184 return 0xDEAD7A37;
185 }
186 return hash_string(cp, len);
187}
188#endif
189
190#define UNPACK_C_SIZE(cs) ((cs->c_size == (PAGE_SIZE-1)) ? PAGE_SIZE : cs->c_size)
191#define PACK_C_SIZE(cs, size) (cs->c_size = ((size == PAGE_SIZE) ? PAGE_SIZE - 1 : size))
192
193
194struct c_sv_hash_entry {
195 union {
196 struct {
197 uint32_t c_sv_he_ref;
198 uint32_t c_sv_he_data;
199 } c_sv_he;
200 uint64_t c_sv_he_record;
201 } c_sv_he_un;
202};
203
204#define he_ref c_sv_he_un.c_sv_he.c_sv_he_ref
205#define he_data c_sv_he_un.c_sv_he.c_sv_he_data
206#define he_record c_sv_he_un.c_sv_he_record
207
208#define C_SV_HASH_MAX_MISS 32
209#define C_SV_HASH_SIZE ((1 << 10))
210#define C_SV_HASH_MASK ((1 << 10) - 1)
211
212#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
213#define C_SV_CSEG_ID ((1 << 21) - 1)
214#else /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
215#define C_SV_CSEG_ID ((1 << 22) - 1)
216#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
217
218
219union c_segu {
220 c_segment_t c_seg;
221 uintptr_t c_segno;
222};
223
224#define C_SLOT_ASSERT_PACKABLE(ptr) \
225 VM_ASSERT_POINTER_PACKABLE((vm_offset_t)(ptr), C_SLOT_PACKED_PTR);
226
227#define C_SLOT_PACK_PTR(ptr) \
228 VM_PACK_POINTER((vm_offset_t)(ptr), C_SLOT_PACKED_PTR)
229
230#define C_SLOT_UNPACK_PTR(cslot) \
231 (c_slot_mapping_t)VM_UNPACK_POINTER((cslot)->c_packed_ptr, C_SLOT_PACKED_PTR)
232
233/* for debugging purposes */
234SECURITY_READ_ONLY_EARLY(vm_packing_params_t) c_slot_packing_params =
235 VM_PACKING_PARAMS(C_SLOT_PACKED_PTR);
236
237uint32_t c_segment_count = 0;
238uint32_t c_segment_count_max = 0;
239
240uint64_t c_generation_id = 0;
241uint64_t c_generation_id_flush_barrier;
242
243
244#define HIBERNATE_FLUSHING_SECS_TO_COMPLETE 120
245
246boolean_t hibernate_no_swapspace = FALSE;
247boolean_t hibernate_flush_timed_out = FALSE;
248clock_sec_t hibernate_flushing_deadline = 0;
249
250#if RECORD_THE_COMPRESSED_DATA
251char *c_compressed_record_sbuf;
252char *c_compressed_record_ebuf;
253char *c_compressed_record_cptr;
254#endif
255
256
257queue_head_t c_age_list_head;
258queue_head_t c_early_swappedin_list_head, c_regular_swappedin_list_head, c_late_swappedin_list_head;
259queue_head_t c_early_swapout_list_head, c_regular_swapout_list_head, c_late_swapout_list_head;
260queue_head_t c_swapio_list_head;
261queue_head_t c_swappedout_list_head;
262queue_head_t c_swappedout_sparse_list_head;
263queue_head_t c_major_list_head;
264queue_head_t c_filling_list_head;
265queue_head_t c_bad_list_head;
266
267uint32_t c_age_count = 0;
268uint32_t c_early_swappedin_count = 0, c_regular_swappedin_count = 0, c_late_swappedin_count = 0;
269uint32_t c_early_swapout_count = 0, c_regular_swapout_count = 0, c_late_swapout_count = 0;
270uint32_t c_swapio_count = 0;
271uint32_t c_swappedout_count = 0;
272uint32_t c_swappedout_sparse_count = 0;
273uint32_t c_major_count = 0;
274uint32_t c_filling_count = 0;
275uint32_t c_empty_count = 0;
276uint32_t c_bad_count = 0;
277
278
279queue_head_t c_minor_list_head;
280uint32_t c_minor_count = 0;
281
282int c_overage_swapped_count = 0;
283int c_overage_swapped_limit = 0;
284
285int c_seg_fixed_array_len;
286union c_segu *c_segments;
287vm_offset_t c_buffers;
288vm_size_t c_buffers_size;
289caddr_t c_segments_next_page;
290boolean_t c_segments_busy;
291uint32_t c_segments_available;
292uint32_t c_segments_limit;
293uint32_t c_segments_nearing_limit;
294
295uint32_t c_segment_svp_in_hash;
296uint32_t c_segment_svp_hash_succeeded;
297uint32_t c_segment_svp_hash_failed;
298uint32_t c_segment_svp_zero_compressions;
299uint32_t c_segment_svp_nonzero_compressions;
300uint32_t c_segment_svp_zero_decompressions;
301uint32_t c_segment_svp_nonzero_decompressions;
302
303uint32_t c_segment_noncompressible_pages;
304
305uint32_t c_segment_pages_compressed = 0; /* Tracks # of uncompressed pages fed into the compressor */
306#if CONFIG_FREEZE
307int32_t c_segment_pages_compressed_incore = 0; /* Tracks # of uncompressed pages fed into the compressor that are in memory */
308int32_t c_segment_pages_compressed_incore_late_swapout = 0; /* Tracks # of uncompressed pages fed into the compressor that are in memory and tagged for swapout */
309uint32_t c_segments_incore_limit = 0; /* Tracks # of segments allowed to be in-core. Based on compressor pool size */
310#endif /* CONFIG_FREEZE */
311
312uint32_t c_segment_pages_compressed_limit;
313uint32_t c_segment_pages_compressed_nearing_limit;
314uint32_t c_free_segno_head = (uint32_t)-1;
315
316uint32_t vm_compressor_minorcompact_threshold_divisor = 10;
317uint32_t vm_compressor_majorcompact_threshold_divisor = 10;
318uint32_t vm_compressor_unthrottle_threshold_divisor = 10;
319uint32_t vm_compressor_catchup_threshold_divisor = 10;
320
321uint32_t vm_compressor_minorcompact_threshold_divisor_overridden = 0;
322uint32_t vm_compressor_majorcompact_threshold_divisor_overridden = 0;
323uint32_t vm_compressor_unthrottle_threshold_divisor_overridden = 0;
324uint32_t vm_compressor_catchup_threshold_divisor_overridden = 0;
325
326#define C_SEGMENTS_PER_PAGE (PAGE_SIZE / sizeof(union c_segu))
327
328LCK_GRP_DECLARE(vm_compressor_lck_grp, "vm_compressor");
329LCK_RW_DECLARE(c_master_lock, &vm_compressor_lck_grp);
330LCK_MTX_DECLARE(c_list_lock_storage, &vm_compressor_lck_grp);
331
332boolean_t decompressions_blocked = FALSE;
333
334zone_t compressor_segment_zone;
335int c_compressor_swap_trigger = 0;
336
337uint32_t compressor_cpus;
338char *compressor_scratch_bufs;
339char *kdp_compressor_scratch_buf;
340char *kdp_compressor_decompressed_page;
341addr64_t kdp_compressor_decompressed_page_paddr;
342ppnum_t kdp_compressor_decompressed_page_ppnum;
343
344clock_sec_t start_of_sample_period_sec = 0;
345clock_nsec_t start_of_sample_period_nsec = 0;
346clock_sec_t start_of_eval_period_sec = 0;
347clock_nsec_t start_of_eval_period_nsec = 0;
348uint32_t sample_period_decompression_count = 0;
349uint32_t sample_period_compression_count = 0;
350uint32_t last_eval_decompression_count = 0;
351uint32_t last_eval_compression_count = 0;
352
353#define DECOMPRESSION_SAMPLE_MAX_AGE (60 * 30)
354
355boolean_t vm_swapout_ripe_segments = FALSE;
356uint32_t vm_ripe_target_age = (60 * 60 * 48);
357
358uint32_t swapout_target_age = 0;
359uint32_t age_of_decompressions_during_sample_period[DECOMPRESSION_SAMPLE_MAX_AGE];
360uint32_t overage_decompressions_during_sample_period = 0;
361
362
363void do_fastwake_warmup(queue_head_t *, boolean_t);
364boolean_t fastwake_warmup = FALSE;
365boolean_t fastwake_recording_in_progress = FALSE;
366clock_sec_t dont_trim_until_ts = 0;
367
368uint64_t c_segment_warmup_count;
369uint64_t first_c_segment_to_warm_generation_id = 0;
370uint64_t last_c_segment_to_warm_generation_id = 0;
371boolean_t hibernate_flushing = FALSE;
372
373int64_t c_segment_input_bytes __attribute__((aligned(8))) = 0;
374int64_t c_segment_compressed_bytes __attribute__((aligned(8))) = 0;
375int64_t compressor_bytes_used __attribute__((aligned(8))) = 0;
376
377/* Keeps track of the most recent timestamp for when major compaction finished. */
378mach_timespec_t major_compact_ts;
379
380struct c_sv_hash_entry c_segment_sv_hash_table[C_SV_HASH_SIZE] __attribute__ ((aligned(8)));
381
382static void vm_compressor_swap_trigger_thread(void);
383static void vm_compressor_do_delayed_compactions(boolean_t);
384static void vm_compressor_compact_and_swap(boolean_t);
385static void vm_compressor_process_regular_swapped_in_segments(boolean_t);
386void vm_compressor_process_special_swapped_in_segments(void);
387static void vm_compressor_process_special_swapped_in_segments_locked(void);
388
389struct vm_compressor_swapper_stats vmcs_stats;
390
391static void vm_compressor_process_major_segments(bool);
392#if XNU_TARGET_OS_OSX
393static void vm_compressor_take_paging_space_action(void);
394#endif /* XNU_TARGET_OS_OSX */
395
396void compute_swapout_target_age(void);
397
398boolean_t c_seg_major_compact(c_segment_t, c_segment_t);
399boolean_t c_seg_major_compact_ok(c_segment_t, c_segment_t);
400
401int c_seg_minor_compaction_and_unlock(c_segment_t, boolean_t);
402int c_seg_do_minor_compaction_and_unlock(c_segment_t, boolean_t, boolean_t, boolean_t);
403void c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg);
404
405void c_seg_move_to_sparse_list(c_segment_t);
406void c_seg_insert_into_q(queue_head_t *, c_segment_t);
407
408uint64_t vm_available_memory(void);
409uint64_t vm_compressor_pages_compressed(void);
410uint32_t vm_compressor_pool_size(void);
411uint32_t vm_compressor_fragmentation_level(void);
412uint32_t vm_compression_ratio(void);
413
414/*
415 * indicate the need to do a major compaction if
416 * the overall set of in-use compression segments
417 * becomes sparse... on systems that support pressure
418 * driven swapping, this will also cause swapouts to
419 * be initiated.
420 */
421static inline bool
422vm_compressor_needs_to_major_compact()
423{
424 uint32_t incore_seg_count;
425
426 incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count;
427
428 if ((c_segment_count >= (c_segments_nearing_limit / 8)) &&
429 ((incore_seg_count * c_seg_max_pages) - VM_PAGE_COMPRESSOR_COUNT) >
430 ((incore_seg_count / 8) * c_seg_max_pages)) {
431 return true;
432 }
433 return false;
434}
435
436
437uint64_t
438vm_available_memory(void)
439{
440 return ((uint64_t)AVAILABLE_NON_COMPRESSED_MEMORY) * PAGE_SIZE_64;
441}
442
443
444uint32_t
445vm_compressor_pool_size(void)
446{
447 return VM_PAGE_COMPRESSOR_COUNT;
448}
449
450uint32_t
451vm_compressor_fragmentation_level(void)
452{
453 const uint32_t incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count;
454 if ((incore_seg_count == 0) || (c_seg_max_pages == 0)) {
455 return 0;
456 }
457 return 100 - (vm_compressor_pool_size() * 100 / (incore_seg_count * c_seg_max_pages));
458}
459
460uint32_t
461vm_compression_ratio(void)
462{
463 if (vm_compressor_pool_size() == 0) {
464 return UINT32_MAX;
465 }
466 return c_segment_pages_compressed / vm_compressor_pool_size();
467}
468
469uint64_t
470vm_compressor_pages_compressed(void)
471{
472 return c_segment_pages_compressed * PAGE_SIZE_64;
473}
474
475bool
476vm_compressor_compressed_pages_nearing_limit(void)
477{
478 uint32_t pages = 0;
479
480#if CONFIG_FREEZE
481 pages = os_atomic_load(&c_segment_pages_compressed_incore, relaxed);
482#else /* CONFIG_FREEZE */
483 pages = c_segment_pages_compressed;
484#endif /* CONFIG_FREEZE */
485
486 return pages > c_segment_pages_compressed_nearing_limit;
487}
488
489static bool
490vm_compressor_segments_nearing_limit(void)
491{
492 uint64_t segments;
493
494#if CONFIG_FREEZE
495 if (freezer_incore_cseg_acct) {
496 if (os_sub_overflow(c_segment_count, c_swappedout_count, &segments)) {
497 segments = 0;
498 }
499 if (os_sub_overflow(segments, c_swappedout_sparse_count, &segments)) {
500 segments = 0;
501 }
502 } else {
503 segments = os_atomic_load(&c_segment_count, relaxed);
504 }
505#else /* CONFIG_FREEZE */
506 segments = c_segment_count;
507#endif /* CONFIG_FREEZE */
508
509 return segments > c_segments_nearing_limit;
510}
511
512boolean_t
513vm_compressor_low_on_space(void)
514{
515 return vm_compressor_compressed_pages_nearing_limit() ||
516 vm_compressor_segments_nearing_limit();
517}
518
519
520boolean_t
521vm_compressor_out_of_space(void)
522{
523#if CONFIG_FREEZE
524 uint64_t incore_seg_count;
525 uint32_t incore_compressed_pages;
526 if (freezer_incore_cseg_acct) {
527 if (os_sub_overflow(c_segment_count, c_swappedout_count, &incore_seg_count)) {
528 incore_seg_count = 0;
529 }
530 if (os_sub_overflow(incore_seg_count, c_swappedout_sparse_count, &incore_seg_count)) {
531 incore_seg_count = 0;
532 }
533 incore_compressed_pages = os_atomic_load(&c_segment_pages_compressed_incore, relaxed);
534 } else {
535 incore_seg_count = os_atomic_load(&c_segment_count, relaxed);
536 incore_compressed_pages = os_atomic_load(&c_segment_pages_compressed_incore, relaxed);
537 }
538
539 if ((incore_compressed_pages >= c_segment_pages_compressed_limit) ||
540 (incore_seg_count > c_segments_incore_limit)) {
541 return TRUE;
542 }
543#else /* CONFIG_FREEZE */
544 if ((c_segment_pages_compressed >= c_segment_pages_compressed_limit) ||
545 (c_segment_count >= c_segments_limit)) {
546 return TRUE;
547 }
548#endif /* CONFIG_FREEZE */
549 return FALSE;
550}
551
552bool
553vm_compressor_is_thrashing()
554{
555 compute_swapout_target_age();
556
557 if (swapout_target_age) {
558 c_segment_t c_seg;
559
560 lck_mtx_lock_spin_always(c_list_lock);
561
562 if (!queue_empty(&c_age_list_head)) {
563 c_seg = (c_segment_t) queue_first(&c_age_list_head);
564
565 if (c_seg->c_creation_ts > swapout_target_age) {
566 swapout_target_age = 0;
567 }
568 }
569 lck_mtx_unlock_always(c_list_lock);
570 }
571
572 return swapout_target_age != 0;
573}
574
575
576int
577vm_wants_task_throttled(task_t task)
578{
579 ledger_amount_t compressed;
580 if (task == kernel_task) {
581 return 0;
582 }
583
584 if (VM_CONFIG_SWAP_IS_ACTIVE) {
585 if ((vm_compressor_low_on_space() || HARD_THROTTLE_LIMIT_REACHED())) {
586 ledger_get_balance(ledger: task->ledger, entry: task_ledgers.internal_compressed, balance: &compressed);
587 compressed >>= VM_MAP_PAGE_SHIFT(map: task->map);
588 if ((unsigned int)compressed > (c_segment_pages_compressed / 4)) {
589 return 1;
590 }
591 }
592 }
593 return 0;
594}
595
596
597#if DEVELOPMENT || DEBUG
598/*
599 * On compressor/swap exhaustion, kill the largest process regardless of
600 * its chosen process policy.
601 */
602TUNABLE(bool, kill_on_no_paging_space, "-kill_on_no_paging_space", false);
603#endif /* DEVELOPMENT || DEBUG */
604
605#if CONFIG_JETSAM
606boolean_t memorystatus_kill_on_VM_compressor_space_shortage(boolean_t);
607void memorystatus_thread_wake(void);
608extern uint32_t jetsam_kill_on_low_swap;
609bool memorystatus_disable_swap(void);
610#if CONFIG_PHANTOM_CACHE
611extern bool memorystatus_phantom_cache_pressure;
612#endif /* CONFIG_PHANTOM_CACHE */
613int compressor_thrashing_induced_jetsam = 0;
614int filecache_thrashing_induced_jetsam = 0;
615static boolean_t vm_compressor_thrashing_detected = FALSE;
616#else /* CONFIG_JETSAM */
617static uint32_t no_paging_space_action_in_progress = 0;
618extern void memorystatus_send_low_swap_note(void);
619#endif /* CONFIG_JETSAM */
620
621static void
622vm_compressor_take_paging_space_action(void)
623{
624#if CONFIG_JETSAM
625 /*
626 * On systems with both swap and jetsam,
627 * just wake up the jetsam thread and have it handle the low swap condition
628 * by killing apps.
629 */
630 if (jetsam_kill_on_low_swap) {
631 memorystatus_thread_wake();
632 }
633#else /* CONFIG_JETSAM */
634 if (no_paging_space_action_in_progress == 0) {
635 if (OSCompareAndSwap(0, 1, (UInt32 *)&no_paging_space_action_in_progress)) {
636 if (no_paging_space_action()) {
637#if DEVELOPMENT || DEBUG
638 if (kill_on_no_paging_space) {
639 /*
640 * Since we are choosing to always kill a process, we don't need the
641 * "out of application memory" dialog box in this mode. And, hence we won't
642 * send the knote.
643 */
644 no_paging_space_action_in_progress = 0;
645 return;
646 }
647#endif /* DEVELOPMENT || DEBUG */
648 memorystatus_send_low_swap_note();
649 }
650
651 no_paging_space_action_in_progress = 0;
652 }
653 }
654#endif /* !CONFIG_JETSAM */
655}
656
657
658void
659vm_decompressor_lock(void)
660{
661 PAGE_REPLACEMENT_ALLOWED(TRUE);
662
663 decompressions_blocked = TRUE;
664
665 PAGE_REPLACEMENT_ALLOWED(FALSE);
666}
667
668void
669vm_decompressor_unlock(void)
670{
671 PAGE_REPLACEMENT_ALLOWED(TRUE);
672
673 decompressions_blocked = FALSE;
674
675 PAGE_REPLACEMENT_ALLOWED(FALSE);
676
677 thread_wakeup((event_t)&decompressions_blocked);
678}
679
680static inline void
681cslot_copy(c_slot_t cdst, c_slot_t csrc)
682{
683#if CHECKSUM_THE_DATA
684 cdst->c_hash_data = csrc->c_hash_data;
685#endif
686#if CHECKSUM_THE_COMPRESSED_DATA
687 cdst->c_hash_compressed_data = csrc->c_hash_compressed_data;
688#endif
689#if POPCOUNT_THE_COMPRESSED_DATA
690 cdst->c_pop_cdata = csrc->c_pop_cdata;
691#endif
692 cdst->c_size = csrc->c_size;
693 cdst->c_packed_ptr = csrc->c_packed_ptr;
694#if defined(__arm64__)
695 cdst->c_codec = csrc->c_codec;
696#endif
697}
698
699#if XNU_TARGET_OS_OSX
700#define VM_COMPRESSOR_MAX_POOL_SIZE (192UL << 30)
701#else
702#define VM_COMPRESSOR_MAX_POOL_SIZE (0)
703#endif
704
705static vm_map_size_t compressor_size;
706static SECURITY_READ_ONLY_LATE(struct mach_vm_range) compressor_range;
707vm_map_t compressor_map;
708uint64_t compressor_pool_max_size;
709uint64_t compressor_pool_size;
710uint32_t compressor_pool_multiplier;
711
712#if DEVELOPMENT || DEBUG
713/*
714 * Compressor segments are write-protected in development/debug
715 * kernels to help debug memory corruption.
716 * In cases where performance is a concern, this can be disabled
717 * via the boot-arg "-disable_cseg_write_protection".
718 */
719boolean_t write_protect_c_segs = TRUE;
720int vm_compressor_test_seg_wp;
721uint32_t vm_ktrace_enabled;
722#endif /* DEVELOPMENT || DEBUG */
723
724#if (XNU_TARGET_OS_OSX && __arm64__)
725
726#include <IOKit/IOPlatformExpert.h>
727#include <sys/random.h>
728
729static const char *csegbufsizeExperimentProperty = "_csegbufsz_experiment";
730static thread_call_t csegbufsz_experiment_thread_call;
731
732extern boolean_t IOServiceWaitForMatchingResource(const char * property, uint64_t timeout);
733static void
734erase_csegbufsz_experiment_property(__unused void *param0, __unused void *param1)
735{
736 // Wait for NVRAM to be writable
737 if (!IOServiceWaitForMatchingResource(property: "IONVRAM", UINT64_MAX)) {
738 printf(format: "csegbufsz_experiment_property: Failed to wait for IONVRAM.");
739 }
740
741 if (!PERemoveNVRAMProperty(symbol: csegbufsizeExperimentProperty)) {
742 printf(format: "csegbufsize_experiment_property: Failed to remove %s from NVRAM.", csegbufsizeExperimentProperty);
743 }
744 thread_call_free(call: csegbufsz_experiment_thread_call);
745}
746
747static void
748erase_csegbufsz_experiment_property_async()
749{
750 csegbufsz_experiment_thread_call = thread_call_allocate_with_priority(
751 func: erase_csegbufsz_experiment_property,
752 NULL,
753 pri: THREAD_CALL_PRIORITY_LOW
754 );
755 if (csegbufsz_experiment_thread_call == NULL) {
756 printf(format: "csegbufsize_experiment_property: Unable to allocate thread call.");
757 } else {
758 thread_call_enter(call: csegbufsz_experiment_thread_call);
759 }
760}
761
762static void
763cleanup_csegbufsz_experiment(__unused void *arg0)
764{
765 char nvram = 0;
766 unsigned int len = sizeof(nvram);
767 if (PEReadNVRAMProperty(symbol: csegbufsizeExperimentProperty, value: &nvram, len: &len)) {
768 erase_csegbufsz_experiment_property_async();
769 }
770}
771
772STARTUP_ARG(EARLY_BOOT, STARTUP_RANK_FIRST, cleanup_csegbufsz_experiment, NULL);
773#endif /* XNU_TARGET_OS_OSX && __arm64__ */
774
775#if CONFIG_JETSAM
776extern unsigned int memorystatus_swap_all_apps;
777#endif /* CONFIG_JETSAM */
778
779TUNABLE_DT(uint64_t, swap_vol_min_capacity, "/defaults", "kern.swap_min_capacity", "kern.swap_min_capacity", 0, TUNABLE_DT_NONE);
780
781static void
782vm_compressor_set_size(void)
783{
784 /*
785 * Note that this function may be called multiple times on systems with app swap
786 * because the value of vm_swap_get_max_configured_space() and memorystatus_swap_all_apps
787 * can change based the size of the swap volume. On these systems, we'll call
788 * this function once early in boot to reserve the maximum amount of VA required
789 * for the compressor submap and then one more time in vm_compressor_init after
790 * determining the swap volume size. We must not return a larger value the second
791 * time around.
792 */
793 vm_size_t c_segments_arr_size = 0;
794 struct c_slot_mapping tmp_slot_ptr;
795
796 /* The segment size can be overwritten by a boot-arg */
797 if (!PE_parse_boot_argn(arg_string: "vm_compressor_segment_buffer_size", arg_ptr: &c_seg_bufsize, max_arg: sizeof(c_seg_bufsize))) {
798#if CONFIG_JETSAM
799 if (memorystatus_swap_all_apps) {
800 c_seg_bufsize = C_SEG_BUFSIZE_ARM_SWAP;
801 } else {
802 c_seg_bufsize = C_SEG_BUFSIZE_DEFAULT;
803 }
804#else
805 c_seg_bufsize = C_SEG_BUFSIZE_DEFAULT;
806#endif /* CONFIG_JETSAM */
807 }
808
809 vm_compressor_swap_init_swap_file_limit();
810 if (vm_compression_limit) {
811 compressor_pool_size = ptoa_64(vm_compression_limit);
812 }
813
814 compressor_pool_max_size = C_SEG_MAX_LIMIT;
815 compressor_pool_max_size *= c_seg_bufsize;
816
817#if XNU_TARGET_OS_OSX
818
819 if (vm_compression_limit == 0) {
820 if (max_mem <= (4ULL * 1024ULL * 1024ULL * 1024ULL)) {
821 compressor_pool_size = 16ULL * max_mem;
822 } else if (max_mem <= (8ULL * 1024ULL * 1024ULL * 1024ULL)) {
823 compressor_pool_size = 8ULL * max_mem;
824 } else if (max_mem <= (32ULL * 1024ULL * 1024ULL * 1024ULL)) {
825 compressor_pool_size = 4ULL * max_mem;
826 } else {
827 compressor_pool_size = 2ULL * max_mem;
828 }
829 }
830 /*
831 * Cap the compressor pool size to a max of 192G
832 */
833 if (compressor_pool_size > VM_COMPRESSOR_MAX_POOL_SIZE) {
834 compressor_pool_size = VM_COMPRESSOR_MAX_POOL_SIZE;
835 }
836 if (max_mem <= (8ULL * 1024ULL * 1024ULL * 1024ULL)) {
837 compressor_pool_multiplier = 1;
838 } else if (max_mem <= (32ULL * 1024ULL * 1024ULL * 1024ULL)) {
839 compressor_pool_multiplier = 2;
840 } else {
841 compressor_pool_multiplier = 4;
842 }
843
844#else
845
846 if (compressor_pool_max_size > max_mem) {
847 compressor_pool_max_size = max_mem;
848 }
849
850 if (vm_compression_limit == 0) {
851 compressor_pool_size = max_mem;
852 }
853
854#if XNU_TARGET_OS_WATCH
855 compressor_pool_multiplier = 2;
856#elif XNU_TARGET_OS_IOS
857 if (max_mem <= (2ULL * 1024ULL * 1024ULL * 1024ULL)) {
858 compressor_pool_multiplier = 2;
859 } else {
860 compressor_pool_multiplier = 1;
861 }
862#else
863 compressor_pool_multiplier = 1;
864#endif
865
866#endif
867
868 PE_parse_boot_argn(arg_string: "kern.compressor_pool_multiplier", arg_ptr: &compressor_pool_multiplier, max_arg: sizeof(compressor_pool_multiplier));
869 if (compressor_pool_multiplier < 1) {
870 compressor_pool_multiplier = 1;
871 }
872
873 if (compressor_pool_size > compressor_pool_max_size) {
874 compressor_pool_size = compressor_pool_max_size;
875 }
876
877 c_seg_max_pages = (c_seg_bufsize / PAGE_SIZE);
878 c_seg_slot_var_array_min_len = c_seg_max_pages;
879
880#if !defined(__x86_64__)
881 c_seg_off_limit = (C_SEG_BYTES_TO_OFFSET((c_seg_bufsize - 512)));
882 c_seg_allocsize = (c_seg_bufsize + PAGE_SIZE);
883#else
884 c_seg_off_limit = (C_SEG_BYTES_TO_OFFSET((c_seg_bufsize - 128)));
885 c_seg_allocsize = c_seg_bufsize;
886#endif /* !defined(__x86_64__) */
887
888 c_segments_limit = (uint32_t)(compressor_pool_size / (vm_size_t)(c_seg_allocsize));
889 tmp_slot_ptr.s_cseg = c_segments_limit;
890 /* Panic on internal configs*/
891 assertf((tmp_slot_ptr.s_cseg == c_segments_limit), "vm_compressor_init: overflowed s_cseg field in c_slot_mapping with c_segno: %d", c_segments_limit);
892
893 if (tmp_slot_ptr.s_cseg != c_segments_limit) {
894 tmp_slot_ptr.s_cseg = -1;
895 c_segments_limit = tmp_slot_ptr.s_cseg - 1; /*limited by segment idx bits in c_slot_mapping*/
896 compressor_pool_size = (c_segments_limit * (vm_size_t)(c_seg_allocsize));
897 }
898
899 c_segments_nearing_limit = (uint32_t)(((uint64_t)c_segments_limit * 98ULL) / 100ULL);
900
901 c_segment_pages_compressed_limit = (c_segments_limit * (c_seg_bufsize / PAGE_SIZE) * compressor_pool_multiplier);
902
903 if (c_segment_pages_compressed_limit < (uint32_t)(max_mem / PAGE_SIZE)) {
904#if defined(XNU_TARGET_OS_WATCH)
905 c_segment_pages_compressed_limit = (uint32_t)(max_mem / PAGE_SIZE);
906#else
907 if (!vm_compression_limit) {
908 c_segment_pages_compressed_limit = (uint32_t)(max_mem / PAGE_SIZE);
909 }
910#endif
911 }
912
913 c_segment_pages_compressed_nearing_limit = (uint32_t)(((uint64_t)c_segment_pages_compressed_limit * 98ULL) / 100ULL);
914
915#if CONFIG_FREEZE
916 /*
917 * Our in-core limits are based on the size of the compressor pool.
918 * The c_segments_nearing_limit is also based on the compressor pool
919 * size and calculated above.
920 */
921 c_segments_incore_limit = c_segments_limit;
922
923 if (freezer_incore_cseg_acct) {
924 /*
925 * Add enough segments to track all frozen c_segs that can be stored in swap.
