1/*
2 * Copyright (c) 2003-2020 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
30/*
31 * todo:
32 * 1) ramesh is looking into how to replace taking a reference on
33 * the user's map (vm_map_reference()) since it is believed that
34 * would not hold the process for us.
35 * 2) david is looking into a way for us to set the priority of the
36 * worker threads to match that of the user's thread when the
37 * async IO was queued.
38 */
39
40
41/*
42 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
43 */
44
45#include <sys/systm.h>
46#include <sys/fcntl.h>
47#include <sys/file_internal.h>
48#include <sys/filedesc.h>
49#include <sys/kernel.h>
50#include <sys/vnode_internal.h>
51#include <sys/kauth.h>
52#include <sys/mount_internal.h>
53#include <sys/param.h>
54#include <sys/proc_internal.h>
55#include <sys/sysctl.h>
56#include <sys/unistd.h>
57#include <sys/user.h>
58
59#include <sys/aio_kern.h>
60#include <sys/sysproto.h>
61
62#include <machine/limits.h>
63
64#include <mach/mach_types.h>
65#include <kern/kern_types.h>
66#include <kern/waitq.h>
67#include <kern/zalloc.h>
68#include <kern/task.h>
69#include <kern/sched_prim.h>
70
71#include <vm/vm_map.h>
72
73#include <os/refcnt.h>
74
75#include <sys/kdebug.h>
76#define AIO_work_queued 1
77#define AIO_worker_wake 2
78#define AIO_completion_sig 3
79#define AIO_completion_cleanup_wait 4
80#define AIO_completion_cleanup_wake 5
81#define AIO_completion_suspend_wake 6
82#define AIO_fsync_delay 7
83#define AIO_cancel 10
84#define AIO_cancel_async_workq 11
85#define AIO_cancel_sync_workq 12
86#define AIO_cancel_activeq 13
87#define AIO_cancel_doneq 14
88#define AIO_fsync 20
89#define AIO_read 30
90#define AIO_write 40
91#define AIO_listio 50
92#define AIO_error 60
93#define AIO_error_val 61
94#define AIO_error_activeq 62
95#define AIO_error_workq 63
96#define AIO_return 70
97#define AIO_return_val 71
98#define AIO_return_activeq 72
99#define AIO_return_workq 73
100#define AIO_exec 80
101#define AIO_exit 90
102#define AIO_exit_sleep 91
103#define AIO_close 100
104#define AIO_close_sleep 101
105#define AIO_suspend 110
106#define AIO_suspend_sleep 111
107#define AIO_worker_thread 120
108
109__options_decl(aio_entry_flags_t, uint32_t, {
110 AIO_READ = 0x00000001, /* a read */
111 AIO_WRITE = 0x00000002, /* a write */
112 AIO_FSYNC = 0x00000004, /* aio_fsync with op = O_SYNC */
113 AIO_DSYNC = 0x00000008, /* aio_fsync with op = O_DSYNC (not supported yet) */
114 AIO_LIO = 0x00000010, /* lio_listio generated IO */
115 AIO_LIO_WAIT = 0x00000020, /* lio_listio is waiting on the leader */
116
117 /*
118 * These flags mean that this entry is blocking either:
119 * - close (AIO_CLOSE_WAIT)
120 * - exit or exec (AIO_EXIT_WAIT)
121 *
122 * These flags are mutually exclusive, and the AIO_EXIT_WAIT variant
123 * will also neuter notifications in do_aio_completion_and_unlock().
124 */
125 AIO_CLOSE_WAIT = 0x00004000,
126 AIO_EXIT_WAIT = 0x00008000,
127});
128
129/*! @struct aio_workq_entry
130 *
131 * @discussion
132 * This represents a piece of aio/lio work.
133 *
134 * The ownership rules go as follows:
135 *
136 * - the "proc" owns one refcount on the entry (from creation), while it is
137 * enqueued on the aio_activeq and then the aio_doneq.
138 *
139 * either aio_return() (user read the status) or _aio_exit() (the process
140 * died) will dequeue the entry and consume this ref.
141 *
142 * - the async workqueue owns one refcount once the work is submitted,
143 * which is consumed in do_aio_completion_and_unlock().
144 *
145 * This ref protects the entry for the the end of
146 * do_aio_completion_and_unlock() (when signal delivery happens).
147 *
148 * - lio_listio() for batches picks one of the entries to be the "leader"
149 * of the batch. Each work item will have a refcount on its leader
150 * so that the accounting of the batch completion can be done on the leader
151 * (to be able to decrement lio_pending).
152 *
153 * This ref is consumed in do_aio_completion_and_unlock() as well.
154 *
155 * - lastly, in lio_listio() when the LIO_WAIT behavior is requested,
156 * an extra ref is taken in this syscall as it needs to keep accessing
157 * the leader "lio_pending" field until it hits 0.
158 */
159struct aio_workq_entry {
160 /* queue lock */
161 TAILQ_ENTRY(aio_workq_entry) aio_workq_link;
162
163 /* Proc lock */
164 TAILQ_ENTRY(aio_workq_entry) aio_proc_link; /* p_aio_activeq or p_aio_doneq */
165 user_ssize_t returnval; /* return value from read / write request */
166 errno_t errorval; /* error value from read / write request */
167 os_refcnt_t aio_refcount;
168 aio_entry_flags_t flags;
169
170 int lio_pending; /* pending I/Os in lio group, only on leader */
171 struct aio_workq_entry *lio_leader; /* pointer to the lio leader, can be self */
172
173 /* Initialized and never changed, safe to access */
174 struct proc *procp; /* user proc that queued this request */
175 user_addr_t uaiocbp; /* pointer passed in from user land */
176 struct user_aiocb aiocb; /* copy of aiocb from user land */
177 struct vfs_context context; /* context which enqueued the request */
178
179 /* Initialized, and possibly freed by aio_work_thread() or at free if cancelled */
180 vm_map_t aio_map; /* user land map we have a reference to */
181};
182
183/*
184 * aio requests queue up on the aio_async_workq or lio_sync_workq (for
185 * lio_listio LIO_WAIT). Requests then move to the per process aio_activeq
186 * (proc.aio_activeq) when one of our worker threads start the IO.
187 * And finally, requests move to the per process aio_doneq (proc.aio_doneq)
188 * when the IO request completes. The request remains on aio_doneq until
189 * user process calls aio_return or the process exits, either way that is our
190 * trigger to release aio resources.
191 */
192typedef struct aio_workq {
193 TAILQ_HEAD(, aio_workq_entry) aioq_entries;
194 lck_spin_t aioq_lock;
195 struct waitq aioq_waitq;
196} *aio_workq_t;
197
198#define AIO_NUM_WORK_QUEUES 1
199struct aio_anchor_cb {
200 os_atomic(int) aio_total_count; /* total extant entries */
201
202 /* Hash table of queues here */
203 int aio_num_workqs;
204 struct aio_workq aio_async_workqs[AIO_NUM_WORK_QUEUES];
205};
206typedef struct aio_anchor_cb aio_anchor_cb;
207
208/*
209 * Notes on aio sleep / wake channels.
210 * We currently pick a couple fields within the proc structure that will allow
211 * us sleep channels that currently do not collide with any other kernel routines.
212 * At this time, for binary compatibility reasons, we cannot create new proc fields.
213 */
214#define AIO_SUSPEND_SLEEP_CHAN p_aio_activeq
215#define AIO_CLEANUP_SLEEP_CHAN p_aio_total_count
216
217#define ASSERT_AIO_FROM_PROC(aiop, theproc) \
218 if ((aiop)->procp != (theproc)) { \
219 panic("AIO on a proc list that does not belong to that proc."); \
220 }
221
222/*
223 * LOCAL PROTOTYPES
224 */
225static void aio_proc_lock(proc_t procp);
226static void aio_proc_lock_spin(proc_t procp);
227static void aio_proc_unlock(proc_t procp);
228static lck_mtx_t *aio_proc_mutex(proc_t procp);
229static bool aio_has_active_requests_for_process(proc_t procp);
230static bool aio_proc_has_active_requests_for_file(proc_t procp, int fd);
231static boolean_t is_already_queued(proc_t procp, user_addr_t aiocbp);
232
233static aio_workq_t aio_entry_workq(aio_workq_entry *entryp);
234static void aio_workq_remove_entry_locked(aio_workq_t queue, aio_workq_entry *entryp);
235static void aio_workq_add_entry_locked(aio_workq_t queue, aio_workq_entry *entryp);
236static void aio_entry_ref(aio_workq_entry *entryp);
237static void aio_entry_unref(aio_workq_entry *entryp);
238static bool aio_entry_try_workq_remove(aio_workq_entry *entryp);
239static boolean_t aio_delay_fsync_request(aio_workq_entry *entryp);
240static void aio_free_request(aio_workq_entry *entryp);
241
242static void aio_workq_init(aio_workq_t wq);
243static void aio_workq_lock_spin(aio_workq_t wq);
244static void aio_workq_unlock(aio_workq_t wq);
245static lck_spin_t *aio_workq_lock(aio_workq_t wq);
246
247static void aio_work_thread(void *arg, wait_result_t wr);
248static aio_workq_entry *aio_get_some_work(void);
249
250static int aio_queue_async_request(proc_t procp, user_addr_t aiocbp, aio_entry_flags_t);
251static int aio_validate(proc_t, aio_workq_entry *entryp);
252
253static int do_aio_cancel_locked(proc_t p, int fd, user_addr_t aiocbp, aio_entry_flags_t);
254static void do_aio_completion_and_unlock(proc_t p, aio_workq_entry *entryp);
255static int do_aio_fsync(aio_workq_entry *entryp);
256static int do_aio_read(aio_workq_entry *entryp);
257static int do_aio_write(aio_workq_entry *entryp);
258static void do_munge_aiocb_user32_to_user(struct user32_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp);
259static void do_munge_aiocb_user64_to_user(struct user64_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp);
260static aio_workq_entry *aio_create_queue_entry(proc_t procp, user_addr_t aiocbp, aio_entry_flags_t);
261static int aio_copy_in_list(proc_t, user_addr_t, user_addr_t *, int);
262
263#define ASSERT_AIO_PROC_LOCK_OWNED(p) LCK_MTX_ASSERT(aio_proc_mutex(p), LCK_MTX_ASSERT_OWNED)
264#define ASSERT_AIO_WORKQ_LOCK_OWNED(q) LCK_SPIN_ASSERT(aio_workq_lock(q), LCK_ASSERT_OWNED)
265
266/*
267 * EXTERNAL PROTOTYPES
268 */
269
270/* in ...bsd/kern/sys_generic.c */
271extern int dofileread(vfs_context_t ctx, struct fileproc *fp,
272 user_addr_t bufp, user_size_t nbyte,
273 off_t offset, int flags, user_ssize_t *retval);
274extern int dofilewrite(vfs_context_t ctx, struct fileproc *fp,
275 user_addr_t bufp, user_size_t nbyte, off_t offset,
276 int flags, user_ssize_t *retval);
277
278/*
279 * aio external global variables.
