2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. John S. Dyson's name may not be used to endorse or promote products
10 * derived from this software without specific prior written permission.
12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
13 * bad that happens because of using this software isn't the responsibility
14 * of the author. This software is distributed AS-IS.
18 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
24 #include "opt_compat.h"
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
40 #include <sys/limits.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
66 #include <machine/atomic.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
78 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
79 * overflow. (XXX will be removed soon.)
81 static u_long jobrefid;
84 * Counter for aio_fsync.
86 static uint64_t jobseqno;
88 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC 32
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
97 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
101 #define MAX_BUF_AIO 16
104 FEATURE(aio, "Asynchronous I/O");
105 SYSCTL_DECL(_p1003_1b);
107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
109 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
110 "Async IO management");
112 static int enable_aio_unsafe = 0;
113 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
114 "Permit asynchronous IO on all file types, not just known-safe types");
116 static unsigned int unsafe_warningcnt = 1;
117 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
118 &unsafe_warningcnt, 0,
119 "Warnings that will be triggered upon failed IO requests on unsafe files");
121 static int max_aio_procs = MAX_AIO_PROCS;
122 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
123 "Maximum number of kernel processes to use for handling async IO ");
125 static int num_aio_procs = 0;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
127 "Number of presently active kernel processes for async IO");
130 * The code will adjust the actual number of AIO processes towards this
131 * number when it gets a chance.
133 static int target_aio_procs = TARGET_AIO_PROCS;
134 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
136 "Preferred number of ready kernel processes for async IO");
138 static int max_queue_count = MAX_AIO_QUEUE;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
140 "Maximum number of aio requests to queue, globally");
142 static int num_queue_count = 0;
143 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
144 "Number of queued aio requests");
146 static int num_buf_aio = 0;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
148 "Number of aio requests presently handled by the buf subsystem");
150 /* Number of async I/O processes in the process of being started */
151 /* XXX This should be local to aio_aqueue() */
152 static int num_aio_resv_start = 0;
154 static int aiod_lifetime;
155 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
156 "Maximum lifetime for idle aiod");
158 static int max_aio_per_proc = MAX_AIO_PER_PROC;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
161 "Maximum active aio requests per process (stored in the process)");
163 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
164 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
165 &max_aio_queue_per_proc, 0,
166 "Maximum queued aio requests per process (stored in the process)");
168 static int max_buf_aio = MAX_BUF_AIO;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
170 "Maximum buf aio requests per process (stored in the process)");
172 static int aio_listio_max = AIO_LISTIO_MAX;
173 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
174 CTLFLAG_RDTUN | CTLFLAG_CAPRD, &aio_listio_max, 0,
175 "Maximum aio requests for a single lio_listio call");
177 #ifdef COMPAT_FREEBSD6
178 typedef struct oaiocb {
179 int aio_fildes; /* File descriptor */
180 off_t aio_offset; /* File offset for I/O */
181 volatile void *aio_buf; /* I/O buffer in process space */
182 size_t aio_nbytes; /* Number of bytes for I/O */
183 struct osigevent aio_sigevent; /* Signal to deliver */
184 int aio_lio_opcode; /* LIO opcode */
185 int aio_reqprio; /* Request priority -- ignored */
186 struct __aiocb_private _aiocb_private;
191 * Below is a key of locks used to protect each member of struct kaiocb
192 * aioliojob and kaioinfo and any backends.
194 * * - need not protected
195 * a - locked by kaioinfo lock
196 * b - locked by backend lock, the backend lock can be null in some cases,
197 * for example, BIO belongs to this type, in this case, proc lock is
199 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
203 * If the routine that services an AIO request blocks while running in an
204 * AIO kernel process it can starve other I/O requests. BIO requests
205 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
206 * processes at all. Socket I/O requests use a separate pool of
207 * kprocs and also force non-blocking I/O. Other file I/O requests
208 * use the generic fo_read/fo_write operations which can block. The
209 * fsync and mlock operations can also block while executing. Ideally
210 * none of these requests would block while executing.
212 * Note that the service routines cannot toggle O_NONBLOCK in the file
213 * structure directly while handling a request due to races with
218 #define KAIOCB_QUEUEING 0x01
219 #define KAIOCB_CANCELLED 0x02
220 #define KAIOCB_CANCELLING 0x04
221 #define KAIOCB_CHECKSYNC 0x08
222 #define KAIOCB_CLEARED 0x10
223 #define KAIOCB_FINISHED 0x20
228 #define AIOP_FREE 0x1 /* proc on free queue */
231 int aioprocflags; /* (c) AIO proc flags */
232 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
233 struct proc *aioproc; /* (*) the AIO proc */
237 * data-structure for lio signal management
240 int lioj_flags; /* (a) listio flags */
241 int lioj_count; /* (a) listio flags */
242 int lioj_finished_count; /* (a) listio flags */
243 struct sigevent lioj_signal; /* (a) signal on all I/O done */
244 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
245 struct knlist klist; /* (a) list of knotes */
246 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
249 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
250 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
251 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
254 * per process aio data structure
257 struct mtx kaio_mtx; /* the lock to protect this struct */
258 int kaio_flags; /* (a) per process kaio flags */
259 int kaio_maxactive_count; /* (*) maximum number of AIOs */
260 int kaio_active_count; /* (c) number of currently used AIOs */
261 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
262 int kaio_count; /* (a) size of AIO queue */
263 int kaio_ballowed_count; /* (*) maximum number of buffers */
264 int kaio_buffer_count; /* (a) number of physio buffers */
265 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */
266 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */
267 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
268 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */
269 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */
270 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */
271 struct task kaio_task; /* (*) task to kick aio processes */
272 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */
275 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
276 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
277 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
278 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
280 #define KAIO_RUNDOWN 0x1 /* process is being run down */
281 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */
284 * Operations used to interact with userland aio control blocks.
285 * Different ABIs provide their own operations.