926 */
927 c_segments_limit += (uint32_t)(vm_swap_get_max_configured_space() / (vm_size_t)(c_seg_allocsize));
928 tmp_slot_ptr.s_cseg = c_segments_limit;
929 /* Panic on internal configs*/
930 assertf((tmp_slot_ptr.s_cseg == c_segments_limit), "vm_compressor_init: freezer reserve overflowed s_cseg field in c_slot_mapping with c_segno: %d", c_segments_limit);
931 }
932#endif
933 /*
934 * Submap needs space for:
935 * - c_segments
936 * - c_buffers
937 * - swap reclaimations -- c_seg_bufsize
938 */
939 c_segments_arr_size = vm_map_round_page((sizeof(union c_segu) * c_segments_limit), VM_MAP_PAGE_MASK(kernel_map));
940 c_buffers_size = vm_map_round_page(((vm_size_t)c_seg_allocsize * (vm_size_t)c_segments_limit), VM_MAP_PAGE_MASK(kernel_map));
941
942 compressor_size = c_segments_arr_size + c_buffers_size + c_seg_bufsize;
943
944#if RECORD_THE_COMPRESSED_DATA
945 c_compressed_record_sbuf_size = (vm_size_t)c_seg_allocsize + (PAGE_SIZE * 2);
946 compressor_size += c_compressed_record_sbuf_size;
947#endif /* RECORD_THE_COMPRESSED_DATA */
948}
949STARTUP(KMEM, STARTUP_RANK_FIRST, vm_compressor_set_size);
950
951KMEM_RANGE_REGISTER_DYNAMIC(compressor, &compressor_range, ^() {
952 return compressor_size;
953});
954
955bool
956osenvironment_is_diagnostics(void)
957{
958 DTEntry chosen;
959 const char *osenvironment;
960 unsigned int size;
961 if (kSuccess == SecureDTLookupEntry(searchPoint: 0, pathName: "/chosen", foundEntry: &chosen)) {
962 if (kSuccess == SecureDTGetProperty(entry: chosen, propertyName: "osenvironment", propertyValue: (void const **) &osenvironment, propertySize: &size)) {
963 return strcmp(s1: osenvironment, s2: "diagnostics") == 0;
964 }
965 }
966 return false;
967}
968
969void
970vm_compressor_init(void)
971{
972 thread_t thread;
973#if RECORD_THE_COMPRESSED_DATA
974 vm_size_t c_compressed_record_sbuf_size = 0;
975#endif /* RECORD_THE_COMPRESSED_DATA */
976
977#if DEVELOPMENT || DEBUG || CONFIG_FREEZE
978 char bootarg_name[32];
979#endif /* DEVELOPMENT || DEBUG || CONFIG_FREEZE */
980 __unused uint64_t early_boot_compressor_size = compressor_size;
981
982#if CONFIG_JETSAM
983 if (memorystatus_swap_all_apps && osenvironment_is_diagnostics()) {
984 printf("osenvironment == \"diagnostics\". Disabling app swap.\n");
985 memorystatus_disable_swap();
986 }
987
988 if (memorystatus_swap_all_apps) {
989 /*
990 * App swap is disabled on devices with small NANDs.
991 * Now that we're no longer in early boot, we can get
992 * the NAND size and re-run vm_compressor_set_size.
993 */
994 int error = vm_swap_vol_get_capacity(SWAP_VOLUME_NAME, &vm_swap_volume_capacity);
995#if DEVELOPMENT || DEBUG
996 if (error != 0) {
997 panic("vm_compressor_init: Unable to get swap volume capacity. error=%d\n", error);
998 }
999#else
1000 if (error != 0) {
1001 os_log_with_startup_serial(OS_LOG_DEFAULT, "vm_compressor_init: Unable to get swap volume capacity. error=%d\n", error);
1002 }
1003#endif /* DEVELOPMENT || DEBUG */
1004 if (vm_swap_volume_capacity < swap_vol_min_capacity) {
1005 memorystatus_disable_swap();
1006 }
1007 /*
1008 * Resize the compressor and swap now that we know the capacity
1009 * of the swap volume.
1010 */
1011 vm_compressor_set_size();
1012 /*
1013 * We reserved a chunk of VA early in boot for the compressor submap.
1014 * We can't allocate more than that.
1015 */
1016 assert(compressor_size <= early_boot_compressor_size);
1017 }
1018#endif /* CONFIG_JETSAM */
1019
1020#if DEVELOPMENT || DEBUG
1021 if (PE_parse_boot_argn("-disable_cseg_write_protection", bootarg_name, sizeof(bootarg_name))) {
1022 write_protect_c_segs = FALSE;
1023 }
1024
1025 int vmcval = 1;
1026#if defined(XNU_TARGET_OS_WATCH)
1027 vmcval = 0;
1028#endif /* XNU_TARGET_OS_WATCH */
1029 PE_parse_boot_argn("vm_compressor_validation", &vmcval, sizeof(vmcval));
1030
1031 if (kern_feature_override(KF_COMPRSV_OVRD)) {
1032 vmcval = 0;
1033 }
1034
1035 if (vmcval == 0) {
1036#if POPCOUNT_THE_COMPRESSED_DATA
1037 popcount_c_segs = FALSE;
1038#endif
1039#if CHECKSUM_THE_DATA || CHECKSUM_THE_COMPRESSED_DATA
1040 checksum_c_segs = FALSE;
1041#endif
1042#if VALIDATE_C_SEGMENTS
1043 validate_c_segs = FALSE;
1044#endif
1045 write_protect_c_segs = FALSE;
1046 }
1047#endif /* DEVELOPMENT || DEBUG */
1048
1049#if CONFIG_FREEZE
1050 if (PE_parse_boot_argn("-disable_freezer_cseg_acct", bootarg_name, sizeof(bootarg_name))) {
1051 freezer_incore_cseg_acct = FALSE;
1052 }
1053#endif /* CONFIG_FREEZE */
1054
1055 assert((C_SEGMENTS_PER_PAGE * sizeof(union c_segu)) == PAGE_SIZE);
1056
1057#if !XNU_TARGET_OS_OSX
1058 vm_compressor_minorcompact_threshold_divisor = 20;
1059 vm_compressor_majorcompact_threshold_divisor = 30;
1060 vm_compressor_unthrottle_threshold_divisor = 40;
1061 vm_compressor_catchup_threshold_divisor = 60;
1062#else /* !XNU_TARGET_OS_OSX */
1063 if (max_mem <= (3ULL * 1024ULL * 1024ULL * 1024ULL)) {
1064 vm_compressor_minorcompact_threshold_divisor = 11;
1065 vm_compressor_majorcompact_threshold_divisor = 13;
1066 vm_compressor_unthrottle_threshold_divisor = 20;
1067 vm_compressor_catchup_threshold_divisor = 35;
1068 } else {
1069 vm_compressor_minorcompact_threshold_divisor = 20;
1070 vm_compressor_majorcompact_threshold_divisor = 25;
1071 vm_compressor_unthrottle_threshold_divisor = 35;
1072 vm_compressor_catchup_threshold_divisor = 50;
1073 }
1074#endif /* !XNU_TARGET_OS_OSX */
1075
1076 queue_init(&c_bad_list_head);
1077 queue_init(&c_age_list_head);
1078 queue_init(&c_minor_list_head);
1079 queue_init(&c_major_list_head);
1080 queue_init(&c_filling_list_head);
1081 queue_init(&c_early_swapout_list_head);
1082 queue_init(&c_regular_swapout_list_head);
1083 queue_init(&c_late_swapout_list_head);
1084 queue_init(&c_swapio_list_head);
1085 queue_init(&c_early_swappedin_list_head);
1086 queue_init(&c_regular_swappedin_list_head);
1087 queue_init(&c_late_swappedin_list_head);
1088 queue_init(&c_swappedout_list_head);
1089 queue_init(&c_swappedout_sparse_list_head);
1090
1091 c_free_segno_head = -1;
1092 c_segments_available = 0;
1093
1094 compressor_map = kmem_suballoc(parent: kernel_map, addr: &compressor_range.min_address,
1095 size: compressor_size, vmc_options: VM_MAP_CREATE_NEVER_FAULTS,
1096 VM_FLAGS_FIXED | VM_FLAGS_OVERWRITE, flags: KMS_NOFAIL | KMS_PERMANENT,
1097 VM_KERN_MEMORY_COMPRESSOR).kmr_submap;
1098
1099 kmem_alloc(map: compressor_map, addrp: (vm_offset_t *)(&c_segments),
1100 size: (sizeof(union c_segu) * c_segments_limit),
1101 flags: KMA_NOFAIL | KMA_KOBJECT | KMA_VAONLY | KMA_PERMANENT,
1102 VM_KERN_MEMORY_COMPRESSOR);
1103 kmem_alloc(map: compressor_map, addrp: &c_buffers, size: c_buffers_size,
1104 flags: KMA_NOFAIL | KMA_COMPRESSOR | KMA_VAONLY | KMA_PERMANENT,
1105 VM_KERN_MEMORY_COMPRESSOR);
1106
1107#if DEVELOPMENT || DEBUG
1108 if (hvg_is_hcall_available(HVG_HCALL_SET_COREDUMP_DATA)) {
1109 hvg_hcall_set_coredump_data();
1110 }
1111#endif
1112
1113 /*
1114 * Pick a good size that will minimize fragmentation in zalloc
1115 * by minimizing the fragmentation in a 16k run.
1116 *
1117 * c_seg_slot_var_array_min_len is larger on 4k systems than 16k ones,
1118 * making the fragmentation in a 4k page terrible. Using 16k for all
1119 * systems matches zalloc() and will minimize fragmentation.
1120 */
1121 uint32_t c_segment_size = sizeof(struct c_segment) + (c_seg_slot_var_array_min_len * sizeof(struct c_slot));
1122 uint32_t cnt = (16 << 10) / c_segment_size;
1123 uint32_t frag = (16 << 10) % c_segment_size;
1124
1125 c_seg_fixed_array_len = c_seg_slot_var_array_min_len;
1126
1127 while (cnt * sizeof(struct c_slot) < frag) {
1128 c_segment_size += sizeof(struct c_slot);
1129 c_seg_fixed_array_len++;
1130 frag -= cnt * sizeof(struct c_slot);
1131 }
1132
1133 compressor_segment_zone = zone_create(name: "compressor_segment",
1134 size: c_segment_size, flags: ZC_PGZ_USE_GUARDS | ZC_NOENCRYPT | ZC_ZFREE_CLEARMEM);
1135
1136 c_segments_busy = FALSE;
1137
1138 c_segments_next_page = (caddr_t)c_segments;
1139 vm_compressor_algorithm_init();
1140
1141 {
1142 host_basic_info_data_t hinfo;
1143 mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT;
1144 size_t bufsize;
1145 char *buf;
1146
1147#define BSD_HOST 1
1148 host_info(host: (host_t)BSD_HOST, HOST_BASIC_INFO, host_info_out: (host_info_t)&hinfo, host_info_outCnt: &count);
1149
1150 compressor_cpus = hinfo.max_cpus;
1151
1152 bufsize = PAGE_SIZE;
1153 bufsize += compressor_cpus * vm_compressor_get_decode_scratch_size();
1154 /* For the KDP path */
1155 bufsize += vm_compressor_get_decode_scratch_size();
1156#if CONFIG_FREEZE
1157 bufsize += vm_compressor_get_encode_scratch_size();
1158#endif
1159#if RECORD_THE_COMPRESSED_DATA
1160 bufsize += c_compressed_record_sbuf_size;
1161#endif
1162
1163 kmem_alloc(map: kernel_map, addrp: (vm_offset_t *)&buf, size: bufsize,
1164 flags: KMA_DATA | KMA_NOFAIL | KMA_KOBJECT | KMA_PERMANENT,
1165 VM_KERN_MEMORY_COMPRESSOR);
1166
1167 /*
1168 * kdp_compressor_decompressed_page must be page aligned because we access
1169 * it through the physical aperture by page number.
1170 */
1171 kdp_compressor_decompressed_page = buf;
1172 kdp_compressor_decompressed_page_paddr = kvtophys(va: (vm_offset_t)kdp_compressor_decompressed_page);
1173 kdp_compressor_decompressed_page_ppnum = (ppnum_t) atop(kdp_compressor_decompressed_page_paddr);
1174 buf += PAGE_SIZE;
1175 bufsize -= PAGE_SIZE;
1176
1177 compressor_scratch_bufs = buf;
1178 buf += compressor_cpus * vm_compressor_get_decode_scratch_size();
1179 bufsize -= compressor_cpus * vm_compressor_get_decode_scratch_size();
1180
1181 kdp_compressor_scratch_buf = buf;
1182 buf += vm_compressor_get_decode_scratch_size();
1183 bufsize -= vm_compressor_get_decode_scratch_size();
1184
1185#if CONFIG_FREEZE
1186 freezer_context_global.freezer_ctx_compressor_scratch_buf = buf;
1187 buf += vm_compressor_get_encode_scratch_size();
1188 bufsize -= vm_compressor_get_encode_scratch_size();
1189#endif
1190
1191#if RECORD_THE_COMPRESSED_DATA
1192 c_compressed_record_sbuf = buf;
1193 c_compressed_record_cptr = buf;
1194 c_compressed_record_ebuf = c_compressed_record_sbuf + c_compressed_record_sbuf_size;
1195 buf += c_compressed_record_sbuf_size;
1196 bufsize -= c_compressed_record_sbuf_size;
1197#endif
1198 assert(bufsize == 0);
1199 }
1200
1201 if (kernel_thread_start_priority(continuation: (thread_continue_t)vm_compressor_swap_trigger_thread, NULL,
1202 BASEPRI_VM, new_thread: &thread) != KERN_SUCCESS) {
1203 panic("vm_compressor_swap_trigger_thread: create failed");
1204 }
1205 thread_deallocate(thread);
1206
1207 if (vm_pageout_internal_start() != KERN_SUCCESS) {
1208 panic("vm_compressor_init: Failed to start the internal pageout thread.");
1209 }
1210 if (VM_CONFIG_SWAP_IS_PRESENT) {
1211 vm_compressor_swap_init();
1212 }
1213
1214 if (VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
1215 vm_compressor_is_active = 1;
1216 }
1217
1218 vm_compressor_available = 1;
1219
1220 vm_page_reactivate_all_throttled();
1221
1222 bzero(s: &vmcs_stats, n: sizeof(struct vm_compressor_swapper_stats));
1223}
1224
1225
1226#if VALIDATE_C_SEGMENTS
1227
1228static void
1229c_seg_validate(c_segment_t c_seg, boolean_t must_be_compact)
1230{
1231 uint16_t c_indx;
1232 int32_t bytes_used;
1233 uint32_t c_rounded_size;
1234 uint32_t c_size;
1235 c_slot_t cs;
1236
1237 if (__probable(validate_c_segs == FALSE)) {
1238 return;
1239 }
1240 if (c_seg->c_firstemptyslot < c_seg->c_nextslot) {
1241 c_indx = c_seg->c_firstemptyslot;
1242 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
1243
1244 if (cs == NULL) {
1245 panic("c_seg_validate: no slot backing c_firstemptyslot");
1246 }
1247
1248 if (cs->c_size) {
1249 panic("c_seg_validate: c_firstemptyslot has non-zero size (%d)", cs->c_size);
1250 }
1251 }
1252 bytes_used = 0;
1253
1254 for (c_indx = 0; c_indx < c_seg->c_nextslot; c_indx++) {
1255 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
1256
1257 c_size = UNPACK_C_SIZE(cs);
1258
1259 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
1260
1261 bytes_used += c_rounded_size;
1262
1263#if CHECKSUM_THE_COMPRESSED_DATA
1264 unsigned csvhash;
1265 if (c_size && cs->c_hash_compressed_data != (csvhash = vmc_hash((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size))) {
1266 addr64_t csvphys = kvtophys((vm_offset_t)&c_seg->c_store.c_buffer[cs->c_offset]);
1267 panic("Compressed data doesn't match original %p phys: 0x%llx %d %p %d %d 0x%x 0x%x", c_seg, csvphys, cs->c_offset, cs, c_indx, c_size, cs->c_hash_compressed_data, csvhash);
1268 }
1269#endif
1270#if POPCOUNT_THE_COMPRESSED_DATA
1271 unsigned csvpop;
1272 if (c_size) {
1273 uintptr_t csvaddr = (uintptr_t) &c_seg->c_store.c_buffer[cs->c_offset];
1274 if (cs->c_pop_cdata != (csvpop = vmc_pop(csvaddr, c_size))) {
1275 panic("Compressed data popcount doesn't match original, bit distance: %d %p (phys: %p) %p %p 0x%llx 0x%x 0x%x 0x%x", (csvpop - cs->c_pop_cdata), (void *)csvaddr, (void *) kvtophys(csvaddr), c_seg, cs, (uint64_t)cs->c_offset, c_size, csvpop, cs->c_pop_cdata);
1276 }
1277 }
1278#endif
1279 }
1280
1281 if (bytes_used != c_seg->c_bytes_used) {
1282 panic("c_seg_validate: bytes_used mismatch - found %d, segment has %d", bytes_used, c_seg->c_bytes_used);
1283 }
1284
1285 if (c_seg->c_bytes_used > C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset)) {
1286 panic("c_seg_validate: c_bytes_used > c_nextoffset - c_nextoffset = %d, c_bytes_used = %d",
1287 (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used);
1288 }
1289
1290 if (must_be_compact) {
1291 if (c_seg->c_bytes_used != C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset)) {
1292 panic("c_seg_validate: c_bytes_used doesn't match c_nextoffset - c_nextoffset = %d, c_bytes_used = %d",
1293 (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used);
1294 }
1295 }
1296}
1297
1298#endif
1299
1300
1301void
1302c_seg_need_delayed_compaction(c_segment_t c_seg, boolean_t c_list_lock_held)
1303{
1304 boolean_t clear_busy = FALSE;
1305
1306 if (c_list_lock_held == FALSE) {
1307 if (!lck_mtx_try_lock_spin_always(c_list_lock)) {
1308 C_SEG_BUSY(c_seg);
1309
1310 lck_mtx_unlock_always(&c_seg->c_lock);
1311 lck_mtx_lock_spin_always(c_list_lock);
1312 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
1313
1314 clear_busy = TRUE;
1315 }
1316 }
1317 assert(c_seg->c_state != C_IS_FILLING);
1318
1319 if (!c_seg->c_on_minorcompact_q && !(C_SEG_IS_ON_DISK_OR_SOQ(c_seg)) && !c_seg->c_has_donated_pages) {
1320 queue_enter(&c_minor_list_head, c_seg, c_segment_t, c_list);
1321 c_seg->c_on_minorcompact_q = 1;
1322 c_minor_count++;
1323 }
1324 if (c_list_lock_held == FALSE) {
1325 lck_mtx_unlock_always(c_list_lock);
1326 }
1327
1328 if (clear_busy == TRUE) {
1329 C_SEG_WAKEUP_DONE(c_seg);
1330 }
1331}
1332
1333
1334unsigned int c_seg_moved_to_sparse_list = 0;
1335
1336void
1337c_seg_move_to_sparse_list(c_segment_t c_seg)
1338{
1339 boolean_t clear_busy = FALSE;
1340
1341 if (!lck_mtx_try_lock_spin_always(c_list_lock)) {
1342 C_SEG_BUSY(c_seg);
1343
1344 lck_mtx_unlock_always(&c_seg->c_lock);
1345 lck_mtx_lock_spin_always(c_list_lock);
1346 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
1347
1348 clear_busy = TRUE;
1349 }
1350 c_seg_switch_state(c_seg, C_ON_SWAPPEDOUTSPARSE_Q, FALSE);
1351
1352 c_seg_moved_to_sparse_list++;
1353
1354 lck_mtx_unlock_always(c_list_lock);
1355
1356 if (clear_busy == TRUE) {
1357 C_SEG_WAKEUP_DONE(c_seg);
1358 }
1359}
1360
1361
1362void
1363c_seg_insert_into_q(queue_head_t *qhead, c_segment_t c_seg)
1364{
1365 c_segment_t c_seg_next;
1366
1367 if (queue_empty(qhead)) {
1368 queue_enter(qhead, c_seg, c_segment_t, c_age_list);
1369 } else {
1370 c_seg_next = (c_segment_t)queue_first(qhead);
1371
1372 while (TRUE) {
1373 if (c_seg->c_generation_id < c_seg_next->c_generation_id) {
1374 queue_insert_before(qhead, c_seg, c_seg_next, c_segment_t, c_age_list);
1375 break;
1376 }
1377 c_seg_next = (c_segment_t) queue_next(&c_seg_next->c_age_list);
1378
1379 if (queue_end(qhead, (queue_entry_t) c_seg_next)) {
1380 queue_enter(qhead, c_seg, c_segment_t, c_age_list);
1381 break;
1382 }
1383 }
1384 }
1385}
1386
1387
1388int try_minor_compaction_failed = 0;
1389int try_minor_compaction_succeeded = 0;
1390
1391void
1392c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg)
1393{
1394 assert(c_seg->c_on_minorcompact_q);
1395 /*
1396 * c_seg is currently on the delayed minor compaction
1397 * queue and we have c_seg locked... if we can get the
1398 * c_list_lock w/o blocking (if we blocked we could deadlock
1399 * because the lock order is c_list_lock then c_seg's lock)
1400 * we'll pull it from the delayed list and free it directly
1401 */
1402 if (!lck_mtx_try_lock_spin_always(c_list_lock)) {
1403 /*
1404 * c_list_lock is held, we need to bail
1405 */
1406 try_minor_compaction_failed++;
1407
1408 lck_mtx_unlock_always(&c_seg->c_lock);
1409 } else {
1410 try_minor_compaction_succeeded++;
1411
1412 C_SEG_BUSY(c_seg);
1413 c_seg_do_minor_compaction_and_unlock(c_seg, TRUE, FALSE, FALSE);
1414 }
1415}
1416
1417
1418int
1419c_seg_do_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy, boolean_t need_list_lock, boolean_t disallow_page_replacement)
1420{
1421 int c_seg_freed;
1422
1423 assert(c_seg->c_busy);
1424 assert(!C_SEG_IS_ON_DISK_OR_SOQ(c_seg));
1425
1426 /*
1427 * check for the case that can occur when we are not swapping
1428 * and this segment has been major compacted in the past
1429 * and moved to the majorcompact q to remove it from further
1430 * consideration... if the occupancy falls too low we need
1431 * to put it back on the age_q so that it will be considered
1432 * in the next major compaction sweep... if we don't do this
1433 * we will eventually run into the c_segments_limit
1434 */
1435 if (c_seg->c_state == C_ON_MAJORCOMPACT_Q && C_SEG_SHOULD_MAJORCOMPACT_NOW(c_seg)) {
1436 c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE);
1437 }
1438 if (!c_seg->c_on_minorcompact_q) {
1439 if (clear_busy == TRUE) {
1440 C_SEG_WAKEUP_DONE(c_seg);
1441 }
1442
1443 lck_mtx_unlock_always(&c_seg->c_lock);
1444
1445 return 0;
1446 }
1447 queue_remove(&c_minor_list_head, c_seg, c_segment_t, c_list);
1448 c_seg->c_on_minorcompact_q = 0;
1449 c_minor_count--;
1450
1451 lck_mtx_unlock_always(c_list_lock);
1452
1453 if (disallow_page_replacement == TRUE) {
1454 lck_mtx_unlock_always(&c_seg->c_lock);
1455
1456 PAGE_REPLACEMENT_DISALLOWED(TRUE);
1457
1458 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
1459 }
1460 c_seg_freed = c_seg_minor_compaction_and_unlock(c_seg, clear_busy);
1461
1462 if (disallow_page_replacement == TRUE) {
1463 PAGE_REPLACEMENT_DISALLOWED(FALSE);
1464 }
1465
1466 if (need_list_lock == TRUE) {
1467 lck_mtx_lock_spin_always(c_list_lock);
1468 }
1469
1470 return c_seg_freed;
1471}
1472
1473void
1474kdp_compressor_busy_find_owner(event64_t wait_event, thread_waitinfo_t *waitinfo)
1475{
1476 c_segment_t c_seg = (c_segment_t) wait_event;
1477
1478 waitinfo->owner = thread_tid(thread: c_seg->c_busy_for_thread);
1479 waitinfo->context = VM_KERNEL_UNSLIDE_OR_PERM(c_seg);
1480}
1481
1482#if DEVELOPMENT || DEBUG
1483int
1484do_cseg_wedge_thread(void)
1485{
1486 struct c_segment c_seg;
1487 c_seg.c_busy_for_thread = current_thread();
1488
1489 debug_cseg_wait_event = (event_t) &c_seg;
1490
1491 thread_set_pending_block_hint(current_thread(), kThreadWaitCompressor);
1492 assert_wait((event_t) (&c_seg), THREAD_INTERRUPTIBLE);
1493
1494 thread_block(THREAD_CONTINUE_NULL);
1495
1496 return 0;
1497}
1498
1499int
1500do_cseg_unwedge_thread(void)
1501{
1502 thread_wakeup(debug_cseg_wait_event);
1503 debug_cseg_wait_event = NULL;
1504
1505 return 0;
1506}
1507#endif /* DEVELOPMENT || DEBUG */
1508
1509void
1510c_seg_wait_on_busy(c_segment_t c_seg)
1511{
1512 c_seg->c_wanted = 1;
1513
1514 thread_set_pending_block_hint(thread: current_thread(), block_hint: kThreadWaitCompressor);
1515 assert_wait(event: (event_t) (c_seg), THREAD_UNINT);
1516
1517 lck_mtx_unlock_always(&c_seg->c_lock);
1518 thread_block(THREAD_CONTINUE_NULL);
1519}
1520
1521#if CONFIG_FREEZE
1522/*
1523 * We don't have the task lock held while updating the task's
1524 * c_seg queues. We can do that because of the following restrictions:
1525 *
1526 * - SINGLE FREEZER CONTEXT:
1527 * We 'insert' c_segs into the task list on the task_freeze path.
1528 * There can only be one such freeze in progress and the task
1529 * isn't disappearing because we have the VM map lock held throughout
1530 * and we have a reference on the proc too.
1531 *
1532 * - SINGLE TASK DISOWN CONTEXT:
1533 * We 'disown' c_segs of a task ONLY from the task_terminate context. So
1534 * we don't need the task lock but we need the c_list_lock and the
1535 * compressor master lock (shared). We also hold the individual
1536 * c_seg locks (exclusive).
1537 *
1538 * If we either:
1539 * - can't get the c_seg lock on a try, then we start again because maybe
1540 * the c_seg is part of a compaction and might get freed. So we can't trust
1541 * that linkage and need to restart our queue traversal.
1542 * - OR, we run into a busy c_seg (say being swapped in or free-ing) we
1543 * drop all locks again and wait and restart our queue traversal.
1544 *
1545 * - The new_owner_task below is currently only the kernel or NULL.
1546 *
1547 */
1548void
1549c_seg_update_task_owner(c_segment_t c_seg, task_t new_owner_task)
1550{
1551 task_t owner_task = c_seg->c_task_owner;
1552 uint64_t uncompressed_bytes = ((c_seg->c_slots_used) * PAGE_SIZE_64);
1553
1554 LCK_MTX_ASSERT(c_list_lock, LCK_MTX_ASSERT_OWNED);
1555 LCK_MTX_ASSERT(&c_seg->c_lock, LCK_MTX_ASSERT_OWNED);
1556
1557 if (owner_task) {
1558 task_update_frozen_to_swap_acct(owner_task, uncompressed_bytes, DEBIT_FROM_SWAP);
1559 queue_remove(&owner_task->task_frozen_cseg_q, c_seg,
1560 c_segment_t, c_task_list_next_cseg);
1561 }
1562
1563 if (new_owner_task) {
1564 queue_enter(&new_owner_task->task_frozen_cseg_q, c_seg,
1565 c_segment_t, c_task_list_next_cseg);
1566 task_update_frozen_to_swap_acct(new_owner_task, uncompressed_bytes, CREDIT_TO_SWAP);
1567 }
1568
1569 c_seg->c_task_owner = new_owner_task;
1570}
1571
1572void
1573task_disown_frozen_csegs(task_t owner_task)
1574{
1575 c_segment_t c_seg = NULL, next_cseg = NULL;
1576
1577again:
1578 PAGE_REPLACEMENT_DISALLOWED(TRUE);
1579 lck_mtx_lock_spin_always(c_list_lock);
1580
1581 for (c_seg = (c_segment_t) queue_first(&owner_task->task_frozen_cseg_q);
1582 !queue_end(&owner_task->task_frozen_cseg_q, (queue_entry_t) c_seg);
1583 c_seg = next_cseg) {
1584 next_cseg = (c_segment_t) queue_next(&c_seg->c_task_list_next_cseg);
1585
1586 if (!lck_mtx_try_lock_spin_always(&c_seg->c_lock)) {
1587 lck_mtx_unlock(c_list_lock);
1588 PAGE_REPLACEMENT_DISALLOWED(FALSE);
1589 goto again;
1590 }
1591
1592 if (c_seg->c_busy) {
1593 lck_mtx_unlock(c_list_lock);
1594 PAGE_REPLACEMENT_DISALLOWED(FALSE);
1595
1596 c_seg_wait_on_busy(c_seg);
1597
1598 goto again;
1599 }
1600 assert(c_seg->c_task_owner == owner_task);
1601 c_seg_update_task_owner(c_seg, kernel_task);
1602 lck_mtx_unlock_always(&c_seg->c_lock);
1603 }
1604
1605 lck_mtx_unlock(c_list_lock);
1606 PAGE_REPLACEMENT_DISALLOWED(FALSE);
1607}
1608#endif /* CONFIG_FREEZE */
1609
1610void
1611c_seg_switch_state(c_segment_t c_seg, int new_state, boolean_t insert_head)
1612{
1613 int old_state = c_seg->c_state;
1614 queue_head_t *donate_swapout_list_head, *donate_swappedin_list_head;
1615 uint32_t *donate_swapout_count, *donate_swappedin_count;
1616
1617 /*
1618 * On macOS the donate queue is swapped first ie the c_early_swapout queue.
1619 * On other swap-capable platforms, we want to swap those out last. So we
1620 * use the c_late_swapout queue.