280 */
281extern int aio_max_requests; /* AIO_MAX - configurable */
282extern int aio_max_requests_per_process; /* AIO_PROCESS_MAX - configurable */
283extern int aio_worker_threads; /* AIO_THREAD_COUNT - configurable */
284
285
286/*
287 * aio static variables.
288 */
289static aio_anchor_cb aio_anchor = {
290 .aio_num_workqs = AIO_NUM_WORK_QUEUES,
291};
292os_refgrp_decl(static, aio_refgrp, "aio", NULL);
293static LCK_GRP_DECLARE(aio_proc_lock_grp, "aio_proc");
294static LCK_GRP_DECLARE(aio_queue_lock_grp, "aio_queue");
295static LCK_MTX_DECLARE(aio_proc_mtx, &aio_proc_lock_grp);
296
297static KALLOC_TYPE_DEFINE(aio_workq_zonep, aio_workq_entry, KT_DEFAULT);
298
299/* Hash */
300static aio_workq_t
301aio_entry_workq(__unused aio_workq_entry *entryp)
302{
303 return &aio_anchor.aio_async_workqs[0];
304}
305
306static void
307aio_workq_init(aio_workq_t wq)
308{
309 TAILQ_INIT(&wq->aioq_entries);
310 lck_spin_init(lck: &wq->aioq_lock, grp: &aio_queue_lock_grp, LCK_ATTR_NULL);
311 waitq_init(waitq: &wq->aioq_waitq, type: WQT_QUEUE, SYNC_POLICY_FIFO);
312}
313
314
315/*
316 * Can be passed a queue which is locked spin.
317 */
318static void
319aio_workq_remove_entry_locked(aio_workq_t queue, aio_workq_entry *entryp)
320{
321 ASSERT_AIO_WORKQ_LOCK_OWNED(queue);
322
323 if (entryp->aio_workq_link.tqe_prev == NULL) {
324 panic("Trying to remove an entry from a work queue, but it is not on a queue");
325 }
326
327 TAILQ_REMOVE(&queue->aioq_entries, entryp, aio_workq_link);
328 entryp->aio_workq_link.tqe_prev = NULL; /* Not on a workq */
329}
330
331static void
332aio_workq_add_entry_locked(aio_workq_t queue, aio_workq_entry *entryp)
333{
334 ASSERT_AIO_WORKQ_LOCK_OWNED(queue);
335
336 TAILQ_INSERT_TAIL(&queue->aioq_entries, entryp, aio_workq_link);
337}
338
339static void
340aio_proc_lock(proc_t procp)
341{
342 lck_mtx_lock(lck: aio_proc_mutex(procp));
343}
344
345static void
346aio_proc_lock_spin(proc_t procp)
347{
348 lck_mtx_lock_spin(lck: aio_proc_mutex(procp));
349}
350
351static bool
352aio_has_any_work(void)
353{
354 return os_atomic_load(&aio_anchor.aio_total_count, relaxed) != 0;
355}
356
357static bool
358aio_try_proc_insert_active_locked(proc_t procp, aio_workq_entry *entryp)
359{
360 int old, new;
361
362 ASSERT_AIO_PROC_LOCK_OWNED(procp);
363
364 if (procp->p_aio_total_count >= aio_max_requests_per_process) {
365 return false;
366 }
367
368 if (is_already_queued(procp, aiocbp: entryp->uaiocbp)) {
369 return false;
370 }
371
372 os_atomic_rmw_loop(&aio_anchor.aio_total_count, old, new, relaxed, {
373 if (old >= aio_max_requests) {
374 os_atomic_rmw_loop_give_up(return false);
375 }
376 new = old + 1;
377 });
378
379 TAILQ_INSERT_TAIL(&procp->p_aio_activeq, entryp, aio_proc_link);
380 procp->p_aio_total_count++;
381 return true;
382}
383
384static void
385aio_proc_move_done_locked(proc_t procp, aio_workq_entry *entryp)
386{
387 TAILQ_REMOVE(&procp->p_aio_activeq, entryp, aio_proc_link);
388 TAILQ_INSERT_TAIL(&procp->p_aio_doneq, entryp, aio_proc_link);
389}
390
391static void
392aio_proc_remove_done_locked(proc_t procp, aio_workq_entry *entryp)
393{
394 TAILQ_REMOVE(&procp->p_aio_doneq, entryp, aio_proc_link);
395 entryp->aio_proc_link.tqe_prev = NULL;
396 if (os_atomic_dec_orig(&aio_anchor.aio_total_count, relaxed) <= 0) {
397 panic("Negative total AIO count!");
398 }
399 if (procp->p_aio_total_count-- <= 0) {
400 panic("proc %p: p_aio_total_count accounting mismatch", procp);
401 }
402}
403
404static void
405aio_proc_unlock(proc_t procp)
406{
407 lck_mtx_unlock(lck: aio_proc_mutex(procp));
408}
409
410static lck_mtx_t*
411aio_proc_mutex(proc_t procp)
412{
413 return &procp->p_mlock;
414}
415
416static void
417aio_entry_ref(aio_workq_entry *entryp)
418{
419 os_ref_retain(rc: &entryp->aio_refcount);
420}
421
422static void
423aio_entry_unref(aio_workq_entry *entryp)
424{
425 if (os_ref_release(rc: &entryp->aio_refcount) == 0) {
426 aio_free_request(entryp);
427 }
428}
429
430static bool
431aio_entry_try_workq_remove(aio_workq_entry *entryp)
432{
433 /* Can only be cancelled if it's still on a work queue */
434 if (entryp->aio_workq_link.tqe_prev != NULL) {
435 aio_workq_t queue;
436
437 /* Will have to check again under the lock */
438 queue = aio_entry_workq(entryp);
439 aio_workq_lock_spin(wq: queue);
440 if (entryp->aio_workq_link.tqe_prev != NULL) {
441 aio_workq_remove_entry_locked(queue, entryp);
442 aio_workq_unlock(wq: queue);
443 return true;
444 } else {
445 aio_workq_unlock(wq: queue);
446 }
447 }
448
449 return false;
450}
451
452static void
453aio_workq_lock_spin(aio_workq_t wq)
454{
455 lck_spin_lock(lck: aio_workq_lock(wq));
456}
457
458static void
459aio_workq_unlock(aio_workq_t wq)
460{
461 lck_spin_unlock(lck: aio_workq_lock(wq));
462}
463
464static lck_spin_t*
465aio_workq_lock(aio_workq_t wq)
466{
467 return &wq->aioq_lock;
468}
469
470/*
471 * aio_cancel - attempt to cancel one or more async IO requests currently
472 * outstanding against file descriptor uap->fd. If uap->aiocbp is not
473 * NULL then only one specific IO is cancelled (if possible). If uap->aiocbp
474 * is NULL then all outstanding async IO request for the given file
475 * descriptor are cancelled (if possible).
476 */
477int
478aio_cancel(proc_t p, struct aio_cancel_args *uap, int *retval)
479{
480 struct user_aiocb my_aiocb;
481 int result;
482
483 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel) | DBG_FUNC_START,
484 VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);
485
486 /* quick check to see if there are any async IO requests queued up */
487 if (!aio_has_any_work()) {
488 result = 0;
489 *retval = AIO_ALLDONE;
490 goto ExitRoutine;
491 }
492
493 *retval = -1;
494 if (uap->aiocbp != USER_ADDR_NULL) {
495 if (proc_is64bit(p)) {
496 struct user64_aiocb aiocb64;
497
498 result = copyin(uap->aiocbp, &aiocb64, sizeof(aiocb64));
499 if (result == 0) {
500 do_munge_aiocb_user64_to_user(my_aiocbp: &aiocb64, the_user_aiocbp: &my_aiocb);
501 }
502 } else {
503 struct user32_aiocb aiocb32;
504
505 result = copyin(uap->aiocbp, &aiocb32, sizeof(aiocb32));
506 if (result == 0) {
507 do_munge_aiocb_user32_to_user(my_aiocbp: &aiocb32, the_user_aiocbp: &my_aiocb);
508 }
509 }
510
511 if (result != 0) {
512 result = EAGAIN;
513 goto ExitRoutine;
514 }
515
516 /* NOTE - POSIX standard says a mismatch between the file */
517 /* descriptor passed in and the file descriptor embedded in */
518 /* the aiocb causes unspecified results. We return EBADF in */
519 /* that situation. */
520 if (uap->fd != my_aiocb.aio_fildes) {
521 result = EBADF;
522 goto ExitRoutine;
523 }
524 }
525
526 aio_proc_lock(procp: p);
527 result = do_aio_cancel_locked(p, fd: uap->fd, aiocbp: uap->aiocbp, 0);
528 ASSERT_AIO_PROC_LOCK_OWNED(p);
529 aio_proc_unlock(procp: p);
530
531 if (result != -1) {
532 *retval = result;
533 result = 0;
534 goto ExitRoutine;
535 }
536
537 result = EBADF;
538
539ExitRoutine:
540 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel) | DBG_FUNC_END,
541 VM_KERNEL_ADDRPERM(p), uap->aiocbp, result, 0, 0);
542
543 return result;
544}
545
546
547/*
548 * _aio_close - internal function used to clean up async IO requests for
549 * a file descriptor that is closing.
550 * THIS MAY BLOCK.