288 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
289 long (*fetch_status)(struct aiocb *ujob);
290 long (*fetch_error)(struct aiocb *ujob);
291 int (*store_status)(struct aiocb *ujob, long status);
292 int (*store_error)(struct aiocb *ujob, long error);
293 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
294 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
297 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */
298 static struct sema aio_newproc_sem;
299 static struct mtx aio_job_mtx;
300 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */
301 static struct unrhdr *aiod_unr;
303 void aio_init_aioinfo(struct proc *p);
304 static int aio_onceonly(void);
305 static int aio_free_entry(struct kaiocb *job);
306 static void aio_process_rw(struct kaiocb *job);
307 static void aio_process_sync(struct kaiocb *job);
308 static void aio_process_mlock(struct kaiocb *job);
309 static void aio_schedule_fsync(void *context, int pending);
310 static int aio_newproc(int *);
311 int aio_aqueue(struct thread *td, struct aiocb *ujob,
312 struct aioliojob *lio, int type, struct aiocb_ops *ops);
313 static int aio_queue_file(struct file *fp, struct kaiocb *job);
314 static void aio_physwakeup(struct bio *bp);
315 static void aio_proc_rundown(void *arg, struct proc *p);
316 static void aio_proc_rundown_exec(void *arg, struct proc *p,
317 struct image_params *imgp);
318 static int aio_qphysio(struct proc *p, struct kaiocb *job);
319 static void aio_daemon(void *param);
320 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
321 static bool aio_clear_cancel_function_locked(struct kaiocb *job);
322 static int aio_kick(struct proc *userp);
323 static void aio_kick_nowait(struct proc *userp);
324 static void aio_kick_helper(void *context, int pending);
325 static int filt_aioattach(struct knote *kn);
326 static void filt_aiodetach(struct knote *kn);
327 static int filt_aio(struct knote *kn, long hint);
328 static int filt_lioattach(struct knote *kn);
329 static void filt_liodetach(struct knote *kn);
330 static int filt_lio(struct knote *kn, long hint);
334 * kaio Per process async io info
335 * aiop async io process data
336 * aiocb async io jobs
337 * aiol list io job pointer - internal to aio_suspend XXX
338 * aiolio list io jobs
340 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
342 /* kqueue filters for aio */
343 static struct filterops aio_filtops = {
345 .f_attach = filt_aioattach,
346 .f_detach = filt_aiodetach,
349 static struct filterops lio_filtops = {
351 .f_attach = filt_lioattach,
352 .f_detach = filt_liodetach,
356 static eventhandler_tag exit_tag, exec_tag;
358 TASKQUEUE_DEFINE_THREAD(aiod_kick);
361 * Main operations function for use as a kernel module.
364 aio_modload(struct module *module, int cmd, void *arg)
381 static moduledata_t aio_mod = {
387 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
388 MODULE_VERSION(aio, 1);
391 * Startup initialization
397 if (aio_listio_max < AIO_LISTIO_MAX)
398 aio_listio_max = AIO_LISTIO_MAX;
399 if (aio_listio_max > MIN(MAX_AIO_QUEUE_PER_PROC, max_queue_count))
400 aio_listio_max = MIN(MAX_AIO_QUEUE_PER_PROC, max_queue_count);
402 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
403 EVENTHANDLER_PRI_ANY);
404 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
405 NULL, EVENTHANDLER_PRI_ANY);
406 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
407 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
408 TAILQ_INIT(&aio_freeproc);
409 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
410 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
411 TAILQ_INIT(&aio_jobs);
412 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
413 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
414 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
415 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
416 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
417 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
418 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
419 aiol_zone = uma_zcreate("AIOL", aio_listio_max * sizeof(intptr_t) ,
420 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
421 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
422 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
423 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
425 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
426 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
427 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
433 * Init the per-process aioinfo structure. The aioinfo limits are set
434 * per-process for user limit (resource) management.
437 aio_init_aioinfo(struct proc *p)
441 ki = uma_zalloc(kaio_zone, M_WAITOK);
442 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
444 ki->kaio_maxactive_count = max_aio_per_proc;
445 ki->kaio_active_count = 0;
446 ki->kaio_qallowed_count = max_aio_queue_per_proc;
448 ki->kaio_ballowed_count = max_buf_aio;
449 ki->kaio_buffer_count = 0;
450 TAILQ_INIT(&ki->kaio_all);
451 TAILQ_INIT(&ki->kaio_done);
452 TAILQ_INIT(&ki->kaio_jobqueue);
453 TAILQ_INIT(&ki->kaio_liojoblist);
454 TAILQ_INIT(&ki->kaio_syncqueue);
455 TAILQ_INIT(&ki->kaio_syncready);
456 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
457 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
459 if (p->p_aioinfo == NULL) {
464 mtx_destroy(&ki->kaio_mtx);
465 uma_zfree(kaio_zone, ki);
468 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
473 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
478 error = sigev_findtd(p, sigev, &td);
482 ksiginfo_set_sigev(ksi, sigev);
483 ksi->ksi_code = SI_ASYNCIO;
484 ksi->ksi_flags |= KSI_EXT | KSI_INS;
485 tdsendsignal(p, td, ksi->ksi_signo, ksi);
492 * Free a job entry. Wait for completion if it is currently active, but don't
493 * delay forever. If we delay, we return a flag that says that we have to
494 * restart the queue scan.
497 aio_free_entry(struct kaiocb *job)
500 struct aioliojob *lj;
508 AIO_LOCK_ASSERT(ki, MA_OWNED);
509 MPASS(job->jobflags & KAIOCB_FINISHED);
511 atomic_subtract_int(&num_queue_count, 1);
514 MPASS(ki->kaio_count >= 0);
516 TAILQ_REMOVE(&ki->kaio_done, job, plist);
517 TAILQ_REMOVE(&ki->kaio_all, job, allist);
522 lj->lioj_finished_count--;
524 if (lj->lioj_count == 0) {
525 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
526 /* lio is going away, we need to destroy any knotes */
527 knlist_delete(&lj->klist, curthread, 1);
529 sigqueue_take(&lj->lioj_ksi);
531 uma_zfree(aiolio_zone, lj);
535 /* job is going away, we need to destroy any knotes */
536 knlist_delete(&job->klist, curthread, 1);
538 sigqueue_take(&job->ksi);
544 * The thread argument here is used to find the owning process
545 * and is also passed to fo_close() which may pass it to various
546 * places such as devsw close() routines. Because of that, we
547 * need a thread pointer from the process owning the job that is
548 * persistent and won't disappear out from under us or move to
551 * Currently, all the callers of this function call it to remove
552 * a kaiocb from the current process' job list either via a
553 * syscall or due to the current process calling exit() or
554 * execve(). Thus, we know that p == curproc. We also know that
555 * curthread can't exit since we are curthread.
557 * Therefore, we use curthread as the thread to pass to
558 * knlist_delete(). This does mean that it is possible for the
559 * thread pointer at close time to differ from the thread pointer
560 * at open time, but this is already true of file descriptors in
561 * a multithreaded process.
564 fdrop(job->fd_file, curthread);
566 uma_zfree(aiocb_zone, job);
573 aio_proc_rundown_exec(void *arg, struct proc *p,
574 struct image_params *imgp __unused)
576 aio_proc_rundown(arg, p);
580 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
582 aio_cancel_fn_t *func;
585 AIO_LOCK_ASSERT(ki, MA_OWNED);
586 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
588 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
589 job->jobflags |= KAIOCB_CANCELLED;
591 func = job->cancel_fn;
594 * If there is no cancel routine, just leave the job marked as
595 * cancelled. The job should be in active use by a caller who
596 * should complete it normally or when it fails to install a
603 * Set the CANCELLING flag so that aio_complete() will defer
604 * completions of this job. This prevents the job from being
605 * freed out from under the cancel callback. After the
606 * callback any deferred completion (whether from the callback
607 * or any other source) will be completed.