1621 */
1622#if XNU_TARGET_OS_OSX
1623#if (DEVELOPMENT || DEBUG)
1624 if (new_state != C_IS_FILLING) {
1625 LCK_MTX_ASSERT(&c_seg->c_lock, LCK_MTX_ASSERT_OWNED);
1626 }
1627 LCK_MTX_ASSERT(c_list_lock, LCK_MTX_ASSERT_OWNED);
1628#endif /* DEVELOPMENT || DEBUG */
1629
1630 donate_swapout_list_head = &c_early_swapout_list_head;
1631 donate_swapout_count = &c_early_swapout_count;
1632 donate_swappedin_list_head = &c_early_swappedin_list_head;
1633 donate_swappedin_count = &c_early_swappedin_count;
1634#else /* XNU_TARGET_OS_OSX */
1635 donate_swapout_list_head = &c_late_swapout_list_head;
1636 donate_swapout_count = &c_late_swapout_count;
1637 donate_swappedin_list_head = &c_late_swappedin_list_head;
1638 donate_swappedin_count = &c_late_swappedin_count;
1639#endif /* XNU_TARGET_OS_OSX */
1640
1641 switch (old_state) {
1642 case C_IS_EMPTY:
1643 assert(new_state == C_IS_FILLING || new_state == C_IS_FREE);
1644
1645 c_empty_count--;
1646 break;
1647
1648 case C_IS_FILLING:
1649 assert(new_state == C_ON_AGE_Q || new_state == C_ON_SWAPOUT_Q);
1650
1651 queue_remove(&c_filling_list_head, c_seg, c_segment_t, c_age_list);
1652 c_filling_count--;
1653 break;
1654
1655 case C_ON_AGE_Q:
1656 assert(new_state == C_ON_SWAPOUT_Q || new_state == C_ON_MAJORCOMPACT_Q ||
1657 new_state == C_IS_FREE);
1658
1659 queue_remove(&c_age_list_head, c_seg, c_segment_t, c_age_list);
1660 c_age_count--;
1661 break;
1662
1663 case C_ON_SWAPPEDIN_Q:
1664 if (c_seg->c_has_donated_pages) {
1665 assert(new_state == C_ON_SWAPOUT_Q || new_state == C_IS_FREE);
1666 queue_remove(donate_swappedin_list_head, c_seg, c_segment_t, c_age_list);
1667 *donate_swappedin_count -= 1;
1668 } else {
1669 assert(new_state == C_ON_AGE_Q || new_state == C_IS_FREE);
1670#if CONFIG_FREEZE
1671 assert(c_seg->c_has_freezer_pages);
1672 queue_remove(&c_early_swappedin_list_head, c_seg, c_segment_t, c_age_list);
1673 c_early_swappedin_count--;
1674#else /* CONFIG_FREEZE */
1675 queue_remove(&c_regular_swappedin_list_head, c_seg, c_segment_t, c_age_list);
1676 c_regular_swappedin_count--;
1677#endif /* CONFIG_FREEZE */
1678 }
1679 break;
1680
1681 case C_ON_SWAPOUT_Q:
1682 assert(new_state == C_ON_AGE_Q || new_state == C_IS_FREE || new_state == C_IS_EMPTY || new_state == C_ON_SWAPIO_Q);
1683
1684#if CONFIG_FREEZE
1685 if (c_seg->c_has_freezer_pages) {
1686 if (c_seg->c_task_owner && (new_state != C_ON_SWAPIO_Q)) {
1687 c_seg_update_task_owner(c_seg, NULL);
1688 }
1689 queue_remove(&c_early_swapout_list_head, c_seg, c_segment_t, c_age_list);
1690 c_early_swapout_count--;
1691 } else
1692#endif /* CONFIG_FREEZE */
1693 {
1694 if (c_seg->c_has_donated_pages) {
1695 queue_remove(donate_swapout_list_head, c_seg, c_segment_t, c_age_list);
1696 *donate_swapout_count -= 1;
1697 } else {
1698 queue_remove(&c_regular_swapout_list_head, c_seg, c_segment_t, c_age_list);
1699 c_regular_swapout_count--;
1700 }
1701 }
1702
1703 if (new_state == C_ON_AGE_Q) {
1704 c_seg->c_has_donated_pages = 0;
1705 }
1706 thread_wakeup((event_t)&compaction_swapper_running);
1707 break;
1708
1709 case C_ON_SWAPIO_Q:
1710#if CONFIG_FREEZE
1711 if (c_seg->c_has_freezer_pages) {
1712 assert(new_state == C_ON_SWAPPEDOUT_Q || new_state == C_ON_SWAPPEDOUTSPARSE_Q || new_state == C_ON_AGE_Q);
1713 } else
1714#endif /* CONFIG_FREEZE */
1715 {
1716 if (c_seg->c_has_donated_pages) {
1717 assert(new_state == C_ON_SWAPPEDOUT_Q || new_state == C_ON_SWAPPEDOUTSPARSE_Q || new_state == C_ON_SWAPPEDIN_Q);
1718 } else {
1719 assert(new_state == C_ON_SWAPPEDOUT_Q || new_state == C_ON_SWAPPEDOUTSPARSE_Q || new_state == C_ON_AGE_Q);
1720 }
1721 }
1722
1723 queue_remove(&c_swapio_list_head, c_seg, c_segment_t, c_age_list);
1724 c_swapio_count--;
1725 break;
1726
1727 case C_ON_SWAPPEDOUT_Q:
1728 assert(new_state == C_ON_SWAPPEDIN_Q || new_state == C_ON_AGE_Q ||
1729 new_state == C_ON_SWAPPEDOUTSPARSE_Q ||
1730 new_state == C_ON_BAD_Q || new_state == C_IS_EMPTY || new_state == C_IS_FREE);
1731
1732 queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
1733 c_swappedout_count--;
1734 break;
1735
1736 case C_ON_SWAPPEDOUTSPARSE_Q:
1737 assert(new_state == C_ON_SWAPPEDIN_Q || new_state == C_ON_AGE_Q ||
1738 new_state == C_ON_BAD_Q || new_state == C_IS_EMPTY || new_state == C_IS_FREE);
1739
1740 queue_remove(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list);
1741 c_swappedout_sparse_count--;
1742 break;
1743
1744 case C_ON_MAJORCOMPACT_Q:
1745 assert(new_state == C_ON_AGE_Q || new_state == C_IS_FREE);
1746
1747 queue_remove(&c_major_list_head, c_seg, c_segment_t, c_age_list);
1748 c_major_count--;
1749 break;
1750
1751 case C_ON_BAD_Q:
1752 assert(new_state == C_IS_FREE);
1753
1754 queue_remove(&c_bad_list_head, c_seg, c_segment_t, c_age_list);
1755 c_bad_count--;
1756 break;
1757
1758 default:
1759 panic("c_seg %p has bad c_state = %d", c_seg, old_state);
1760 }
1761
1762 switch (new_state) {
1763 case C_IS_FREE:
1764 assert(old_state != C_IS_FILLING);
1765
1766 break;
1767
1768 case C_IS_EMPTY:
1769 assert(old_state == C_ON_SWAPOUT_Q || old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q);
1770
1771 c_empty_count++;
1772 break;
1773
1774 case C_IS_FILLING:
1775 assert(old_state == C_IS_EMPTY);
1776
1777 queue_enter(&c_filling_list_head, c_seg, c_segment_t, c_age_list);
1778 c_filling_count++;
1779 break;
1780
1781 case C_ON_AGE_Q:
1782 assert(old_state == C_IS_FILLING || old_state == C_ON_SWAPPEDIN_Q ||
1783 old_state == C_ON_SWAPOUT_Q || old_state == C_ON_SWAPIO_Q ||
1784 old_state == C_ON_MAJORCOMPACT_Q || old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q);
1785
1786 assert(!c_seg->c_has_donated_pages);
1787 if (old_state == C_IS_FILLING) {
1788 queue_enter(&c_age_list_head, c_seg, c_segment_t, c_age_list);
1789 } else {
1790 if (!queue_empty(&c_age_list_head)) {
1791 c_segment_t c_first;
1792
1793 c_first = (c_segment_t)queue_first(&c_age_list_head);
1794 c_seg->c_creation_ts = c_first->c_creation_ts;
1795 }
1796 queue_enter_first(&c_age_list_head, c_seg, c_segment_t, c_age_list);
1797 }
1798 c_age_count++;
1799 break;
1800
1801 case C_ON_SWAPPEDIN_Q:
1802 {
1803 queue_head_t *list_head;
1804
1805 assert(old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q || old_state == C_ON_SWAPIO_Q);
1806 if (c_seg->c_has_donated_pages) {
1807 /* Error in swapouts could happen while the c_seg is still on the swapio queue */
1808 list_head = donate_swappedin_list_head;
1809 *donate_swappedin_count += 1;
1810 } else {
1811#if CONFIG_FREEZE
1812 assert(c_seg->c_has_freezer_pages);
1813 list_head = &c_early_swappedin_list_head;
1814 c_early_swappedin_count++;
1815#else /* CONFIG_FREEZE */
1816 list_head = &c_regular_swappedin_list_head;
1817 c_regular_swappedin_count++;
1818#endif /* CONFIG_FREEZE */
1819 }
1820
1821 if (insert_head == TRUE) {
1822 queue_enter_first(list_head, c_seg, c_segment_t, c_age_list);
1823 } else {
1824 queue_enter(list_head, c_seg, c_segment_t, c_age_list);
1825 }
1826 break;
1827 }
1828
1829 case C_ON_SWAPOUT_Q:
1830 {
1831 queue_head_t *list_head;
1832
1833#if CONFIG_FREEZE
1834 /*
1835 * A segment with both identities of frozen + donated pages
1836 * will be put on early swapout Q ie the frozen identity wins.
1837 * This is because when both identities are set, the donation bit
1838 * is added on after in the c_current_seg_filled path for accounting
1839 * purposes.
1840 */
1841 if (c_seg->c_has_freezer_pages) {
1842 assert(old_state == C_ON_AGE_Q || old_state == C_IS_FILLING);
1843 list_head = &c_early_swapout_list_head;
1844 c_early_swapout_count++;
1845 } else
1846#endif
1847 {
1848 if (c_seg->c_has_donated_pages) {
1849 assert(old_state == C_ON_SWAPPEDIN_Q || old_state == C_IS_FILLING);
1850 list_head = donate_swapout_list_head;
1851 *donate_swapout_count += 1;
1852 } else {
1853 assert(old_state == C_ON_AGE_Q || old_state == C_IS_FILLING);
1854 list_head = &c_regular_swapout_list_head;
1855 c_regular_swapout_count++;
1856 }
1857 }
1858
1859 if (insert_head == TRUE) {
1860 queue_enter_first(list_head, c_seg, c_segment_t, c_age_list);
1861 } else {
1862 queue_enter(list_head, c_seg, c_segment_t, c_age_list);
1863 }
1864 break;
1865 }
1866
1867 case C_ON_SWAPIO_Q:
1868 assert(old_state == C_ON_SWAPOUT_Q);
1869
1870 if (insert_head == TRUE) {
1871 queue_enter_first(&c_swapio_list_head, c_seg, c_segment_t, c_age_list);
1872 } else {
1873 queue_enter(&c_swapio_list_head, c_seg, c_segment_t, c_age_list);
1874 }
1875 c_swapio_count++;
1876 break;
1877
1878 case C_ON_SWAPPEDOUT_Q:
1879 assert(old_state == C_ON_SWAPIO_Q);
1880
1881 if (insert_head == TRUE) {
1882 queue_enter_first(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
1883 } else {
1884 queue_enter(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
1885 }
1886 c_swappedout_count++;
1887 break;
1888
1889 case C_ON_SWAPPEDOUTSPARSE_Q:
1890 assert(old_state == C_ON_SWAPIO_Q || old_state == C_ON_SWAPPEDOUT_Q);
1891
1892 if (insert_head == TRUE) {
1893 queue_enter_first(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list);
1894 } else {
1895 queue_enter(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list);
1896 }
1897
1898 c_swappedout_sparse_count++;
1899 break;
1900
1901 case C_ON_MAJORCOMPACT_Q:
1902 assert(old_state == C_ON_AGE_Q);
1903 assert(!c_seg->c_has_donated_pages);
1904
1905 if (insert_head == TRUE) {
1906 queue_enter_first(&c_major_list_head, c_seg, c_segment_t, c_age_list);
1907 } else {
1908 queue_enter(&c_major_list_head, c_seg, c_segment_t, c_age_list);
1909 }
1910 c_major_count++;
1911 break;
1912
1913 case C_ON_BAD_Q:
1914 assert(old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q);
1915
1916 if (insert_head == TRUE) {
1917 queue_enter_first(&c_bad_list_head, c_seg, c_segment_t, c_age_list);
1918 } else {
1919 queue_enter(&c_bad_list_head, c_seg, c_segment_t, c_age_list);
1920 }
1921 c_bad_count++;
1922 break;
1923
1924 default:
1925 panic("c_seg %p requesting bad c_state = %d", c_seg, new_state);
1926 }
1927 c_seg->c_state = new_state;
1928}
1929
1930
1931
1932void
1933c_seg_free(c_segment_t c_seg)
1934{
1935 assert(c_seg->c_busy);
1936
1937 lck_mtx_unlock_always(&c_seg->c_lock);
1938 lck_mtx_lock_spin_always(c_list_lock);
1939 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
1940
1941 c_seg_free_locked(c_seg);
1942}
1943
1944
1945void
1946c_seg_free_locked(c_segment_t c_seg)
1947{
1948 int segno;
1949 int pages_populated = 0;
1950 int32_t *c_buffer = NULL;
1951 uint64_t c_swap_handle = 0;
1952
1953 assert(c_seg->c_busy);
1954 assert(c_seg->c_slots_used == 0);
1955 assert(!c_seg->c_on_minorcompact_q);
1956 assert(!c_seg->c_busy_swapping);
1957
1958 if (c_seg->c_overage_swap == TRUE) {
1959 c_overage_swapped_count--;
1960 c_seg->c_overage_swap = FALSE;
1961 }
1962 if (!(C_SEG_IS_ONDISK(c_seg))) {
1963 c_buffer = c_seg->c_store.c_buffer;
1964 } else {
1965 c_swap_handle = c_seg->c_store.c_swap_handle;
1966 }
1967
1968 c_seg_switch_state(c_seg, C_IS_FREE, FALSE);
1969
1970 if (c_buffer) {
1971 pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE;
1972 c_seg->c_store.c_buffer = NULL;
1973 } else {
1974#if CONFIG_FREEZE
1975 c_seg_update_task_owner(c_seg, NULL);
1976#endif /* CONFIG_FREEZE */
1977
1978 c_seg->c_store.c_swap_handle = (uint64_t)-1;
1979 }
1980
1981 lck_mtx_unlock_always(&c_seg->c_lock);
1982
1983 lck_mtx_unlock_always(c_list_lock);
1984
1985 if (c_buffer) {
1986 if (pages_populated) {
1987 kernel_memory_depopulate(addr: (vm_offset_t)c_buffer,
1988 ptoa(pages_populated), flags: KMA_COMPRESSOR,
1989 VM_KERN_MEMORY_COMPRESSOR);
1990 }
1991 } else if (c_swap_handle) {
1992 /*
1993 * Free swap space on disk.
1994 */
1995 vm_swap_free(c_swap_handle);
1996 }
1997 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
1998 /*
1999 * c_seg must remain busy until
2000 * after the call to vm_swap_free
2001 */
2002 C_SEG_WAKEUP_DONE(c_seg);
2003 lck_mtx_unlock_always(&c_seg->c_lock);
2004
2005 segno = c_seg->c_mysegno;
2006
2007 lck_mtx_lock_spin_always(c_list_lock);
2008 /*
2009 * because the c_buffer is now associated with the segno,
2010 * we can't put the segno back on the free list until
2011 * after we have depopulated the c_buffer range, or
2012 * we run the risk of depopulating a range that is
2013 * now being used in one of the compressor heads
2014 */
2015 c_segments[segno].c_segno = c_free_segno_head;
2016 c_free_segno_head = segno;
2017 c_segment_count--;
2018
2019 lck_mtx_unlock_always(c_list_lock);
2020
2021 lck_mtx_destroy(lck: &c_seg->c_lock, grp: &vm_compressor_lck_grp);
2022
2023 if (c_seg->c_slot_var_array_len) {
2024 kfree_type(struct c_slot, c_seg->c_slot_var_array_len,
2025 c_seg->c_slot_var_array);
2026 }
2027
2028 zfree(compressor_segment_zone, c_seg);
2029}
2030
2031#if DEVELOPMENT || DEBUG
2032int c_seg_trim_page_count = 0;
2033#endif
2034
2035void
2036c_seg_trim_tail(c_segment_t c_seg)
2037{
2038 c_slot_t cs;
2039 uint32_t c_size;
2040 uint32_t c_offset;
2041 uint32_t c_rounded_size;
2042 uint16_t current_nextslot;
2043 uint32_t current_populated_offset;
2044
2045 if (c_seg->c_bytes_used == 0) {
2046 return;
2047 }
2048 current_nextslot = c_seg->c_nextslot;
2049 current_populated_offset = c_seg->c_populated_offset;
2050
2051 while (c_seg->c_nextslot) {
2052 cs = C_SEG_SLOT_FROM_INDEX(c_seg, (c_seg->c_nextslot - 1));
2053
2054 c_size = UNPACK_C_SIZE(cs);
2055
2056 if (c_size) {
2057 if (current_nextslot != c_seg->c_nextslot) {
2058 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
2059 c_offset = cs->c_offset + C_SEG_BYTES_TO_OFFSET(c_rounded_size);
2060
2061 c_seg->c_nextoffset = c_offset;
2062 c_seg->c_populated_offset = (c_offset + (C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1)) &
2063 ~(C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1);
2064
2065 if (c_seg->c_firstemptyslot > c_seg->c_nextslot) {
2066 c_seg->c_firstemptyslot = c_seg->c_nextslot;
2067 }
2068#if DEVELOPMENT || DEBUG
2069 c_seg_trim_page_count += ((round_page_32(C_SEG_OFFSET_TO_BYTES(current_populated_offset)) -
2070 round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) /
2071 PAGE_SIZE);
2072#endif
2073 }
2074 break;
2075 }
2076 c_seg->c_nextslot--;
2077 }
2078 assert(c_seg->c_nextslot);
2079}
2080
2081
2082int
2083c_seg_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy)
2084{
2085 c_slot_mapping_t slot_ptr;
2086 uint32_t c_offset = 0;
2087 uint32_t old_populated_offset;
2088 uint32_t c_rounded_size;
2089 uint32_t c_size;
2090 uint16_t c_indx = 0;
2091 int i;
2092 c_slot_t c_dst;
2093 c_slot_t c_src;
2094
2095 assert(c_seg->c_busy);
2096
2097#if VALIDATE_C_SEGMENTS
2098 c_seg_validate(c_seg, FALSE);
2099#endif
2100 if (c_seg->c_bytes_used == 0) {
2101 c_seg_free(c_seg);
2102 return 1;
2103 }
2104 lck_mtx_unlock_always(&c_seg->c_lock);
2105
2106 if (c_seg->c_firstemptyslot >= c_seg->c_nextslot || C_SEG_UNUSED_BYTES(c_seg) < PAGE_SIZE) {
2107 goto done;
2108 }
2109
2110/* TODO: assert first emptyslot's c_size is actually 0 */
2111
2112#if DEVELOPMENT || DEBUG
2113 C_SEG_MAKE_WRITEABLE(c_seg);
2114#endif
2115
2116#if VALIDATE_C_SEGMENTS
2117 c_seg->c_was_minor_compacted++;
2118#endif
2119 c_indx = c_seg->c_firstemptyslot;
2120 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
2121
2122 old_populated_offset = c_seg->c_populated_offset;
2123 c_offset = c_dst->c_offset;
2124
2125 for (i = c_indx + 1; i < c_seg->c_nextslot && c_offset < c_seg->c_nextoffset; i++) {
2126 c_src = C_SEG_SLOT_FROM_INDEX(c_seg, i);
2127
2128 c_size = UNPACK_C_SIZE(c_src);
2129
2130 if (c_size == 0) {
2131 continue;
2132 }
2133
2134 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
2135/* N.B.: This memcpy may be an overlapping copy */
2136 memcpy(dst: &c_seg->c_store.c_buffer[c_offset], src: &c_seg->c_store.c_buffer[c_src->c_offset], n: c_rounded_size);
2137
2138 cslot_copy(cdst: c_dst, csrc: c_src);
2139 c_dst->c_offset = c_offset;
2140
2141 slot_ptr = C_SLOT_UNPACK_PTR(c_dst);
2142 slot_ptr->s_cindx = c_indx;
2143
2144 c_offset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
2145 PACK_C_SIZE(c_src, 0);
2146 c_indx++;
2147
2148 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
2149 }
2150 c_seg->c_firstemptyslot = c_indx;
2151 c_seg->c_nextslot = c_indx;
2152 c_seg->c_nextoffset = c_offset;
2153 c_seg->c_populated_offset = (c_offset + (C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1)) & ~(C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1);
2154 c_seg->c_bytes_unused = 0;
2155
2156#if VALIDATE_C_SEGMENTS
2157 c_seg_validate(c_seg, TRUE);
2158#endif
2159 if (old_populated_offset > c_seg->c_populated_offset) {
2160 uint32_t gc_size;
2161 int32_t *gc_ptr;
2162
2163 gc_size = C_SEG_OFFSET_TO_BYTES(old_populated_offset - c_seg->c_populated_offset);
2164 gc_ptr = &c_seg->c_store.c_buffer[c_seg->c_populated_offset];
2165
2166 kernel_memory_depopulate(addr: (vm_offset_t)gc_ptr, size: gc_size,
2167 flags: KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR);
2168 }
2169
2170#if DEVELOPMENT || DEBUG
2171 C_SEG_WRITE_PROTECT(c_seg);
2172#endif
2173
2174done:
2175 if (clear_busy == TRUE) {
2176 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
2177 C_SEG_WAKEUP_DONE(c_seg);
2178 lck_mtx_unlock_always(&c_seg->c_lock);
2179 }
2180 return 0;
2181}
2182
2183
2184static void
2185c_seg_alloc_nextslot(c_segment_t c_seg)
2186{
2187 struct c_slot *old_slot_array = NULL;
2188 struct c_slot *new_slot_array = NULL;
2189 int newlen;
2190 int oldlen;
2191
2192 if (c_seg->c_nextslot < c_seg_fixed_array_len) {
2193 return;
2194 }
2195
2196 if ((c_seg->c_nextslot - c_seg_fixed_array_len) >= c_seg->c_slot_var_array_len) {
2197 oldlen = c_seg->c_slot_var_array_len;
2198 old_slot_array = c_seg->c_slot_var_array;
2199
2200 if (oldlen == 0) {
2201 newlen = c_seg_slot_var_array_min_len;
2202 } else {
2203 newlen = oldlen * 2;
2204 }
2205
2206 new_slot_array = kalloc_type(struct c_slot, newlen, Z_WAITOK);
2207
2208 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
2209
2210 if (old_slot_array) {
2211 memcpy(dst: new_slot_array, src: old_slot_array,
2212 n: sizeof(struct c_slot) * oldlen);
2213 }
2214
2215 c_seg->c_slot_var_array_len = newlen;
2216 c_seg->c_slot_var_array = new_slot_array;
2217
2218 lck_mtx_unlock_always(&c_seg->c_lock);
2219
2220 kfree_type(struct c_slot, oldlen, old_slot_array);
2221 }
2222}
2223
2224
2225#define C_SEG_MAJOR_COMPACT_STATS_MAX (30)
2226
2227struct {
2228 uint64_t asked_permission;
2229 uint64_t compactions;
2230 uint64_t moved_slots;
2231 uint64_t moved_bytes;
2232 uint64_t wasted_space_in_swapouts;
2233 uint64_t count_of_swapouts;
2234 uint64_t count_of_freed_segs;
2235 uint64_t bailed_compactions;
2236 uint64_t bytes_freed_rate_us;
2237} c_seg_major_compact_stats[C_SEG_MAJOR_COMPACT_STATS_MAX];
2238
2239int c_seg_major_compact_stats_now = 0;
2240
2241
2242#define C_MAJOR_COMPACTION_SIZE_APPROPRIATE ((c_seg_bufsize * 90) / 100)
2243
2244
2245boolean_t
2246c_seg_major_compact_ok(
2247 c_segment_t c_seg_dst,
2248 c_segment_t c_seg_src)
2249{
2250 c_seg_major_compact_stats[c_seg_major_compact_stats_now].asked_permission++;
2251
2252 if (c_seg_src->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE &&
2253 c_seg_dst->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE) {
2254 return FALSE;
2255 }
2256
2257 if (c_seg_dst->c_nextoffset >= c_seg_off_limit || c_seg_dst->c_nextslot >= C_SLOT_MAX_INDEX) {
2258 /*
2259 * destination segment is full... can't compact
2260 */
2261 return FALSE;
2262 }
2263
2264 return TRUE;
2265}
2266
2267
2268boolean_t
2269c_seg_major_compact(
2270 c_segment_t c_seg_dst,
2271 c_segment_t c_seg_src)
2272{
2273 c_slot_mapping_t slot_ptr;
2274 uint32_t c_rounded_size;
2275 uint32_t c_size;
2276 uint16_t dst_slot;
2277 int i;
2278 c_slot_t c_dst;
2279 c_slot_t c_src;
2280 boolean_t keep_compacting = TRUE;
2281
2282 /*
2283 * segments are not locked but they are both marked c_busy
2284 * which keeps c_decompress from working on them...
2285 * we can safely allocate new pages, move compressed data
2286 * from c_seg_src to c_seg_dst and update both c_segment's
2287 * state w/o holding the master lock
2288 */
2289#if DEVELOPMENT || DEBUG
2290 C_SEG_MAKE_WRITEABLE(c_seg_dst);
2291#endif
2292
2293#if VALIDATE_C_SEGMENTS
2294 c_seg_dst->c_was_major_compacted++;
2295 c_seg_src->c_was_major_donor++;
2296#endif
2297 assertf(c_seg_dst->c_has_donated_pages == c_seg_src->c_has_donated_pages, "Mismatched donation status Dst: %p, Src: %p\n", c_seg_dst, c_seg_src);
2298 c_seg_major_compact_stats[c_seg_major_compact_stats_now].compactions++;
2299
2300 dst_slot = c_seg_dst->c_nextslot;
2301
2302 for (i = 0; i < c_seg_src->c_nextslot; i++) {
2303 c_src = C_SEG_SLOT_FROM_INDEX(c_seg_src, i);
2304
2305 c_size = UNPACK_C_SIZE(c_src);
2306
2307 if (c_size == 0) {
2308 /* BATCH: move what we have so far; */
2309 continue;
2310 }
2311
2312 if (C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset - c_seg_dst->c_nextoffset) < (unsigned) c_size) {
2313 int size_to_populate;
2314
2315 /* doesn't fit */
2316 size_to_populate = c_seg_bufsize - C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset);
2317
2318 if (size_to_populate == 0) {
2319 /* can't fit */
2320 keep_compacting = FALSE;
2321 break;
2322 }
2323 if (size_to_populate > C_SEG_MAX_POPULATE_SIZE) {
2324 size_to_populate = C_SEG_MAX_POPULATE_SIZE;
2325 }
2326
2327 kernel_memory_populate(
2328 addr: (vm_offset_t) &c_seg_dst->c_store.c_buffer[c_seg_dst->c_populated_offset],
2329 size: size_to_populate,
2330 flags: KMA_NOFAIL | KMA_COMPRESSOR,
2331 VM_KERN_MEMORY_COMPRESSOR);
2332
2333 c_seg_dst->c_populated_offset += C_SEG_BYTES_TO_OFFSET(size_to_populate);
2334 assert(C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset) <= c_seg_bufsize);
2335 }
2336 c_seg_alloc_nextslot(c_seg: c_seg_dst);
2337
2338 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot);
2339
2340 memcpy(dst: &c_seg_dst->c_store.c_buffer[c_seg_dst->c_nextoffset], src: &c_seg_src->c_store.c_buffer[c_src->c_offset], n: c_size);
2341
2342 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
2343
2344 c_seg_major_compact_stats[c_seg_major_compact_stats_now].moved_slots++;
2345 c_seg_major_compact_stats[c_seg_major_compact_stats_now].moved_bytes += c_size;
2346
2347 cslot_copy(cdst: c_dst, csrc: c_src);
2348 c_dst->c_offset = c_seg_dst->c_nextoffset;
2349
2350 if (c_seg_dst->c_firstemptyslot == c_seg_dst->c_nextslot) {
2351 c_seg_dst->c_firstemptyslot++;
2352 }
2353 c_seg_dst->c_slots_used++;
2354 c_seg_dst->c_nextslot++;
2355 c_seg_dst->c_bytes_used += c_rounded_size;
2356 c_seg_dst->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
2357
2358 PACK_C_SIZE(c_src, 0);
2359
2360 c_seg_src->c_bytes_used -= c_rounded_size;
2361 c_seg_src->c_bytes_unused += c_rounded_size;
2362 c_seg_src->c_firstemptyslot = 0;
2363
2364 assert(c_seg_src->c_slots_used);
2365 c_seg_src->c_slots_used--;
2366
2367 if (!c_seg_src->c_swappedin) {
2368 /* Pessimistically lose swappedin status when non-swappedin pages are added. */
2369 c_seg_dst->c_swappedin = false;
2370 }
2371
2372 if (c_seg_dst->c_nextoffset >= c_seg_off_limit || c_seg_dst->c_nextslot >= C_SLOT_MAX_INDEX) {
2373 /* dest segment is now full */
2374 keep_compacting = FALSE;
2375 break;
2376 }
2377 }
2378#if DEVELOPMENT || DEBUG
2379 C_SEG_WRITE_PROTECT(c_seg_dst);
2380#endif
2381 if (dst_slot < c_seg_dst->c_nextslot) {
2382 PAGE_REPLACEMENT_ALLOWED(TRUE);
2383 /*
2384 * we've now locked out c_decompress from
2385 * converting the slot passed into it into
2386 * a c_segment_t which allows us to use
2387 * the backptr to change which c_segment and
2388 * index the slot points to
2389 */
2390 while (dst_slot < c_seg_dst->c_nextslot) {
2391 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, dst_slot);
2392
2393 slot_ptr = C_SLOT_UNPACK_PTR(c_dst);
2394 /* <csegno=0,indx=0> would mean "empty slot", so use csegno+1 */
2395 slot_ptr->s_cseg = c_seg_dst->c_mysegno + 1;
2396 slot_ptr->s_cindx = dst_slot++;
2397 }
2398 PAGE_REPLACEMENT_ALLOWED(FALSE);
2399 }
2400 return keep_compacting;
2401}
2402
2403
2404uint64_t
2405vm_compressor_compute_elapsed_msecs(clock_sec_t end_sec, clock_nsec_t end_nsec, clock_sec_t start_sec, clock_nsec_t start_nsec)
2406{
2407 uint64_t end_msecs;
2408 uint64_t start_msecs;
2409
2410 end_msecs = (end_sec * 1000) + end_nsec / 1000000;
2411 start_msecs = (start_sec * 1000) + start_nsec / 1000000;
2412
2413 return end_msecs - start_msecs;
2414}
2415
2416
2417
2418uint32_t compressor_eval_period_in_msecs = 250;
2419uint32_t compressor_sample_min_in_msecs = 500;
2420uint32_t compressor_sample_max_in_msecs = 10000;
2421uint32_t compressor_thrashing_threshold_per_10msecs = 50;
2422uint32_t compressor_thrashing_min_per_10msecs = 20;
2423
2424/* When true, reset sample data next chance we get. */
2425static boolean_t compressor_need_sample_reset = FALSE;
2426
2427
2428void
2429compute_swapout_target_age(void)
2430{
2431 clock_sec_t cur_ts_sec;
2432 clock_nsec_t cur_ts_nsec;
2433 uint32_t min_operations_needed_in_this_sample;
2434 uint64_t elapsed_msecs_in_eval;
2435 uint64_t elapsed_msecs_in_sample;
2436 boolean_t need_eval_reset = FALSE;
2437
2438 clock_get_system_nanotime(secs: &cur_ts_sec, nanosecs: &cur_ts_nsec);
2439
2440 elapsed_msecs_in_sample = vm_compressor_compute_elapsed_msecs(end_sec: cur_ts_sec, end_nsec: cur_ts_nsec, start_sec: start_of_sample_period_sec, start_nsec: start_of_sample_period_nsec);
2441
2442 if (compressor_need_sample_reset ||
2443 elapsed_msecs_in_sample >= compressor_sample_max_in_msecs) {
2444 compressor_need_sample_reset = TRUE;
2445 need_eval_reset = TRUE;
2446 goto done;
2447 }
2448 elapsed_msecs_in_eval = vm_compressor_compute_elapsed_msecs(end_sec: cur_ts_sec, end_nsec: cur_ts_nsec, start_sec: start_of_eval_period_sec, start_nsec: start_of_eval_period_nsec);
2449
2450 if (elapsed_msecs_in_eval < compressor_eval_period_in_msecs) {
2451 goto done;
2452 }
2453 need_eval_reset = TRUE;
2454
2455 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_START, elapsed_msecs_in_eval, sample_period_compression_count, sample_period_decompression_count, 0, 0);
2456
2457 min_operations_needed_in_this_sample = (compressor_thrashing_min_per_10msecs * (uint32_t)elapsed_msecs_in_eval) / 10;
2458
2459 if ((sample_period_compression_count - last_eval_compression_count) < min_operations_needed_in_this_sample ||
2460 (sample_period_decompression_count - last_eval_decompression_count) < min_operations_needed_in_this_sample) {
2461 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, sample_period_compression_count - last_eval_compression_count,
2462 sample_period_decompression_count - last_eval_decompression_count, 0, 1, 0);
2463
2464 swapout_target_age = 0;
2465
2466 compressor_need_sample_reset = TRUE;
2467 need_eval_reset = TRUE;
2468 goto done;
2469 }
2470 last_eval_compression_count = sample_period_compression_count;
2471 last_eval_decompression_count = sample_period_decompression_count;
2472
2473 if (elapsed_msecs_in_sample < compressor_sample_min_in_msecs) {
2474 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, swapout_target_age, 0, 0, 5, 0);
2475 goto done;
2476 }
2477 if (sample_period_decompression_count > ((compressor_thrashing_threshold_per_10msecs * elapsed_msecs_in_sample) / 10)) {
2478 uint64_t running_total;
2479 uint64_t working_target;
2480 uint64_t aging_target;
2481 uint32_t oldest_age_of_csegs_sampled = 0;
2482 uint64_t working_set_approximation = 0;
2483
2484 swapout_target_age = 0;
2485
2486 working_target = (sample_period_decompression_count / 100) * 95; /* 95 percent */
2487 aging_target = (sample_period_decompression_count / 100) * 1; /* 1 percent */
2488 running_total = 0;
2489
2490 for (oldest_age_of_csegs_sampled = 0; oldest_age_of_csegs_sampled < DECOMPRESSION_SAMPLE_MAX_AGE; oldest_age_of_csegs_sampled++) {
2491 running_total += age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled];
2492
2493 working_set_approximation += oldest_age_of_csegs_sampled * age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled];
2494
2495 if (running_total >= working_target) {
2496 break;
2497 }
2498 }
2499 if (oldest_age_of_csegs_sampled < DECOMPRESSION_SAMPLE_MAX_AGE) {
2500 working_set_approximation = (working_set_approximation * 1000) / elapsed_msecs_in_sample;
2501
2502 if (working_set_approximation < VM_PAGE_COMPRESSOR_COUNT) {
2503 running_total = overage_decompressions_during_sample_period;
2504
2505 for (oldest_age_of_csegs_sampled = DECOMPRESSION_SAMPLE_MAX_AGE - 1; oldest_age_of_csegs_sampled; oldest_age_of_csegs_sampled--) {
2506 running_total += age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled];
2507
2508 if (running_total >= aging_target) {
2509 break;
2510 }
2511 }
2512 swapout_target_age = (uint32_t)cur_ts_sec - oldest_age_of_csegs_sampled;
2513
2514 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, swapout_target_age, working_set_approximation, VM_PAGE_COMPRESSOR_COUNT, 2, 0);
2515 } else {
2516 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, working_set_approximation, VM_PAGE_COMPRESSOR_COUNT, 0, 3, 0);
2517 }
2518 } else {
2519 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, working_target, running_total, 0, 4, 0);
2520 }
2521
2522 compressor_need_sample_reset = TRUE;
2523 need_eval_reset = TRUE;
2524 } else {
2525 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, sample_period_decompression_count, (compressor_thrashing_threshold_per_10msecs * elapsed_msecs_in_sample) / 10, 0, 6, 0);
2526 }
2527done:
2528 if (compressor_need_sample_reset == TRUE) {
2529 bzero(s: age_of_decompressions_during_sample_period, n: sizeof(age_of_decompressions_during_sample_period));
2530 overage_decompressions_during_sample_period = 0;
2531
2532 start_of_sample_period_sec = cur_ts_sec;
2533 start_of_sample_period_nsec = cur_ts_nsec;
2534 sample_period_decompression_count = 0;
2535 sample_period_compression_count = 0;
2536 last_eval_decompression_count = 0;
2537 last_eval_compression_count = 0;
2538 compressor_need_sample_reset = FALSE;
2539 }
2540 if (need_eval_reset == TRUE) {
2541 start_of_eval_period_sec = cur_ts_sec;
2542 start_of_eval_period_nsec = cur_ts_nsec;
2543 }
2544}
2545
2546
2547int compaction_swapper_init_now = 0;
2548int compaction_swapper_running = 0;
2549int compaction_swapper_awakened = 0;
2550int compaction_swapper_abort = 0;
2551
2552bool
2553vm_compressor_swapout_is_ripe()
2554{
2555 bool is_ripe = false;
2556 if (vm_swapout_ripe_segments == TRUE && c_overage_swapped_count < c_overage_swapped_limit) {
2557 c_segment_t c_seg;
2558 clock_sec_t now;
2559 clock_sec_t age;
2560 clock_nsec_t nsec;
2561
2562 clock_get_system_nanotime(secs: &now, nanosecs: &nsec);
2563 age = 0;
2564
2565 lck_mtx_lock_spin_always(c_list_lock);
2566
2567 if (!queue_empty(&c_age_list_head)) {
2568 c_seg = (c_segment_t) queue_first(&c_age_list_head);
2569
2570 age = now - c_seg->c_creation_ts;
2571 }
2572 lck_mtx_unlock_always(c_list_lock);
2573
2574 if (age >= vm_ripe_target_age) {
2575 is_ripe = true;
2576 }
2577 }
2578 return is_ripe;
2579}
2580
2581static bool
2582compressor_swapout_conditions_met(void)
2583{
2584 bool should_swap = false;
2585 if (COMPRESSOR_NEEDS_TO_SWAP()) {
2586 should_swap = true;
2587 vmcs_stats.compressor_swap_threshold_exceeded++;
2588 }
2589 if (VM_PAGE_Q_THROTTLED(&vm_pageout_queue_external) && vm_page_anonymous_count < (vm_page_inactive_count / 20)) {
2590 should_swap = true;
2591 vmcs_stats.external_q_throttled++;
2592 }
2593 if (vm_page_free_count < (vm_page_free_reserved - (COMPRESSOR_FREE_RESERVED_LIMIT * 2))) {
2594 should_swap = true;
2595 vmcs_stats.free_count_below_reserve++;
2596 }
2597 return should_swap;
2598}
2599
2600static bool
2601compressor_needs_to_swap()
2602{
2603 bool should_swap = false;
2604 if (vm_compressor_swapout_is_ripe()) {
2605 should_swap = true;
2606 goto check_if_low_space;
2607 }
2608
2609 if (VM_CONFIG_SWAP_IS_ACTIVE) {
2610 should_swap = compressor_swapout_conditions_met();
2611 if (should_swap) {
2612 goto check_if_low_space;
2613 }
2614 }
2615
2616#if (XNU_TARGET_OS_OSX && __arm64__)
2617 /*
2618 * Thrashing detection disabled.