551 */
552__private_extern__ void
553_aio_close(proc_t p, int fd)
554{
555 int error;
556
557 /* quick check to see if there are any async IO requests queued up */
558 if (!aio_has_any_work()) {
559 return;
560 }
561
562 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_close) | DBG_FUNC_START,
563 VM_KERNEL_ADDRPERM(p), fd, 0, 0, 0);
564
565 /* cancel all async IO requests on our todo queues for this file descriptor */
566 aio_proc_lock(procp: p);
567 error = do_aio_cancel_locked(p, fd, USER_ADDR_NULL, AIO_CLOSE_WAIT);
568 ASSERT_AIO_PROC_LOCK_OWNED(p);
569 if (error == AIO_NOTCANCELED) {
570 /*
571 * AIO_NOTCANCELED is returned when we find an aio request for this process
572 * and file descriptor on the active async IO queue. Active requests cannot
573 * be cancelled so we must wait for them to complete. We will get a special
574 * wake up call on our channel used to sleep for ALL active requests to
575 * complete. This sleep channel (proc.AIO_CLEANUP_SLEEP_CHAN) is only used
576 * when we must wait for all active aio requests.
577 */
578
579 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_close_sleep) | DBG_FUNC_NONE,
580 VM_KERNEL_ADDRPERM(p), fd, 0, 0, 0);
581
582 while (aio_proc_has_active_requests_for_file(procp: p, fd)) {
583 msleep(chan: &p->AIO_CLEANUP_SLEEP_CHAN, mtx: aio_proc_mutex(procp: p), PRIBIO, wmesg: "aio_close", ts: 0);
584 }
585 }
586
587 aio_proc_unlock(procp: p);
588
589 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_close) | DBG_FUNC_END,
590 VM_KERNEL_ADDRPERM(p), fd, 0, 0, 0);
591}
592
593
594/*
595 * aio_error - return the error status associated with the async IO
596 * request referred to by uap->aiocbp. The error status is the errno
597 * value that would be set by the corresponding IO request (read, wrtie,
598 * fdatasync, or sync).
599 */
600int
601aio_error(proc_t p, struct aio_error_args *uap, int *retval)
602{
603 aio_workq_entry *entryp;
604 int error;
605
606 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error) | DBG_FUNC_START,
607 VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);
608
609 /* see if there are any aios to check */
610 if (!aio_has_any_work()) {
611 return EINVAL;
612 }
613
614 aio_proc_lock(procp: p);
615
616 /* look for a match on our queue of async IO requests that have completed */
617 TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
618 if (entryp->uaiocbp == uap->aiocbp) {
619 ASSERT_AIO_FROM_PROC(entryp, p);
620
621 *retval = entryp->errorval;
622 error = 0;
623
624 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error_val) | DBG_FUNC_NONE,
625 VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
626 goto ExitRoutine;
627 }
628 }
629
630 /* look for a match on our queue of active async IO requests */
631 TAILQ_FOREACH(entryp, &p->p_aio_activeq, aio_proc_link) {
632 if (entryp->uaiocbp == uap->aiocbp) {
633 ASSERT_AIO_FROM_PROC(entryp, p);
634 *retval = EINPROGRESS;
635 error = 0;
636 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error_activeq) | DBG_FUNC_NONE,
637 VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
638 goto ExitRoutine;
639 }
640 }
641
642 error = EINVAL;
643
644ExitRoutine:
645 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error) | DBG_FUNC_END,
646 VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);
647 aio_proc_unlock(procp: p);
648
649 return error;
650}
651
652
653/*
654 * aio_fsync - asynchronously force all IO operations associated
655 * with the file indicated by the file descriptor (uap->aiocbp->aio_fildes) and
656 * queued at the time of the call to the synchronized completion state.
657 * NOTE - we do not support op O_DSYNC at this point since we do not support the
658 * fdatasync() call.
659 */
660int
661aio_fsync(proc_t p, struct aio_fsync_args *uap, int *retval)
662{
663 aio_entry_flags_t fsync_kind;
664 int error;
665
666 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync) | DBG_FUNC_START,
667 VM_KERNEL_ADDRPERM(p), uap->aiocbp, uap->op, 0, 0);
668
669 *retval = 0;
670 /* 0 := O_SYNC for binary backward compatibility with Panther */
671 if (uap->op == O_SYNC || uap->op == 0) {
672 fsync_kind = AIO_FSYNC;
673 } else if (uap->op == O_DSYNC) {
674 fsync_kind = AIO_DSYNC;
675 } else {
676 *retval = -1;
677 error = EINVAL;
678 goto ExitRoutine;
679 }
680
681 error = aio_queue_async_request(procp: p, aiocbp: uap->aiocbp, fsync_kind);
682 if (error != 0) {
683 *retval = -1;
684 }
685
686ExitRoutine:
687 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync) | DBG_FUNC_END,
688 VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);
689
690 return error;
691}
692
693
694/* aio_read - asynchronously read uap->aiocbp->aio_nbytes bytes from the
695 * file descriptor (uap->aiocbp->aio_fildes) into the buffer
696 * (uap->aiocbp->aio_buf).
697 */
698int
699aio_read(proc_t p, struct aio_read_args *uap, int *retval)
700{
701 int error;
702
703 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_read) | DBG_FUNC_START,
704 VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);
705
706 *retval = 0;
707
708 error = aio_queue_async_request(procp: p, aiocbp: uap->aiocbp, AIO_READ);
709 if (error != 0) {
710 *retval = -1;
711 }
712
713 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_read) | DBG_FUNC_END,
714 VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);
715
716 return error;
717}
718
719
720/*
721 * aio_return - return the return status associated with the async IO
722 * request referred to by uap->aiocbp. The return status is the value
723 * that would be returned by corresponding IO request (read, write,
724 * fdatasync, or sync). This is where we release kernel resources
725 * held for async IO call associated with the given aiocb pointer.
726 */
727int
728aio_return(proc_t p, struct aio_return_args *uap, user_ssize_t *retval)
729{
730 aio_workq_entry *entryp;
731 int error = EINVAL;
732
733 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return) | DBG_FUNC_START,
734 VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);
735
736 /* See if there are any entries to check */
737 if (!aio_has_any_work()) {
738 goto ExitRoutine;
739 }
740
741 aio_proc_lock(procp: p);
742 *retval = 0;
743
744 /* look for a match on our queue of async IO requests that have completed */
745 TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
746 ASSERT_AIO_FROM_PROC(entryp, p);
747 if (entryp->uaiocbp == uap->aiocbp) {
748 /* Done and valid for aio_return(), pull it off the list */
749 aio_proc_remove_done_locked(procp: p, entryp);
750
751 *retval = entryp->returnval;
752 error = 0;
753 aio_proc_unlock(procp: p);
754
755 aio_entry_unref(entryp);
756
757 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return_val) | DBG_FUNC_NONE,
758 VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
759 goto ExitRoutine;
760 }
761 }
762
763 /* look for a match on our queue of active async IO requests */
764 TAILQ_FOREACH(entryp, &p->p_aio_activeq, aio_proc_link) {
765 ASSERT_AIO_FROM_PROC(entryp, p);
766 if (entryp->uaiocbp == uap->aiocbp) {
767 error = EINPROGRESS;
768 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return_activeq) | DBG_FUNC_NONE,
769 VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
770 break;
771 }
772 }
773
774 aio_proc_unlock(procp: p);
775
776ExitRoutine:
777 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return) | DBG_FUNC_END,
778 VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);
779
780 return error;
781}
782
783
784/*
785 * _aio_exec - internal function used to clean up async IO requests for
786 * a process that is going away due to exec(). We cancel any async IOs
787 * we can and wait for those already active. We also disable signaling
788 * for cancelled or active aio requests that complete.
789 * This routine MAY block!
790 */
791__private_extern__ void
792_aio_exec(proc_t p)
793{
794 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exec) | DBG_FUNC_START,
795 VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
796
797 _aio_exit(p);
798
799 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exec) | DBG_FUNC_END,
800 VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
801}
802
803
804/*
805 * _aio_exit - internal function used to clean up async IO requests for
806 * a process that is terminating (via exit() or exec()). We cancel any async IOs
807 * we can and wait for those already active. We also disable signaling
808 * for cancelled or active aio requests that complete. This routine MAY block!
809 */
810__private_extern__ void
811_aio_exit(proc_t p)
812{
813 TAILQ_HEAD(, aio_workq_entry) tofree = TAILQ_HEAD_INITIALIZER(tofree);
814 aio_workq_entry *entryp, *tmp;
815 int error;
816
817 /* quick check to see if there are any async IO requests queued up */
818 if (!aio_has_any_work()) {
819 return;
820 }
821
822 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exit) | DBG_FUNC_START,
823 VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
824
825 aio_proc_lock(procp: p);
826
827 /*
828 * cancel async IO requests on the todo work queue and wait for those
829 * already active to complete.
830 */
831 error = do_aio_cancel_locked(p, fd: -1, USER_ADDR_NULL, AIO_EXIT_WAIT);
832 ASSERT_AIO_PROC_LOCK_OWNED(p);
833 if (error == AIO_NOTCANCELED) {
834 /*
835 * AIO_NOTCANCELED is returned when we find an aio request for this process
836 * on the active async IO queue. Active requests cannot be cancelled so we
837 * must wait for them to complete. We will get a special wake up call on
838 * our channel used to sleep for ALL active requests to complete. This sleep
839 * channel (proc.AIO_CLEANUP_SLEEP_CHAN) is only used when we must wait for all
840 * active aio requests.
841 */
842
843 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exit_sleep) | DBG_FUNC_NONE,
844 VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
845
846 while (aio_has_active_requests_for_process(procp: p)) {
847 msleep(chan: &p->AIO_CLEANUP_SLEEP_CHAN, mtx: aio_proc_mutex(procp: p), PRIBIO, wmesg: "aio_exit", ts: 0);
848 }
849 }
850
851 assert(!aio_has_active_requests_for_process(p));
852
853 /* release all aio resources used by this process */
854 TAILQ_FOREACH_SAFE(entryp, &p->p_aio_doneq, aio_proc_link, tmp) {
855 ASSERT_AIO_FROM_PROC(entryp, p);
856
857 aio_proc_remove_done_locked(procp: p, entryp);
858 TAILQ_INSERT_TAIL(&tofree, entryp, aio_proc_link);
859 }
860
861 aio_proc_unlock(procp: p);
862
863 /* free all the entries outside of the aio_proc_lock() */
864 TAILQ_FOREACH_SAFE(entryp, &tofree, aio_proc_link, tmp) {
865 entryp->aio_proc_link.tqe_prev = NULL;
866 aio_entry_unref(entryp);
867 }
868
869 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exit) | DBG_FUNC_END,
870 VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
871}
872
873
874static bool
875should_cancel(aio_workq_entry *entryp, int fd, user_addr_t aiocbp,
876 aio_entry_flags_t reason)
877{
878 if (reason & AIO_EXIT_WAIT) {
879 /* caller is _aio_exit() */
880 return true;
881 }
882 if (fd != entryp->aiocb.aio_fildes) {
883 /* not the file we're looking for */
884 return false;
885 }
886 /*
887 * aio_cancel() or _aio_close() cancel
888 * everything for a given fd when aiocbp is NULL
889 */
890 return aiocbp == USER_ADDR_NULL || entryp->uaiocbp == aiocbp;
891}
892
893/*
894 * do_aio_cancel_locked - cancel async IO requests (if possible). We get called by
895 * aio_cancel, close, and at exit.