609 job->jobflags |= KAIOCB_CANCELLING;
613 job->jobflags &= ~KAIOCB_CANCELLING;
614 if (job->jobflags & KAIOCB_FINISHED) {
615 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
616 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
617 aio_bio_done_notify(p, job);
620 * The cancel callback might have scheduled an
621 * operation to cancel this request, but it is
622 * only counted as cancelled if the request is
623 * cancelled when the callback returns.
631 * Rundown the jobs for a given process.
634 aio_proc_rundown(void *arg, struct proc *p)
637 struct aioliojob *lj;
638 struct kaiocb *job, *jobn;
640 KASSERT(curthread->td_proc == p,
641 ("%s: called on non-curproc", __func__));
647 ki->kaio_flags |= KAIO_RUNDOWN;
652 * Try to cancel all pending requests. This code simulates
653 * aio_cancel on all pending I/O requests.
655 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
656 aio_cancel_job(p, ki, job);
659 /* Wait for all running I/O to be finished */
660 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
661 ki->kaio_flags |= KAIO_WAKEUP;
662 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
666 /* Free all completed I/O requests. */
667 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
670 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
671 if (lj->lioj_count == 0) {
672 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
673 knlist_delete(&lj->klist, curthread, 1);
675 sigqueue_take(&lj->lioj_ksi);
677 uma_zfree(aiolio_zone, lj);
679 panic("LIO job not cleaned up: C:%d, FC:%d\n",
680 lj->lioj_count, lj->lioj_finished_count);
684 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
685 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
686 mtx_destroy(&ki->kaio_mtx);
687 uma_zfree(kaio_zone, ki);
692 * Select a job to run (called by an AIO daemon).
694 static struct kaiocb *
695 aio_selectjob(struct aioproc *aiop)
701 mtx_assert(&aio_job_mtx, MA_OWNED);
703 TAILQ_FOREACH(job, &aio_jobs, list) {
704 userp = job->userproc;
705 ki = userp->p_aioinfo;
707 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
708 TAILQ_REMOVE(&aio_jobs, job, list);
709 if (!aio_clear_cancel_function(job))
712 /* Account for currently active jobs. */
713 ki->kaio_active_count++;
721 * Move all data to a permanent storage device. This code
722 * simulates the fsync syscall.
725 aio_fsync_vnode(struct thread *td, struct vnode *vp)
730 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
732 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
733 if (vp->v_object != NULL) {
734 VM_OBJECT_WLOCK(vp->v_object);
735 vm_object_page_clean(vp->v_object, 0, 0, 0);
736 VM_OBJECT_WUNLOCK(vp->v_object);
738 error = VOP_FSYNC(vp, MNT_WAIT, td);
741 vn_finished_write(mp);
747 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
748 * does the I/O request for the non-physio version of the operations. The
749 * normal vn operations are used, and this code should work in all instances
750 * for every type of file, including pipes, sockets, fifos, and regular files.
752 * XXX I don't think it works well for socket, pipe, and fifo.
755 aio_process_rw(struct kaiocb *job)
757 struct ucred *td_savedcred;
764 long msgsnd_st, msgsnd_end;
765 long msgrcv_st, msgrcv_end;
766 long oublock_st, oublock_end;
767 long inblock_st, inblock_end;
770 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
771 job->uaiocb.aio_lio_opcode == LIO_WRITE,
772 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
774 aio_switch_vmspace(job);
776 td_savedcred = td->td_ucred;
777 td->td_ucred = job->cred;
781 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
782 aiov.iov_len = cb->aio_nbytes;
784 auio.uio_iov = &aiov;
786 auio.uio_offset = cb->aio_offset;
787 auio.uio_resid = cb->aio_nbytes;
788 cnt = cb->aio_nbytes;
789 auio.uio_segflg = UIO_USERSPACE;
792 msgrcv_st = td->td_ru.ru_msgrcv;
793 msgsnd_st = td->td_ru.ru_msgsnd;
794 inblock_st = td->td_ru.ru_inblock;
795 oublock_st = td->td_ru.ru_oublock;
798 * aio_aqueue() acquires a reference to the file that is
799 * released in aio_free_entry().
801 if (cb->aio_lio_opcode == LIO_READ) {
802 auio.uio_rw = UIO_READ;
803 if (auio.uio_resid == 0)
806 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
808 if (fp->f_type == DTYPE_VNODE)
810 auio.uio_rw = UIO_WRITE;
811 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
813 msgrcv_end = td->td_ru.ru_msgrcv;
814 msgsnd_end = td->td_ru.ru_msgsnd;
815 inblock_end = td->td_ru.ru_inblock;
816 oublock_end = td->td_ru.ru_oublock;
818 job->msgrcv = msgrcv_end - msgrcv_st;
819 job->msgsnd = msgsnd_end - msgsnd_st;
820 job->inblock = inblock_end - inblock_st;
821 job->outblock = oublock_end - oublock_st;
823 if ((error) && (auio.uio_resid != cnt)) {
824 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
826 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
827 PROC_LOCK(job->userproc);
828 kern_psignal(job->userproc, SIGPIPE);
829 PROC_UNLOCK(job->userproc);
833 cnt -= auio.uio_resid;
834 td->td_ucred = td_savedcred;
836 aio_complete(job, -1, error);
838 aio_complete(job, cnt, 0);
842 aio_process_sync(struct kaiocb *job)
844 struct thread *td = curthread;
845 struct ucred *td_savedcred = td->td_ucred;
846 struct file *fp = job->fd_file;
849 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
850 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
852 td->td_ucred = job->cred;
853 if (fp->f_vnode != NULL)
854 error = aio_fsync_vnode(td, fp->f_vnode);
855 td->td_ucred = td_savedcred;
857 aio_complete(job, -1, error);
859 aio_complete(job, 0, 0);
863 aio_process_mlock(struct kaiocb *job)
865 struct aiocb *cb = &job->uaiocb;
868 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
869 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
871 aio_switch_vmspace(job);
872 error = kern_mlock(job->userproc, job->cred,
873 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
874 aio_complete(job, error != 0 ? -1 : 0, error);
878 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
880 struct aioliojob *lj;
882 struct kaiocb *sjob, *sjobn;
886 ki = userp->p_aioinfo;
887 AIO_LOCK_ASSERT(ki, MA_OWNED);
891 lj->lioj_finished_count++;
892 if (lj->lioj_count == lj->lioj_finished_count)
895 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
896 MPASS(job->jobflags & KAIOCB_FINISHED);
898 if (ki->kaio_flags & KAIO_RUNDOWN)
899 goto notification_done;
901 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
902 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
903 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
905 KNOTE_LOCKED(&job->klist, 1);
908 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
909 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
910 KNOTE_LOCKED(&lj->klist, 1);
912 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
914 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
915 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
916 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
917 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
922 if (job->jobflags & KAIOCB_CHECKSYNC) {
923 schedule_fsync = false;
924 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
925 if (job->fd_file != sjob->fd_file ||
926 job->seqno >= sjob->seqno)
928 if (--sjob->pending > 0)
930 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
931 if (!aio_clear_cancel_function_locked(sjob))
933 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
934 schedule_fsync = true;
937 taskqueue_enqueue(taskqueue_aiod_kick,
938 &ki->kaio_sync_task);
940 if (ki->kaio_flags & KAIO_WAKEUP) {
941 ki->kaio_flags &= ~KAIO_WAKEUP;
942 wakeup(&userp->p_aioinfo);
947 aio_schedule_fsync(void *context, int pending)
954 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
955 job = TAILQ_FIRST(&ki->kaio_syncready);
956 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
958 aio_schedule(job, aio_process_sync);
965 aio_cancel_cleared(struct kaiocb *job)
970 * The caller should hold the same queue lock held when
971 * aio_clear_cancel_function() was called and set this flag
972 * ensuring this check sees an up-to-date value. However,
973 * there is no way to assert that.