2619 */
2620#else /* (XNU_TARGET_OS_OSX && __arm64__) */
2621
2622 if (vm_compressor_is_thrashing()) {
2623 should_swap = true;
2624 vmcs_stats.thrashing_detected++;
2625 }
2626
2627#if CONFIG_PHANTOM_CACHE
2628 if (vm_phantom_cache_check_pressure()) {
2629 os_atomic_store(&memorystatus_phantom_cache_pressure, true, release);
2630 should_swap = true;
2631 }
2632#endif
2633 if (swapout_target_age) {
2634 should_swap = true;
2635 }
2636#endif /* (XNU_TARGET_OS_OSX && __arm64__) */
2637
2638check_if_low_space:
2639
2640#if CONFIG_JETSAM
2641 if (should_swap || vm_compressor_low_on_space() == TRUE) {
2642 if (vm_compressor_thrashing_detected == FALSE) {
2643 vm_compressor_thrashing_detected = TRUE;
2644
2645 if (swapout_target_age) {
2646 compressor_thrashing_induced_jetsam++;
2647 } else if (vm_compressor_low_on_space() == TRUE) {
2648 compressor_thrashing_induced_jetsam++;
2649 } else {
2650 filecache_thrashing_induced_jetsam++;
2651 }
2652 /*
2653 * Wake up the memorystatus thread so that it can return
2654 * the system to a healthy state (by killing processes).
2655 */
2656 memorystatus_thread_wake();
2657 }
2658 /*
2659 * let the jetsam take precedence over
2660 * any major compactions we might have
2661 * been able to do... otherwise we run
2662 * the risk of doing major compactions
2663 * on segments we're about to free up
2664 * due to the jetsam activity.
2665 */
2666 should_swap = false;
2667 if (memorystatus_swap_all_apps && vm_swap_low_on_space()) {
2668 vm_compressor_take_paging_space_action();
2669 }
2670 }
2671
2672#else /* CONFIG_JETSAM */
2673 if (should_swap && vm_swap_low_on_space()) {
2674 vm_compressor_take_paging_space_action();
2675 }
2676#endif /* CONFIG_JETSAM */
2677
2678 if (should_swap == false) {
2679 /*
2680 * vm_compressor_needs_to_major_compact returns true only if we're
2681 * about to run out of available compressor segments... in this
2682 * case, we absolutely need to run a major compaction even if
2683 * we've just kicked off a jetsam or we don't otherwise need to
2684 * swap... terminating objects releases
2685 * pages back to the uncompressed cache, but does not guarantee
2686 * that we will free up even a single compression segment
2687 */
2688 should_swap = vm_compressor_needs_to_major_compact();
2689 if (should_swap) {
2690 vmcs_stats.fragmentation_detected++;
2691 }
2692 }
2693
2694 /*
2695 * returning TRUE when swap_supported == FALSE
2696 * will cause the major compaction engine to
2697 * run, but will not trigger any swapping...
2698 * segments that have been major compacted
2699 * will be moved to the majorcompact queue
2700 */
2701 return should_swap;
2702}
2703
2704#if CONFIG_JETSAM
2705/*
2706 * This function is called from the jetsam thread after killing something to
2707 * mitigate thrashing.
2708 *
2709 * We need to restart our thrashing detection heuristics since memory pressure
2710 * has potentially changed significantly, and we don't want to detect on old
2711 * data from before the jetsam.
2712 */
2713void
2714vm_thrashing_jetsam_done(void)
2715{
2716 vm_compressor_thrashing_detected = FALSE;
2717
2718 /* Were we compressor-thrashing or filecache-thrashing? */
2719 if (swapout_target_age) {
2720 swapout_target_age = 0;
2721 compressor_need_sample_reset = TRUE;
2722 }
2723#if CONFIG_PHANTOM_CACHE
2724 else {
2725 vm_phantom_cache_restart_sample();
2726 }
2727#endif
2728}
2729#endif /* CONFIG_JETSAM */
2730
2731uint32_t vm_wake_compactor_swapper_calls = 0;
2732uint32_t vm_run_compactor_already_running = 0;
2733uint32_t vm_run_compactor_empty_minor_q = 0;
2734uint32_t vm_run_compactor_did_compact = 0;
2735uint32_t vm_run_compactor_waited = 0;
2736
2737void
2738vm_run_compactor(void)
2739{
2740 if (c_segment_count == 0) {
2741 return;
2742 }
2743
2744 lck_mtx_lock_spin_always(c_list_lock);
2745
2746 if (c_minor_count == 0) {
2747 vm_run_compactor_empty_minor_q++;
2748
2749 lck_mtx_unlock_always(c_list_lock);
2750 return;
2751 }
2752 if (compaction_swapper_running) {
2753 if (vm_pageout_state.vm_restricted_to_single_processor == FALSE) {
2754 vm_run_compactor_already_running++;
2755
2756 lck_mtx_unlock_always(c_list_lock);
2757 return;
2758 }
2759 vm_run_compactor_waited++;
2760
2761 assert_wait(event: (event_t)&compaction_swapper_running, THREAD_UNINT);
2762
2763 lck_mtx_unlock_always(c_list_lock);
2764
2765 thread_block(THREAD_CONTINUE_NULL);
2766
2767 return;
2768 }
2769 vm_run_compactor_did_compact++;
2770
2771 fastwake_warmup = FALSE;
2772 compaction_swapper_running = 1;
2773
2774 vm_compressor_do_delayed_compactions(FALSE);
2775
2776 compaction_swapper_running = 0;
2777
2778 lck_mtx_unlock_always(c_list_lock);
2779
2780 thread_wakeup((event_t)&compaction_swapper_running);
2781}
2782
2783
2784void
2785vm_wake_compactor_swapper(void)
2786{
2787 if (compaction_swapper_running || compaction_swapper_awakened || c_segment_count == 0) {
2788 return;
2789 }
2790
2791 if (c_minor_count || vm_compressor_needs_to_major_compact()) {
2792 lck_mtx_lock_spin_always(c_list_lock);
2793
2794 fastwake_warmup = FALSE;
2795
2796 if (compaction_swapper_running == 0 && compaction_swapper_awakened == 0) {
2797 vm_wake_compactor_swapper_calls++;
2798
2799 compaction_swapper_awakened = 1;
2800 thread_wakeup((event_t)&c_compressor_swap_trigger);
2801 }
2802 lck_mtx_unlock_always(c_list_lock);
2803 }
2804}
2805
2806
2807void
2808vm_consider_swapping()
2809{
2810 assert(VM_CONFIG_SWAP_IS_PRESENT);
2811
2812 lck_mtx_lock_spin_always(c_list_lock);
2813
2814 compaction_swapper_abort = 1;
2815
2816 while (compaction_swapper_running) {
2817 assert_wait(event: (event_t)&compaction_swapper_running, THREAD_UNINT);
2818
2819 lck_mtx_unlock_always(c_list_lock);
2820
2821 thread_block(THREAD_CONTINUE_NULL);
2822
2823 lck_mtx_lock_spin_always(c_list_lock);
2824 }
2825 compaction_swapper_abort = 0;
2826 compaction_swapper_running = 1;
2827
2828 vm_swapout_ripe_segments = TRUE;
2829
2830 vm_compressor_process_major_segments(vm_swapout_ripe_segments);
2831
2832 vm_compressor_compact_and_swap(FALSE);
2833
2834 compaction_swapper_running = 0;
2835
2836 vm_swapout_ripe_segments = FALSE;
2837
2838 lck_mtx_unlock_always(c_list_lock);
2839
2840 thread_wakeup((event_t)&compaction_swapper_running);
2841}
2842
2843
2844void
2845vm_consider_waking_compactor_swapper(void)
2846{
2847 boolean_t need_wakeup = FALSE;
2848
2849 if (c_segment_count == 0) {
2850 return;
2851 }
2852
2853 if (compaction_swapper_running || compaction_swapper_awakened) {
2854 return;
2855 }
2856
2857 if (!compaction_swapper_inited && !compaction_swapper_init_now) {
2858 compaction_swapper_init_now = 1;
2859 need_wakeup = TRUE;
2860 }
2861
2862 if (c_minor_count && (COMPRESSOR_NEEDS_TO_MINOR_COMPACT())) {
2863 need_wakeup = TRUE;
2864 } else if (compressor_needs_to_swap()) {
2865 need_wakeup = TRUE;
2866 } else if (c_minor_count) {
2867 uint64_t total_bytes;
2868
2869 total_bytes = compressor_object->resident_page_count * PAGE_SIZE_64;
2870
2871 if ((total_bytes - compressor_bytes_used) > total_bytes / 10) {
2872 need_wakeup = TRUE;
2873 }
2874 }
2875 if (need_wakeup == TRUE) {
2876 lck_mtx_lock_spin_always(c_list_lock);
2877
2878 fastwake_warmup = FALSE;
2879
2880 if (compaction_swapper_running == 0 && compaction_swapper_awakened == 0) {
2881 memoryshot(VM_WAKEUP_COMPACTOR_SWAPPER, DBG_FUNC_NONE);
2882
2883 compaction_swapper_awakened = 1;
2884 thread_wakeup((event_t)&c_compressor_swap_trigger);
2885 }
2886 lck_mtx_unlock_always(c_list_lock);
2887 }
2888}
2889
2890
2891#define C_SWAPOUT_LIMIT 4
2892#define DELAYED_COMPACTIONS_PER_PASS 30
2893
2894void
2895vm_compressor_do_delayed_compactions(boolean_t flush_all)
2896{
2897 c_segment_t c_seg;
2898 int number_compacted = 0;
2899 boolean_t needs_to_swap = FALSE;
2900 uint32_t c_swapout_count = 0;
2901
2902
2903 VM_DEBUG_CONSTANT_EVENT(vm_compressor_do_delayed_compactions, VM_COMPRESSOR_DO_DELAYED_COMPACTIONS, DBG_FUNC_START, c_minor_count, flush_all, 0, 0);
2904
2905#if XNU_TARGET_OS_OSX
2906 LCK_MTX_ASSERT(c_list_lock, LCK_MTX_ASSERT_OWNED);
2907#endif /* XNU_TARGET_OS_OSX */
2908
2909 while (!queue_empty(&c_minor_list_head) && needs_to_swap == FALSE) {
2910 c_seg = (c_segment_t)queue_first(&c_minor_list_head);
2911
2912 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
2913
2914 if (c_seg->c_busy) {
2915 lck_mtx_unlock_always(c_list_lock);
2916 c_seg_wait_on_busy(c_seg);
2917 lck_mtx_lock_spin_always(c_list_lock);
2918
2919 continue;
2920 }
2921 C_SEG_BUSY(c_seg);
2922
2923 c_seg_do_minor_compaction_and_unlock(c_seg, TRUE, FALSE, TRUE);
2924
2925 c_swapout_count = c_early_swapout_count + c_regular_swapout_count + c_late_swapout_count;
2926 if (VM_CONFIG_SWAP_IS_ACTIVE && (number_compacted++ > DELAYED_COMPACTIONS_PER_PASS)) {
2927 if ((flush_all == TRUE || compressor_needs_to_swap()) && c_swapout_count < C_SWAPOUT_LIMIT) {
2928 needs_to_swap = TRUE;
2929 }
2930
2931 number_compacted = 0;
2932 }
2933 lck_mtx_lock_spin_always(c_list_lock);
2934 }
2935
2936 VM_DEBUG_CONSTANT_EVENT(vm_compressor_do_delayed_compactions, VM_COMPRESSOR_DO_DELAYED_COMPACTIONS, DBG_FUNC_END, c_minor_count, number_compacted, needs_to_swap, 0);
2937}
2938
2939int min_csegs_per_major_compaction = DELAYED_COMPACTIONS_PER_PASS;
2940
2941static bool
2942vm_compressor_major_compact_cseg(c_segment_t c_seg, uint32_t* c_seg_considered, bool* bail_wanted_cseg, uint64_t* total_bytes_freed)
2943{
2944 /*
2945 * Major compaction
2946 */
2947 bool keep_compacting = true, fully_compacted = true;
2948 queue_head_t *list_head = NULL;
2949 c_segment_t c_seg_next;
2950 uint64_t bytes_to_free = 0, bytes_freed = 0;
2951 uint32_t number_considered = 0;
2952
2953 if (c_seg->c_state == C_ON_AGE_Q) {
2954 assert(!c_seg->c_has_donated_pages);
2955 list_head = &c_age_list_head;
2956 } else if (c_seg->c_state == C_ON_SWAPPEDIN_Q) {
2957 assert(c_seg->c_has_donated_pages);
2958 list_head = &c_late_swappedin_list_head;
2959 }
2960
2961 while (keep_compacting == TRUE) {
2962 assert(c_seg->c_busy);
2963
2964 /* look for another segment to consolidate */
2965
2966 c_seg_next = (c_segment_t) queue_next(&c_seg->c_age_list);
2967
2968 if (queue_end(list_head, (queue_entry_t)c_seg_next)) {
2969 break;
2970 }
2971
2972 assert(c_seg_next->c_state == c_seg->c_state);
2973
2974 number_considered++;
2975
2976 if (c_seg_major_compact_ok(c_seg_dst: c_seg, c_seg_src: c_seg_next) == FALSE) {
2977 break;
2978 }
2979
2980 lck_mtx_lock_spin_always(lck: &c_seg_next->c_lock);
2981
2982 if (c_seg_next->c_busy) {
2983 /*
2984 * We are going to block for our neighbor.
2985 * If our c_seg is wanted, we should unbusy
2986 * it because we don't know how long we might
2987 * have to block here.
2988 */
2989 if (c_seg->c_wanted) {
2990 lck_mtx_unlock_always(&c_seg_next->c_lock);
2991 fully_compacted = false;
2992 c_seg_major_compact_stats[c_seg_major_compact_stats_now].bailed_compactions++;
2993 *bail_wanted_cseg = true;
2994 break;
2995 }
2996
2997 lck_mtx_unlock_always(c_list_lock);
2998
2999 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 8, (void*) VM_KERNEL_ADDRPERM(c_seg_next), 0, 0);
3000
3001 c_seg_wait_on_busy(c_seg: c_seg_next);
3002 lck_mtx_lock_spin_always(c_list_lock);
3003
3004 continue;
3005 }
3006 /* grab that segment */
3007 C_SEG_BUSY(c_seg_next);
3008
3009 bytes_to_free = C_SEG_OFFSET_TO_BYTES(c_seg_next->c_populated_offset);
3010 if (c_seg_do_minor_compaction_and_unlock(c_seg: c_seg_next, FALSE, TRUE, TRUE)) {
3011 /*
3012 * found an empty c_segment and freed it
3013 * so we can't continue to use c_seg_next
3014 */
3015 bytes_freed += bytes_to_free;
3016 c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_freed_segs++;
3017 continue;
3018 }
3019
3020 /* unlock the list ... */
3021 lck_mtx_unlock_always(c_list_lock);
3022
3023 /* do the major compaction */
3024
3025 keep_compacting = c_seg_major_compact(c_seg_dst: c_seg, c_seg_src: c_seg_next);
3026
3027 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 9, keep_compacting, 0, 0);
3028
3029 PAGE_REPLACEMENT_DISALLOWED(TRUE);
3030
3031 lck_mtx_lock_spin_always(lck: &c_seg_next->c_lock);
3032 /*
3033 * run a minor compaction on the donor segment
3034 * since we pulled at least some of it's
3035 * data into our target... if we've emptied
3036 * it, now is a good time to free it which
3037 * c_seg_minor_compaction_and_unlock also takes care of
3038 *
3039 * by passing TRUE, we ask for c_busy to be cleared
3040 * and c_wanted to be taken care of
3041 */
3042 bytes_to_free = C_SEG_OFFSET_TO_BYTES(c_seg_next->c_populated_offset);
3043 if (c_seg_minor_compaction_and_unlock(c_seg: c_seg_next, TRUE)) {
3044 bytes_freed += bytes_to_free;
3045 c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_freed_segs++;
3046 } else {
3047 bytes_to_free -= C_SEG_OFFSET_TO_BYTES(c_seg_next->c_populated_offset);
3048 bytes_freed += bytes_to_free;
3049 }
3050
3051 PAGE_REPLACEMENT_DISALLOWED(FALSE);
3052
3053 /* relock the list */
3054 lck_mtx_lock_spin_always(c_list_lock);
3055
3056 if (c_seg->c_wanted) {
3057 /*
3058 * Our c_seg is in demand. Let's
3059 * unbusy it and wakeup the waiters
3060 * instead of continuing the compaction
3061 * because we could be in this loop
3062 * for a while.
3063 */
3064 fully_compacted = false;
3065 *bail_wanted_cseg = true;
3066 c_seg_major_compact_stats[c_seg_major_compact_stats_now].bailed_compactions++;
3067 break;
3068 }
3069 } /* major compaction */
3070
3071 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 10, number_considered, *bail_wanted_cseg, 0);
3072
3073 *c_seg_considered += number_considered;
3074 *total_bytes_freed += bytes_freed;
3075
3076 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3077 return fully_compacted;
3078}
3079
3080#define TIME_SUB(rsecs, secs, rfrac, frac, unit) \
3081 MACRO_BEGIN \
3082 if ((int)((rfrac) -= (frac)) < 0) { \
3083 (rfrac) += (unit); \
3084 (rsecs) -= 1; \
3085 } \
3086 (rsecs) -= (secs); \
3087 MACRO_END
3088
3089clock_nsec_t c_process_major_report_over_ms = 9; /* report if over 9 ms */
3090int c_process_major_yield_after = 1000; /* yield after moving 1,000 segments */
3091uint64_t c_process_major_reports = 0;
3092clock_sec_t c_process_major_max_sec = 0;
3093clock_nsec_t c_process_major_max_nsec = 0;
3094uint32_t c_process_major_peak_segcount = 0;
3095static void
3096vm_compressor_process_major_segments(bool ripe_age_only)
3097{
3098 c_segment_t c_seg = NULL;
3099 int count = 0, total = 0, breaks = 0;
3100 clock_sec_t start_sec, end_sec;
3101 clock_nsec_t start_nsec, end_nsec;
3102 clock_nsec_t report_over_ns;
3103
3104 if (queue_empty(&c_major_list_head)) {
3105 return;
3106 }
3107
3108 // printf("%s: starting to move segments from MAJORQ to AGEQ\n", __FUNCTION__);
3109 if (c_process_major_report_over_ms != 0) {
3110 report_over_ns = c_process_major_report_over_ms * NSEC_PER_MSEC;
3111 } else {
3112 report_over_ns = (clock_nsec_t)-1;
3113 }
3114
3115 if (ripe_age_only) {
3116 if (c_overage_swapped_count >= c_overage_swapped_limit) {
3117 /*
3118 * Return while we wait for the overage segments
3119 * in our queue to get pushed out first.
3120 */
3121 return;
3122 }
3123 }
3124
3125 clock_get_system_nanotime(secs: &start_sec, nanosecs: &start_nsec);
3126 while (!queue_empty(&c_major_list_head)) {
3127 if (!ripe_age_only) {
3128 /*
3129 * Start from the end to preserve aging order. The newer
3130 * segments are at the tail and so need to be inserted in
3131 * the aging queue in this way so we have the older segments
3132 * at the end of the AGE_Q.
3133 */
3134 c_seg = (c_segment_t)queue_last(&c_major_list_head);
3135 } else {
3136 c_seg = (c_segment_t)queue_first(&c_major_list_head);
3137 if ((start_sec - c_seg->c_creation_ts) < vm_ripe_target_age) {
3138 /*
3139 * We have found the first segment in our queue that is not ripe. Segments after it
3140 * will be the same. So let's bail here. Return with c_list_lock held.
3141 */
3142 break;
3143 }
3144 }
3145
3146 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3147 c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE);
3148 lck_mtx_unlock_always(&c_seg->c_lock);
3149
3150 count++;
3151 if (count == c_process_major_yield_after ||
3152 queue_empty(&c_major_list_head)) {
3153 /* done or time to take a break */
3154 } else {
3155 /* keep going */
3156 continue;
3157 }
3158
3159 total += count;
3160 clock_get_system_nanotime(secs: &end_sec, nanosecs: &end_nsec);
3161 TIME_SUB(end_sec, start_sec, end_nsec, start_nsec, NSEC_PER_SEC);
3162 if (end_sec > c_process_major_max_sec) {
3163 c_process_major_max_sec = end_sec;
3164 c_process_major_max_nsec = end_nsec;
3165 } else if (end_sec == c_process_major_max_sec &&
3166 end_nsec > c_process_major_max_nsec) {
3167 c_process_major_max_nsec = end_nsec;
3168 }
3169 if (total > c_process_major_peak_segcount) {
3170 c_process_major_peak_segcount = total;
3171 }
3172 if (end_sec > 0 ||
3173 end_nsec >= report_over_ns) {
3174 /* we used more than expected */
3175 c_process_major_reports++;
3176 printf(format: "%s: moved %d/%d segments from MAJORQ to AGEQ in %lu.%09u seconds and %d breaks\n",
3177 __FUNCTION__, count, total,
3178 end_sec, end_nsec, breaks);
3179 }
3180 if (queue_empty(&c_major_list_head)) {
3181 /* done */
3182 break;
3183 }
3184 /* take a break to allow someone else to grab the lock */
3185 lck_mtx_unlock_always(c_list_lock);
3186 mutex_pause(0); /* 10 microseconds */
3187 lck_mtx_lock_spin_always(c_list_lock);
3188 /* start again */
3189 clock_get_system_nanotime(secs: &start_sec, nanosecs: &start_nsec);
3190 count = 0;
3191 breaks++;
3192 }
3193}
3194
3195/*
3196 * macOS special swappable csegs -> early_swapin queue
3197 * non-macOS special swappable+non-freezer csegs -> late_swapin queue
3198 * Processing special csegs means minor compacting each cseg and then
3199 * major compacting it and putting them on the early or late
3200 * (depending on platform) swapout queue.
3201 */
3202static void
3203vm_compressor_process_special_swapped_in_segments_locked(void)
3204{
3205 c_segment_t c_seg = NULL;
3206 bool switch_state = true, bail_wanted_cseg = false;
3207 unsigned int number_considered = 0, yield_after_considered_per_pass = 0;
3208 uint64_t bytes_freed = 0;
3209 queue_head_t *special_swappedin_list_head;
3210
3211#if XNU_TARGET_OS_OSX
3212 special_swappedin_list_head = &c_early_swappedin_list_head;
3213#else /* XNU_TARGET_OS_OSX */
3214 if (memorystatus_swap_all_apps) {
3215 special_swappedin_list_head = &c_late_swappedin_list_head;
3216 } else {
3217 /* called on unsupported config*/
3218 return;
3219 }
3220#endif /* XNU_TARGET_OS_OSX */
3221
3222 yield_after_considered_per_pass = MAX(min_csegs_per_major_compaction, DELAYED_COMPACTIONS_PER_PASS);
3223 while (!queue_empty(special_swappedin_list_head)) {
3224 c_seg = (c_segment_t)queue_first(special_swappedin_list_head);
3225
3226 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3227
3228 if (c_seg->c_busy) {
3229 lck_mtx_unlock_always(c_list_lock);
3230 c_seg_wait_on_busy(c_seg);
3231 lck_mtx_lock_spin_always(c_list_lock);
3232 continue;
3233 }
3234
3235 C_SEG_BUSY(c_seg);
3236 lck_mtx_unlock_always(&c_seg->c_lock);
3237 lck_mtx_unlock_always(c_list_lock);
3238
3239 PAGE_REPLACEMENT_DISALLOWED(TRUE);
3240
3241 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3242
3243 if (c_seg_minor_compaction_and_unlock(c_seg, FALSE /*clear busy?*/)) {
3244 /*
3245 * found an empty c_segment and freed it
3246 * so go grab the next guy in the queue
3247 */
3248 PAGE_REPLACEMENT_DISALLOWED(FALSE);
3249 lck_mtx_lock_spin_always(c_list_lock);
3250 continue;
3251 }
3252
3253 PAGE_REPLACEMENT_DISALLOWED(FALSE);
3254 lck_mtx_lock_spin_always(c_list_lock);
3255
3256 switch_state = vm_compressor_major_compact_cseg(c_seg, c_seg_considered: &number_considered, bail_wanted_cseg: &bail_wanted_cseg, total_bytes_freed: &bytes_freed);
3257 assert(c_seg->c_busy);
3258 assert(!c_seg->c_on_minorcompact_q);
3259
3260 if (switch_state) {
3261 if (VM_CONFIG_SWAP_IS_ACTIVE || VM_CONFIG_FREEZER_SWAP_IS_ACTIVE) {
3262 /*
3263 * Ordinarily we let swapped in segments age out + get
3264 * major compacted with the rest of the c_segs on the ageQ.