896 * There are three modes of operation: 1) cancel all async IOs for a process -
897 * fd is 0 and aiocbp is NULL 2) cancel all async IOs for file descriptor - fd
898 * is > 0 and aiocbp is NULL 3) cancel one async IO associated with the given
899 * aiocbp.
900 * Returns -1 if no matches were found, AIO_CANCELED when we cancelled all
901 * target async IO requests, AIO_NOTCANCELED if we could not cancel all
902 * target async IO requests, and AIO_ALLDONE if all target async IO requests
903 * were already complete.
904 * WARNING - do not deference aiocbp in this routine, it may point to user
905 * land data that has not been copied in (when called from aio_cancel())
906 *
907 * Called with proc locked, and returns the same way.
908 */
909static int
910do_aio_cancel_locked(proc_t p, int fd, user_addr_t aiocbp,
911 aio_entry_flags_t reason)
912{
913 bool multiple_matches = (aiocbp == USER_ADDR_NULL);
914 aio_workq_entry *entryp, *tmp;
915 int result;
916
917 ASSERT_AIO_PROC_LOCK_OWNED(p);
918
919 /* look for a match on our queue of async todo work. */
920again:
921 result = -1;
922 TAILQ_FOREACH_SAFE(entryp, &p->p_aio_activeq, aio_proc_link, tmp) {
923 ASSERT_AIO_FROM_PROC(entryp, p);
924
925 if (!should_cancel(entryp, fd, aiocbp, reason)) {
926 continue;
927 }
928
929 if (reason) {
930 /* mark the entry as blocking close or exit/exec */
931 entryp->flags |= reason;
932 if ((entryp->flags & AIO_EXIT_WAIT) && (entryp->flags & AIO_CLOSE_WAIT)) {
933 panic("Close and exit flags set at the same time");
934 }
935 }
936
937 /* Can only be cancelled if it's still on a work queue */
938 if (aio_entry_try_workq_remove(entryp)) {
939 entryp->errorval = ECANCELED;
940 entryp->returnval = -1;
941
942 /* Now it's officially cancelled. Do the completion */
943 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_async_workq) | DBG_FUNC_NONE,
944 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
945 fd, 0, 0);
946 do_aio_completion_and_unlock(p, entryp);
947
948 aio_proc_lock(procp: p);
949
950 if (multiple_matches) {
951 /*
952 * Restart from the head of the proc active queue since it
953 * may have been changed while we were away doing completion
954 * processing.
955 *
956 * Note that if we found an uncancellable AIO before, we will
957 * either find it again or discover that it's been completed,
958 * so resetting the result will not cause us to return success
959 * despite outstanding AIOs.
960 */
961 goto again;
962 }
963
964 return AIO_CANCELED;
965 }
966
967 /*
968 * It's been taken off the active queue already, i.e. is in flight.
969 * All we can do is ask for notification.
970 */
971 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_activeq) | DBG_FUNC_NONE,
972 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
973 fd, 0, 0);
974
975 result = AIO_NOTCANCELED;
976 if (!multiple_matches) {
977 return result;
978 }
979 }
980
981 /*
982 * if we didn't find any matches on the todo or active queues then look for a
983 * match on our queue of async IO requests that have completed and if found
984 * return AIO_ALLDONE result.
985 *
986 * Proc AIO lock is still held.
987 */
988 if (result == -1) {
989 TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
990 ASSERT_AIO_FROM_PROC(entryp, p);
991 if (should_cancel(entryp, fd, aiocbp, reason)) {
992 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_doneq) | DBG_FUNC_NONE,
993 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
994 fd, 0, 0);
995
996 result = AIO_ALLDONE;
997 if (!multiple_matches) {
998 return result;
999 }
1000 }
1001 }
1002 }
1003
1004 return result;
1005}
1006
1007
1008/*
1009 * aio_suspend - suspend the calling thread until at least one of the async
1010 * IO operations referenced by uap->aiocblist has completed, until a signal
1011 * interrupts the function, or uap->timeoutp time interval (optional) has
1012 * passed.
1013 * Returns 0 if one or more async IOs have completed else -1 and errno is
1014 * set appropriately - EAGAIN if timeout elapses or EINTR if an interrupt
1015 * woke us up.
1016 */
1017int
1018aio_suspend(proc_t p, struct aio_suspend_args *uap, int *retval)
1019{
1020 __pthread_testcancel(presyscall: 1);
1021 return aio_suspend_nocancel(p, (struct aio_suspend_nocancel_args *)uap, retval);
1022}
1023
1024
1025int
1026aio_suspend_nocancel(proc_t p, struct aio_suspend_nocancel_args *uap, int *retval)
1027{
1028 int error;
1029 int i;
1030 uint64_t abstime;
1031 struct user_timespec ts;
1032 aio_workq_entry *entryp;
1033 user_addr_t *aiocbpp;
1034 size_t aiocbpp_size;
1035
1036 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_suspend) | DBG_FUNC_START,
1037 VM_KERNEL_ADDRPERM(p), uap->nent, 0, 0, 0);
1038
1039 *retval = -1;
1040 abstime = 0;
1041 aiocbpp = NULL;
1042
1043 if (!aio_has_any_work()) {
1044 error = EINVAL;
1045 goto ExitThisRoutine;
1046 }
1047
1048 if (uap->nent < 1 || uap->nent > aio_max_requests_per_process ||
1049 os_mul_overflow(sizeof(user_addr_t), uap->nent, &aiocbpp_size)) {
1050 error = EINVAL;
1051 goto ExitThisRoutine;
1052 }
1053
1054 if (uap->timeoutp != USER_ADDR_NULL) {
1055 if (proc_is64bit(p)) {
1056 struct user64_timespec temp;
1057 error = copyin(uap->timeoutp, &temp, sizeof(temp));
1058 if (error == 0) {
1059 ts.tv_sec = (user_time_t)temp.tv_sec;
1060 ts.tv_nsec = (user_long_t)temp.tv_nsec;
1061 }
1062 } else {
1063 struct user32_timespec temp;
1064 error = copyin(uap->timeoutp, &temp, sizeof(temp));
1065 if (error == 0) {
1066 ts.tv_sec = temp.tv_sec;
1067 ts.tv_nsec = temp.tv_nsec;
1068 }
1069 }
1070 if (error != 0) {
1071 error = EAGAIN;
1072 goto ExitThisRoutine;
1073 }
1074
1075 if (ts.tv_sec < 0 || ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) {
1076 error = EINVAL;
1077 goto ExitThisRoutine;
1078 }
1079
1080 nanoseconds_to_absolutetime(nanoseconds: (uint64_t)ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec,
1081 result: &abstime);
1082 clock_absolutetime_interval_to_deadline(abstime, result: &abstime);
1083 }
1084
1085 aiocbpp = (user_addr_t *)kalloc_data(aiocbpp_size, Z_WAITOK);
1086 if (aiocbpp == NULL || aio_copy_in_list(p, uap->aiocblist, aiocbpp, uap->nent)) {
1087 error = EAGAIN;
1088 goto ExitThisRoutine;
1089 }
1090
1091 /* check list of aio requests to see if any have completed */
1092check_for_our_aiocbp:
1093 aio_proc_lock_spin(procp: p);
1094 for (i = 0; i < uap->nent; i++) {
1095 user_addr_t aiocbp;
1096
1097 /* NULL elements are legal so check for 'em */
1098 aiocbp = *(aiocbpp + i);
1099 if (aiocbp == USER_ADDR_NULL) {
1100 continue;
1101 }
1102
1103 /* return immediately if any aio request in the list is done */
1104 TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
1105 ASSERT_AIO_FROM_PROC(entryp, p);
1106 if (entryp->uaiocbp == aiocbp) {
1107 aio_proc_unlock(procp: p);
1108 *retval = 0;
1109 error = 0;
1110 goto ExitThisRoutine;
1111 }
1112 }
1113 }
1114
1115 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_suspend_sleep) | DBG_FUNC_NONE,
1116 VM_KERNEL_ADDRPERM(p), uap->nent, 0, 0, 0);
1117
1118 /*
1119 * wait for an async IO to complete or a signal fires or timeout expires.
1120 * we return EAGAIN (35) for timeout expiration and EINTR (4) when a signal
1121 * interrupts us. If an async IO completes before a signal fires or our
1122 * timeout expires, we get a wakeup call from aio_work_thread().
1123 */
1124
1125 error = msleep1(chan: &p->AIO_SUSPEND_SLEEP_CHAN, mtx: aio_proc_mutex(procp: p),
1126 PCATCH | PWAIT | PDROP, wmesg: "aio_suspend", timo: abstime);
1127 if (error == 0) {
1128 /*
1129 * got our wakeup call from aio_work_thread().
1130 * Since we can get a wakeup on this channel from another thread in the
1131 * same process we head back up to make sure this is for the correct aiocbp.
1132 * If it is the correct aiocbp we will return from where we do the check
1133 * (see entryp->uaiocbp == aiocbp after check_for_our_aiocbp label)
1134 * else we will fall out and just sleep again.
1135 */
1136 goto check_for_our_aiocbp;
1137 } else if (error == EWOULDBLOCK) {
1138 /* our timeout expired */
1139 error = EAGAIN;
1140 } else {
1141 /* we were interrupted */
1142 error = EINTR;
1143 }
1144
1145ExitThisRoutine:
1146 if (aiocbpp != NULL) {
1147 kfree_data(aiocbpp, aiocbpp_size);
1148 }
1149
1150 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_suspend) | DBG_FUNC_END,
1151 VM_KERNEL_ADDRPERM(p), uap->nent, error, 0, 0);
1152
1153 return error;
1154}
1155
1156
1157/* aio_write - asynchronously write uap->aiocbp->aio_nbytes bytes to the
1158 * file descriptor (uap->aiocbp->aio_fildes) from the buffer
1159 * (uap->aiocbp->aio_buf).