975 ki = job->userproc->p_aioinfo;
976 return ((job->jobflags & KAIOCB_CLEARED) != 0);
980 aio_clear_cancel_function_locked(struct kaiocb *job)
983 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
984 MPASS(job->cancel_fn != NULL);
985 if (job->jobflags & KAIOCB_CANCELLING) {
986 job->jobflags |= KAIOCB_CLEARED;
989 job->cancel_fn = NULL;
994 aio_clear_cancel_function(struct kaiocb *job)
999 ki = job->userproc->p_aioinfo;
1001 ret = aio_clear_cancel_function_locked(job);
1007 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1010 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1011 if (job->jobflags & KAIOCB_CANCELLED)
1013 job->cancel_fn = func;
1018 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1020 struct kaioinfo *ki;
1023 ki = job->userproc->p_aioinfo;
1025 ret = aio_set_cancel_function_locked(job, func);
1031 aio_complete(struct kaiocb *job, long status, int error)
1033 struct kaioinfo *ki;
1036 job->uaiocb._aiocb_private.error = error;
1037 job->uaiocb._aiocb_private.status = status;
1039 userp = job->userproc;
1040 ki = userp->p_aioinfo;
1043 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1044 ("duplicate aio_complete"));
1045 job->jobflags |= KAIOCB_FINISHED;
1046 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1047 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1048 aio_bio_done_notify(userp, job);
1054 aio_cancel(struct kaiocb *job)
1057 aio_complete(job, -1, ECANCELED);
1061 aio_switch_vmspace(struct kaiocb *job)
1064 vmspace_switch_aio(job->userproc->p_vmspace);
1068 * The AIO daemon, most of the actual work is done in aio_process_*,
1069 * but the setup (and address space mgmt) is done in this routine.
1072 aio_daemon(void *_id)
1075 struct aioproc *aiop;
1076 struct kaioinfo *ki;
1078 struct vmspace *myvm;
1079 struct thread *td = curthread;
1080 int id = (intptr_t)_id;
1083 * Grab an extra reference on the daemon's vmspace so that it
1084 * doesn't get freed by jobs that switch to a different
1088 myvm = vmspace_acquire_ref(p);
1090 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1093 * Allocate and ready the aio control info. There is one aiop structure
1096 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1098 aiop->aioprocflags = 0;
1101 * Wakeup parent process. (Parent sleeps to keep from blasting away
1102 * and creating too many daemons.)
1104 sema_post(&aio_newproc_sem);
1106 mtx_lock(&aio_job_mtx);
1109 * Take daemon off of free queue
1111 if (aiop->aioprocflags & AIOP_FREE) {
1112 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1113 aiop->aioprocflags &= ~AIOP_FREE;
1119 while ((job = aio_selectjob(aiop)) != NULL) {
1120 mtx_unlock(&aio_job_mtx);
1122 ki = job->userproc->p_aioinfo;
1123 job->handle_fn(job);
1125 mtx_lock(&aio_job_mtx);
1126 /* Decrement the active job count. */
1127 ki->kaio_active_count--;
1131 * Disconnect from user address space.
1133 if (p->p_vmspace != myvm) {
1134 mtx_unlock(&aio_job_mtx);
1135 vmspace_switch_aio(myvm);
1136 mtx_lock(&aio_job_mtx);
1138 * We have to restart to avoid race, we only sleep if
1139 * no job can be selected.
1144 mtx_assert(&aio_job_mtx, MA_OWNED);
1146 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1147 aiop->aioprocflags |= AIOP_FREE;
1150 * If daemon is inactive for a long time, allow it to exit,
1151 * thereby freeing resources.
1153 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1154 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1155 (aiop->aioprocflags & AIOP_FREE) &&
1156 num_aio_procs > target_aio_procs)
1159 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1161 mtx_unlock(&aio_job_mtx);
1162 uma_zfree(aiop_zone, aiop);
1163 free_unr(aiod_unr, id);
1166 KASSERT(p->p_vmspace == myvm,
1167 ("AIOD: bad vmspace for exiting daemon"));
1168 KASSERT(myvm->vm_refcnt > 1,
1169 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1174 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1175 * AIO daemon modifies its environment itself.
1178 aio_newproc(int *start)
1184 id = alloc_unr(aiod_unr);
1185 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1186 RFNOWAIT, 0, "aiod%d", id);
1189 * Wait until daemon is started.
1191 sema_wait(&aio_newproc_sem);
1192 mtx_lock(&aio_job_mtx);
1196 mtx_unlock(&aio_job_mtx);
1198 free_unr(aiod_unr, id);
1204 * Try the high-performance, low-overhead physio method for eligible
1205 * VCHR devices. This method doesn't use an aio helper thread, and
1206 * thus has very low overhead.
1208 * Assumes that the caller, aio_aqueue(), has incremented the file
1209 * structure's reference count, preventing its deallocation for the
1210 * duration of this call.
1213 aio_qphysio(struct proc *p, struct kaiocb *job)
1222 struct kaioinfo *ki;
1223 int error, ref, poff;
1229 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1233 if (vp->v_type != VCHR)
1235 if (vp->v_bufobj.bo_bsize == 0)
1237 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1241 csw = devvn_refthread(vp, &dev, &ref);
1245 if ((csw->d_flags & D_DISK) == 0) {
1249 if (cb->aio_nbytes > dev->si_iosize_max) {
1255 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1256 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1257 if (cb->aio_nbytes > MAXPHYS) {
1264 if (cb->aio_nbytes > MAXPHYS - poff) {
1268 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1273 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1276 ki->kaio_buffer_count++;
1279 job->bp = bp = g_alloc_bio();
1281 bp->bio_length = cb->aio_nbytes;
1282 bp->bio_bcount = cb->aio_nbytes;
1283 bp->bio_done = aio_physwakeup;
1284 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1285 bp->bio_offset = cb->aio_offset;
1286 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1288 bp->bio_caller1 = (void *)job;
1290 prot = VM_PROT_READ;
1291 if (cb->aio_lio_opcode == LIO_READ)
1292 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1293 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1294 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1295 nitems(job->pages));
1296 if (job->npages < 0) {
1301 pmap_qenter((vm_offset_t)pbuf->b_data,
1302 job->pages, job->npages);
1303 bp->bio_data = pbuf->b_data + poff;
1304 atomic_add_int(&num_buf_aio, 1);
1306 bp->bio_ma = job->pages;
1307 bp->bio_ma_n = job->npages;
1308 bp->bio_ma_offset = poff;
1309 bp->bio_data = unmapped_buf;
1310 bp->bio_flags |= BIO_UNMAPPED;
1313 /* Perform transfer. */
1314 csw->d_strategy(bp);
1315 dev_relthread(dev, ref);
1321 ki->kaio_buffer_count--;
1323 relpbuf(pbuf, NULL);
1329 dev_relthread(dev, ref);
1333 #ifdef COMPAT_FREEBSD6
1335 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1339 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1340 * supported by AIO with the old sigevent structure.