3265 * But the early donated c_segs, if well compacted, should be
3266 * kept ready to be swapped out if needed. These are typically
3267 * describing memory belonging to a leaky app (macOS) or a swap-
3268 * capable app (iPadOS) and for the latter we can keep these
3269 * around longer because we control the triggers in the memorystatus
3270 * subsystem
3271 */
3272 c_seg_switch_state(c_seg, C_ON_SWAPOUT_Q, FALSE);
3273 }
3274 }
3275
3276 C_SEG_WAKEUP_DONE(c_seg);
3277
3278 lck_mtx_unlock_always(&c_seg->c_lock);
3279
3280 if (number_considered >= yield_after_considered_per_pass) {
3281 if (bail_wanted_cseg) {
3282 /*
3283 * We stopped major compactions on a c_seg
3284 * that is wanted. We don't know the priority
3285 * of the waiter unfortunately but we are at
3286 * a very high priority and so, just in case
3287 * the waiter is a critical system daemon or
3288 * UI thread, let's give up the CPU in case
3289 * the system is running a few CPU intensive
3290 * tasks.
3291 */
3292 bail_wanted_cseg = false;
3293 lck_mtx_unlock_always(c_list_lock);
3294
3295 mutex_pause(2); /* 100us yield */
3296
3297 lck_mtx_lock_spin_always(c_list_lock);
3298 }
3299
3300 number_considered = 0;
3301 }
3302 }
3303}
3304
3305void
3306vm_compressor_process_special_swapped_in_segments(void)
3307{
3308 lck_mtx_lock_spin_always(c_list_lock);
3309 vm_compressor_process_special_swapped_in_segments_locked();
3310 lck_mtx_unlock_always(c_list_lock);
3311}
3312
3313#define C_SEGMENT_SWAPPEDIN_AGE_LIMIT 10
3314/*
3315 * Processing regular csegs means aging them.
3316 */
3317static void
3318vm_compressor_process_regular_swapped_in_segments(boolean_t flush_all)
3319{
3320 c_segment_t c_seg;
3321 clock_sec_t now;
3322 clock_nsec_t nsec;
3323
3324 clock_get_system_nanotime(secs: &now, nanosecs: &nsec);
3325
3326 while (!queue_empty(&c_regular_swappedin_list_head)) {
3327 c_seg = (c_segment_t)queue_first(&c_regular_swappedin_list_head);
3328
3329 if (flush_all == FALSE && (now - c_seg->c_swappedin_ts) < C_SEGMENT_SWAPPEDIN_AGE_LIMIT) {
3330 break;
3331 }
3332
3333 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3334
3335 c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE);
3336 c_seg->c_agedin_ts = (uint32_t) now;
3337
3338 lck_mtx_unlock_always(&c_seg->c_lock);
3339 }
3340}
3341
3342
3343extern int vm_num_swap_files;
3344extern int vm_num_pinned_swap_files;
3345extern int vm_swappin_enabled;
3346
3347extern unsigned int vm_swapfile_total_segs_used;
3348extern unsigned int vm_swapfile_total_segs_alloced;
3349
3350
3351void
3352vm_compressor_flush(void)
3353{
3354 uint64_t vm_swap_put_failures_at_start;
3355 wait_result_t wait_result = 0;
3356 AbsoluteTime startTime, endTime;
3357 clock_sec_t now_sec;
3358 clock_nsec_t now_nsec;
3359 uint64_t nsec;
3360 c_segment_t c_seg, c_seg_next;
3361
3362 HIBLOG("vm_compressor_flush - starting\n");
3363
3364 clock_get_uptime(result: &startTime);
3365
3366 lck_mtx_lock_spin_always(c_list_lock);
3367
3368 fastwake_warmup = FALSE;
3369 compaction_swapper_abort = 1;
3370
3371 while (compaction_swapper_running) {
3372 assert_wait(event: (event_t)&compaction_swapper_running, THREAD_UNINT);
3373
3374 lck_mtx_unlock_always(c_list_lock);
3375
3376 thread_block(THREAD_CONTINUE_NULL);
3377
3378 lck_mtx_lock_spin_always(c_list_lock);
3379 }
3380 compaction_swapper_abort = 0;
3381 compaction_swapper_running = 1;
3382
3383 hibernate_flushing = TRUE;
3384 hibernate_no_swapspace = FALSE;
3385 hibernate_flush_timed_out = FALSE;
3386 c_generation_id_flush_barrier = c_generation_id + 1000;
3387
3388 clock_get_system_nanotime(secs: &now_sec, nanosecs: &now_nsec);
3389 hibernate_flushing_deadline = now_sec + HIBERNATE_FLUSHING_SECS_TO_COMPLETE;
3390
3391 vm_swap_put_failures_at_start = vm_swap_put_failures;
3392
3393 /*
3394 * We are about to hibernate and so we want all segments flushed to disk.
3395 * Segments that are on the major compaction queue won't be considered in
3396 * the vm_compressor_compact_and_swap() pass. So we need to bring them to
3397 * the ageQ for consideration.
3398 */
3399 if (!queue_empty(&c_major_list_head)) {
3400 c_seg = (c_segment_t)queue_first(&c_major_list_head);
3401
3402 while (!queue_end(&c_major_list_head, (queue_entry_t)c_seg)) {
3403 c_seg_next = (c_segment_t) queue_next(&c_seg->c_age_list);
3404 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3405 c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE);
3406 lck_mtx_unlock_always(&c_seg->c_lock);
3407 c_seg = c_seg_next;
3408 }
3409 }
3410 vm_compressor_compact_and_swap(TRUE);
3411
3412 while (!queue_empty(&c_early_swapout_list_head) || !queue_empty(&c_regular_swapout_list_head) || !queue_empty(&c_late_swapout_list_head)) {
3413 assert_wait_timeout(event: (event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, interval: 5000, scale_factor: 1000 * NSEC_PER_USEC);
3414
3415 lck_mtx_unlock_always(c_list_lock);
3416
3417 wait_result = thread_block(THREAD_CONTINUE_NULL);
3418
3419 lck_mtx_lock_spin_always(c_list_lock);
3420
3421 if (wait_result == THREAD_TIMED_OUT) {
3422 break;
3423 }
3424 }
3425 hibernate_flushing = FALSE;
3426 compaction_swapper_running = 0;
3427
3428 if (vm_swap_put_failures > vm_swap_put_failures_at_start) {
3429 HIBLOG("vm_compressor_flush failed to clean %llu segments - vm_page_compressor_count(%d)\n",
3430 vm_swap_put_failures - vm_swap_put_failures_at_start, VM_PAGE_COMPRESSOR_COUNT);
3431 }
3432
3433 lck_mtx_unlock_always(c_list_lock);
3434
3435 thread_wakeup((event_t)&compaction_swapper_running);
3436
3437 clock_get_uptime(result: &endTime);
3438 SUB_ABSOLUTETIME(&endTime, &startTime);
3439 absolutetime_to_nanoseconds(abstime: endTime, result: &nsec);
3440
3441 HIBLOG("vm_compressor_flush completed - took %qd msecs - vm_num_swap_files = %d, vm_num_pinned_swap_files = %d, vm_swappin_enabled = %d\n",
3442 nsec / 1000000ULL, vm_num_swap_files, vm_num_pinned_swap_files, vm_swappin_enabled);
3443}
3444
3445
3446int compaction_swap_trigger_thread_awakened = 0;
3447
3448static void
3449vm_compressor_swap_trigger_thread(void)
3450{
3451 current_thread()->options |= TH_OPT_VMPRIV;
3452
3453 /*
3454 * compaction_swapper_init_now is set when the first call to
3455 * vm_consider_waking_compactor_swapper is made from
3456 * vm_pageout_scan... since this function is called upon
3457 * thread creation, we want to make sure to delay adjusting
3458 * the tuneables until we are awakened via vm_pageout_scan
3459 * so that we are at a point where the vm_swapfile_open will
3460 * be operating on the correct directory (in case the default
3461 * of using the VM volume is overridden by the dynamic_pager)
3462 */
3463 if (compaction_swapper_init_now) {
3464 vm_compaction_swapper_do_init();
3465
3466 if (vm_pageout_state.vm_restricted_to_single_processor == TRUE) {
3467 thread_vm_bind_group_add();
3468 }
3469#if CONFIG_THREAD_GROUPS
3470 thread_group_vm_add();
3471#endif
3472 thread_set_thread_name(th: current_thread(), name: "VM_cswap_trigger");
3473 compaction_swapper_init_now = 0;
3474 }
3475 lck_mtx_lock_spin_always(c_list_lock);
3476
3477 compaction_swap_trigger_thread_awakened++;
3478 compaction_swapper_awakened = 0;
3479
3480 if (compaction_swapper_running == 0) {
3481 compaction_swapper_running = 1;
3482
3483 vm_compressor_compact_and_swap(FALSE);
3484
3485 compaction_swapper_running = 0;
3486 }
3487 assert_wait(event: (event_t)&c_compressor_swap_trigger, THREAD_UNINT);
3488
3489 if (compaction_swapper_running == 0) {
3490 thread_wakeup((event_t)&compaction_swapper_running);
3491 }
3492
3493 lck_mtx_unlock_always(c_list_lock);
3494
3495 thread_block(continuation: (thread_continue_t)vm_compressor_swap_trigger_thread);
3496
3497 /* NOTREACHED */
3498}
3499
3500
3501void
3502vm_compressor_record_warmup_start(void)
3503{
3504 c_segment_t c_seg;
3505
3506 lck_mtx_lock_spin_always(c_list_lock);
3507
3508 if (first_c_segment_to_warm_generation_id == 0) {
3509 if (!queue_empty(&c_age_list_head)) {
3510 c_seg = (c_segment_t)queue_last(&c_age_list_head);
3511
3512 first_c_segment_to_warm_generation_id = c_seg->c_generation_id;
3513 } else {
3514 first_c_segment_to_warm_generation_id = 0;
3515 }
3516
3517 fastwake_recording_in_progress = TRUE;
3518 }
3519 lck_mtx_unlock_always(c_list_lock);
3520}
3521
3522
3523void
3524vm_compressor_record_warmup_end(void)
3525{
3526 c_segment_t c_seg;
3527
3528 lck_mtx_lock_spin_always(c_list_lock);
3529
3530 if (fastwake_recording_in_progress == TRUE) {
3531 if (!queue_empty(&c_age_list_head)) {
3532 c_seg = (c_segment_t)queue_last(&c_age_list_head);
3533
3534 last_c_segment_to_warm_generation_id = c_seg->c_generation_id;
3535 } else {
3536 last_c_segment_to_warm_generation_id = first_c_segment_to_warm_generation_id;
3537 }
3538
3539 fastwake_recording_in_progress = FALSE;
3540
3541 HIBLOG("vm_compressor_record_warmup (%qd - %qd)\n", first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id);
3542 }
3543 lck_mtx_unlock_always(c_list_lock);
3544}
3545
3546
3547#define DELAY_TRIM_ON_WAKE_SECS 25
3548
3549void
3550vm_compressor_delay_trim(void)
3551{
3552 clock_sec_t sec;
3553 clock_nsec_t nsec;
3554
3555 clock_get_system_nanotime(secs: &sec, nanosecs: &nsec);
3556 dont_trim_until_ts = sec + DELAY_TRIM_ON_WAKE_SECS;
3557}
3558
3559
3560void
3561vm_compressor_do_warmup(void)
3562{
3563 lck_mtx_lock_spin_always(c_list_lock);
3564
3565 if (first_c_segment_to_warm_generation_id == last_c_segment_to_warm_generation_id) {
3566 first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0;
3567
3568 lck_mtx_unlock_always(c_list_lock);
3569 return;
3570 }
3571
3572 if (compaction_swapper_running == 0 && compaction_swapper_awakened == 0) {
3573 fastwake_warmup = TRUE;
3574
3575 compaction_swapper_awakened = 1;
3576 thread_wakeup((event_t)&c_compressor_swap_trigger);
3577 }
3578 lck_mtx_unlock_always(c_list_lock);
3579}
3580
3581void
3582do_fastwake_warmup_all(void)
3583{
3584 lck_mtx_lock_spin_always(c_list_lock);
3585
3586 if (queue_empty(&c_swappedout_list_head) && queue_empty(&c_swappedout_sparse_list_head)) {
3587 lck_mtx_unlock_always(c_list_lock);
3588 return;
3589 }
3590
3591 fastwake_warmup = TRUE;
3592
3593 do_fastwake_warmup(&c_swappedout_list_head, TRUE);
3594
3595 do_fastwake_warmup(&c_swappedout_sparse_list_head, TRUE);
3596
3597 fastwake_warmup = FALSE;
3598
3599 lck_mtx_unlock_always(c_list_lock);
3600}
3601
3602void
3603do_fastwake_warmup(queue_head_t *c_queue, boolean_t consider_all_cseg)
3604{
3605 c_segment_t c_seg = NULL;
3606 AbsoluteTime startTime, endTime;
3607 uint64_t nsec;
3608
3609
3610 HIBLOG("vm_compressor_fastwake_warmup (%qd - %qd) - starting\n", first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id);
3611
3612 clock_get_uptime(result: &startTime);
3613
3614 lck_mtx_unlock_always(c_list_lock);
3615
3616 proc_set_thread_policy(thread: current_thread(),
3617 TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER2);
3618
3619 PAGE_REPLACEMENT_DISALLOWED(TRUE);
3620
3621 lck_mtx_lock_spin_always(c_list_lock);
3622
3623 while (!queue_empty(c_queue) && fastwake_warmup == TRUE) {
3624 c_seg = (c_segment_t) queue_first(c_queue);
3625
3626 if (consider_all_cseg == FALSE) {
3627 if (c_seg->c_generation_id < first_c_segment_to_warm_generation_id ||
3628 c_seg->c_generation_id > last_c_segment_to_warm_generation_id) {
3629 break;
3630 }
3631
3632 if (vm_page_free_count < (AVAILABLE_MEMORY / 4)) {
3633 break;
3634 }
3635 }
3636
3637 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3638 lck_mtx_unlock_always(c_list_lock);
3639
3640 if (c_seg->c_busy) {
3641 PAGE_REPLACEMENT_DISALLOWED(FALSE);
3642 c_seg_wait_on_busy(c_seg);
3643 PAGE_REPLACEMENT_DISALLOWED(TRUE);
3644 } else {
3645 if (c_seg_swapin(c_seg, TRUE, FALSE) == 0) {
3646 lck_mtx_unlock_always(&c_seg->c_lock);
3647 }
3648 c_segment_warmup_count++;
3649
3650 PAGE_REPLACEMENT_DISALLOWED(FALSE);
3651 vm_pageout_io_throttle();
3652 PAGE_REPLACEMENT_DISALLOWED(TRUE);
3653 }
3654 lck_mtx_lock_spin_always(c_list_lock);
3655 }
3656 lck_mtx_unlock_always(c_list_lock);
3657
3658 PAGE_REPLACEMENT_DISALLOWED(FALSE);
3659
3660 proc_set_thread_policy(thread: current_thread(),
3661 TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER0);
3662
3663 clock_get_uptime(result: &endTime);
3664 SUB_ABSOLUTETIME(&endTime, &startTime);
3665 absolutetime_to_nanoseconds(abstime: endTime, result: &nsec);
3666
3667 HIBLOG("vm_compressor_fastwake_warmup completed - took %qd msecs\n", nsec / 1000000ULL);
3668
3669 lck_mtx_lock_spin_always(c_list_lock);
3670
3671 if (consider_all_cseg == FALSE) {
3672 first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0;
3673 }
3674}
3675
3676extern bool vm_swapout_thread_running;
3677extern boolean_t compressor_store_stop_compaction;
3678
3679void
3680vm_compressor_compact_and_swap(boolean_t flush_all)
3681{
3682 c_segment_t c_seg;
3683 bool switch_state, bail_wanted_cseg = false;
3684 clock_sec_t now;
3685 clock_nsec_t nsec;
3686 mach_timespec_t start_ts, end_ts;
3687 unsigned int number_considered, wanted_cseg_found, yield_after_considered_per_pass, number_yields;
3688 uint64_t bytes_freed, delta_usec;
3689 uint32_t c_swapout_count = 0;
3690
3691 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_START, c_age_count, c_minor_count, c_major_count, vm_page_free_count);
3692
3693 if (fastwake_warmup == TRUE) {
3694 uint64_t starting_warmup_count;
3695
3696 starting_warmup_count = c_segment_warmup_count;
3697
3698 KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 11) | DBG_FUNC_START, c_segment_warmup_count,
3699 first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id, 0, 0);
3700 do_fastwake_warmup(c_queue: &c_swappedout_list_head, FALSE);
3701 KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 11) | DBG_FUNC_END, c_segment_warmup_count, c_segment_warmup_count - starting_warmup_count, 0, 0, 0);
3702
3703 fastwake_warmup = FALSE;
3704 }
3705
3706#if (XNU_TARGET_OS_OSX && __arm64__)
3707 /*
3708 * Re-considering major csegs showed benefits on all platforms by
3709 * significantly reducing fragmentation and getting back memory.
3710 * However, on smaller devices, eg watch, there was increased power
3711 * use for the additional compactions. And the turnover in csegs on
3712 * those smaller platforms is high enough in the decompression/free
3713 * path that we can skip reconsidering them here because we already
3714 * consider them for major compaction in those paths.
3715 */
3716 vm_compressor_process_major_segments(false /*all segments and not just the ripe-aged ones*/);
3717#endif /* (XNU_TARGET_OS_OSX && __arm64__) */
3718
3719 /*
3720 * it's possible for the c_age_list_head to be empty if we
3721 * hit our limits for growing the compressor pool and we subsequently
3722 * hibernated... on the next hibernation we could see the queue as
3723 * empty and not proceeed even though we have a bunch of segments on
3724 * the swapped in queue that need to be dealt with.
3725 */
3726 vm_compressor_do_delayed_compactions(flush_all);
3727 vm_compressor_process_special_swapped_in_segments_locked();
3728 vm_compressor_process_regular_swapped_in_segments(flush_all);
3729
3730 /*
3731 * we only need to grab the timestamp once per
3732 * invocation of this function since the
3733 * timescale we're interested in is measured
3734 * in days
3735 */
3736 clock_get_system_nanotime(secs: &now, nanosecs: &nsec);
3737
3738 start_ts.tv_sec = (int) now;
3739 start_ts.tv_nsec = nsec;
3740 delta_usec = 0;
3741 number_considered = 0;
3742 wanted_cseg_found = 0;
3743 number_yields = 0;
3744 bytes_freed = 0;
3745 yield_after_considered_per_pass = MAX(min_csegs_per_major_compaction, DELAYED_COMPACTIONS_PER_PASS);
3746
3747#if 0
3748 /**
3749 * SW: Need to figure out how to properly rate limit this log because it is currently way too
3750 * noisy. rdar://99379414 (Figure out how to rate limit the fragmentation level logging)
3751 */
3752 os_log(OS_LOG_DEFAULT, "memorystatus: before compaction fragmentation level %u\n", vm_compressor_fragmentation_level());
3753#endif
3754
3755 while (!queue_empty(&c_age_list_head) && !compaction_swapper_abort && !compressor_store_stop_compaction) {
3756 if (hibernate_flushing == TRUE) {
3757 clock_sec_t sec;
3758
3759 if (hibernate_should_abort()) {
3760 HIBLOG("vm_compressor_flush - hibernate_should_abort returned TRUE\n");
3761 break;
3762 }
3763 if (hibernate_no_swapspace == TRUE) {
3764 HIBLOG("vm_compressor_flush - out of swap space\n");
3765 break;
3766 }
3767 if (vm_swap_files_pinned() == FALSE) {
3768 HIBLOG("vm_compressor_flush - unpinned swap files\n");
3769 break;
3770 }
3771 if (hibernate_in_progress_with_pinned_swap == TRUE &&
3772 (vm_swapfile_total_segs_alloced == vm_swapfile_total_segs_used)) {
3773 HIBLOG("vm_compressor_flush - out of pinned swap space\n");
3774 break;
3775 }
3776 clock_get_system_nanotime(secs: &sec, nanosecs: &nsec);
3777
3778 if (sec > hibernate_flushing_deadline) {
3779 hibernate_flush_timed_out = TRUE;
3780 HIBLOG("vm_compressor_flush - failed to finish before deadline\n");
3781 break;
3782 }
3783 }
3784
3785 c_swapout_count = c_early_swapout_count + c_regular_swapout_count + c_late_swapout_count;
3786 if (VM_CONFIG_SWAP_IS_ACTIVE && !vm_swap_out_of_space() && c_swapout_count >= C_SWAPOUT_LIMIT) {
3787 assert_wait_timeout(event: (event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, interval: 100, scale_factor: 1000 * NSEC_PER_USEC);
3788
3789 if (!vm_swapout_thread_running) {
3790 thread_wakeup((event_t)&vm_swapout_thread);
3791 }
3792
3793 lck_mtx_unlock_always(c_list_lock);
3794
3795 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 1, c_swapout_count, 0, 0);
3796
3797 thread_block(THREAD_CONTINUE_NULL);
3798
3799 lck_mtx_lock_spin_always(c_list_lock);
3800 }
3801 /*
3802 * Minor compactions
3803 */
3804 vm_compressor_do_delayed_compactions(flush_all);
3805
3806 /*
3807 * vm_compressor_process_early_swapped_in_segments()
3808 * might be too aggressive. So OFF for now.
3809 */
3810 vm_compressor_process_regular_swapped_in_segments(flush_all);
3811
3812 /* Recompute because we dropped the c_list_lock above*/
3813 c_swapout_count = c_early_swapout_count + c_regular_swapout_count + c_late_swapout_count;
3814 if (VM_CONFIG_SWAP_IS_ACTIVE && !vm_swap_out_of_space() && c_swapout_count >= C_SWAPOUT_LIMIT) {
3815 /*
3816 * we timed out on the above thread_block
3817 * let's loop around and try again
3818 * the timeout allows us to continue
3819 * to do minor compactions to make
3820 * more memory available
3821 */
3822 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 2, c_swapout_count, 0, 0);
3823
3824 continue;
3825 }
3826
3827 /*
3828 * Swap out segments?
3829 */
3830 if (flush_all == FALSE) {
3831 bool needs_to_swap;
3832
3833 lck_mtx_unlock_always(c_list_lock);
3834
3835 needs_to_swap = compressor_needs_to_swap();
3836
3837 lck_mtx_lock_spin_always(c_list_lock);
3838
3839 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 3, needs_to_swap, 0, 0);
3840
3841 if (!needs_to_swap) {
3842 break;
3843 }
3844 }
3845 if (queue_empty(&c_age_list_head)) {
3846 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 4, c_age_count, 0, 0);
3847 break;
3848 }
3849 c_seg = (c_segment_t) queue_first(&c_age_list_head);
3850
3851 assert(c_seg->c_state == C_ON_AGE_Q);
3852
3853 if (flush_all == TRUE && c_seg->c_generation_id > c_generation_id_flush_barrier) {
3854 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 5, 0, 0, 0);
3855 break;
3856 }
3857
3858 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
3859
3860 if (c_seg->c_busy) {
3861 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 6, (void*) VM_KERNEL_ADDRPERM(c_seg), 0, 0);
3862
3863 lck_mtx_unlock_always(c_list_lock);
3864 c_seg_wait_on_busy(c_seg);
3865 lck_mtx_lock_spin_always(c_list_lock);
3866
3867 continue;
3868 }
3869 C_SEG_BUSY(c_seg);
3870
3871 if (c_seg_do_minor_compaction_and_unlock(c_seg, FALSE, TRUE, TRUE)) {
3872 /*
3873 * found an empty c_segment and freed it
3874 * so go grab the next guy in the queue
3875 */
3876 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 7, 0, 0, 0);
3877 c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_freed_segs++;
3878 continue;
3879 }
3880
3881 switch_state = vm_compressor_major_compact_cseg(c_seg, c_seg_considered: &number_considered, bail_wanted_cseg: &bail_wanted_cseg, total_bytes_freed: &bytes_freed);
3882 if (bail_wanted_cseg) {
3883 wanted_cseg_found++;
3884 bail_wanted_cseg = false;
3885 }
3886
3887 assert(c_seg->c_busy);
3888 assert(!c_seg->c_on_minorcompact_q);
3889
3890 if (switch_state) {
3891 if (VM_CONFIG_SWAP_IS_ACTIVE) {
3892 int new_state = C_ON_SWAPOUT_Q;
3893#if (XNU_TARGET_OS_OSX && __arm64__)
3894 if (flush_all == false && compressor_swapout_conditions_met() == false) {
3895 new_state = C_ON_MAJORCOMPACT_Q;
3896 }
3897#endif /* (XNU_TARGET_OS_OSX && __arm64__) */
3898
3899 if (new_state == C_ON_SWAPOUT_Q) {
3900 /*
3901 * This mode of putting a generic c_seg on the swapout list is
3902 * only supported when we have general swapping enabled
3903 */
3904 clock_sec_t lnow;
3905 clock_nsec_t lnsec;
3906 clock_get_system_nanotime(secs: &lnow, nanosecs: &lnsec);
3907 if (c_seg->c_agedin_ts && (lnow - c_seg->c_agedin_ts) < 30) {
3908 vmcs_stats.unripe_under_30s++;
3909 } else if (c_seg->c_agedin_ts && (lnow - c_seg->c_agedin_ts) < 60) {
3910 vmcs_stats.unripe_under_60s++;
3911 } else if (c_seg->c_agedin_ts && (lnow - c_seg->c_agedin_ts) < 300) {
3912 vmcs_stats.unripe_under_300s++;
3913 }
3914 }
3915
3916 c_seg_switch_state(c_seg, new_state, FALSE);
3917 } else {
3918 if ((vm_swapout_ripe_segments == TRUE && c_overage_swapped_count < c_overage_swapped_limit)) {
3919 assert(VM_CONFIG_SWAP_IS_PRESENT);
3920 /*
3921 * we are running compressor sweeps with swap-behind
3922 * make sure the c_seg has aged enough before swapping it
3923 * out...
3924 */
3925 if ((now - c_seg->c_creation_ts) >= vm_ripe_target_age) {
3926 c_seg->c_overage_swap = TRUE;
3927 c_overage_swapped_count++;
3928 c_seg_switch_state(c_seg, C_ON_SWAPOUT_Q, FALSE);
3929 }
3930 }
3931 }
3932 if (c_seg->c_state == C_ON_AGE_Q) {
3933 /*
3934 * this c_seg didn't get moved to the swapout queue
3935 * so we need to move it out of the way...
3936 * we just did a major compaction on it so put it
3937 * on that queue
3938 */
3939 c_seg_switch_state(c_seg, C_ON_MAJORCOMPACT_Q, FALSE);
3940 } else {
3941 c_seg_major_compact_stats[c_seg_major_compact_stats_now].wasted_space_in_swapouts += c_seg_bufsize - c_seg->c_bytes_used;
3942 c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_swapouts++;
3943 }
3944 }
3945
3946 C_SEG_WAKEUP_DONE(c_seg);
3947
3948 lck_mtx_unlock_always(&c_seg->c_lock);
3949
3950 /*
3951 * On systems _with_ general swap, regardless of jetsam, we wake up the swapout thread here.
3952 * On systems _without_ general swap, it's the responsibility of the memorystatus
3953 * subsystem to wake up the swapper.
3954 * TODO: When we have full jetsam support on a swap enabled system, we will need to revisit
3955 * this policy.
3956 */
3957 if (VM_CONFIG_SWAP_IS_ACTIVE && c_swapout_count) {
3958 /*
3959 * We don't pause/yield here because we will either
3960 * yield below or at the top of the loop with the
3961 * assert_wait_timeout.
3962 */
3963 if (!vm_swapout_thread_running) {
3964 thread_wakeup((event_t)&vm_swapout_thread);
3965 }
3966 }
3967
3968 if (number_considered >= yield_after_considered_per_pass) {
3969 if (wanted_cseg_found) {
3970 /*
3971 * We stopped major compactions on a c_seg
3972 * that is wanted. We don't know the priority
3973 * of the waiter unfortunately but we are at
3974 * a very high priority and so, just in case
3975 * the waiter is a critical system daemon or
3976 * UI thread, let's give up the CPU in case
3977 * the system is running a few CPU intensive
3978 * tasks.