1160 */
1161
1162int
1163aio_write(proc_t p, struct aio_write_args *uap, int *retval __unused)
1164{
1165 int error;
1166
1167 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_write) | DBG_FUNC_START,
1168 VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);
1169
1170 error = aio_queue_async_request(procp: p, aiocbp: uap->aiocbp, AIO_WRITE);
1171
1172 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_write) | DBG_FUNC_END,
1173 VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);
1174
1175 return error;
1176}
1177
1178
1179static int
1180aio_copy_in_list(proc_t procp, user_addr_t aiocblist, user_addr_t *aiocbpp,
1181 int nent)
1182{
1183 int result;
1184
1185 /* copyin our aiocb pointers from list */
1186 result = copyin(aiocblist, aiocbpp,
1187 proc_is64bit(procp) ? (nent * sizeof(user64_addr_t))
1188 : (nent * sizeof(user32_addr_t)));
1189 if (result) {
1190 return result;
1191 }
1192
1193 /*
1194 * We depend on a list of user_addr_t's so we need to
1195 * munge and expand when these pointers came from a
1196 * 32-bit process
1197 */
1198 if (!proc_is64bit(procp)) {
1199 /* copy from last to first to deal with overlap */
1200 user32_addr_t *my_ptrp = ((user32_addr_t *)aiocbpp) + (nent - 1);
1201 user_addr_t *my_addrp = aiocbpp + (nent - 1);
1202
1203 for (int i = 0; i < nent; i++, my_ptrp--, my_addrp--) {
1204 *my_addrp = (user_addr_t) (*my_ptrp);
1205 }
1206 }
1207
1208 return 0;
1209}
1210
1211
1212static int
1213aio_copy_in_sigev(proc_t procp, user_addr_t sigp, struct user_sigevent *sigev)
1214{
1215 int result = 0;
1216
1217 if (sigp == USER_ADDR_NULL) {
1218 goto out;
1219 }
1220
1221 /*
1222 * We need to munge aio_sigevent since it contains pointers.
1223 * Since we do not know if sigev_value is an int or a ptr we do
1224 * NOT cast the ptr to a user_addr_t. This means if we send
1225 * this info back to user space we need to remember sigev_value
1226 * was not expanded for the 32-bit case.
1227 *
1228 * Notes: This does NOT affect us since we don't support
1229 * sigev_value yet in the aio context.
1230 */
1231 if (proc_is64bit(procp)) {
1232#if __LP64__
1233 struct user64_sigevent sigevent64;
1234
1235 result = copyin(sigp, &sigevent64, sizeof(sigevent64));
1236 if (result == 0) {
1237 sigev->sigev_notify = sigevent64.sigev_notify;
1238 sigev->sigev_signo = sigevent64.sigev_signo;
1239 sigev->sigev_value.size_equivalent.sival_int = sigevent64.sigev_value.size_equivalent.sival_int;
1240 sigev->sigev_notify_function = sigevent64.sigev_notify_function;
1241 sigev->sigev_notify_attributes = sigevent64.sigev_notify_attributes;
1242 }
1243#else
1244 panic("64bit process on 32bit kernel is not supported");
1245#endif
1246 } else {
1247 struct user32_sigevent sigevent32;
1248
1249 result = copyin(sigp, &sigevent32, sizeof(sigevent32));
1250 if (result == 0) {
1251 sigev->sigev_notify = sigevent32.sigev_notify;
1252 sigev->sigev_signo = sigevent32.sigev_signo;
1253 sigev->sigev_value.size_equivalent.sival_int = sigevent32.sigev_value.sival_int;
1254 sigev->sigev_notify_function = CAST_USER_ADDR_T(sigevent32.sigev_notify_function);
1255 sigev->sigev_notify_attributes = CAST_USER_ADDR_T(sigevent32.sigev_notify_attributes);
1256 }
1257 }
1258
1259 if (result != 0) {
1260 result = EAGAIN;
1261 }
1262
1263out:
1264 return result;
1265}
1266
1267/*
1268 * validate user_sigevent. at this point we only support
1269 * sigev_notify equal to SIGEV_SIGNAL or SIGEV_NONE. this means
1270 * sigev_value, sigev_notify_function, and sigev_notify_attributes
1271 * are ignored, since SIGEV_THREAD is unsupported. This is consistent
1272 * with no [RTS] (RalTime Signal) option group support.
1273 */
1274static int
1275aio_sigev_validate(const struct user_sigevent *sigev)
1276{
1277 switch (sigev->sigev_notify) {
1278 case SIGEV_SIGNAL:
1279 {
1280 int signum;
1281
1282 /* make sure we have a valid signal number */
1283 signum = sigev->sigev_signo;
1284 if (signum <= 0 || signum >= NSIG ||
1285 signum == SIGKILL || signum == SIGSTOP) {
1286 return EINVAL;
1287 }
1288 }
1289 break;
1290
1291 case SIGEV_NONE:
1292 break;
1293
1294 case SIGEV_THREAD:
1295 /* Unsupported [RTS] */
1296
1297 default:
1298 return EINVAL;
1299 }
1300
1301 return 0;
1302}
1303
1304
1305/*
1306 * aio_try_enqueue_work_locked
1307 *
1308 * Queue up the entry on the aio asynchronous work queue in priority order
1309 * based on the relative priority of the request. We calculate the relative
1310 * priority using the nice value of the caller and the value
1311 *
1312 * Parameters: procp Process queueing the I/O
1313 * entryp The work queue entry being queued
1314 * leader The work leader if any
1315 *
1316 * Returns: Wether the enqueue was successful
1317 *
1318 * Notes: This function is used for both lio_listio and aio
1319 *
1320 * XXX: At some point, we may have to consider thread priority
1321 * rather than process priority, but we don't maintain the
1322 * adjusted priority for threads the POSIX way.
1323 *
1324 * Called with proc locked.
1325 */
1326static bool
1327aio_try_enqueue_work_locked(proc_t procp, aio_workq_entry *entryp,
1328 aio_workq_entry *leader)
1329{
1330 aio_workq_t queue = aio_entry_workq(entryp);
1331
1332 ASSERT_AIO_PROC_LOCK_OWNED(procp);
1333
1334 /* Onto proc queue */
1335 if (!aio_try_proc_insert_active_locked(procp, entryp)) {
1336 return false;
1337 }
1338
1339 if (leader) {
1340 aio_entry_ref(entryp: leader); /* consumed in do_aio_completion_and_unlock */
1341 leader->lio_pending++;
1342 entryp->lio_leader = leader;
1343 }
1344
1345 /* And work queue */
1346 aio_entry_ref(entryp); /* consumed in do_aio_completion_and_unlock */
1347 aio_workq_lock_spin(wq: queue);
1348 aio_workq_add_entry_locked(queue, entryp);
1349 waitq_wakeup64_one(waitq: &queue->aioq_waitq, CAST_EVENT64_T(queue),
1350 THREAD_AWAKENED, flags: WAITQ_WAKEUP_DEFAULT);
1351 aio_workq_unlock(wq: queue);
1352
1353 KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_AIO, AIO_work_queued) | DBG_FUNC_START,
1354 VM_KERNEL_ADDRPERM(procp), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1355 entryp->flags, entryp->aiocb.aio_fildes, 0);
1356 KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_AIO, AIO_work_queued) | DBG_FUNC_END,
1357 entryp->aiocb.aio_offset, 0, entryp->aiocb.aio_nbytes, 0, 0);
1358 return true;
1359}
1360
1361
1362/*
1363 * lio_listio - initiate a list of IO requests. We process the list of
1364 * aiocbs either synchronously (mode == LIO_WAIT) or asynchronously
1365 * (mode == LIO_NOWAIT).
1366 *
1367 * The caller gets error and return status for each aiocb in the list
1368 * via aio_error and aio_return. We must keep completed requests until
1369 * released by the aio_return call.
1370 */
1371int
1372lio_listio(proc_t p, struct lio_listio_args *uap, int *retval __unused)
1373{
1374 aio_workq_entry *entries[AIO_LISTIO_MAX] = { };
1375 user_addr_t aiocbpp[AIO_LISTIO_MAX];
1376 struct user_sigevent aiosigev = { };
1377 int result = 0;
1378 int lio_count = 0;
1379
1380 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_listio) | DBG_FUNC_START,
1381 VM_KERNEL_ADDRPERM(p), uap->nent, uap->mode, 0, 0);
1382
1383 if (!(uap->mode == LIO_NOWAIT || uap->mode == LIO_WAIT)) {
1384 result = EINVAL;
1385 goto ExitRoutine;
1386 }
1387
1388 if (uap->nent < 1 || uap->nent > AIO_LISTIO_MAX) {
1389 result = EINVAL;
1390 goto ExitRoutine;
1391 }
1392
1393 /*
1394 * Use sigevent passed in to lio_listio for each of our calls, but
1395 * only do completion notification after the last request completes.