1342 nsig->sigev_notify = osig->sigev_notify;
1343 switch (nsig->sigev_notify) {
1347 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1350 nsig->sigev_notify_kqueue =
1351 osig->__sigev_u.__sigev_notify_kqueue;
1352 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1361 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1363 struct oaiocb *ojob;
1366 bzero(kjob, sizeof(struct aiocb));
1367 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1370 ojob = (struct oaiocb *)kjob;
1371 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1376 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1379 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1383 aiocb_fetch_status(struct aiocb *ujob)
1386 return (fuword(&ujob->_aiocb_private.status));
1390 aiocb_fetch_error(struct aiocb *ujob)
1393 return (fuword(&ujob->_aiocb_private.error));
1397 aiocb_store_status(struct aiocb *ujob, long status)
1400 return (suword(&ujob->_aiocb_private.status, status));
1404 aiocb_store_error(struct aiocb *ujob, long error)
1407 return (suword(&ujob->_aiocb_private.error, error));
1411 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1414 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1418 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1421 return (suword(ujobp, (long)ujob));
1424 static struct aiocb_ops aiocb_ops = {
1425 .copyin = aiocb_copyin,
1426 .fetch_status = aiocb_fetch_status,
1427 .fetch_error = aiocb_fetch_error,
1428 .store_status = aiocb_store_status,
1429 .store_error = aiocb_store_error,
1430 .store_kernelinfo = aiocb_store_kernelinfo,
1431 .store_aiocb = aiocb_store_aiocb,
1434 #ifdef COMPAT_FREEBSD6
1435 static struct aiocb_ops aiocb_ops_osigevent = {
1436 .copyin = aiocb_copyin_old_sigevent,
1437 .fetch_status = aiocb_fetch_status,
1438 .fetch_error = aiocb_fetch_error,
1439 .store_status = aiocb_store_status,
1440 .store_error = aiocb_store_error,
1441 .store_kernelinfo = aiocb_store_kernelinfo,
1442 .store_aiocb = aiocb_store_aiocb,
1447 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1448 * technique is done in this code.
1451 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1452 int type, struct aiocb_ops *ops)
1454 struct proc *p = td->td_proc;
1455 cap_rights_t rights;
1458 struct kaioinfo *ki;
1466 if (p->p_aioinfo == NULL)
1467 aio_init_aioinfo(p);
1471 ops->store_status(ujob, -1);
1472 ops->store_error(ujob, 0);
1473 ops->store_kernelinfo(ujob, -1);
1475 if (num_queue_count >= max_queue_count ||
1476 ki->kaio_count >= ki->kaio_qallowed_count) {
1477 ops->store_error(ujob, EAGAIN);
1481 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1482 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1484 error = ops->copyin(ujob, &job->uaiocb);
1486 ops->store_error(ujob, error);
1487 uma_zfree(aiocb_zone, job);
1491 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1492 uma_zfree(aiocb_zone, job);
1496 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1497 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1498 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1499 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1500 ops->store_error(ujob, EINVAL);
1501 uma_zfree(aiocb_zone, job);
1505 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1506 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1507 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1508 uma_zfree(aiocb_zone, job);
1512 ksiginfo_init(&job->ksi);
1514 /* Save userspace address of the job info. */
1517 /* Get the opcode. */
1518 if (type != LIO_NOP)
1519 job->uaiocb.aio_lio_opcode = type;
1520 opcode = job->uaiocb.aio_lio_opcode;
1523 * Validate the opcode and fetch the file object for the specified
1526 * XXXRW: Moved the opcode validation up here so that we don't
1527 * retrieve a file descriptor without knowing what the capabiltity
1530 fd = job->uaiocb.aio_fildes;
1533 error = fget_write(td, fd,
1534 cap_rights_init(&rights, CAP_PWRITE), &fp);
1537 error = fget_read(td, fd,
1538 cap_rights_init(&rights, CAP_PREAD), &fp);
1541 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1547 error = fget(td, fd, cap_rights_init(&rights), &fp);
1553 uma_zfree(aiocb_zone, job);
1554 ops->store_error(ujob, error);
1558 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1563 if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1564 job->uaiocb.aio_offset < 0 &&
1565 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1572 mtx_lock(&aio_job_mtx);
1574 job->seqno = jobseqno++;
1575 mtx_unlock(&aio_job_mtx);
1576 error = ops->store_kernelinfo(ujob, jid);
1581 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1583 if (opcode == LIO_NOP) {
1585 uma_zfree(aiocb_zone, job);
1589 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1591 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1592 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1596 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1597 kev.ident = (uintptr_t)job->ujob;
1598 kev.filter = EVFILT_AIO;
1599 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1600 kev.data = (intptr_t)job;
1601 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1602 error = kqfd_register(kqfd, &kev, td, 1);
1608 ops->store_error(ujob, EINPROGRESS);
1609 job->uaiocb._aiocb_private.error = EINPROGRESS;
1611 job->cred = crhold(td->td_ucred);
1612 job->jobflags = KAIOCB_QUEUEING;
1615 if (opcode == LIO_MLOCK) {
1616 aio_schedule(job, aio_process_mlock);
1618 } else if (fp->f_ops->fo_aio_queue == NULL)
1619 error = aio_queue_file(fp, job);
1621 error = fo_aio_queue(fp, job);
1626 job->jobflags &= ~KAIOCB_QUEUEING;
1627 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1631 atomic_add_int(&num_queue_count, 1);
1632 if (job->jobflags & KAIOCB_FINISHED) {
1634 * The queue callback completed the request synchronously.