3979 */
3980 lck_mtx_unlock_always(c_list_lock);
3981
3982 mutex_pause(2); /* 100us yield */
3983
3984 number_yields++;
3985
3986 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 11, number_considered, number_yields, 0);
3987
3988 lck_mtx_lock_spin_always(c_list_lock);
3989 }
3990
3991 number_considered = 0;
3992 wanted_cseg_found = 0;
3993 }
3994 }
3995 clock_get_system_nanotime(secs: &now, nanosecs: &nsec);
3996
3997 end_ts = major_compact_ts = (mach_timespec_t){.tv_sec = (int)now, .tv_nsec = nsec};
3998
3999 SUB_MACH_TIMESPEC(&end_ts, &start_ts);
4000
4001 delta_usec = (end_ts.tv_sec * USEC_PER_SEC) + (end_ts.tv_nsec / NSEC_PER_USEC) - (number_yields * 100);
4002
4003 delta_usec = MAX(1, delta_usec); /* we could have 0 usec run if conditions weren't right */
4004
4005 c_seg_major_compact_stats[c_seg_major_compact_stats_now].bytes_freed_rate_us = (bytes_freed / delta_usec);
4006
4007 if ((c_seg_major_compact_stats_now + 1) == C_SEG_MAJOR_COMPACT_STATS_MAX) {
4008 c_seg_major_compact_stats_now = 0;
4009 } else {
4010 c_seg_major_compact_stats_now++;
4011 }
4012
4013 assert(c_seg_major_compact_stats_now < C_SEG_MAJOR_COMPACT_STATS_MAX);
4014
4015 VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_END, c_age_count, c_minor_count, c_major_count, vm_page_free_count);
4016}
4017
4018
4019static c_segment_t
4020c_seg_allocate(c_segment_t *current_chead)
4021{
4022 c_segment_t c_seg;
4023 int min_needed;
4024 int size_to_populate;
4025 c_segment_t *donate_queue_head;
4026
4027#if XNU_TARGET_OS_OSX
4028 if (vm_compressor_low_on_space()) {
4029 vm_compressor_take_paging_space_action();
4030 }
4031#endif /* XNU_TARGET_OS_OSX */
4032
4033 if ((c_seg = *current_chead) == NULL) {
4034 uint32_t c_segno;
4035
4036 lck_mtx_lock_spin_always(c_list_lock);
4037
4038 while (c_segments_busy == TRUE) {
4039 assert_wait(event: (event_t) (&c_segments_busy), THREAD_UNINT);
4040
4041 lck_mtx_unlock_always(c_list_lock);
4042
4043 thread_block(THREAD_CONTINUE_NULL);
4044
4045 lck_mtx_lock_spin_always(c_list_lock);
4046 }
4047 if (c_free_segno_head == (uint32_t)-1) {
4048 uint32_t c_segments_available_new;
4049 uint32_t compressed_pages;
4050
4051#if CONFIG_FREEZE
4052 if (freezer_incore_cseg_acct) {
4053 compressed_pages = c_segment_pages_compressed_incore;
4054 } else {
4055 compressed_pages = c_segment_pages_compressed;
4056 }
4057#else
4058 compressed_pages = c_segment_pages_compressed;
4059#endif /* CONFIG_FREEZE */
4060
4061 if (c_segments_available >= c_segments_limit || compressed_pages >= c_segment_pages_compressed_limit) {
4062 lck_mtx_unlock_always(c_list_lock);
4063
4064 return NULL;
4065 }
4066 c_segments_busy = TRUE;
4067 lck_mtx_unlock_always(c_list_lock);
4068
4069 kernel_memory_populate(addr: (vm_offset_t)c_segments_next_page,
4070 PAGE_SIZE, flags: KMA_NOFAIL | KMA_KOBJECT,
4071 VM_KERN_MEMORY_COMPRESSOR);
4072 c_segments_next_page += PAGE_SIZE;
4073
4074 c_segments_available_new = c_segments_available + C_SEGMENTS_PER_PAGE;
4075
4076 if (c_segments_available_new > c_segments_limit) {
4077 c_segments_available_new = c_segments_limit;
4078 }
4079
4080 for (c_segno = c_segments_available + 1; c_segno < c_segments_available_new; c_segno++) {
4081 c_segments[c_segno - 1].c_segno = c_segno;
4082 }
4083
4084 lck_mtx_lock_spin_always(c_list_lock);
4085
4086 c_segments[c_segno - 1].c_segno = c_free_segno_head;
4087 c_free_segno_head = c_segments_available;
4088 c_segments_available = c_segments_available_new;
4089
4090 c_segments_busy = FALSE;
4091 thread_wakeup((event_t) (&c_segments_busy));
4092 }
4093 c_segno = c_free_segno_head;
4094 assert(c_segno >= 0 && c_segno < c_segments_limit);
4095
4096 c_free_segno_head = (uint32_t)c_segments[c_segno].c_segno;
4097
4098 /*
4099 * do the rest of the bookkeeping now while we're still behind
4100 * the list lock and grab our generation id now into a local
4101 * so that we can install it once we have the c_seg allocated
4102 */
4103 c_segment_count++;
4104 if (c_segment_count > c_segment_count_max) {
4105 c_segment_count_max = c_segment_count;
4106 }
4107
4108 lck_mtx_unlock_always(c_list_lock);
4109
4110 c_seg = zalloc_flags(compressor_segment_zone, Z_WAITOK | Z_ZERO);
4111
4112 c_seg->c_store.c_buffer = (int32_t *)C_SEG_BUFFER_ADDRESS(c_segno);
4113
4114 lck_mtx_init(lck: &c_seg->c_lock, grp: &vm_compressor_lck_grp, LCK_ATTR_NULL);
4115
4116 c_seg->c_state = C_IS_EMPTY;
4117 c_seg->c_firstemptyslot = C_SLOT_MAX_INDEX;
4118 c_seg->c_mysegno = c_segno;
4119
4120 lck_mtx_lock_spin_always(c_list_lock);
4121 c_empty_count++;
4122 c_seg_switch_state(c_seg, C_IS_FILLING, FALSE);
4123 c_segments[c_segno].c_seg = c_seg;
4124 assert(c_segments[c_segno].c_segno > c_segments_available);
4125 lck_mtx_unlock_always(c_list_lock);
4126
4127 for (int i = 0; i < vm_pageout_state.vm_compressor_thread_count; i++) {
4128#if XNU_TARGET_OS_OSX
4129 donate_queue_head = (c_segment_t*) &(pgo_iothread_internal_state[i].current_early_swapout_chead);
4130#else /* XNU_TARGET_OS_OSX */
4131 if (memorystatus_swap_all_apps) {
4132 donate_queue_head = (c_segment_t*) &(pgo_iothread_internal_state[i].current_late_swapout_chead);
4133 } else {
4134 donate_queue_head = NULL;
4135 }
4136#endif /* XNU_TARGET_OS_OSX */
4137
4138 if (current_chead == donate_queue_head) {
4139 c_seg->c_has_donated_pages = 1;
4140 break;
4141 }
4142 }
4143
4144 *current_chead = c_seg;
4145
4146#if DEVELOPMENT || DEBUG
4147 C_SEG_MAKE_WRITEABLE(c_seg);
4148#endif
4149 }
4150 c_seg_alloc_nextslot(c_seg);
4151
4152 size_to_populate = c_seg_allocsize - C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset);
4153
4154 if (size_to_populate) {
4155 min_needed = PAGE_SIZE + (c_seg_allocsize - c_seg_bufsize);
4156
4157 if (C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset - c_seg->c_nextoffset) < (unsigned) min_needed) {
4158 if (size_to_populate > C_SEG_MAX_POPULATE_SIZE) {
4159 size_to_populate = C_SEG_MAX_POPULATE_SIZE;
4160 }
4161
4162 OSAddAtomic64(size_to_populate / PAGE_SIZE, &vm_pageout_vminfo.vm_compressor_pages_grabbed);
4163
4164 kernel_memory_populate(
4165 addr: (vm_offset_t) &c_seg->c_store.c_buffer[c_seg->c_populated_offset],
4166 size: size_to_populate,
4167 flags: KMA_NOFAIL | KMA_COMPRESSOR,
4168 VM_KERN_MEMORY_COMPRESSOR);
4169 } else {
4170 size_to_populate = 0;
4171 }
4172 }
4173 PAGE_REPLACEMENT_DISALLOWED(TRUE);
4174
4175 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
4176
4177 if (size_to_populate) {
4178 c_seg->c_populated_offset += C_SEG_BYTES_TO_OFFSET(size_to_populate);
4179 }
4180
4181 return c_seg;
4182}
4183
4184#if DEVELOPMENT || DEBUG
4185#if CONFIG_FREEZE
4186extern boolean_t memorystatus_freeze_to_memory;
4187#endif /* CONFIG_FREEZE */
4188#endif /* DEVELOPMENT || DEBUG */
4189uint64_t c_seg_total_donated_bytes = 0; /* For testing/debugging only for now. Remove and add new counters for vm_stat.*/
4190
4191uint64_t c_seg_filled_no_contention = 0;
4192uint64_t c_seg_filled_contention = 0;
4193clock_sec_t c_seg_filled_contention_sec_max = 0;
4194clock_nsec_t c_seg_filled_contention_nsec_max = 0;
4195
4196static void
4197c_current_seg_filled(c_segment_t c_seg, c_segment_t *current_chead)
4198{
4199 uint32_t unused_bytes;
4200 uint32_t offset_to_depopulate;
4201 int new_state = C_ON_AGE_Q;
4202 clock_sec_t sec;
4203 clock_nsec_t nsec;
4204 bool head_insert = false, wakeup_swapout_thread = false;
4205
4206 unused_bytes = trunc_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset - c_seg->c_nextoffset));
4207
4208 if (unused_bytes) {
4209 offset_to_depopulate = C_SEG_BYTES_TO_OFFSET(round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_nextoffset)));
4210
4211 /*
4212 * release the extra physical page(s) at the end of the segment
4213 */
4214 lck_mtx_unlock_always(&c_seg->c_lock);
4215
4216 kernel_memory_depopulate(
4217 addr: (vm_offset_t) &c_seg->c_store.c_buffer[offset_to_depopulate],
4218 size: unused_bytes,
4219 flags: KMA_COMPRESSOR,
4220 VM_KERN_MEMORY_COMPRESSOR);
4221
4222 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
4223
4224 c_seg->c_populated_offset = offset_to_depopulate;
4225 }
4226 assert(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset) <= c_seg_bufsize);
4227
4228#if DEVELOPMENT || DEBUG
4229 {
4230 boolean_t c_seg_was_busy = FALSE;
4231
4232 if (!c_seg->c_busy) {
4233 C_SEG_BUSY(c_seg);
4234 } else {
4235 c_seg_was_busy = TRUE;
4236 }
4237
4238 lck_mtx_unlock_always(&c_seg->c_lock);
4239
4240 C_SEG_WRITE_PROTECT(c_seg);
4241
4242 lck_mtx_lock_spin_always(&c_seg->c_lock);
4243
4244 if (c_seg_was_busy == FALSE) {
4245 C_SEG_WAKEUP_DONE(c_seg);
4246 }
4247 }
4248#endif
4249
4250#if CONFIG_FREEZE
4251 if (current_chead == (c_segment_t*) &(freezer_context_global.freezer_ctx_chead) &&
4252 VM_CONFIG_SWAP_IS_PRESENT &&
4253 VM_CONFIG_FREEZER_SWAP_IS_ACTIVE
4254#if DEVELOPMENT || DEBUG
4255 && !memorystatus_freeze_to_memory
4256#endif /* DEVELOPMENT || DEBUG */
4257 ) {
4258 new_state = C_ON_SWAPOUT_Q;
4259 wakeup_swapout_thread = true;
4260 }
4261#endif /* CONFIG_FREEZE */
4262
4263 if (vm_darkwake_mode == TRUE) {
4264 new_state = C_ON_SWAPOUT_Q;
4265 head_insert = true;
4266 wakeup_swapout_thread = true;
4267 } else {
4268 c_segment_t *donate_queue_head;
4269 for (int i = 0; i < vm_pageout_state.vm_compressor_thread_count; i++) {
4270#if XNU_TARGET_OS_OSX
4271 donate_queue_head = (c_segment_t*) &(pgo_iothread_internal_state[i].current_early_swapout_chead);
4272#else /* XNU_TARGET_OS_OSX */
4273 donate_queue_head = (c_segment_t*) &(pgo_iothread_internal_state[i].current_late_swapout_chead);
4274#endif /* XNU_TARGET_OS_OSX */
4275
4276 if (current_chead == donate_queue_head) {
4277 assert(c_seg->c_has_donated_pages);
4278 new_state = C_ON_SWAPOUT_Q;
4279 c_seg_total_donated_bytes += c_seg->c_bytes_used;
4280 break;
4281 }
4282 }
4283 }
4284
4285 clock_get_system_nanotime(secs: &sec, nanosecs: &nsec);
4286 c_seg->c_creation_ts = (uint32_t)sec;
4287
4288 if (!lck_mtx_try_lock_spin_always(c_list_lock)) {
4289 clock_sec_t sec2;
4290 clock_nsec_t nsec2;
4291
4292 lck_mtx_lock_spin_always(c_list_lock);
4293 clock_get_system_nanotime(secs: &sec2, nanosecs: &nsec2);
4294 TIME_SUB(sec2, sec, nsec2, nsec, NSEC_PER_SEC);
4295 // printf("FBDP %s: head %p waited for c_list_lock for %lu.%09u seconds\n", __FUNCTION__, current_chead, sec2, nsec2);
4296 if (sec2 > c_seg_filled_contention_sec_max) {
4297 c_seg_filled_contention_sec_max = sec2;
4298 c_seg_filled_contention_nsec_max = nsec2;
4299 } else if (sec2 == c_seg_filled_contention_sec_max &&
4300 nsec2 > c_seg_filled_contention_nsec_max) {
4301 c_seg_filled_contention_nsec_max = nsec2;
4302 }
4303 c_seg_filled_contention++;
4304 } else {
4305 c_seg_filled_no_contention++;
4306 }
4307
4308#if CONFIG_FREEZE
4309 if (current_chead == (c_segment_t*) &(freezer_context_global.freezer_ctx_chead)) {
4310 if (freezer_context_global.freezer_ctx_task->donates_own_pages) {
4311 assert(!c_seg->c_has_donated_pages);
4312 c_seg->c_has_donated_pages = 1;
4313 OSAddAtomic(c_seg->c_slots_used, &c_segment_pages_compressed_incore_late_swapout);
4314 }
4315 c_seg->c_has_freezer_pages = 1;
4316 }
4317#endif /* CONFIG_FREEZE */
4318
4319 c_seg->c_generation_id = c_generation_id++;
4320 c_seg_switch_state(c_seg, new_state, insert_head: head_insert);
4321
4322#if CONFIG_FREEZE
4323 /*
4324 * Donated segments count as frozen to swap if we go through the freezer.
4325 * TODO: What we need is a new ledger and cseg state that can describe
4326 * a frozen cseg from a donated task so we can accurately decrement it on
4327 * swapins.
4328 */
4329 if (current_chead == (c_segment_t*) &(freezer_context_global.freezer_ctx_chead) && (c_seg->c_state == C_ON_SWAPOUT_Q)) {
4330 /*
4331 * darkwake and freezer can't co-exist together
4332 * We'll need to fix this accounting as a start.
4333 * And early donation c_segs are separate from frozen c_segs.
4334 */
4335 assert(vm_darkwake_mode == FALSE);
4336 c_seg_update_task_owner(c_seg, freezer_context_global.freezer_ctx_task);
4337 freezer_context_global.freezer_ctx_swapped_bytes += c_seg->c_bytes_used;
4338 }
4339#endif /* CONFIG_FREEZE */
4340
4341 if (c_seg->c_state == C_ON_AGE_Q && C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) {
4342#if CONFIG_FREEZE
4343 assert(c_seg->c_task_owner == NULL);
4344#endif /* CONFIG_FREEZE */
4345 c_seg_need_delayed_compaction(c_seg, TRUE);
4346 }
4347
4348 lck_mtx_unlock_always(c_list_lock);
4349
4350 if (wakeup_swapout_thread) {
4351 /*
4352 * Darkwake and Freeze configs always
4353 * wake up the swapout thread because
4354 * the compactor thread that normally handles
4355 * it may not be running as much in these
4356 * configs.
4357 */
4358 thread_wakeup((event_t)&vm_swapout_thread);
4359 }
4360
4361 *current_chead = NULL;
4362}
4363
4364/*
4365 * returns with c_seg locked
4366 */
4367void
4368c_seg_swapin_requeue(c_segment_t c_seg, boolean_t has_data, boolean_t minor_compact_ok, boolean_t age_on_swapin_q)
4369{
4370 clock_sec_t sec;
4371 clock_nsec_t nsec;
4372
4373 clock_get_system_nanotime(secs: &sec, nanosecs: &nsec);
4374
4375 lck_mtx_lock_spin_always(c_list_lock);
4376 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
4377
4378 assert(c_seg->c_busy_swapping);
4379 assert(c_seg->c_busy);
4380
4381 c_seg->c_busy_swapping = 0;
4382
4383 if (c_seg->c_overage_swap == TRUE) {
4384 c_overage_swapped_count--;
4385 c_seg->c_overage_swap = FALSE;
4386 }
4387 if (has_data == TRUE) {
4388 if (age_on_swapin_q == TRUE || c_seg->c_has_donated_pages) {
4389#if CONFIG_FREEZE
4390 /*
4391 * If a segment has both identities, frozen and donated bits set, the donated
4392 * bit wins on the swapin path. This is because the segment is being swapped back
4393 * in and so is in demand and should be given more time to spend in memory before
4394 * being swapped back out under pressure.
4395 */
4396 if (c_seg->c_has_donated_pages) {
4397 c_seg->c_has_freezer_pages = 0;
4398 }
4399#endif /* CONFIG_FREEZE */
4400 c_seg_switch_state(c_seg, C_ON_SWAPPEDIN_Q, FALSE);
4401 } else {
4402 c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE);
4403 }
4404
4405 if (minor_compact_ok == TRUE && !c_seg->c_on_minorcompact_q && C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) {
4406 c_seg_need_delayed_compaction(c_seg, TRUE);
4407 }
4408 } else {
4409 c_seg->c_store.c_buffer = (int32_t*) NULL;
4410 c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0);
4411
4412 c_seg_switch_state(c_seg, C_ON_BAD_Q, FALSE);
4413 }
4414 c_seg->c_swappedin_ts = (uint32_t)sec;
4415 c_seg->c_swappedin = true;
4416
4417 lck_mtx_unlock_always(c_list_lock);
4418}
4419
4420
4421
4422/*
4423 * c_seg has to be locked and is returned locked if the c_seg isn't freed
4424 * PAGE_REPLACMENT_DISALLOWED has to be TRUE on entry and is returned TRUE
4425 * c_seg_swapin returns 1 if the c_seg was freed, 0 otherwise
4426 */
4427
4428int
4429c_seg_swapin(c_segment_t c_seg, boolean_t force_minor_compaction, boolean_t age_on_swapin_q)
4430{
4431 vm_offset_t addr = 0;
4432 uint32_t io_size = 0;
4433 uint64_t f_offset;
4434 thread_pri_floor_t token;
4435
4436 assert(C_SEG_IS_ONDISK(c_seg));
4437
4438#if !CHECKSUM_THE_SWAP
4439 c_seg_trim_tail(c_seg);
4440#endif
4441 io_size = round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset));
4442 f_offset = c_seg->c_store.c_swap_handle;
4443
4444 C_SEG_BUSY(c_seg);
4445 c_seg->c_busy_swapping = 1;
4446
4447 /*
4448 * This thread is likely going to block for I/O.
4449 * Make sure it is ready to run when the I/O completes because
4450 * it needs to clear the busy bit on the c_seg so that other
4451 * waiting threads can make progress too.
4452 */
4453 token = thread_priority_floor_start();
4454 lck_mtx_unlock_always(&c_seg->c_lock);
4455
4456 PAGE_REPLACEMENT_DISALLOWED(FALSE);
4457
4458 addr = (vm_offset_t)C_SEG_BUFFER_ADDRESS(c_seg->c_mysegno);
4459 c_seg->c_store.c_buffer = (int32_t*) addr;
4460
4461 kernel_memory_populate(addr, size: io_size, flags: KMA_NOFAIL | KMA_COMPRESSOR,
4462 VM_KERN_MEMORY_COMPRESSOR);
4463
4464 if (vm_swap_get(c_seg, f_offset, io_size) != KERN_SUCCESS) {
4465 PAGE_REPLACEMENT_DISALLOWED(TRUE);
4466
4467 kernel_memory_depopulate(addr, size: io_size, flags: KMA_COMPRESSOR,
4468 VM_KERN_MEMORY_COMPRESSOR);
4469
4470 c_seg_swapin_requeue(c_seg, FALSE, TRUE, age_on_swapin_q);
4471 } else {
4472#if ENCRYPTED_SWAP
4473 vm_swap_decrypt(c_seg);
4474#endif /* ENCRYPTED_SWAP */
4475
4476#if CHECKSUM_THE_SWAP
4477 if (c_seg->cseg_swap_size != io_size) {
4478 panic("swapin size doesn't match swapout size");
4479 }
4480
4481 if (c_seg->cseg_hash != vmc_hash((char*) c_seg->c_store.c_buffer, (int)io_size)) {
4482 panic("c_seg_swapin - Swap hash mismatch");
4483 }
4484#endif /* CHECKSUM_THE_SWAP */
4485
4486 PAGE_REPLACEMENT_DISALLOWED(TRUE);
4487
4488 c_seg_swapin_requeue(c_seg, TRUE, minor_compact_ok: force_minor_compaction == TRUE ? FALSE : TRUE, age_on_swapin_q);
4489
4490#if CONFIG_FREEZE
4491 /*
4492 * c_seg_swapin_requeue() returns with the c_seg lock held.
4493 */
4494 if (!lck_mtx_try_lock_spin_always(c_list_lock)) {
4495 assert(c_seg->c_busy);
4496
4497 lck_mtx_unlock_always(&c_seg->c_lock);
4498 lck_mtx_lock_spin_always(c_list_lock);
4499 lck_mtx_lock_spin_always(&c_seg->c_lock);
4500 }
4501
4502 if (c_seg->c_task_owner) {
4503 c_seg_update_task_owner(c_seg, NULL);
4504 }
4505
4506 lck_mtx_unlock_always(c_list_lock);
4507
4508 OSAddAtomic(c_seg->c_slots_used, &c_segment_pages_compressed_incore);
4509 if (c_seg->c_has_donated_pages) {
4510 OSAddAtomic(c_seg->c_slots_used, &c_segment_pages_compressed_incore_late_swapout);
4511 }
4512#endif /* CONFIG_FREEZE */
4513
4514 OSAddAtomic64(c_seg->c_bytes_used, &compressor_bytes_used);
4515
4516 if (force_minor_compaction == TRUE) {
4517 if (c_seg_minor_compaction_and_unlock(c_seg, FALSE)) {
4518 /*
4519 * c_seg was completely empty so it was freed,
4520 * so be careful not to reference it again
4521 *
4522 * Drop the boost so that the thread priority
4523 * is returned back to where it is supposed to be.
4524 */
4525 thread_priority_floor_end(token: &token);
4526 return 1;
4527 }
4528
4529 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
4530 }
4531 }
4532 C_SEG_WAKEUP_DONE(c_seg);
4533
4534 /*
4535 * Drop the boost so that the thread priority
4536 * is returned back to where it is supposed to be.
4537 */
4538 thread_priority_floor_end(token: &token);
4539
4540 return 0;
4541}
4542
4543
4544static void
4545c_segment_sv_hash_drop_ref(int hash_indx)
4546{
4547 struct c_sv_hash_entry o_sv_he, n_sv_he;
4548
4549 while (1) {
4550 o_sv_he.he_record = c_segment_sv_hash_table[hash_indx].he_record;
4551
4552 n_sv_he.he_ref = o_sv_he.he_ref - 1;
4553 n_sv_he.he_data = o_sv_he.he_data;
4554
4555 if (OSCompareAndSwap64((UInt64)o_sv_he.he_record, (UInt64)n_sv_he.he_record, (UInt64 *) &c_segment_sv_hash_table[hash_indx].he_record) == TRUE) {
4556 if (n_sv_he.he_ref == 0) {
4557 OSAddAtomic(-1, &c_segment_svp_in_hash);
4558 }
4559 break;
4560 }
4561 }
4562}
4563
4564
4565static int
4566c_segment_sv_hash_insert(uint32_t data)
4567{
4568 int hash_sindx;
4569 int misses;
4570 struct c_sv_hash_entry o_sv_he, n_sv_he;
4571 boolean_t got_ref = FALSE;
4572
4573 if (data == 0) {
4574 OSAddAtomic(1, &c_segment_svp_zero_compressions);
4575 } else {
4576 OSAddAtomic(1, &c_segment_svp_nonzero_compressions);
4577 }
4578
4579 hash_sindx = data & C_SV_HASH_MASK;
4580
4581 for (misses = 0; misses < C_SV_HASH_MAX_MISS; misses++) {
4582 o_sv_he.he_record = c_segment_sv_hash_table[hash_sindx].he_record;
4583
4584 while (o_sv_he.he_data == data || o_sv_he.he_ref == 0) {
4585 n_sv_he.he_ref = o_sv_he.he_ref + 1;
4586 n_sv_he.he_data = data;
4587
4588 if (OSCompareAndSwap64((UInt64)o_sv_he.he_record, (UInt64)n_sv_he.he_record, (UInt64 *) &c_segment_sv_hash_table[hash_sindx].he_record) == TRUE) {
4589 if (n_sv_he.he_ref == 1) {
4590 OSAddAtomic(1, &c_segment_svp_in_hash);
4591 }
4592 got_ref = TRUE;
4593 break;
4594 }
4595 o_sv_he.he_record = c_segment_sv_hash_table[hash_sindx].he_record;
4596 }
4597 if (got_ref == TRUE) {
4598 break;
4599 }
4600 hash_sindx++;
4601
4602 if (hash_sindx == C_SV_HASH_SIZE) {
4603 hash_sindx = 0;
4604 }
4605 }
4606 if (got_ref == FALSE) {
4607 return -1;
4608 }
4609
4610 return hash_sindx;
4611}
4612
4613
4614#if RECORD_THE_COMPRESSED_DATA
4615
4616static void
4617c_compressed_record_data(char *src, int c_size)
4618{
4619 if ((c_compressed_record_cptr + c_size + 4) >= c_compressed_record_ebuf) {
4620 panic("c_compressed_record_cptr >= c_compressed_record_ebuf");
4621 }
4622
4623 *(int *)((void *)c_compressed_record_cptr) = c_size;
4624
4625 c_compressed_record_cptr += 4;
4626
4627 memcpy(c_compressed_record_cptr, src, c_size);
4628 c_compressed_record_cptr += c_size;
4629}
4630#endif
4631
4632
4633static int
4634c_compress_page(char *src, c_slot_mapping_t slot_ptr, c_segment_t *current_chead, char *scratch_buf)
4635{
4636 int c_size = -1;
4637 int c_rounded_size = 0;
4638 int max_csize;
4639 c_slot_t cs;
4640 c_segment_t c_seg;
4641 bool single_value = false;
4642
4643 KERNEL_DEBUG(0xe0400000 | DBG_FUNC_START, *current_chead, 0, 0, 0, 0);
4644retry:
4645 if ((c_seg = c_seg_allocate(current_chead)) == NULL) {
4646 return 1;
4647 }
4648 /*
4649 * returns with c_seg lock held
4650 * and PAGE_REPLACEMENT_DISALLOWED(TRUE)...
4651 * c_nextslot has been allocated and
4652 * c_store.c_buffer populated
4653 */
4654 assert(c_seg->c_state == C_IS_FILLING);
4655
4656 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_seg->c_nextslot);
4657
4658 C_SLOT_ASSERT_PACKABLE(slot_ptr);
4659 cs->c_packed_ptr = C_SLOT_PACK_PTR(slot_ptr);
4660
4661 cs->c_offset = c_seg->c_nextoffset;
4662
4663 max_csize = c_seg_bufsize - C_SEG_OFFSET_TO_BYTES((int32_t)cs->c_offset);
4664
4665 if (max_csize > PAGE_SIZE) {
4666 max_csize = PAGE_SIZE;
4667 }
4668
4669#if CHECKSUM_THE_DATA
4670 cs->c_hash_data = vmc_hash(src, PAGE_SIZE);
4671#endif
4672 boolean_t incomp_copy = FALSE;
4673 int max_csize_adj = (max_csize - 4);
4674
4675 if (vm_compressor_algorithm() != VM_COMPRESSOR_DEFAULT_CODEC) {
4676#if defined(__arm64__)
4677 uint16_t ccodec = CINVALID;
4678 uint32_t inline_popcount;
4679 if (max_csize >= C_SEG_OFFSET_ALIGNMENT_BOUNDARY) {
4680 c_size = metacompressor(in: (const uint8_t *) src,
4681 cdst: (uint8_t *) &c_seg->c_store.c_buffer[cs->c_offset],
4682 outbufsz: max_csize_adj, codec: &ccodec,
4683 cscratch: scratch_buf, &incomp_copy, pop_count_p: &inline_popcount);
4684 assert(inline_popcount == C_SLOT_NO_POPCOUNT);
4685
4686#if C_SEG_OFFSET_ALIGNMENT_BOUNDARY > 4
4687 if (c_size > max_csize_adj) {
4688 c_size = -1;
4689 }
4690#endif
4691 } else {
4692 c_size = -1;
4693 }
4694 assert(ccodec == CCWK || ccodec == CCLZ4);
4695 cs->c_codec = ccodec;
4696#endif
4697 } else {
4698#if defined(__arm64__)
4699 cs->c_codec = CCWK;
4700 __unreachable_ok_push
4701 if (PAGE_SIZE == 4096) {
4702 c_size = WKdm_compress_4k(src_buf: (WK_word *)(uintptr_t)src, dest_buf: (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
4703 scratch: (WK_word *)(uintptr_t)scratch_buf, limit: max_csize_adj);
4704 } else {
4705 c_size = WKdm_compress_16k(src_buf: (WK_word *)(uintptr_t)src, dest_buf: (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
4706 scratch: (WK_word *)(uintptr_t)scratch_buf, limit: max_csize_adj);
4707 }
4708 __unreachable_ok_pop
4709#else
4710 c_size = WKdm_compress_new((const WK_word *)(uintptr_t)src, (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
4711 (WK_word *)(uintptr_t)scratch_buf, max_csize_adj);
4712#endif
4713 }
4714 assertf(((c_size <= max_csize_adj) && (c_size >= -1)),
4715 "c_size invalid (%d, %d), cur compressions: %d", c_size, max_csize_adj, c_segment_pages_compressed);
4716
4717 if (c_size == -1) {
4718 if (max_csize < PAGE_SIZE) {
4719 c_current_seg_filled(c_seg, current_chead);
4720 assert(*current_chead == NULL);
4721
4722 lck_mtx_unlock_always(&c_seg->c_lock);
4723 /* TODO: it may be worth requiring codecs to distinguish
4724 * between incompressible inputs and failures due to
4725 * budget exhaustion.
4726 */
4727 PAGE_REPLACEMENT_DISALLOWED(FALSE);
4728 goto retry;
4729 }
4730 c_size = PAGE_SIZE;
4731
4732 if (incomp_copy == FALSE) {
4733 memcpy(dst: &c_seg->c_store.c_buffer[cs->c_offset], src, n: c_size);
4734 }
4735
4736 OSAddAtomic(1, &c_segment_noncompressible_pages);
4737 } else if (c_size == 0) {
4738 int hash_index;
4739
4740 /*
4741 * special case - this is a page completely full of a single 32 bit value
4742 */
4743 single_value = true;
4744 hash_index = c_segment_sv_hash_insert(data: *(uint32_t *)(uintptr_t)src);
4745
4746 if (hash_index != -1) {
4747 slot_ptr->s_cindx = hash_index;
4748 slot_ptr->s_cseg = C_SV_CSEG_ID;
4749#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
4750 slot_ptr->s_uncompressed = 0;
4751#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
4752
4753 OSAddAtomic(1, &c_segment_svp_hash_succeeded);
4754#if RECORD_THE_COMPRESSED_DATA
4755 c_compressed_record_data(src, 4);
4756#endif
4757 goto sv_compression;
4758 }
4759 c_size = 4;
4760
4761 memcpy(dst: &c_seg->c_store.c_buffer[cs->c_offset], src, n: c_size);
4762
4763 OSAddAtomic(1, &c_segment_svp_hash_failed);
4764 }
4765
4766#if RECORD_THE_COMPRESSED_DATA
4767 c_compressed_record_data((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size);
4768#endif
4769#if CHECKSUM_THE_COMPRESSED_DATA
4770 cs->c_hash_compressed_data = vmc_hash((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size);
4771#endif
4772#if POPCOUNT_THE_COMPRESSED_DATA
4773 cs->c_pop_cdata = vmc_pop((uintptr_t) &c_seg->c_store.c_buffer[cs->c_offset], c_size);
4774#endif
4775 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
4776
4777 PACK_C_SIZE(cs, c_size);
4778 c_seg->c_bytes_used += c_rounded_size;
4779 c_seg->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
4780 c_seg->c_slots_used++;
4781
4782#if CONFIG_FREEZE
4783 /* TODO: should c_segment_pages_compressed be up here too? See 88598046 for details */
4784 OSAddAtomic(1, &c_segment_pages_compressed_incore);
4785 if (c_seg->c_has_donated_pages) {
4786 OSAddAtomic(1, &c_segment_pages_compressed_incore_late_swapout);
4787 }
4788#endif /* CONFIG_FREEZE */
4789
4790 slot_ptr->s_cindx = c_seg->c_nextslot++;
4791 /* <csegno=0,indx=0> would mean "empty slot", so use csegno+1 */
4792 slot_ptr->s_cseg = c_seg->c_mysegno + 1;
4793
4794#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
4795 slot_ptr->s_uncompressed = 0;
4796#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
4797
4798sv_compression:
4799 if (c_seg->c_nextoffset >= c_seg_off_limit || c_seg->c_nextslot >= C_SLOT_MAX_INDEX) {
4800 c_current_seg_filled(c_seg, current_chead);
4801 assert(*current_chead == NULL);
4802 }
4803
4804 lck_mtx_unlock_always(&c_seg->c_lock);
4805
4806 PAGE_REPLACEMENT_DISALLOWED(FALSE);
4807
4808#if RECORD_THE_COMPRESSED_DATA
4809 if ((c_compressed_record_cptr - c_compressed_record_sbuf) >= c_seg_allocsize) {
4810 c_compressed_record_write(c_compressed_record_sbuf, (int)(c_compressed_record_cptr - c_compressed_record_sbuf));
4811 c_compressed_record_cptr = c_compressed_record_sbuf;
4812 }
4813#endif
4814 if (c_size) {
4815 OSAddAtomic64(c_size, &c_segment_compressed_bytes);
4816 OSAddAtomic64(c_rounded_size, &compressor_bytes_used);
4817 }
4818 OSAddAtomic64(PAGE_SIZE, &c_segment_input_bytes);
4819
4820 OSAddAtomic(1, &c_segment_pages_compressed);
4821#if DEVELOPMENT || DEBUG
4822 if (!compressor_running_perf_test) {
4823 /*
4824 * The perf_compressor benchmark should not be able to trigger
4825 * compressor thrashing jetsams.