1396 */
1397 if (uap->sigp != USER_ADDR_NULL) {
1398 result = aio_copy_in_sigev(procp: p, sigp: uap->sigp, sigev: &aiosigev);
1399 if (result) {
1400 goto ExitRoutine;
1401 }
1402 result = aio_sigev_validate(sigev: &aiosigev);
1403 if (result) {
1404 goto ExitRoutine;
1405 }
1406 }
1407
1408 if (aio_copy_in_list(procp: p, aiocblist: uap->aiocblist, aiocbpp, nent: uap->nent)) {
1409 result = EAGAIN;
1410 goto ExitRoutine;
1411 }
1412
1413 /*
1414 * allocate/parse all entries
1415 */
1416 for (int i = 0; i < uap->nent; i++) {
1417 aio_workq_entry *entryp;
1418
1419 /* NULL elements are legal so check for 'em */
1420 if (aiocbpp[i] == USER_ADDR_NULL) {
1421 continue;
1422 }
1423
1424 entryp = aio_create_queue_entry(procp: p, aiocbp: aiocbpp[i], AIO_LIO);
1425 if (entryp == NULL) {
1426 result = EAGAIN;
1427 goto ExitRoutine;
1428 }
1429
1430 /*
1431 * This refcount is cleaned up on exit if the entry
1432 * isn't submitted
1433 */
1434 entries[lio_count++] = entryp;
1435 if (uap->mode == LIO_NOWAIT) {
1436 /* Set signal hander, if any */
1437 entryp->aiocb.aio_sigevent = aiosigev;
1438 }
1439 }
1440
1441 if (lio_count == 0) {
1442 /* There's nothing to submit */
1443 goto ExitRoutine;
1444 }
1445
1446 /*
1447 * Past this point we're commited and will not bail out
1448 *
1449 * - keep a reference on the leader for LIO_WAIT
1450 * - perform the submissions and optionally wait
1451 */
1452
1453 aio_workq_entry *leader = entries[0];
1454 if (uap->mode == LIO_WAIT) {
1455 aio_entry_ref(entryp: leader); /* consumed below */
1456 }
1457
1458 aio_proc_lock_spin(procp: p);
1459
1460 for (int i = 0; i < lio_count; i++) {
1461 if (aio_try_enqueue_work_locked(procp: p, entryp: entries[i], leader)) {
1462 entries[i] = NULL; /* the entry was submitted */
1463 } else {
1464 result = EAGAIN;
1465 }
1466 }
1467
1468 if (uap->mode == LIO_WAIT && result == 0) {
1469 leader->flags |= AIO_LIO_WAIT;
1470
1471 while (leader->lio_pending) {
1472 /* If we were interrupted, fail out (even if all finished) */
1473 if (msleep(chan: leader, mtx: aio_proc_mutex(procp: p),
1474 PCATCH | PRIBIO | PSPIN, wmesg: "lio_listio", ts: 0) != 0) {
1475 result = EINTR;
1476 break;
1477 }
1478 }
1479
1480 leader->flags &= ~AIO_LIO_WAIT;
1481 }
1482
1483 aio_proc_unlock(procp: p);
1484
1485 if (uap->mode == LIO_WAIT) {
1486 aio_entry_unref(entryp: leader);
1487 }
1488
1489ExitRoutine:
1490 /* Consume unsubmitted entries */
1491 for (int i = 0; i < lio_count; i++) {
1492 if (entries[i]) {
1493 aio_entry_unref(entryp: entries[i]);
1494 }
1495 }
1496
1497 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_listio) | DBG_FUNC_END,
1498 VM_KERNEL_ADDRPERM(p), result, 0, 0, 0);
1499
1500 return result;
1501}
1502
1503
1504/*
1505 * aio worker thread. this is where all the real work gets done.
1506 * we get a wake up call on sleep channel &aio_anchor.aio_async_workq
1507 * after new work is queued up.
1508 */
1509__attribute__((noreturn))
1510static void
1511aio_work_thread(void *arg __unused, wait_result_t wr __unused)
1512{
1513 aio_workq_entry *entryp;
1514 int error;
1515 vm_map_t currentmap;
1516 vm_map_t oldmap = VM_MAP_NULL;
1517 task_t oldaiotask = TASK_NULL;
1518 struct uthread *uthreadp = NULL;
1519 proc_t p = NULL;
1520
1521 for (;;) {
1522 /*
1523 * returns with the entry ref'ed.
1524 * sleeps until work is available.
1525 */
1526 entryp = aio_get_some_work();
1527 p = entryp->procp;
1528
1529 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_worker_thread) | DBG_FUNC_START,
1530 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1531 entryp->flags, 0, 0);
1532
1533 /*
1534 * Assume the target's address space identity for the duration
1535 * of the IO. Note: don't need to have the entryp locked,
1536 * because the proc and map don't change until it's freed.
1537 */
1538 currentmap = get_task_map(proc_task(current_proc()));
1539 if (currentmap != entryp->aio_map) {
1540 uthreadp = (struct uthread *) current_uthread();
1541 oldaiotask = uthreadp->uu_aio_task;
1542 /*
1543 * workq entries at this stage cause _aio_exec() and _aio_exit() to
1544 * block until we hit `do_aio_completion_and_unlock()` below,
1545 * which means that it is safe to dereference p->task without
1546 * holding a lock or taking references.
1547 */
1548 uthreadp->uu_aio_task = proc_task(p);
1549 oldmap = vm_map_switch(map: entryp->aio_map);
1550 }
1551
1552 if ((entryp->flags & AIO_READ) != 0) {
1553 error = do_aio_read(entryp);
1554 } else if ((entryp->flags & AIO_WRITE) != 0) {
1555 uthreadp = (struct uthread *)current_uthread();
1556 uthread_t context_uthreadp = get_bsdthread_info(vfs_context_thread(ctx: &entryp->context));
1557
1558 if ((context_uthreadp && (context_uthreadp->uu_flag & UT_FS_BLKSIZE_NOCACHE_WRITES)) ||
1559 os_atomic_load(&p->p_vfs_iopolicy, relaxed) & P_VFS_IOPOLICY_NOCACHE_WRITE_FS_BLKSIZE) {
1560 uthreadp->uu_flag |= UT_FS_BLKSIZE_NOCACHE_WRITES;
1561 }
1562
1563 error = do_aio_write(entryp);
1564
1565 uthreadp->uu_flag &= ~UT_FS_BLKSIZE_NOCACHE_WRITES;
1566 } else if ((entryp->flags & (AIO_FSYNC | AIO_DSYNC)) != 0) {
1567 error = do_aio_fsync(entryp);
1568 } else {
1569 error = EINVAL;
1570 }
1571
1572 /* Restore old map */
1573 if (currentmap != entryp->aio_map) {
1574 vm_map_switch(map: oldmap);
1575 uthreadp->uu_aio_task = oldaiotask;
1576 }
1577
1578 /* liberate unused map */
1579 vm_map_deallocate(map: entryp->aio_map);
1580 entryp->aio_map = VM_MAP_NULL;
1581
1582 KERNEL_DEBUG(SDDBG_CODE(DBG_BSD_AIO, AIO_worker_thread) | DBG_FUNC_END,
1583 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1584 entryp->errorval, entryp->returnval, 0);
1585
1586 /* we're done with the IO request so pop it off the active queue and */
1587 /* push it on the done queue */
1588 aio_proc_lock(procp: p);
1589 entryp->errorval = error;
1590 do_aio_completion_and_unlock(p, entryp);
1591 }
1592}
1593
1594
1595/*
1596 * aio_get_some_work - get the next async IO request that is ready to be executed.
1597 * aio_fsync complicates matters a bit since we cannot do the fsync until all async
1598 * IO requests at the time the aio_fsync call came in have completed.
1599 * NOTE - AIO_LOCK must be held by caller
1600 */
1601static aio_workq_entry *
1602aio_get_some_work(void)
1603{
1604 aio_workq_entry *entryp = NULL;
1605 aio_workq_t queue = NULL;
1606
1607 /* Just one queue for the moment. In the future there will be many. */
1608 queue = &aio_anchor.aio_async_workqs[0];
1609 aio_workq_lock_spin(wq: queue);
1610
1611 /*
1612 * Hold the queue lock.
1613 *
1614 * pop some work off the work queue and add to our active queue
1615 * Always start with the queue lock held.
1616 */
1617 while ((entryp = TAILQ_FIRST(&queue->aioq_entries))) {
1618 /*
1619 * Pull of of work queue. Once it's off, it can't be cancelled,
1620 * so we can take our ref once we drop the queue lock.
1621 */
1622
1623 aio_workq_remove_entry_locked(queue, entryp);
1624
1625 aio_workq_unlock(wq: queue);
1626
1627 /*
1628 * Check if it's an fsync that must be delayed. No need to lock the entry;
1629 * that flag would have been set at initialization.
1630 */
1631 if ((entryp->flags & AIO_FSYNC) != 0) {
1632 /*
1633 * Check for unfinished operations on the same file
1634 * in this proc's queue.
1635 */
1636 aio_proc_lock_spin(procp: entryp->procp);
1637 if (aio_delay_fsync_request(entryp)) {
1638 /* It needs to be delayed. Put it back on the end of the work queue */
1639 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync_delay) | DBG_FUNC_NONE,
1640 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1641 0, 0, 0);
1642
1643 aio_proc_unlock(procp: entryp->procp);
1644
1645 aio_workq_lock_spin(wq: queue);
1646 aio_workq_add_entry_locked(queue, entryp);
1647 continue;
1648 }
1649 aio_proc_unlock(procp: entryp->procp);
1650 }
1651
1652 return entryp;
1653 }
1654
1655 /* We will wake up when someone enqueues something */
1656 waitq_assert_wait64(waitq: &queue->aioq_waitq, CAST_EVENT64_T(queue), THREAD_UNINT, deadline: 0);
1657 aio_workq_unlock(wq: queue);
1658 thread_block(continuation: aio_work_thread);
1659
1660 __builtin_unreachable();
1661}
1662
1663/*
1664 * aio_delay_fsync_request - look to see if this aio_fsync request should be delayed.
1665 * A big, simple hammer: only send it off if it's the most recently filed IO which has
1666 * not been completed.
1667 */
1668static boolean_t
1669aio_delay_fsync_request(aio_workq_entry *entryp)
1670{
1671 if (proc_in_teardown(entryp->procp)) {
1672 /*
1673 * we can't delay FSYNCS when in teardown as it will confuse _aio_exit,
1674 * if it was dequeued, then we must now commit to it
1675 */
1676 return FALSE;
1677 }
1678
1679 if (entryp == TAILQ_FIRST(&entryp->procp->p_aio_activeq)) {
1680 return FALSE;
1681 }
1682
1683 return TRUE;
1684}
1685
1686static aio_workq_entry *
1687aio_create_queue_entry(proc_t procp, user_addr_t aiocbp, aio_entry_flags_t flags)
1688{
1689 aio_workq_entry *entryp;
1690
1691 entryp = zalloc_flags(aio_workq_zonep, Z_WAITOK | Z_ZERO);
1692 entryp->procp = procp;
1693 entryp->uaiocbp = aiocbp;
1694 entryp->flags = flags;
1695 /* consumed in aio_return or _aio_exit */
1696 os_ref_init(&entryp->aio_refcount, &aio_refgrp);
1697
1698 if (proc_is64bit(procp)) {
1699 struct user64_aiocb aiocb64;
1700
1701 if (copyin(aiocbp, &aiocb64, sizeof(aiocb64)) != 0) {
1702 goto error_exit;
1703 }
1704 do_munge_aiocb_user64_to_user(my_aiocbp: &aiocb64, the_user_aiocbp: &entryp->aiocb);
1705 } else {
1706 struct user32_aiocb aiocb32;
1707
1708 if (copyin(aiocbp, &aiocb32, sizeof(aiocb32)) != 0) {
1709 goto error_exit;
1710 }
1711 do_munge_aiocb_user32_to_user(my_aiocbp: &aiocb32, the_user_aiocbp: &entryp->aiocb);
1712 }
1713
1714 /* do some more validation on the aiocb and embedded file descriptor */
1715 if (aio_validate(procp, entryp) != 0) {
1716 goto error_exit;
1717 }
1718
1719 /* get a reference to the user land map in order to keep it around */
1720 entryp->aio_map = get_task_map(proc_task(procp));
1721 vm_map_reference(map: entryp->aio_map);
1722
1723 /* get a reference on the current_thread, which is passed in vfs_context. */
1724 entryp->context = *vfs_context_current();
1725 thread_reference(thread: entryp->context.vc_thread);
1726 kauth_cred_ref(cred: entryp->context.vc_ucred);
1727 return entryp;
1728
1729error_exit:
1730 zfree(aio_workq_zonep, entryp);
1731 return NULL;
1732}
1733
1734
1735/*
1736 * aio_queue_async_request - queue up an async IO request on our work queue then
1737 * wake up one of our worker threads to do the actual work. We get a reference
1738 * to our caller's user land map in order to keep it around while we are
1739 * processing the request.