1635 * The bulk of the completion is deferred in that case
1638 aio_bio_done_notify(p, job);
1640 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1645 knlist_delete(&job->klist, curthread, 0);
1648 uma_zfree(aiocb_zone, job);
1649 ops->store_error(ujob, error);
1654 aio_cancel_daemon_job(struct kaiocb *job)
1657 mtx_lock(&aio_job_mtx);
1658 if (!aio_cancel_cleared(job))
1659 TAILQ_REMOVE(&aio_jobs, job, list);
1660 mtx_unlock(&aio_job_mtx);
1665 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1668 mtx_lock(&aio_job_mtx);
1669 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1670 mtx_unlock(&aio_job_mtx);
1674 job->handle_fn = func;
1675 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1676 aio_kick_nowait(job->userproc);
1677 mtx_unlock(&aio_job_mtx);
1681 aio_cancel_sync(struct kaiocb *job)
1683 struct kaioinfo *ki;
1685 ki = job->userproc->p_aioinfo;
1687 if (!aio_cancel_cleared(job))
1688 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1694 aio_queue_file(struct file *fp, struct kaiocb *job)
1696 struct aioliojob *lj;
1697 struct kaioinfo *ki;
1698 struct kaiocb *job2;
1705 ki = job->userproc->p_aioinfo;
1706 opcode = job->uaiocb.aio_lio_opcode;
1707 if (opcode == LIO_SYNC)
1710 if ((error = aio_qphysio(job->userproc, job)) == 0)
1714 * XXX: This means qphysio() failed with EFAULT. The current
1715 * behavior is to retry the operation via fo_read/fo_write.
1716 * Wouldn't it be better to just complete the request with an
1724 if (fp->f_type == DTYPE_VNODE) {
1726 if (vp->v_type == VREG || vp->v_type == VDIR) {
1727 mp = fp->f_vnode->v_mount;
1728 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1732 if (!(safe || enable_aio_unsafe)) {
1733 counted_warning(&unsafe_warningcnt,
1734 "is attempting to use unsafe AIO requests");
1735 return (EOPNOTSUPP);
1738 if (opcode == LIO_SYNC) {
1740 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1741 if (job2->fd_file == job->fd_file &&
1742 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1743 job2->seqno < job->seqno) {
1744 job2->jobflags |= KAIOCB_CHECKSYNC;
1748 if (job->pending != 0) {
1749 if (!aio_set_cancel_function_locked(job,
1755 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1765 aio_schedule(job, aio_process_rw);
1769 aio_schedule(job, aio_process_sync);
1780 aio_kick_nowait(struct proc *userp)
1782 struct kaioinfo *ki = userp->p_aioinfo;
1783 struct aioproc *aiop;
1785 mtx_assert(&aio_job_mtx, MA_OWNED);
1786 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1787 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1788 aiop->aioprocflags &= ~AIOP_FREE;
1789 wakeup(aiop->aioproc);
1790 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1791 ki->kaio_active_count + num_aio_resv_start <
1792 ki->kaio_maxactive_count) {
1793 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1798 aio_kick(struct proc *userp)
1800 struct kaioinfo *ki = userp->p_aioinfo;
1801 struct aioproc *aiop;
1804 mtx_assert(&aio_job_mtx, MA_OWNED);
1806 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1807 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1808 aiop->aioprocflags &= ~AIOP_FREE;
1809 wakeup(aiop->aioproc);
1810 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1811 ki->kaio_active_count + num_aio_resv_start <
1812 ki->kaio_maxactive_count) {
1813 num_aio_resv_start++;
1814 mtx_unlock(&aio_job_mtx);
1815 error = aio_newproc(&num_aio_resv_start);
1816 mtx_lock(&aio_job_mtx);
1818 num_aio_resv_start--;
1828 aio_kick_helper(void *context, int pending)
1830 struct proc *userp = context;
1832 mtx_lock(&aio_job_mtx);
1833 while (--pending >= 0) {
1834 if (aio_kick(userp))
1837 mtx_unlock(&aio_job_mtx);
1841 * Support the aio_return system call, as a side-effect, kernel resources are
1845 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1847 struct proc *p = td->td_proc;
1849 struct kaioinfo *ki;
1856 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1857 if (job->ujob == ujob)
1861 MPASS(job->jobflags & KAIOCB_FINISHED);
1862 status = job->uaiocb._aiocb_private.status;
1863 error = job->uaiocb._aiocb_private.error;
1864 td->td_retval[0] = status;
1865 td->td_ru.ru_oublock += job->outblock;
1866 td->td_ru.ru_inblock += job->inblock;
1867 td->td_ru.ru_msgsnd += job->msgsnd;
1868 td->td_ru.ru_msgrcv += job->msgrcv;
1869 aio_free_entry(job);
1871 ops->store_error(ujob, error);
1872 ops->store_status(ujob, status);
1881 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1884 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1888 * Allow a process to wakeup when any of the I/O requests are completed.
1891 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1892 struct timespec *ts)
1894 struct proc *p = td->td_proc;
1896 struct kaioinfo *ki;
1897 struct kaiocb *firstjob, *job;
1902 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1905 TIMESPEC_TO_TIMEVAL(&atv, ts);
1906 if (itimerfix(&atv))
1908 timo = tvtohz(&atv);
1922 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1923 for (i = 0; i < njoblist; i++) {
1924 if (job->ujob == ujoblist[i]) {
1925 if (firstjob == NULL)
1927 if (job->jobflags & KAIOCB_FINISHED)
1932 /* All tasks were finished. */
1933 if (firstjob == NULL)
1936 ki->kaio_flags |= KAIO_WAKEUP;
1937 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1939 if (error == ERESTART)
1950 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1952 struct timespec ts, *tsp;
1953 struct aiocb **ujoblist;
1956 if (uap->nent < 0 || uap->nent > aio_listio_max)
1960 /* Get timespec struct. */
1961 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1967 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1968 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1970 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1971 uma_zfree(aiol_zone, ujoblist);
1976 * aio_cancel cancels any non-physio aio operations not currently in
1980 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1982 struct proc *p = td->td_proc;
1983 struct kaioinfo *ki;
1984 struct kaiocb *job, *jobn;
1986 cap_rights_t rights;
1989 int notcancelled = 0;
1992 /* Lookup file object. */
1993 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
2001 if (fp->f_type == DTYPE_VNODE) {
2003 if (vn_isdisk(vp, &error)) {
2005 td->td_retval[0] = AIO_NOTCANCELED;
2011 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2012 if ((uap->fd == job->uaiocb.aio_fildes) &&
2013 ((uap->aiocbp == NULL) ||
2014 (uap->aiocbp == job->ujob))) {
2015 if (aio_cancel_job(p, ki, job)) {
2020 if (uap->aiocbp != NULL)
2029 if (uap->aiocbp != NULL) {
2031 td->td_retval[0] = AIO_CANCELED;
2037 td->td_retval[0] = AIO_NOTCANCELED;
2042 td->td_retval[0] = AIO_CANCELED;
2046 td->td_retval[0] = AIO_ALLDONE;
2052 * aio_error is implemented in the kernel level for compatibility purposes
2053 * only. For a user mode async implementation, it would be best to do it in
2054 * a userland subroutine.
2057 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2059 struct proc *p = td->td_proc;
2061 struct kaioinfo *ki;
2066 td->td_retval[0] = EINVAL;
2071 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2072 if (job->ujob == ujob) {
2073 if (job->jobflags & KAIOCB_FINISHED)
2075 job->uaiocb._aiocb_private.error;
2077 td->td_retval[0] = EINPROGRESS;
2085 * Hack for failure of aio_aqueue.