4826 */
4827 OSAddAtomic(1, &sample_period_compression_count);
4828 }
4829#else /* DEVELOPMENT || DEBUG */
4830 OSAddAtomic(1, &sample_period_compression_count);
4831#endif /* DEVELOPMENT || DEBUG */
4832
4833 KERNEL_DEBUG(0xe0400000 | DBG_FUNC_END, *current_chead, c_size, c_segment_input_bytes, c_segment_compressed_bytes, 0);
4834
4835 return 0;
4836}
4837
4838static inline void
4839sv_decompress(int32_t *ddst, int32_t pattern)
4840{
4841// assert(__builtin_constant_p(PAGE_SIZE) != 0);
4842#if defined(__x86_64__)
4843 memset_word(ddst, pattern, PAGE_SIZE / sizeof(int32_t));
4844#elif defined(__arm64__)
4845 assert((PAGE_SIZE % 128) == 0);
4846 if (pattern == 0) {
4847 fill32_dczva((addr64_t)ddst, PAGE_SIZE);
4848 } else {
4849 fill32_nt((addr64_t)ddst, PAGE_SIZE, pattern);
4850 }
4851#else
4852 size_t i;
4853
4854 /* Unroll the pattern fill loop 4x to encourage the
4855 * compiler to emit NEON stores, cf.
4856 * <rdar://problem/25839866> Loop autovectorization
4857 * anomalies.
4858 */
4859 /* * We use separate loops for each PAGE_SIZE
4860 * to allow the autovectorizer to engage, as PAGE_SIZE
4861 * may not be a constant.
4862 */
4863
4864 __unreachable_ok_push
4865 if (PAGE_SIZE == 4096) {
4866 for (i = 0; i < (4096U / sizeof(int32_t)); i += 4) {
4867 *ddst++ = pattern;
4868 *ddst++ = pattern;
4869 *ddst++ = pattern;
4870 *ddst++ = pattern;
4871 }
4872 } else {
4873 assert(PAGE_SIZE == 16384);
4874 for (i = 0; i < (int)(16384U / sizeof(int32_t)); i += 4) {
4875 *ddst++ = pattern;
4876 *ddst++ = pattern;
4877 *ddst++ = pattern;
4878 *ddst++ = pattern;
4879 }
4880 }
4881 __unreachable_ok_pop
4882#endif
4883}
4884
4885static int
4886c_decompress_page(char *dst, volatile c_slot_mapping_t slot_ptr, vm_compressor_options_t flags, int *zeroslot)
4887{
4888 c_slot_t cs;
4889 c_segment_t c_seg;
4890 uint32_t c_segno;
4891 uint16_t c_indx;
4892 int c_rounded_size;
4893 uint32_t c_size;
4894 int retval = 0;
4895 boolean_t need_unlock = TRUE;
4896 boolean_t consider_defragmenting = FALSE;
4897 boolean_t kdp_mode = FALSE;
4898
4899 if (__improbable(flags & C_KDP)) {
4900 if (not_in_kdp) {
4901 panic("C_KDP passed to decompress page from outside of debugger context");
4902 }
4903
4904 assert((flags & C_KEEP) == C_KEEP);
4905 assert((flags & C_DONT_BLOCK) == C_DONT_BLOCK);
4906
4907 if ((flags & (C_DONT_BLOCK | C_KEEP)) != (C_DONT_BLOCK | C_KEEP)) {
4908 return -2;
4909 }
4910
4911 kdp_mode = TRUE;
4912 *zeroslot = 0;
4913 }
4914
4915ReTry:
4916 if (__probable(!kdp_mode)) {
4917 PAGE_REPLACEMENT_DISALLOWED(TRUE);
4918 } else {
4919 if (kdp_lck_rw_lock_is_acquired_exclusive(lck: &c_master_lock)) {
4920 return -2;
4921 }
4922 }
4923
4924#if HIBERNATION
4925 /*
4926 * if hibernation is enabled, it indicates (via a call
4927 * to 'vm_decompressor_lock' that no further
4928 * decompressions are allowed once it reaches
4929 * the point of flushing all of the currently dirty
4930 * anonymous memory through the compressor and out
4931 * to disk... in this state we allow freeing of compressed
4932 * pages and must honor the C_DONT_BLOCK case
4933 */
4934 if (__improbable(dst && decompressions_blocked == TRUE)) {
4935 if (flags & C_DONT_BLOCK) {
4936 if (__probable(!kdp_mode)) {
4937 PAGE_REPLACEMENT_DISALLOWED(FALSE);
4938 }
4939
4940 *zeroslot = 0;
4941 return -2;
4942 }
4943 /*
4944 * it's safe to atomically assert and block behind the
4945 * lock held in shared mode because "decompressions_blocked" is
4946 * only set and cleared and the thread_wakeup done when the lock
4947 * is held exclusively
4948 */
4949 assert_wait((event_t)&decompressions_blocked, THREAD_UNINT);
4950
4951 PAGE_REPLACEMENT_DISALLOWED(FALSE);
4952
4953 thread_block(THREAD_CONTINUE_NULL);
4954
4955 goto ReTry;
4956 }
4957#endif
4958 /* s_cseg is actually "segno+1" */
4959 c_segno = slot_ptr->s_cseg - 1;
4960
4961 if (__improbable(c_segno >= c_segments_available)) {
4962 panic("c_decompress_page: c_segno %d >= c_segments_available %d, slot_ptr(%p), slot_data(%x)",
4963 c_segno, c_segments_available, slot_ptr, *(int *)((void *)slot_ptr));
4964 }
4965
4966 if (__improbable(c_segments[c_segno].c_segno < c_segments_available)) {
4967 panic("c_decompress_page: c_segno %d is free, slot_ptr(%p), slot_data(%x)",
4968 c_segno, slot_ptr, *(int *)((void *)slot_ptr));
4969 }
4970
4971 c_seg = c_segments[c_segno].c_seg;
4972
4973 if (__probable(!kdp_mode)) {
4974 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
4975 } else {
4976 if (kdp_lck_mtx_lock_spin_is_acquired(lck: &c_seg->c_lock)) {
4977 return -2;
4978 }
4979 }
4980
4981 assert(c_seg->c_state != C_IS_EMPTY && c_seg->c_state != C_IS_FREE);
4982
4983 if (dst == NULL && c_seg->c_busy_swapping) {
4984 assert(c_seg->c_busy);
4985
4986 goto bypass_busy_check;
4987 }
4988 if (flags & C_DONT_BLOCK) {
4989 if (c_seg->c_busy || (C_SEG_IS_ONDISK(c_seg) && dst)) {
4990 *zeroslot = 0;
4991
4992 retval = -2;
4993 goto done;
4994 }
4995 }
4996 if (c_seg->c_busy) {
4997 PAGE_REPLACEMENT_DISALLOWED(FALSE);
4998
4999 c_seg_wait_on_busy(c_seg);
5000
5001 goto ReTry;
5002 }
5003bypass_busy_check:
5004
5005 c_indx = slot_ptr->s_cindx;
5006
5007 if (__improbable(c_indx >= c_seg->c_nextslot)) {
5008 panic("c_decompress_page: c_indx %d >= c_nextslot %d, c_seg(%p), slot_ptr(%p), slot_data(%x)",
5009 c_indx, c_seg->c_nextslot, c_seg, slot_ptr, *(int *)((void *)slot_ptr));
5010 }
5011
5012 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
5013
5014 c_size = UNPACK_C_SIZE(cs);
5015
5016 if (__improbable(c_size == 0)) {
5017 panic("c_decompress_page: c_size == 0, c_seg(%p), slot_ptr(%p), slot_data(%x)",
5018 c_seg, slot_ptr, *(int *)((void *)slot_ptr));
5019 }
5020
5021 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
5022
5023 if (dst) {
5024 uint32_t age_of_cseg;
5025 clock_sec_t cur_ts_sec;
5026 clock_nsec_t cur_ts_nsec;
5027
5028 if (C_SEG_IS_ONDISK(c_seg)) {
5029#if CONFIG_FREEZE
5030 if (freezer_incore_cseg_acct) {
5031 if ((c_seg->c_slots_used + c_segment_pages_compressed_incore) >= c_segment_pages_compressed_nearing_limit) {
5032 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5033 lck_mtx_unlock_always(&c_seg->c_lock);
5034
5035 memorystatus_kill_on_VM_compressor_space_shortage(FALSE /* async */);
5036
5037 goto ReTry;
5038 }
5039
5040 uint32_t incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count;
5041 if ((incore_seg_count + 1) >= c_segments_nearing_limit) {
5042 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5043 lck_mtx_unlock_always(&c_seg->c_lock);
5044
5045 memorystatus_kill_on_VM_compressor_space_shortage(FALSE /* async */);
5046
5047 goto ReTry;
5048 }
5049 }
5050#endif /* CONFIG_FREEZE */
5051 assert(kdp_mode == FALSE);
5052 retval = c_seg_swapin(c_seg, FALSE, TRUE);
5053 assert(retval == 0);
5054
5055 retval = 1;
5056 }
5057 if (c_seg->c_state == C_ON_BAD_Q) {
5058 assert(c_seg->c_store.c_buffer == NULL);
5059 *zeroslot = 0;
5060
5061 retval = -1;
5062 goto done;
5063 }
5064
5065#if POPCOUNT_THE_COMPRESSED_DATA
5066 unsigned csvpop;
5067 uintptr_t csvaddr = (uintptr_t) &c_seg->c_store.c_buffer[cs->c_offset];
5068 if (cs->c_pop_cdata != (csvpop = vmc_pop(csvaddr, c_size))) {
5069 panic("Compressed data popcount doesn't match original, bit distance: %d %p (phys: %p) %p %p 0x%x 0x%x 0x%x 0x%x", (csvpop - cs->c_pop_cdata), (void *)csvaddr, (void *) kvtophys(csvaddr), c_seg, cs, cs->c_offset, c_size, csvpop, cs->c_pop_cdata);
5070 }
5071#endif
5072
5073#if CHECKSUM_THE_COMPRESSED_DATA
5074 unsigned csvhash;
5075 if (cs->c_hash_compressed_data != (csvhash = vmc_hash((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size))) {
5076 panic("Compressed data doesn't match original %p %p %u %u %u", c_seg, cs, c_size, cs->c_hash_compressed_data, csvhash);
5077 }
5078#endif
5079 if (c_rounded_size == PAGE_SIZE) {
5080 /*
5081 * page wasn't compressible... just copy it out
5082 */
5083 memcpy(dst, src: &c_seg->c_store.c_buffer[cs->c_offset], PAGE_SIZE);
5084 } else if (c_size == 4) {
5085 int32_t data;
5086 int32_t *dptr;
5087
5088 /*
5089 * page was populated with a single value
5090 * that didn't fit into our fast hash
5091 * so we packed it in as a single non-compressed value
5092 * that we need to populate the page with
5093 */
5094 dptr = (int32_t *)(uintptr_t)dst;
5095 data = *(int32_t *)(&c_seg->c_store.c_buffer[cs->c_offset]);
5096 sv_decompress(ddst: dptr, pattern: data);
5097 } else {
5098 uint32_t my_cpu_no;
5099 char *scratch_buf;
5100
5101 if (__probable(!kdp_mode)) {
5102 /*
5103 * we're behind the c_seg lock held in spin mode
5104 * which means pre-emption is disabled... therefore
5105 * the following sequence is atomic and safe
5106 */
5107 my_cpu_no = cpu_number();
5108
5109 assert(my_cpu_no < compressor_cpus);
5110
5111 scratch_buf = &compressor_scratch_bufs[my_cpu_no * vm_compressor_get_decode_scratch_size()];
5112 } else {
5113 scratch_buf = kdp_compressor_scratch_buf;
5114 }
5115
5116 if (vm_compressor_algorithm() != VM_COMPRESSOR_DEFAULT_CODEC) {
5117#if defined(__arm64__)
5118 uint16_t c_codec = cs->c_codec;
5119 uint32_t inline_popcount;
5120 if (!metadecompressor(source: (const uint8_t *) &c_seg->c_store.c_buffer[cs->c_offset],
5121 dest: (uint8_t *)dst, csize: c_size, ccodec: c_codec, compressor_dscratch: (void *)scratch_buf, pop_count_p: &inline_popcount)) {
5122 retval = -1;
5123 } else {
5124 assert(inline_popcount == C_SLOT_NO_POPCOUNT);
5125 }
5126#endif
5127 } else {
5128#if defined(__arm64__)
5129 __unreachable_ok_push
5130 if (PAGE_SIZE == 4096) {
5131 WKdm_decompress_4k(src_buf: (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
5132 dest_buf: (WK_word *)(uintptr_t)dst, scratch: (WK_word *)(uintptr_t)scratch_buf, bytes: c_size);
5133 } else {
5134 WKdm_decompress_16k(src_buf: (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
5135 dest_buf: (WK_word *)(uintptr_t)dst, scratch: (WK_word *)(uintptr_t)scratch_buf, bytes: c_size);
5136 }
5137 __unreachable_ok_pop
5138#else
5139 WKdm_decompress_new((WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
5140 (WK_word *)(uintptr_t)dst, (WK_word *)(uintptr_t)scratch_buf, c_size);
5141#endif
5142 }
5143 }
5144
5145#if CHECKSUM_THE_DATA
5146 if (cs->c_hash_data != vmc_hash(dst, PAGE_SIZE)) {
5147#if defined(__arm64__)
5148 int32_t *dinput = &c_seg->c_store.c_buffer[cs->c_offset];
5149 panic("decompressed data doesn't match original cs: %p, hash: 0x%x, offset: %d, c_size: %d, c_rounded_size: %d, codec: %d, header: 0x%x 0x%x 0x%x", cs, cs->c_hash_data, cs->c_offset, c_size, c_rounded_size, cs->c_codec, *dinput, *(dinput + 1), *(dinput + 2));
5150#else
5151 panic("decompressed data doesn't match original cs: %p, hash: %d, offset: 0x%x, c_size: %d", cs, cs->c_hash_data, cs->c_offset, c_size);
5152#endif
5153 }
5154#endif
5155 if (c_seg->c_swappedin_ts == 0 && !kdp_mode) {
5156 clock_get_system_nanotime(secs: &cur_ts_sec, nanosecs: &cur_ts_nsec);
5157
5158 age_of_cseg = (uint32_t)cur_ts_sec - c_seg->c_creation_ts;
5159 if (age_of_cseg < DECOMPRESSION_SAMPLE_MAX_AGE) {
5160 OSAddAtomic(1, &age_of_decompressions_during_sample_period[age_of_cseg]);
5161 } else {
5162 OSAddAtomic(1, &overage_decompressions_during_sample_period);
5163 }
5164
5165 OSAddAtomic(1, &sample_period_decompression_count);
5166 }
5167 }
5168#if CONFIG_FREEZE
5169 else {
5170 /*
5171 * We are freeing an uncompressed page from this c_seg and so balance the ledgers.
5172 */
5173 if (C_SEG_IS_ONDISK(c_seg)) {
5174 /*
5175 * The compression sweep feature will push out anonymous pages to disk
5176 * without going through the freezer path and so those c_segs, while
5177 * swapped out, won't have an owner.
5178 */
5179 if (c_seg->c_task_owner) {
5180 task_update_frozen_to_swap_acct(c_seg->c_task_owner, PAGE_SIZE_64, DEBIT_FROM_SWAP);
5181 }
5182
5183 /*
5184 * We are freeing a page in swap without swapping it in. We bump the in-core
5185 * count here to simulate a swapin of a page so that we can accurately
5186 * decrement it below.
5187 */
5188 OSAddAtomic(1, &c_segment_pages_compressed_incore);
5189 if (c_seg->c_has_donated_pages) {
5190 OSAddAtomic(1, &c_segment_pages_compressed_incore_late_swapout);
5191 }
5192 } else if (c_seg->c_state == C_ON_BAD_Q) {
5193 assert(c_seg->c_store.c_buffer == NULL);
5194 *zeroslot = 0;
5195
5196 retval = -1;
5197 goto done;
5198 }
5199 }
5200#endif /* CONFIG_FREEZE */
5201
5202 if (flags & C_KEEP) {
5203 *zeroslot = 0;
5204 goto done;
5205 }
5206 assert(kdp_mode == FALSE);
5207
5208 c_seg->c_bytes_unused += c_rounded_size;
5209 c_seg->c_bytes_used -= c_rounded_size;
5210
5211 assert(c_seg->c_slots_used);
5212 c_seg->c_slots_used--;
5213 if (dst && c_seg->c_swappedin) {
5214 task_t task = current_task();
5215 if (task) {
5216 ledger_credit(ledger: task->ledger, entry: task_ledgers.swapins, PAGE_SIZE);
5217 }
5218 }
5219
5220 PACK_C_SIZE(cs, 0);
5221
5222 if (c_indx < c_seg->c_firstemptyslot) {
5223 c_seg->c_firstemptyslot = c_indx;
5224 }
5225
5226 OSAddAtomic(-1, &c_segment_pages_compressed);
5227#if CONFIG_FREEZE
5228 OSAddAtomic(-1, &c_segment_pages_compressed_incore);
5229 assertf(c_segment_pages_compressed_incore >= 0, "-ve incore count %p 0x%x", c_seg, c_segment_pages_compressed_incore);
5230 if (c_seg->c_has_donated_pages) {
5231 OSAddAtomic(-1, &c_segment_pages_compressed_incore_late_swapout);
5232 assertf(c_segment_pages_compressed_incore_late_swapout >= 0, "-ve lateswapout count %p 0x%x", c_seg, c_segment_pages_compressed_incore_late_swapout);
5233 }
5234#endif /* CONFIG_FREEZE */
5235
5236 if (c_seg->c_state != C_ON_BAD_Q && !(C_SEG_IS_ONDISK(c_seg))) {
5237 /*
5238 * C_SEG_IS_ONDISK == TRUE can occur when we're doing a
5239 * free of a compressed page (i.e. dst == NULL)
5240 */
5241 OSAddAtomic64(-c_rounded_size, &compressor_bytes_used);
5242 }
5243 if (c_seg->c_busy_swapping) {
5244 /*
5245 * bypass case for c_busy_swapping...
5246 * let the swapin/swapout paths deal with putting
5247 * the c_seg on the minor compaction queue if needed
5248 */
5249 assert(c_seg->c_busy);
5250 goto done;
5251 }
5252 assert(!c_seg->c_busy);
5253
5254 if (c_seg->c_state != C_IS_FILLING) {
5255 if (c_seg->c_bytes_used == 0) {
5256 if (!(C_SEG_IS_ONDISK(c_seg))) {
5257 int pages_populated;
5258
5259 pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE;
5260 c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0);
5261
5262 if (pages_populated) {
5263 assert(c_seg->c_state != C_ON_BAD_Q);
5264 assert(c_seg->c_store.c_buffer != NULL);
5265
5266 C_SEG_BUSY(c_seg);
5267 lck_mtx_unlock_always(&c_seg->c_lock);
5268
5269 kernel_memory_depopulate(
5270 addr: (vm_offset_t) c_seg->c_store.c_buffer,
5271 ptoa(pages_populated),
5272 flags: KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR);
5273
5274 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
5275 C_SEG_WAKEUP_DONE(c_seg);
5276 }
5277 if (!c_seg->c_on_minorcompact_q && c_seg->c_state != C_ON_SWAPIO_Q) {
5278 if (c_seg->c_state == C_ON_SWAPOUT_Q) {
5279 bool clear_busy = false;
5280 if (!lck_mtx_try_lock_spin_always(c_list_lock)) {
5281 C_SEG_BUSY(c_seg);
5282
5283 lck_mtx_unlock_always(&c_seg->c_lock);
5284 lck_mtx_lock_spin_always(c_list_lock);
5285 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
5286 clear_busy = true;
5287 }
5288 c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE);
5289 if (clear_busy) {
5290 C_SEG_WAKEUP_DONE(c_seg);
5291 clear_busy = false;
5292 }
5293 lck_mtx_unlock_always(c_list_lock);
5294 }
5295 c_seg_need_delayed_compaction(c_seg, FALSE);
5296 }
5297 } else {
5298 if (c_seg->c_state != C_ON_SWAPPEDOUTSPARSE_Q) {
5299 c_seg_move_to_sparse_list(c_seg);
5300 consider_defragmenting = TRUE;
5301 }
5302 }
5303 } else if (c_seg->c_on_minorcompact_q) {
5304 assert(c_seg->c_state != C_ON_BAD_Q);
5305 assert(!C_SEG_IS_ON_DISK_OR_SOQ(c_seg));
5306
5307 if (C_SEG_SHOULD_MINORCOMPACT_NOW(c_seg)) {
5308 c_seg_try_minor_compaction_and_unlock(c_seg);
5309 need_unlock = FALSE;
5310 }
5311 } else if (!(C_SEG_IS_ONDISK(c_seg))) {
5312 if (c_seg->c_state != C_ON_BAD_Q && c_seg->c_state != C_ON_SWAPOUT_Q && c_seg->c_state != C_ON_SWAPIO_Q &&
5313 C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) {
5314 c_seg_need_delayed_compaction(c_seg, FALSE);
5315 }
5316 } else if (c_seg->c_state != C_ON_SWAPPEDOUTSPARSE_Q && C_SEG_ONDISK_IS_SPARSE(c_seg)) {
5317 c_seg_move_to_sparse_list(c_seg);
5318 consider_defragmenting = TRUE;
5319 }
5320 }
5321done:
5322 if (__improbable(kdp_mode)) {
5323 return retval;
5324 }
5325
5326 if (need_unlock == TRUE) {
5327 lck_mtx_unlock_always(&c_seg->c_lock);
5328 }
5329
5330 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5331
5332 if (consider_defragmenting == TRUE) {
5333 vm_swap_consider_defragmenting(VM_SWAP_FLAGS_NONE);
5334 }
5335
5336#if !XNU_TARGET_OS_OSX
5337 if ((c_minor_count && COMPRESSOR_NEEDS_TO_MINOR_COMPACT()) || vm_compressor_needs_to_major_compact()) {
5338 vm_wake_compactor_swapper();
5339 }
5340#endif /* !XNU_TARGET_OS_OSX */
5341
5342 return retval;
5343}
5344
5345
5346inline bool
5347vm_compressor_is_slot_compressed(int *slot)
5348{
5349#if !CONFIG_TRACK_UNMODIFIED_ANON_PAGES
5350#pragma unused(slot)
5351 return true;
5352#else /* !CONFIG_TRACK_UNMODIFIED_ANON_PAGES*/
5353 c_slot_mapping_t slot_ptr = (c_slot_mapping_t)slot;
5354 return !slot_ptr->s_uncompressed;
5355#endif /* !CONFIG_TRACK_UNMODIFIED_ANON_PAGES*/
5356}
5357
5358int
5359vm_compressor_get(ppnum_t pn, int *slot, vm_compressor_options_t flags)
5360{
5361 c_slot_mapping_t slot_ptr;
5362 char *dst;
5363 int zeroslot = 1;
5364 int retval;
5365
5366#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
5367 if (flags & C_PAGE_UNMODIFIED) {
5368 retval = vm_uncompressed_get(pn, slot, flags | C_KEEP);
5369 if (retval == 0) {
5370 os_atomic_inc(&compressor_ro_uncompressed_get, relaxed);
5371 }
5372
5373 return retval;
5374 }
5375#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
5376
5377 dst = pmap_map_compressor_page(pn);
5378 slot_ptr = (c_slot_mapping_t)slot;
5379
5380 assert(dst != NULL);
5381
5382 if (slot_ptr->s_cseg == C_SV_CSEG_ID) {
5383 int32_t data;
5384 int32_t *dptr;
5385
5386 /*
5387 * page was populated with a single value
5388 * that found a home in our hash table
5389 * grab that value from the hash and populate the page
5390 * that we need to populate the page with
5391 */
5392 dptr = (int32_t *)(uintptr_t)dst;
5393 data = c_segment_sv_hash_table[slot_ptr->s_cindx].he_data;
5394 sv_decompress(ddst: dptr, pattern: data);
5395 if (!(flags & C_KEEP)) {
5396 c_segment_sv_hash_drop_ref(hash_indx: slot_ptr->s_cindx);
5397
5398 OSAddAtomic(-1, &c_segment_pages_compressed);
5399 *slot = 0;
5400 }
5401 if (data) {
5402 OSAddAtomic(1, &c_segment_svp_nonzero_decompressions);
5403 } else {
5404 OSAddAtomic(1, &c_segment_svp_zero_decompressions);
5405 }
5406
5407 pmap_unmap_compressor_page(pn, dst);
5408 return 0;
5409 }
5410
5411 retval = c_decompress_page(dst, slot_ptr, flags, zeroslot: &zeroslot);
5412
5413 /*
5414 * zeroslot will be set to 0 by c_decompress_page if (flags & C_KEEP)
5415 * or (flags & C_DONT_BLOCK) and we found 'c_busy' or 'C_SEG_IS_ONDISK' to be TRUE
5416 */
5417 if (zeroslot) {
5418 *slot = 0;
5419 }
5420
5421 pmap_unmap_compressor_page(pn, dst);
5422
5423 /*
5424 * returns 0 if we successfully decompressed a page from a segment already in memory
5425 * returns 1 if we had to first swap in the segment, before successfully decompressing the page
5426 * returns -1 if we encountered an error swapping in the segment - decompression failed
5427 * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' or 'C_SEG_IS_ONDISK' to be true
5428 */
5429 return retval;
5430}
5431
5432int
5433vm_compressor_free(int *slot, vm_compressor_options_t flags)
5434{
5435 bool slot_is_compressed = vm_compressor_is_slot_compressed(slot);
5436
5437 if (slot_is_compressed) {
5438 c_slot_mapping_t slot_ptr;
5439 int zeroslot = 1;
5440 int retval = 0;
5441
5442 assert(flags == 0 || flags == C_DONT_BLOCK);
5443
5444 slot_ptr = (c_slot_mapping_t)slot;
5445
5446 if (slot_ptr->s_cseg == C_SV_CSEG_ID) {
5447 c_segment_sv_hash_drop_ref(hash_indx: slot_ptr->s_cindx);
5448 OSAddAtomic(-1, &c_segment_pages_compressed);
5449
5450 *slot = 0;
5451 return 0;
5452 }
5453 retval = c_decompress_page(NULL, slot_ptr, flags, zeroslot: &zeroslot);
5454 /*
5455 * returns 0 if we successfully freed the specified compressed page
5456 * returns -1 if we encountered an error swapping in the segment - decompression failed
5457 * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' set
5458 */
5459
5460 if (retval == 0) {
5461 *slot = 0;
5462 }
5463
5464 return retval;
5465 }
5466#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
5467 else {
5468 if ((flags & C_PAGE_UNMODIFIED) == 0) {
5469 /* moving from uncompressed state to compressed. Free it.*/
5470 vm_uncompressed_free(slot, 0);
5471 assert(*slot == 0);
5472 }
5473 }
5474#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
5475 return KERN_SUCCESS;
5476}
5477
5478int
5479vm_compressor_put(ppnum_t pn, int *slot, void **current_chead, char *scratch_buf, bool unmodified)
5480{
5481 char *src;
5482 int retval = 0;
5483
5484#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
5485 if (unmodified) {
5486 if (*slot) {
5487 os_atomic_inc(&compressor_ro_uncompressed_skip_returned, relaxed);
5488 return retval;
5489 } else {
5490 retval = vm_uncompressed_put(pn, slot);
5491 if (retval == KERN_SUCCESS) {
5492 os_atomic_inc(&compressor_ro_uncompressed_put, relaxed);
5493 return retval;
5494 }
5495 }
5496 }
5497#else /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
5498#pragma unused(unmodified)
5499#endif /* CONFIG_TRACK_UNMODIFIED_ANON_PAGES */
5500
5501 src = pmap_map_compressor_page(pn);
5502 assert(src != NULL);
5503
5504 retval = c_compress_page(src, slot_ptr: (c_slot_mapping_t)slot, current_chead: (c_segment_t *)current_chead, scratch_buf);
5505 pmap_unmap_compressor_page(pn, src);
5506
5507 return retval;
5508}
5509
5510void
5511vm_compressor_transfer(
5512 int *dst_slot_p,
5513 int *src_slot_p)
5514{
5515 c_slot_mapping_t dst_slot, src_slot;
5516 c_segment_t c_seg;
5517 uint16_t c_indx;
5518 c_slot_t cs;
5519
5520 src_slot = (c_slot_mapping_t) src_slot_p;
5521
5522 if (src_slot->s_cseg == C_SV_CSEG_ID || !vm_compressor_is_slot_compressed(slot: src_slot_p)) {
5523 *dst_slot_p = *src_slot_p;
5524 *src_slot_p = 0;
5525 return;
5526 }
5527 dst_slot = (c_slot_mapping_t) dst_slot_p;
5528Retry:
5529 PAGE_REPLACEMENT_DISALLOWED(TRUE);
5530 /* get segment for src_slot */
5531 c_seg = c_segments[src_slot->s_cseg - 1].c_seg;
5532 /* lock segment */
5533 lck_mtx_lock_spin_always(lck: &c_seg->c_lock);
5534 /* wait if it's busy */
5535 if (c_seg->c_busy && !c_seg->c_busy_swapping) {
5536 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5537 c_seg_wait_on_busy(c_seg);
5538 goto Retry;
5539 }
5540 /* find the c_slot */
5541 c_indx = src_slot->s_cindx;
5542 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
5543 /* point the c_slot back to dst_slot instead of src_slot */
5544 C_SLOT_ASSERT_PACKABLE(dst_slot);
5545 cs->c_packed_ptr = C_SLOT_PACK_PTR(dst_slot);
5546 /* transfer */
5547 *dst_slot_p = *src_slot_p;
5548 *src_slot_p = 0;
5549 lck_mtx_unlock_always(&c_seg->c_lock);
5550 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5551}
5552
5553#if defined(__arm64__)
5554extern clock_sec_t vm_swapfile_last_failed_to_create_ts;
5555__attribute__((noreturn))
5556void
5557vm_panic_hibernate_write_image_failed(int err)
5558{
5559 panic("hibernate_write_image encountered error 0x%x - %u, %u, %d, %d, %d, %d, %d, %d, %d, %d, %llu, %d, %d, %d\n",
5560 err,
5561 VM_PAGE_COMPRESSOR_COUNT, vm_page_wire_count,
5562 c_age_count, c_major_count, c_minor_count, (c_early_swapout_count + c_regular_swapout_count + c_late_swapout_count), c_swappedout_sparse_count,
5563 vm_num_swap_files, vm_num_pinned_swap_files, vm_swappin_enabled, vm_swap_put_failures,
5564 (vm_swapfile_last_failed_to_create_ts ? 1:0), hibernate_no_swapspace, hibernate_flush_timed_out);
5565}
5566#endif /*(__arm64__)*/
5567
5568#if CONFIG_FREEZE
5569
5570int freezer_finished_filling = 0;
5571
5572void
5573vm_compressor_finished_filling(
5574 void **current_chead)
5575{
5576 c_segment_t c_seg;
5577
5578 if ((c_seg = *(c_segment_t *)current_chead) == NULL) {
5579 return;
5580 }
5581
5582 assert(c_seg->c_state == C_IS_FILLING);
5583
5584 lck_mtx_lock_spin_always(&c_seg->c_lock);
5585
5586 c_current_seg_filled(c_seg, (c_segment_t *)current_chead);
5587
5588 lck_mtx_unlock_always(&c_seg->c_lock);
5589
5590 freezer_finished_filling++;
5591}
5592
5593
5594/*
5595 * This routine is used to transfer the compressed chunks from
5596 * the c_seg/cindx pointed to by slot_p into a new c_seg headed
5597 * by the current_chead and a new cindx within that c_seg.