1740 */
1741static int
1742aio_queue_async_request(proc_t procp, user_addr_t aiocbp,
1743 aio_entry_flags_t flags)
1744{
1745 aio_workq_entry *entryp;
1746 int result;
1747
1748 entryp = aio_create_queue_entry(procp, aiocbp, flags);
1749 if (entryp == NULL) {
1750 result = EAGAIN;
1751 goto error_noalloc;
1752 }
1753
1754 aio_proc_lock_spin(procp);
1755 if (!aio_try_enqueue_work_locked(procp, entryp, NULL)) {
1756 result = EAGAIN;
1757 goto error_exit;
1758 }
1759 aio_proc_unlock(procp);
1760 return 0;
1761
1762error_exit:
1763 /*
1764 * This entry has not been queued up so no worries about
1765 * unlocked state and aio_map
1766 */
1767 aio_proc_unlock(procp);
1768 aio_free_request(entryp);
1769error_noalloc:
1770 return result;
1771}
1772
1773
1774/*
1775 * aio_free_request - remove our reference on the user land map and
1776 * free the work queue entry resources. The entry is off all lists
1777 * and has zero refcount, so no one can have a pointer to it.
1778 */
1779static void
1780aio_free_request(aio_workq_entry *entryp)
1781{
1782 if (entryp->aio_proc_link.tqe_prev || entryp->aio_workq_link.tqe_prev) {
1783 panic("aio_workq_entry %p being freed while still enqueued", entryp);
1784 }
1785
1786 /* remove our reference to the user land map. */
1787 if (VM_MAP_NULL != entryp->aio_map) {
1788 vm_map_deallocate(map: entryp->aio_map);
1789 }
1790
1791 /* remove our reference to thread which enqueued the request */
1792 if (entryp->context.vc_thread) {
1793 thread_deallocate(thread: entryp->context.vc_thread);
1794 }
1795 kauth_cred_unref(&entryp->context.vc_ucred);
1796
1797 zfree(aio_workq_zonep, entryp);
1798}
1799
1800
1801/*
1802 * aio_validate
1803 *
1804 * validate the aiocb passed in by one of the aio syscalls.
1805 */
1806static int
1807aio_validate(proc_t p, aio_workq_entry *entryp)
1808{
1809 struct fileproc *fp;
1810 int flag;
1811 int result;
1812
1813 result = 0;
1814
1815 if ((entryp->flags & AIO_LIO) != 0) {
1816 if (entryp->aiocb.aio_lio_opcode == LIO_READ) {
1817 entryp->flags |= AIO_READ;
1818 } else if (entryp->aiocb.aio_lio_opcode == LIO_WRITE) {
1819 entryp->flags |= AIO_WRITE;
1820 } else if (entryp->aiocb.aio_lio_opcode == LIO_NOP) {
1821 return 0;
1822 } else {
1823 return EINVAL;
1824 }
1825 }
1826
1827 flag = FREAD;
1828 if ((entryp->flags & (AIO_WRITE | AIO_FSYNC | AIO_DSYNC)) != 0) {
1829 flag = FWRITE;
1830 }
1831
1832 if ((entryp->flags & (AIO_READ | AIO_WRITE)) != 0) {
1833 if (entryp->aiocb.aio_nbytes > INT_MAX ||
1834 entryp->aiocb.aio_buf == USER_ADDR_NULL ||
1835 entryp->aiocb.aio_offset < 0) {
1836 return EINVAL;
1837 }
1838 }
1839
1840 result = aio_sigev_validate(sigev: &entryp->aiocb.aio_sigevent);
1841 if (result) {
1842 return result;
1843 }
1844
1845 /* validate the file descriptor and that the file was opened
1846 * for the appropriate read / write access.
1847 */
1848 proc_fdlock(p);
1849
1850 fp = fp_get_noref_locked(p, fd: entryp->aiocb.aio_fildes);
1851 if (fp == NULL) {
1852 result = EBADF;
1853 } else if ((fp->fp_glob->fg_flag & flag) == 0) {
1854 /* we don't have read or write access */
1855 result = EBADF;
1856 } else if (FILEGLOB_DTYPE(fp->fp_glob) != DTYPE_VNODE) {
1857 /* this is not a file */
1858 result = ESPIPE;
1859 } else {
1860 fp->fp_flags |= FP_AIOISSUED;
1861 }
1862
1863 proc_fdunlock(p);
1864
1865 return result;
1866}
1867
1868/*
1869 * do_aio_completion_and_unlock. Handle async IO completion.
1870 */
1871static void
1872do_aio_completion_and_unlock(proc_t p, aio_workq_entry *entryp)
1873{
1874 aio_workq_entry *leader = entryp->lio_leader;
1875 int lio_pending = 0;
1876 bool do_signal = false;
1877
1878 ASSERT_AIO_PROC_LOCK_OWNED(p);
1879
1880 aio_proc_move_done_locked(procp: p, entryp);
1881
1882 if (leader) {
1883 lio_pending = --leader->lio_pending;
1884 if (lio_pending < 0) {
1885 panic("lio_pending accounting mistake");
1886 }
1887 if (lio_pending == 0 && (leader->flags & AIO_LIO_WAIT)) {
1888 wakeup(chan: leader);
1889 }
1890 entryp->lio_leader = NULL; /* no dangling pointers please */
1891 }
1892
1893 /*
1894 * need to handle case where a process is trying to exit, exec, or
1895 * close and is currently waiting for active aio requests to complete.
1896 * If AIO_CLEANUP_WAIT is set then we need to look to see if there are any
1897 * other requests in the active queue for this process. If there are
1898 * none then wakeup using the AIO_CLEANUP_SLEEP_CHAN tsleep channel.
1899 * If there are some still active then do nothing - we only want to
1900 * wakeup when all active aio requests for the process are complete.
1901 */
1902 if (__improbable(entryp->flags & AIO_EXIT_WAIT)) {
1903 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wait) | DBG_FUNC_NONE,
1904 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1905 0, 0, 0);
1906
1907 if (!aio_has_active_requests_for_process(procp: p)) {
1908 /*
1909 * no active aio requests for this process, continue exiting. In this
1910 * case, there should be no one else waiting ont he proc in AIO...
1911 */
1912 wakeup_one(chan: (caddr_t)&p->AIO_CLEANUP_SLEEP_CHAN);
1913
1914 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wake) | DBG_FUNC_NONE,
1915 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1916 0, 0, 0);
1917 }
1918 } else if (entryp->aiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
1919 /*
1920 * If this was the last request in the group, or not part of
1921 * a group, and that a signal is desired, send one.
1922 */
1923 do_signal = (lio_pending == 0);
1924 }
1925
1926 if (__improbable(entryp->flags & AIO_CLOSE_WAIT)) {
1927 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wait) | DBG_FUNC_NONE,
1928 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1929 0, 0, 0);
1930
1931 if (!aio_proc_has_active_requests_for_file(procp: p, fd: entryp->aiocb.aio_fildes)) {
1932 /* Can't wakeup_one(); multiple closes might be in progress. */
1933 wakeup(chan: &p->AIO_CLEANUP_SLEEP_CHAN);
1934
1935 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wake) | DBG_FUNC_NONE,
1936 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1937 0, 0, 0);
1938 }
1939 }
1940
1941 aio_proc_unlock(procp: p);
1942
1943 if (do_signal) {
1944 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_sig) | DBG_FUNC_NONE,
1945 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
1946 entryp->aiocb.aio_sigevent.sigev_signo, 0, 0);
1947
1948 psignal(p, sig: entryp->aiocb.aio_sigevent.sigev_signo);
1949 }
1950
1951 /*
1952 * A thread in aio_suspend() wants to known about completed IOs. If it checked
1953 * the done list before we moved our AIO there, then it already asserted its wait,
1954 * and we can wake it up without holding the lock. If it checked the list after
1955 * we did our move, then it already has seen the AIO that we moved. Herego, we
1956 * can do our wakeup without holding the lock.