2087 status = ops->fetch_status(ujob);
2089 td->td_retval[0] = ops->fetch_error(ujob);
2093 td->td_retval[0] = EINVAL;
2098 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2101 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2104 /* syscall - asynchronous read from a file (REALTIME) */
2105 #ifdef COMPAT_FREEBSD6
2107 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2110 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2111 &aiocb_ops_osigevent));
2116 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2119 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2122 /* syscall - asynchronous write to a file (REALTIME) */
2123 #ifdef COMPAT_FREEBSD6
2125 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2128 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2129 &aiocb_ops_osigevent));
2134 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2137 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2141 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2144 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2148 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2149 struct aiocb **acb_list, int nent, struct sigevent *sig,
2150 struct aiocb_ops *ops)
2152 struct proc *p = td->td_proc;
2154 struct kaioinfo *ki;
2155 struct aioliojob *lj;
2161 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2164 if (nent < 0 || nent > aio_listio_max)
2167 if (p->p_aioinfo == NULL)
2168 aio_init_aioinfo(p);
2172 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2175 lj->lioj_finished_count = 0;
2176 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2177 ksiginfo_init(&lj->lioj_ksi);
2182 if (sig && (mode == LIO_NOWAIT)) {
2183 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2184 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2185 /* Assume only new style KEVENT */
2186 kev.filter = EVFILT_LIO;
2187 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2188 kev.ident = (uintptr_t)uacb_list; /* something unique */
2189 kev.data = (intptr_t)lj;
2190 /* pass user defined sigval data */
2191 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2192 error = kqfd_register(
2193 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2195 uma_zfree(aiolio_zone, lj);
2198 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2200 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2201 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2202 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2203 uma_zfree(aiolio_zone, lj);
2206 lj->lioj_flags |= LIOJ_SIGNAL;
2208 uma_zfree(aiolio_zone, lj);
2214 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2216 * Add extra aiocb count to avoid the lio to be freed
2217 * by other threads doing aio_waitcomplete or aio_return,
2218 * and prevent event from being sent until we have queued
2225 * Get pointers to the list of I/O requests.
2228 for (i = 0; i < nent; i++) {
2231 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2239 if (mode == LIO_WAIT) {
2240 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2241 ki->kaio_flags |= KAIO_WAKEUP;
2242 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2243 PRIBIO | PCATCH, "aiospn", 0);
2244 if (error == ERESTART)
2250 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2251 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2252 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2253 KNOTE_LOCKED(&lj->klist, 1);
2255 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2257 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2258 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2259 aio_sendsig(p, &lj->lioj_signal,
2261 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2266 if (lj->lioj_count == 0) {
2267 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2268 knlist_delete(&lj->klist, curthread, 1);
2270 sigqueue_take(&lj->lioj_ksi);
2273 uma_zfree(aiolio_zone, lj);
2282 /* syscall - list directed I/O (REALTIME) */
2283 #ifdef COMPAT_FREEBSD6
2285 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2287 struct aiocb **acb_list;
2288 struct sigevent *sigp, sig;
2289 struct osigevent osig;
2292 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2296 if (nent < 0 || nent > aio_listio_max)
2299 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2300 error = copyin(uap->sig, &osig, sizeof(osig));
2303 error = convert_old_sigevent(&osig, &sig);
2310 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2311 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2313 error = kern_lio_listio(td, uap->mode,
2314 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2315 &aiocb_ops_osigevent);
2316 free(acb_list, M_LIO);
2321 /* syscall - list directed I/O (REALTIME) */
2323 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2325 struct aiocb **acb_list;
2326 struct sigevent *sigp, sig;
2329 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2333 if (nent < 0 || nent > aio_listio_max)
2336 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2337 error = copyin(uap->sig, &sig, sizeof(sig));
2344 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2345 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2347 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2348 nent, sigp, &aiocb_ops);
2349 free(acb_list, M_LIO);
2354 aio_physwakeup(struct bio *bp)
2356 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2358 struct kaioinfo *ki;
2362 /* Release mapping into kernel space. */
2363 userp = job->userproc;
2364 ki = userp->p_aioinfo;
2366 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2367 relpbuf(job->pbuf, NULL);
2369 atomic_subtract_int(&num_buf_aio, 1);
2371 ki->kaio_buffer_count--;
2374 vm_page_unhold_pages(job->pages, job->npages);
2378 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2380 if (bp->bio_flags & BIO_ERROR)
2381 error = bp->bio_error;
2382 nblks = btodb(nbytes);
2383 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2384 job->outblock += nblks;
2386 job->inblock += nblks;
2389 aio_complete(job, -1, error);
2391 aio_complete(job, nbytes, 0);
2396 /* syscall - wait for the next completion of an aio request */
2398 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2399 struct timespec *ts, struct aiocb_ops *ops)
2401 struct proc *p = td->td_proc;
2403 struct kaioinfo *ki;
2409 ops->store_aiocb(ujobp, NULL);
2413 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2416 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2419 TIMESPEC_TO_TIMEVAL(&atv, ts);
2420 if (itimerfix(&atv))
2422 timo = tvtohz(&atv);
2425 if (p->p_aioinfo == NULL)
2426 aio_init_aioinfo(p);
2432 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2434 error = EWOULDBLOCK;
2437 ki->kaio_flags |= KAIO_WAKEUP;
2438 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2440 if (timo && error == ERESTART)
2447 MPASS(job->jobflags & KAIOCB_FINISHED);
2449 status = job->uaiocb._aiocb_private.status;
2450 error = job->uaiocb._aiocb_private.error;
2451 td->td_retval[0] = status;
2452 td->td_ru.ru_oublock += job->outblock;