5598 *
5599 * Currently, this routine is only used by the "freezer backed by
5600 * compressor with swap" mode to create a series of c_segs that
5601 * only contain compressed data belonging to one task. So, we
5602 * move a task's previously compressed data into a set of new
5603 * c_segs which will also hold the task's yet to be compressed data.
5604 */
5605
5606kern_return_t
5607vm_compressor_relocate(
5608 void **current_chead,
5609 int *slot_p)
5610{
5611 c_slot_mapping_t slot_ptr;
5612 c_slot_mapping_t src_slot;
5613 uint32_t c_rounded_size;
5614 uint32_t c_size;
5615 uint16_t dst_slot;
5616 c_slot_t c_dst;
5617 c_slot_t c_src;
5618 uint16_t c_indx;
5619 c_segment_t c_seg_dst = NULL;
5620 c_segment_t c_seg_src = NULL;
5621 kern_return_t kr = KERN_SUCCESS;
5622
5623
5624 src_slot = (c_slot_mapping_t) slot_p;
5625
5626 if (src_slot->s_cseg == C_SV_CSEG_ID) {
5627 /*
5628 * no need to relocate... this is a page full of a single
5629 * value which is hashed to a single entry not contained
5630 * in a c_segment_t
5631 */
5632 return kr;
5633 }
5634
5635 if (vm_compressor_is_slot_compressed((int *)src_slot) == false) {
5636 /*
5637 * Unmodified anonymous pages are sitting uncompressed on disk.
5638 * So don't pull them back in again.
5639 */
5640 return kr;
5641 }
5642
5643Relookup_dst:
5644 c_seg_dst = c_seg_allocate((c_segment_t *)current_chead);
5645 /*
5646 * returns with c_seg lock held
5647 * and PAGE_REPLACEMENT_DISALLOWED(TRUE)...
5648 * c_nextslot has been allocated and
5649 * c_store.c_buffer populated
5650 */
5651 if (c_seg_dst == NULL) {
5652 /*
5653 * Out of compression segments?
5654 */
5655 kr = KERN_RESOURCE_SHORTAGE;
5656 goto out;
5657 }
5658
5659 assert(c_seg_dst->c_busy == 0);
5660
5661 C_SEG_BUSY(c_seg_dst);
5662
5663 dst_slot = c_seg_dst->c_nextslot;
5664
5665 lck_mtx_unlock_always(&c_seg_dst->c_lock);
5666
5667Relookup_src:
5668 c_seg_src = c_segments[src_slot->s_cseg - 1].c_seg;
5669
5670 assert(c_seg_dst != c_seg_src);
5671
5672 lck_mtx_lock_spin_always(&c_seg_src->c_lock);
5673
5674 if (C_SEG_IS_ON_DISK_OR_SOQ(c_seg_src) ||
5675 c_seg_src->c_state == C_IS_FILLING) {
5676 /*
5677 * Skip this page if :-
5678 * a) the src c_seg is already on-disk (or on its way there)
5679 * A "thaw" can mark a process as eligible for
5680 * another freeze cycle without bringing any of
5681 * its swapped out c_segs back from disk (because
5682 * that is done on-demand).
5683 * Or, this page may be mapped elsewhere in the task's map,
5684 * and we may have marked it for swap already.
5685 *
5686 * b) Or, the src c_seg is being filled by the compressor
5687 * thread. We don't want the added latency of waiting for
5688 * this c_seg in the freeze path and so we skip it.
5689 */
5690
5691 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5692
5693 lck_mtx_unlock_always(&c_seg_src->c_lock);
5694
5695 c_seg_src = NULL;
5696
5697 goto out;
5698 }
5699
5700 if (c_seg_src->c_busy) {
5701 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5702 c_seg_wait_on_busy(c_seg_src);
5703
5704 c_seg_src = NULL;
5705
5706 PAGE_REPLACEMENT_DISALLOWED(TRUE);
5707
5708 goto Relookup_src;
5709 }
5710
5711 C_SEG_BUSY(c_seg_src);
5712
5713 lck_mtx_unlock_always(&c_seg_src->c_lock);
5714
5715 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5716
5717 /* find the c_slot */
5718 c_indx = src_slot->s_cindx;
5719
5720 c_src = C_SEG_SLOT_FROM_INDEX(c_seg_src, c_indx);
5721
5722 c_size = UNPACK_C_SIZE(c_src);
5723
5724 assert(c_size);
5725
5726 if (c_size > (uint32_t)(c_seg_bufsize - C_SEG_OFFSET_TO_BYTES((int32_t)c_seg_dst->c_nextoffset))) {
5727 /*
5728 * This segment is full. We need a new one.
5729 */
5730
5731 PAGE_REPLACEMENT_DISALLOWED(TRUE);
5732
5733 lck_mtx_lock_spin_always(&c_seg_src->c_lock);
5734 C_SEG_WAKEUP_DONE(c_seg_src);
5735 lck_mtx_unlock_always(&c_seg_src->c_lock);
5736
5737 c_seg_src = NULL;
5738
5739 lck_mtx_lock_spin_always(&c_seg_dst->c_lock);
5740
5741 assert(c_seg_dst->c_busy);
5742 assert(c_seg_dst->c_state == C_IS_FILLING);
5743 assert(!c_seg_dst->c_on_minorcompact_q);
5744
5745 c_current_seg_filled(c_seg_dst, (c_segment_t *)current_chead);
5746 assert(*current_chead == NULL);
5747
5748 C_SEG_WAKEUP_DONE(c_seg_dst);
5749
5750 lck_mtx_unlock_always(&c_seg_dst->c_lock);
5751
5752 c_seg_dst = NULL;
5753
5754 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5755
5756 goto Relookup_dst;
5757 }
5758
5759 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot);
5760
5761 memcpy(&c_seg_dst->c_store.c_buffer[c_seg_dst->c_nextoffset], &c_seg_src->c_store.c_buffer[c_src->c_offset], c_size);
5762 /*
5763 * Is platform alignment actually necessary since wkdm aligns its output?
5764 */
5765 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
5766
5767 cslot_copy(c_dst, c_src);
5768 c_dst->c_offset = c_seg_dst->c_nextoffset;
5769
5770 if (c_seg_dst->c_firstemptyslot == c_seg_dst->c_nextslot) {
5771 c_seg_dst->c_firstemptyslot++;
5772 }
5773
5774 c_seg_dst->c_slots_used++;
5775 c_seg_dst->c_nextslot++;
5776 c_seg_dst->c_bytes_used += c_rounded_size;
5777 c_seg_dst->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
5778
5779
5780 PACK_C_SIZE(c_src, 0);
5781
5782 c_seg_src->c_bytes_used -= c_rounded_size;
5783 c_seg_src->c_bytes_unused += c_rounded_size;
5784
5785 assert(c_seg_src->c_slots_used);
5786 c_seg_src->c_slots_used--;
5787
5788 if (!c_seg_src->c_swappedin) {
5789 /* Pessimistically lose swappedin status when non-swappedin pages are added. */
5790 c_seg_dst->c_swappedin = false;
5791 }
5792
5793 if (c_indx < c_seg_src->c_firstemptyslot) {
5794 c_seg_src->c_firstemptyslot = c_indx;
5795 }
5796
5797 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, dst_slot);
5798
5799 PAGE_REPLACEMENT_ALLOWED(TRUE);
5800 slot_ptr = C_SLOT_UNPACK_PTR(c_dst);
5801 /* <csegno=0,indx=0> would mean "empty slot", so use csegno+1 */
5802 slot_ptr->s_cseg = c_seg_dst->c_mysegno + 1;
5803 slot_ptr->s_cindx = dst_slot;
5804
5805 PAGE_REPLACEMENT_ALLOWED(FALSE);
5806
5807out:
5808 if (c_seg_src) {
5809 lck_mtx_lock_spin_always(&c_seg_src->c_lock);
5810
5811 C_SEG_WAKEUP_DONE(c_seg_src);
5812
5813 if (c_seg_src->c_bytes_used == 0 && c_seg_src->c_state != C_IS_FILLING) {
5814 if (!c_seg_src->c_on_minorcompact_q) {
5815 c_seg_need_delayed_compaction(c_seg_src, FALSE);
5816 }
5817 }
5818
5819 lck_mtx_unlock_always(&c_seg_src->c_lock);
5820 }
5821
5822 if (c_seg_dst) {
5823 PAGE_REPLACEMENT_DISALLOWED(TRUE);
5824
5825 lck_mtx_lock_spin_always(&c_seg_dst->c_lock);
5826
5827 if (c_seg_dst->c_nextoffset >= c_seg_off_limit || c_seg_dst->c_nextslot >= C_SLOT_MAX_INDEX) {
5828 /*
5829 * Nearing or exceeded maximum slot and offset capacity.
5830 */
5831 assert(c_seg_dst->c_busy);
5832 assert(c_seg_dst->c_state == C_IS_FILLING);
5833 assert(!c_seg_dst->c_on_minorcompact_q);
5834
5835 c_current_seg_filled(c_seg_dst, (c_segment_t *)current_chead);
5836 assert(*current_chead == NULL);
5837 }
5838
5839 C_SEG_WAKEUP_DONE(c_seg_dst);
5840
5841 lck_mtx_unlock_always(&c_seg_dst->c_lock);
5842
5843 c_seg_dst = NULL;
5844
5845 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5846 }
5847
5848 return kr;
5849}
5850#endif /* CONFIG_FREEZE */
5851
5852#if DEVELOPMENT || DEBUG
5853
5854void
5855vm_compressor_inject_error(int *slot)
5856{
5857 c_slot_mapping_t slot_ptr = (c_slot_mapping_t)slot;
5858
5859 /* No error detection for single-value compression. */
5860 if (slot_ptr->s_cseg == C_SV_CSEG_ID) {
5861 printf("%s(): cannot inject errors in SV-compressed pages\n", __func__ );
5862 return;
5863 }
5864
5865 /* s_cseg is actually "segno+1" */
5866 const uint32_t c_segno = slot_ptr->s_cseg - 1;
5867
5868 assert(c_segno < c_segments_available);
5869 assert(c_segments[c_segno].c_segno >= c_segments_available);
5870
5871 const c_segment_t c_seg = c_segments[c_segno].c_seg;
5872
5873 PAGE_REPLACEMENT_DISALLOWED(TRUE);
5874
5875 lck_mtx_lock_spin_always(&c_seg->c_lock);
5876 assert(c_seg->c_state != C_IS_EMPTY && c_seg->c_state != C_IS_FREE);
5877
5878 const uint16_t c_indx = slot_ptr->s_cindx;
5879 assert(c_indx < c_seg->c_nextslot);
5880
5881 /*
5882 * To safely make this segment temporarily writable, we need to mark
5883 * the segment busy, which allows us to release the segment lock.
5884 */
5885 while (c_seg->c_busy) {
5886 c_seg_wait_on_busy(c_seg);
5887 lck_mtx_lock_spin_always(&c_seg->c_lock);
5888 }
5889 C_SEG_BUSY(c_seg);
5890
5891 bool already_writable = (c_seg->c_state == C_IS_FILLING);
5892 if (!already_writable) {
5893 /*
5894 * Protection update must be performed preemptibly, so temporarily drop
5895 * the lock. Having set c_busy will prevent most other concurrent
5896 * operations.
5897 */
5898 lck_mtx_unlock_always(&c_seg->c_lock);
5899 C_SEG_MAKE_WRITEABLE(c_seg);
5900 lck_mtx_lock_spin_always(&c_seg->c_lock);
5901 }
5902
5903 /*
5904 * Once we've released the lock following our c_state == C_IS_FILLING check,
5905 * c_current_seg_filled() can (re-)write-protect the segment. However, it
5906 * will transition from C_IS_FILLING before releasing the c_seg lock, so we
5907 * can detect this by re-checking after we've reobtained the lock.
5908 */
5909 if (already_writable && c_seg->c_state != C_IS_FILLING) {
5910 lck_mtx_unlock_always(&c_seg->c_lock);
5911 C_SEG_MAKE_WRITEABLE(c_seg);
5912 lck_mtx_lock_spin_always(&c_seg->c_lock);
5913 already_writable = false;
5914 /* Segment can't be freed while c_busy is set. */
5915 assert(c_seg->c_state != C_IS_FILLING);
5916 }
5917
5918 /*
5919 * Skip if the segment is on disk. This check can only be performed after
5920 * the final acquisition of the segment lock before we attempt to write to
5921 * the segment.
5922 */
5923 if (!C_SEG_IS_ON_DISK_OR_SOQ(c_seg)) {
5924 c_slot_t cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
5925 int32_t *data = &c_seg->c_store.c_buffer[cs->c_offset];
5926 /* assume that the compressed data holds at least one int32_t */
5927 assert(UNPACK_C_SIZE(cs) > sizeof(*data));
5928 /*
5929 * This bit is known to be in the payload of a MISS packet resulting from
5930 * the pattern used in the test pattern from decompression_failure.c.
5931 * Flipping it should result in many corrupted bits in the test page.
5932 */
5933 data[0] ^= 0x00000100;
5934 }
5935
5936 if (!already_writable) {
5937 lck_mtx_unlock_always(&c_seg->c_lock);
5938 C_SEG_WRITE_PROTECT(c_seg);
5939 lck_mtx_lock_spin_always(&c_seg->c_lock);
5940 }
5941
5942 C_SEG_WAKEUP_DONE(c_seg);
5943 lck_mtx_unlock_always(&c_seg->c_lock);
5944
5945 PAGE_REPLACEMENT_DISALLOWED(FALSE);
5946}
5947
5948#endif /* DEVELOPMENT || DEBUG */
5949
5950
5951#if CONFIG_TRACK_UNMODIFIED_ANON_PAGES
5952
5953struct vnode;
5954extern void vm_swapfile_open(const char *path, struct vnode **vp);
5955extern int vm_swapfile_preallocate(struct vnode *vp, uint64_t *size, boolean_t *pin);
5956
5957struct vnode *uncompressed_vp0 = NULL;
5958struct vnode *uncompressed_vp1 = NULL;
5959uint32_t uncompressed_file0_free_pages = 0, uncompressed_file1_free_pages = 0;
5960uint64_t uncompressed_file0_free_offset = 0, uncompressed_file1_free_offset = 0;
5961
5962uint64_t compressor_ro_uncompressed = 0;
5963uint64_t compressor_ro_uncompressed_total_returned = 0;
5964uint64_t compressor_ro_uncompressed_skip_returned = 0;
5965uint64_t compressor_ro_uncompressed_get = 0;
5966uint64_t compressor_ro_uncompressed_put = 0;
5967uint64_t compressor_ro_uncompressed_swap_usage = 0;
5968
5969extern void vnode_put(struct vnode* vp);
5970extern int vnode_getwithref(struct vnode* vp);
5971extern int vm_swapfile_io(struct vnode *vp, uint64_t offset, uint64_t start, int npages, int flags, void *upl_ctx);
5972
5973#define MAX_OFFSET_PAGES (255)
5974uint64_t uncompressed_file0_space_bitmap[MAX_OFFSET_PAGES];
5975uint64_t uncompressed_file1_space_bitmap[MAX_OFFSET_PAGES];
5976
5977#define UNCOMPRESSED_FILEIDX_OFFSET_MASK (((uint32_t)1<<31ull) - 1)
5978#define UNCOMPRESSED_FILEIDX_SHIFT (29)
5979#define UNCOMPRESSED_FILEIDX_MASK (3)
5980#define UNCOMPRESSED_OFFSET_SHIFT (29)
5981#define UNCOMPRESSED_OFFSET_MASK (7)
5982
5983static uint32_t
5984vm_uncompressed_extract_swap_file(int slot)
5985{
5986 uint32_t fileidx = (((uint32_t)slot & UNCOMPRESSED_FILEIDX_OFFSET_MASK) >> UNCOMPRESSED_FILEIDX_SHIFT) & UNCOMPRESSED_FILEIDX_MASK;
5987 return fileidx;
5988}
5989
5990static uint32_t
5991vm_uncompressed_extract_swap_offset(int slot)
5992{
5993 return slot & (uint32_t)(~(UNCOMPRESSED_OFFSET_MASK << UNCOMPRESSED_OFFSET_SHIFT));
5994}
5995
5996static void
5997vm_uncompressed_return_space_to_swap(int slot)
5998{
5999 PAGE_REPLACEMENT_ALLOWED(TRUE);
6000 uint32_t fileidx = vm_uncompressed_extract_swap_file(slot);
6001 if (fileidx == 1) {
6002 uint32_t free_offset = vm_uncompressed_extract_swap_offset(slot);
6003 uint64_t pgidx = free_offset / PAGE_SIZE_64;
6004 uint64_t chunkidx = pgidx / 64;
6005 uint64_t chunkoffset = pgidx % 64;
6006#if DEVELOPMENT || DEBUG
6007 uint64_t vaddr = (uint64_t)&uncompressed_file0_space_bitmap[chunkidx];
6008 uint64_t maxvaddr = (uint64_t)&uncompressed_file0_space_bitmap[MAX_OFFSET_PAGES];
6009 assertf(vaddr < maxvaddr, "0x%llx 0x%llx", vaddr, maxvaddr);
6010#endif /*DEVELOPMENT || DEBUG*/
6011 assertf((uncompressed_file0_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)),
6012 "0x%x %llu %llu", slot, chunkidx, chunkoffset);
6013 uncompressed_file0_space_bitmap[chunkidx] &= ~((uint64_t)1 << chunkoffset);
6014 assertf(!(uncompressed_file0_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)),
6015 "0x%x %llu %llu", slot, chunkidx, chunkoffset);
6016
6017 uncompressed_file0_free_pages++;
6018 } else {
6019 uint32_t free_offset = vm_uncompressed_extract_swap_offset(slot);
6020 uint64_t pgidx = free_offset / PAGE_SIZE_64;
6021 uint64_t chunkidx = pgidx / 64;
6022 uint64_t chunkoffset = pgidx % 64;
6023 assertf((uncompressed_file1_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)),
6024 "%llu %llu", chunkidx, chunkoffset);
6025 uncompressed_file1_space_bitmap[chunkidx] &= ~((uint64_t)1 << chunkoffset);
6026
6027 uncompressed_file1_free_pages++;
6028 }
6029 compressor_ro_uncompressed_swap_usage--;
6030 PAGE_REPLACEMENT_ALLOWED(FALSE);
6031}
6032
6033static int
6034vm_uncompressed_reserve_space_in_swap()
6035{
6036 int slot = 0;
6037 if (uncompressed_file0_free_pages == 0 && uncompressed_file1_free_pages == 0) {
6038 return -1;
6039 }
6040
6041 PAGE_REPLACEMENT_ALLOWED(TRUE);
6042 if (uncompressed_file0_free_pages) {
6043 uint64_t chunkidx = 0;
6044 uint64_t chunkoffset = 0;
6045 while (uncompressed_file0_space_bitmap[chunkidx] == 0xffffffffffffffff) {
6046 chunkidx++;
6047 }
6048 while (uncompressed_file0_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)) {
6049 chunkoffset++;
6050 }
6051
6052 assertf((uncompressed_file0_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)) == 0,
6053 "%llu %llu", chunkidx, chunkoffset);
6054#if DEVELOPMENT || DEBUG
6055 uint64_t vaddr = (uint64_t)&uncompressed_file0_space_bitmap[chunkidx];
6056 uint64_t maxvaddr = (uint64_t)&uncompressed_file0_space_bitmap[MAX_OFFSET_PAGES];
6057 assertf(vaddr < maxvaddr, "0x%llx 0x%llx", vaddr, maxvaddr);
6058#endif /*DEVELOPMENT || DEBUG*/
6059 uncompressed_file0_space_bitmap[chunkidx] |= ((uint64_t)1 << chunkoffset);
6060 uncompressed_file0_free_offset = ((chunkidx * 64) + chunkoffset) * PAGE_SIZE_64;
6061 assertf((uncompressed_file0_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)),
6062 "%llu %llu", chunkidx, chunkoffset);
6063
6064 assert(uncompressed_file0_free_offset <= (1 << UNCOMPRESSED_OFFSET_SHIFT));
6065 slot = (int)((1 << UNCOMPRESSED_FILEIDX_SHIFT) + uncompressed_file0_free_offset);
6066 uncompressed_file0_free_pages--;
6067 } else {
6068 uint64_t chunkidx = 0;
6069 uint64_t chunkoffset = 0;
6070 while (uncompressed_file1_space_bitmap[chunkidx] == 0xFFFFFFFFFFFFFFFF) {
6071 chunkidx++;
6072 }
6073 while (uncompressed_file1_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)) {
6074 chunkoffset++;
6075 }
6076 assert((uncompressed_file1_space_bitmap[chunkidx] & ((uint64_t)1 << chunkoffset)) == 0);
6077 uncompressed_file1_space_bitmap[chunkidx] |= ((uint64_t)1 << chunkoffset);
6078 uncompressed_file1_free_offset = ((chunkidx * 64) + chunkoffset) * PAGE_SIZE_64;
6079 slot = (int)((2 << UNCOMPRESSED_FILEIDX_SHIFT) + uncompressed_file1_free_offset);
6080 uncompressed_file1_free_pages--;
6081 }
6082 compressor_ro_uncompressed_swap_usage++;
6083 PAGE_REPLACEMENT_ALLOWED(FALSE);
6084 return slot;
6085}
6086
6087#define MAX_IO_REQ (16)
6088struct _uncompressor_io_req {
6089 uint64_t addr;
6090 bool inuse;
6091} uncompressor_io_req[MAX_IO_REQ];
6092
6093int
6094vm_uncompressed_put(ppnum_t pn, int *slot)
6095{
6096 int retval = 0;
6097 struct vnode *uncompressed_vp = NULL;
6098 uint64_t uncompress_offset = 0;
6099
6100again:
6101 if (uncompressed_vp0 == NULL) {
6102 PAGE_REPLACEMENT_ALLOWED(TRUE);
6103 if (uncompressed_vp0 == NULL) {
6104 uint64_t size = (MAX_OFFSET_PAGES * 1024 * 1024ULL);
6105 vm_swapfile_open("/private/var/vm/uncompressedswap0", &uncompressed_vp0);
6106 if (uncompressed_vp0 == NULL) {
6107 PAGE_REPLACEMENT_ALLOWED(FALSE);
6108 return KERN_NO_ACCESS;
6109 }
6110 vm_swapfile_preallocate(uncompressed_vp0, &size, NULL);
6111 uncompressed_file0_free_pages = (uint32_t)atop(size);
6112 bzero(uncompressed_file0_space_bitmap, sizeof(uint64_t) * MAX_OFFSET_PAGES);
6113
6114 int i = 0;
6115 for (; i < MAX_IO_REQ; i++) {
6116 kmem_alloc(kernel_map, (vm_offset_t*)&uncompressor_io_req[i].addr, PAGE_SIZE_64, KMA_NOFAIL | KMA_KOBJECT, VM_KERN_MEMORY_COMPRESSOR);
6117 uncompressor_io_req[i].inuse = false;
6118 }
6119
6120 vm_swapfile_open("/private/var/vm/uncompressedswap1", &uncompressed_vp1);
6121 assert(uncompressed_vp1);
6122 vm_swapfile_preallocate(uncompressed_vp1, &size, NULL);
6123 uncompressed_file1_free_pages = (uint32_t)atop(size);
6124 bzero(uncompressed_file1_space_bitmap, sizeof(uint64_t) * MAX_OFFSET_PAGES);
6125 PAGE_REPLACEMENT_ALLOWED(FALSE);
6126 } else {
6127 PAGE_REPLACEMENT_ALLOWED(FALSE);
6128 delay(100);
6129 goto again;
6130 }
6131 }
6132
6133 int swapinfo = vm_uncompressed_reserve_space_in_swap();
6134 if (swapinfo == -1) {
6135 *slot = 0;
6136 return KERN_RESOURCE_SHORTAGE;
6137 }
6138
6139 if (vm_uncompressed_extract_swap_file(swapinfo) == 1) {
6140 uncompressed_vp = uncompressed_vp0;
6141 } else {
6142 uncompressed_vp = uncompressed_vp1;
6143 }
6144 uncompress_offset = vm_uncompressed_extract_swap_offset(swapinfo);
6145 if ((retval = vnode_getwithref(uncompressed_vp)) != 0) {
6146 os_log_error_with_startup_serial(OS_LOG_DEFAULT, "vm_uncompressed_put: vnode_getwithref on swapfile failed with %d\n", retval);
6147 } else {
6148 int i = 0;
6149retry:
6150 PAGE_REPLACEMENT_ALLOWED(TRUE);
6151 for (i = 0; i < MAX_IO_REQ; i++) {
6152 if (uncompressor_io_req[i].inuse == false) {
6153 uncompressor_io_req[i].inuse = true;
6154 break;
6155 }
6156 }
6157 if (i == MAX_IO_REQ) {
6158 assert_wait((event_t)&uncompressor_io_req, THREAD_UNINT);
6159 PAGE_REPLACEMENT_ALLOWED(FALSE);
6160 thread_block(THREAD_CONTINUE_NULL);
6161 goto retry;
6162 }
6163 PAGE_REPLACEMENT_ALLOWED(FALSE);
6164 void *addr = pmap_map_compressor_page(pn);
6165 memcpy((void*)uncompressor_io_req[i].addr, addr, PAGE_SIZE_64);
6166 pmap_unmap_compressor_page(pn, addr);
6167
6168 retval = vm_swapfile_io(uncompressed_vp, uncompress_offset, (uint64_t)uncompressor_io_req[i].addr, 1, SWAP_WRITE, NULL);
6169 if (retval) {
6170 *slot = 0;
6171 } else {
6172 *slot = (int)swapinfo;
6173 ((c_slot_mapping_t)(slot))->s_uncompressed = 1;
6174 }
6175 vnode_put(uncompressed_vp);
6176 PAGE_REPLACEMENT_ALLOWED(TRUE);
6177 uncompressor_io_req[i].inuse = false;
6178 thread_wakeup((event_t)&uncompressor_io_req);
6179 PAGE_REPLACEMENT_ALLOWED(FALSE);
6180 }
6181 return retval;
6182}
6183
6184int
6185vm_uncompressed_get(ppnum_t pn, int *slot, __unused vm_compressor_options_t flags)
6186{
6187 int retval = 0;
6188 struct vnode *uncompressed_vp = NULL;
6189 uint32_t fileidx = vm_uncompressed_extract_swap_file(*slot);
6190 uint64_t uncompress_offset = vm_uncompressed_extract_swap_offset(*slot);
6191
6192 if (__improbable(flags & C_KDP)) {
6193 return -2;
6194 }
6195
6196 if (fileidx == 1) {
6197 uncompressed_vp = uncompressed_vp0;
6198 } else {
6199 uncompressed_vp = uncompressed_vp1;
6200 }
6201
6202 if ((retval = vnode_getwithref(uncompressed_vp)) != 0) {
6203 os_log_error_with_startup_serial(OS_LOG_DEFAULT, "vm_uncompressed_put: vnode_getwithref on swapfile failed with %d\n", retval);
6204 } else {
6205 int i = 0;
6206retry:
6207 PAGE_REPLACEMENT_ALLOWED(TRUE);
6208 for (i = 0; i < MAX_IO_REQ; i++) {
6209 if (uncompressor_io_req[i].inuse == false) {
6210 uncompressor_io_req[i].inuse = true;
6211 break;
6212 }
6213 }
6214 if (i == MAX_IO_REQ) {
6215 assert_wait((event_t)&uncompressor_io_req, THREAD_UNINT);
6216 PAGE_REPLACEMENT_ALLOWED(FALSE);
6217 thread_block(THREAD_CONTINUE_NULL);
6218 goto retry;
6219 }
6220 PAGE_REPLACEMENT_ALLOWED(FALSE);
6221 retval = vm_swapfile_io(uncompressed_vp, uncompress_offset, (uint64_t)uncompressor_io_req[i].addr, 1, SWAP_READ, NULL);
6222 vnode_put(uncompressed_vp);
6223 void *addr = pmap_map_compressor_page(pn);
6224 memcpy(addr, (void*)uncompressor_io_req[i].addr, PAGE_SIZE_64);
6225 pmap_unmap_compressor_page(pn, addr);
6226 PAGE_REPLACEMENT_ALLOWED(TRUE);
6227 uncompressor_io_req[i].inuse = false;
6228 thread_wakeup((event_t)&uncompressor_io_req);
6229 PAGE_REPLACEMENT_ALLOWED(FALSE);
6230 }
6231 return retval;
6232}
6233
6234int
6235vm_uncompressed_free(int *slot, __unused vm_compressor_options_t flags)
6236{
6237 vm_uncompressed_return_space_to_swap(*slot);
6238 *slot = 0;
6239 return 0;
6240}
6241
6242#endif /*CONFIG_TRACK_UNMODIFIED_ANON_PAGES*/
6243