1957 */
1958 wakeup(chan: &p->AIO_SUSPEND_SLEEP_CHAN);
1959 KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_suspend_wake) | DBG_FUNC_NONE,
1960 VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp), 0, 0, 0);
1961
1962 aio_entry_unref(entryp); /* see aio_try_enqueue_work_locked */
1963 if (leader) {
1964 aio_entry_unref(entryp: leader); /* see lio_listio */
1965 }
1966}
1967
1968
1969/*
1970 * do_aio_read
1971 */
1972static int
1973do_aio_read(aio_workq_entry *entryp)
1974{
1975 struct proc *p = entryp->procp;
1976 struct fileproc *fp;
1977 int error;
1978
1979 if ((error = fp_lookup(p, fd: entryp->aiocb.aio_fildes, resultfp: &fp, locked: 0))) {
1980 return error;
1981 }
1982
1983 if (fp->fp_glob->fg_flag & FREAD) {
1984 error = dofileread(ctx: &entryp->context, fp,
1985 bufp: entryp->aiocb.aio_buf,
1986 nbyte: entryp->aiocb.aio_nbytes,
1987 offset: entryp->aiocb.aio_offset, FOF_OFFSET,
1988 retval: &entryp->returnval);
1989 } else {
1990 error = EBADF;
1991 }
1992
1993 fp_drop(p, fd: entryp->aiocb.aio_fildes, fp, locked: 0);
1994 return error;
1995}
1996
1997
1998/*
1999 * do_aio_write
2000 */
2001static int
2002do_aio_write(aio_workq_entry *entryp)
2003{
2004 struct proc *p = entryp->procp;
2005 struct fileproc *fp;
2006 int error;
2007
2008 if ((error = fp_lookup(p, fd: entryp->aiocb.aio_fildes, resultfp: &fp, locked: 0))) {
2009 return error;
2010 }
2011
2012 if (fp->fp_glob->fg_flag & FWRITE) {
2013 int flags = 0;
2014
2015 if ((fp->fp_glob->fg_flag & O_APPEND) == 0) {
2016 flags |= FOF_OFFSET;
2017 }
2018
2019 /* NB: tell dofilewrite the offset, and to use the proc cred */
2020 error = dofilewrite(ctx: &entryp->context,
2021 fp,
2022 bufp: entryp->aiocb.aio_buf,
2023 nbyte: entryp->aiocb.aio_nbytes,
2024 offset: entryp->aiocb.aio_offset,
2025 flags,
2026 retval: &entryp->returnval);
2027 } else {
2028 error = EBADF;
2029 }
2030
2031 fp_drop(p, fd: entryp->aiocb.aio_fildes, fp, locked: 0);
2032 return error;
2033}
2034
2035
2036/*
2037 * aio_has_active_requests_for_process - return whether the process has active
2038 * requests pending.
2039 */
2040static bool
2041aio_has_active_requests_for_process(proc_t procp)
2042{
2043 return !TAILQ_EMPTY(&procp->p_aio_activeq);
2044}
2045
2046/*
2047 * Called with the proc locked.
2048 */
2049static bool
2050aio_proc_has_active_requests_for_file(proc_t procp, int fd)
2051{
2052 aio_workq_entry *entryp;
2053
2054 TAILQ_FOREACH(entryp, &procp->p_aio_activeq, aio_proc_link) {
2055 if (entryp->aiocb.aio_fildes == fd) {
2056 return true;
2057 }
2058 }
2059
2060 return false;
2061}
2062
2063
2064/*
2065 * do_aio_fsync
2066 */
2067static int
2068do_aio_fsync(aio_workq_entry *entryp)
2069{
2070 struct proc *p = entryp->procp;
2071 struct vnode *vp;
2072 struct fileproc *fp;
2073 int sync_flag;
2074 int error;
2075
2076 /*
2077 * We are never called unless either AIO_FSYNC or AIO_DSYNC are set.
2078 *
2079 * If AIO_DSYNC is set, we can tell the lower layers that it is OK
2080 * to mark for update the metadata not strictly necessary for data
2081 * retrieval, rather than forcing it to disk.
2082 *
2083 * If AIO_FSYNC is set, we have to also wait for metadata not really
2084 * necessary to data retrival are committed to stable storage (e.g.
2085 * atime, mtime, ctime, etc.).
2086 *
2087 * Metadata necessary for data retrieval ust be committed to stable
2088 * storage in either case (file length, etc.).
2089 */
2090 if (entryp->flags & AIO_FSYNC) {
2091 sync_flag = MNT_WAIT;
2092 } else {
2093 sync_flag = MNT_DWAIT;
2094 }
2095
2096 error = fp_get_ftype(p, fd: entryp->aiocb.aio_fildes, ftype: DTYPE_VNODE, ENOTSUP, fpp: &fp);
2097 if (error != 0) {
2098 entryp->returnval = -1;
2099 return error;
2100 }
2101 vp = fp_get_data(fp);
2102
2103 if ((error = vnode_getwithref(vp)) == 0) {
2104 error = VNOP_FSYNC(vp, waitfor: sync_flag, ctx: &entryp->context);
2105
2106 (void)vnode_put(vp);
2107 } else {
2108 entryp->returnval = -1;
2109 }
2110
2111 fp_drop(p, fd: entryp->aiocb.aio_fildes, fp, locked: 0);
2112 return error;
2113}
2114
2115
2116/*
2117 * is_already_queued - runs through our queues to see if the given
2118 * aiocbp / process is there. Returns TRUE if there is a match
2119 * on any of our aio queues.
2120 *
2121 * Called with proc aio lock held (can be held spin)
2122 */
2123static boolean_t
2124is_already_queued(proc_t procp, user_addr_t aiocbp)
2125{
2126 aio_workq_entry *entryp;
2127 boolean_t result;
2128
2129 result = FALSE;
2130
2131 /* look for matches on our queue of async IO requests that have completed */
2132 TAILQ_FOREACH(entryp, &procp->p_aio_doneq, aio_proc_link) {
2133 if (aiocbp == entryp->uaiocbp) {
2134 result = TRUE;
2135 goto ExitThisRoutine;
2136 }
2137 }
2138
2139 /* look for matches on our queue of active async IO requests */
2140 TAILQ_FOREACH(entryp, &procp->p_aio_activeq, aio_proc_link) {
2141 if (aiocbp == entryp->uaiocbp) {
2142 result = TRUE;
2143 goto ExitThisRoutine;
2144 }
2145 }
2146
2147ExitThisRoutine:
2148 return result;
2149}
2150
2151
2152/*
2153 * aio initialization
2154 */
2155__private_extern__ void
2156aio_init(void)
2157{
2158 for (int i = 0; i < AIO_NUM_WORK_QUEUES; i++) {
2159 aio_workq_init(wq: &aio_anchor.aio_async_workqs[i]);
2160 }
2161
2162 _aio_create_worker_threads(num: aio_worker_threads);
2163}
2164
2165
2166/*
2167 * aio worker threads created here.
2168 */
2169__private_extern__ void
2170_aio_create_worker_threads(int num)
2171{
2172 int i;
2173
2174 /* create some worker threads to handle the async IO requests */
2175 for (i = 0; i < num; i++) {
2176 thread_t myThread;
2177
2178 if (KERN_SUCCESS != kernel_thread_start(continuation: aio_work_thread, NULL, new_thread: &myThread)) {
2179 printf("%s - failed to create a work thread \n", __FUNCTION__);
2180 } else {
2181 thread_deallocate(thread: myThread);
2182 }
2183 }
2184}
2185
2186/*
2187 * Return the current activation utask
2188 */
2189task_t
2190get_aiotask(void)
2191{
2192 return current_uthread()->uu_aio_task;
2193}
2194
2195
2196/*
2197 * In the case of an aiocb from a
2198 * 32-bit process we need to expand some longs and pointers to the correct
2199 * sizes in order to let downstream code always work on the same type of
2200 * aiocb (in our case that is a user_aiocb)
2201 */
2202static void
2203do_munge_aiocb_user32_to_user(struct user32_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp)
2204{
2205 the_user_aiocbp->aio_fildes = my_aiocbp->aio_fildes;
2206 the_user_aiocbp->aio_offset = my_aiocbp->aio_offset;
2207 the_user_aiocbp->aio_buf = CAST_USER_ADDR_T(my_aiocbp->aio_buf);
2208 the_user_aiocbp->aio_nbytes = my_aiocbp->aio_nbytes;
2209 the_user_aiocbp->aio_reqprio = my_aiocbp->aio_reqprio;
2210 the_user_aiocbp->aio_lio_opcode = my_aiocbp->aio_lio_opcode;
2211
2212 /* special case here. since we do not know if sigev_value is an */
2213 /* int or a ptr we do NOT cast the ptr to a user_addr_t. This */
2214 /* means if we send this info back to user space we need to remember */
2215 /* sigev_value was not expanded for the 32-bit case. */
2216 /* NOTE - this does NOT affect us since we don't support sigev_value */
2217 /* yet in the aio context. */
2218 //LP64
2219 the_user_aiocbp->aio_sigevent.sigev_notify = my_aiocbp->aio_sigevent.sigev_notify;
2220 the_user_aiocbp->aio_sigevent.sigev_signo = my_aiocbp->aio_sigevent.sigev_signo;
2221 the_user_aiocbp->aio_sigevent.sigev_value.size_equivalent.sival_int =
2222 my_aiocbp->aio_sigevent.sigev_value.sival_int;
2223 the_user_aiocbp->aio_sigevent.sigev_notify_function =
2224 CAST_USER_ADDR_T(my_aiocbp->aio_sigevent.sigev_notify_function);
2225 the_user_aiocbp->aio_sigevent.sigev_notify_attributes =
2226 CAST_USER_ADDR_T(my_aiocbp->aio_sigevent.sigev_notify_attributes);
2227}
2228
2229/* Similar for 64-bit user process, so that we don't need to satisfy
2230 * the alignment constraints of the original user64_aiocb
2231 */
2232#if !__LP64__
2233__dead2
2234#endif
2235static void
2236do_munge_aiocb_user64_to_user(struct user64_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp)
2237{
2238#if __LP64__
2239 the_user_aiocbp->aio_fildes = my_aiocbp->aio_fildes;
2240 the_user_aiocbp->aio_offset = my_aiocbp->aio_offset;
2241 the_user_aiocbp->aio_buf = my_aiocbp->aio_buf;
2242 the_user_aiocbp->aio_nbytes = my_aiocbp->aio_nbytes;
2243 the_user_aiocbp->aio_reqprio = my_aiocbp->aio_reqprio;
2244 the_user_aiocbp->aio_lio_opcode = my_aiocbp->aio_lio_opcode;
2245
2246 the_user_aiocbp->aio_sigevent.sigev_notify = my_aiocbp->aio_sigevent.sigev_notify;
2247 the_user_aiocbp->aio_sigevent.sigev_signo = my_aiocbp->aio_sigevent.sigev_signo;
2248 the_user_aiocbp->aio_sigevent.sigev_value.size_equivalent.sival_int =
2249 my_aiocbp->aio_sigevent.sigev_value.size_equivalent.sival_int;
2250 the_user_aiocbp->aio_sigevent.sigev_notify_function =
2251 my_aiocbp->aio_sigevent.sigev_notify_function;
2252 the_user_aiocbp->aio_sigevent.sigev_notify_attributes =
2253 my_aiocbp->aio_sigevent.sigev_notify_attributes;
2254#else
2255#pragma unused(my_aiocbp, the_user_aiocbp)
2256 panic("64bit process on 32bit kernel is not supported");
2257#endif
2258}
2259