2453 td->td_ru.ru_inblock += job->inblock;
2454 td->td_ru.ru_msgsnd += job->msgsnd;
2455 td->td_ru.ru_msgrcv += job->msgrcv;
2456 aio_free_entry(job);
2458 ops->store_aiocb(ujobp, ujob);
2459 ops->store_error(ujob, error);
2460 ops->store_status(ujob, status);
2468 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2470 struct timespec ts, *tsp;
2474 /* Get timespec struct. */
2475 error = copyin(uap->timeout, &ts, sizeof(ts));
2482 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2486 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2487 struct aiocb_ops *ops)
2490 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2492 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2496 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2499 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2502 /* kqueue attach function */
2504 filt_aioattach(struct knote *kn)
2508 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2511 * The job pointer must be validated before using it, so
2512 * registration is restricted to the kernel; the user cannot
2515 if ((kn->kn_flags & EV_FLAG1) == 0)
2517 kn->kn_ptr.p_aio = job;
2518 kn->kn_flags &= ~EV_FLAG1;
2520 knlist_add(&job->klist, kn, 0);
2525 /* kqueue detach function */
2527 filt_aiodetach(struct knote *kn)
2531 knl = &kn->kn_ptr.p_aio->klist;
2532 knl->kl_lock(knl->kl_lockarg);
2533 if (!knlist_empty(knl))
2534 knlist_remove(knl, kn, 1);
2535 knl->kl_unlock(knl->kl_lockarg);
2538 /* kqueue filter function */
2541 filt_aio(struct knote *kn, long hint)
2543 struct kaiocb *job = kn->kn_ptr.p_aio;
2545 kn->kn_data = job->uaiocb._aiocb_private.error;
2546 if (!(job->jobflags & KAIOCB_FINISHED))
2548 kn->kn_flags |= EV_EOF;
2552 /* kqueue attach function */
2554 filt_lioattach(struct knote *kn)
2556 struct aioliojob *lj;
2558 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2561 * The aioliojob pointer must be validated before using it, so
2562 * registration is restricted to the kernel; the user cannot
2565 if ((kn->kn_flags & EV_FLAG1) == 0)
2567 kn->kn_ptr.p_lio = lj;
2568 kn->kn_flags &= ~EV_FLAG1;
2570 knlist_add(&lj->klist, kn, 0);
2575 /* kqueue detach function */
2577 filt_liodetach(struct knote *kn)
2581 knl = &kn->kn_ptr.p_lio->klist;
2582 knl->kl_lock(knl->kl_lockarg);
2583 if (!knlist_empty(knl))
2584 knlist_remove(knl, kn, 1);
2585 knl->kl_unlock(knl->kl_lockarg);
2588 /* kqueue filter function */
2591 filt_lio(struct knote *kn, long hint)
2593 struct aioliojob * lj = kn->kn_ptr.p_lio;
2595 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2598 #ifdef COMPAT_FREEBSD32
2599 #include <sys/mount.h>
2600 #include <sys/socket.h>
2601 #include <compat/freebsd32/freebsd32.h>
2602 #include <compat/freebsd32/freebsd32_proto.h>
2603 #include <compat/freebsd32/freebsd32_signal.h>
2604 #include <compat/freebsd32/freebsd32_syscall.h>
2605 #include <compat/freebsd32/freebsd32_util.h>
2607 struct __aiocb_private32 {
2610 uint32_t kernelinfo;
2613 #ifdef COMPAT_FREEBSD6
2614 typedef struct oaiocb32 {
2615 int aio_fildes; /* File descriptor */
2616 uint64_t aio_offset __packed; /* File offset for I/O */
2617 uint32_t aio_buf; /* I/O buffer in process space */
2618 uint32_t aio_nbytes; /* Number of bytes for I/O */
2619 struct osigevent32 aio_sigevent; /* Signal to deliver */
2620 int aio_lio_opcode; /* LIO opcode */
2621 int aio_reqprio; /* Request priority -- ignored */
2622 struct __aiocb_private32 _aiocb_private;
2626 typedef struct aiocb32 {
2627 int32_t aio_fildes; /* File descriptor */
2628 uint64_t aio_offset __packed; /* File offset for I/O */
2629 uint32_t aio_buf; /* I/O buffer in process space */
2630 uint32_t aio_nbytes; /* Number of bytes for I/O */
2632 uint32_t __spare2__;
2633 int aio_lio_opcode; /* LIO opcode */
2634 int aio_reqprio; /* Request priority -- ignored */
2635 struct __aiocb_private32 _aiocb_private;
2636 struct sigevent32 aio_sigevent; /* Signal to deliver */
2639 #ifdef COMPAT_FREEBSD6
2641 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2645 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2646 * supported by AIO with the old sigevent structure.
2648 CP(*osig, *nsig, sigev_notify);
2649 switch (nsig->sigev_notify) {
2653 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2656 nsig->sigev_notify_kqueue =
2657 osig->__sigev_u.__sigev_notify_kqueue;
2658 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2667 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2669 struct oaiocb32 job32;
2672 bzero(kjob, sizeof(struct aiocb));
2673 error = copyin(ujob, &job32, sizeof(job32));
2677 CP(job32, *kjob, aio_fildes);
2678 CP(job32, *kjob, aio_offset);
2679 PTRIN_CP(job32, *kjob, aio_buf);
2680 CP(job32, *kjob, aio_nbytes);
2681 CP(job32, *kjob, aio_lio_opcode);
2682 CP(job32, *kjob, aio_reqprio);
2683 CP(job32, *kjob, _aiocb_private.status);
2684 CP(job32, *kjob, _aiocb_private.error);
2685 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2686 return (convert_old_sigevent32(&job32.aio_sigevent,
2687 &kjob->aio_sigevent));
2692 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2694 struct aiocb32 job32;
2697 error = copyin(ujob, &job32, sizeof(job32));
2700 CP(job32, *kjob, aio_fildes);
2701 CP(job32, *kjob, aio_offset);
2702 PTRIN_CP(job32, *kjob, aio_buf);
2703 CP(job32, *kjob, aio_nbytes);
2704 CP(job32, *kjob, aio_lio_opcode);
2705 CP(job32, *kjob, aio_reqprio);
2706 CP(job32, *kjob, _aiocb_private.status);
2707 CP(job32, *kjob, _aiocb_private.error);
2708 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2709 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2713 aiocb32_fetch_status(struct aiocb *ujob)
2715 struct aiocb32 *ujob32;
2717 ujob32 = (struct aiocb32 *)ujob;
2718 return (fuword32(&ujob32->_aiocb_private.status));
2722 aiocb32_fetch_error(struct aiocb *ujob)
2724 struct aiocb32 *ujob32;
2726 ujob32 = (struct aiocb32 *)ujob;
2727 return (fuword32(&ujob32->_aiocb_private.error));
2731 aiocb32_store_status(struct aiocb *ujob, long status)
2733 struct aiocb32 *ujob32;
2735 ujob32 = (struct aiocb32 *)ujob;
2736 return (suword32(&ujob32->_aiocb_private.status, status));
2740 aiocb32_store_error(struct aiocb *ujob, long error)
2742 struct aiocb32 *ujob32;
2744 ujob32 = (struct aiocb32 *)ujob;
2745 return (suword32(&ujob32->_aiocb_private.error, error));
2749 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2751 struct aiocb32 *ujob32;
2753 ujob32 = (struct aiocb32 *)ujob;
2754 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2758 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2761 return (suword32(ujobp, (long)ujob));
2764 static struct aiocb_ops aiocb32_ops = {
2765 .copyin = aiocb32_copyin,
2766 .fetch_status = aiocb32_fetch_status,
2767 .fetch_error = aiocb32_fetch_error,
2768 .store_status = aiocb32_store_status,
2769 .store_error = aiocb32_store_error,
2770 .store_kernelinfo = aiocb32_store_kernelinfo,
2771 .store_aiocb = aiocb32_store_aiocb,
2774 #ifdef COMPAT_FREEBSD6
2775 static struct aiocb_ops aiocb32_ops_osigevent = {
2776 .copyin = aiocb32_copyin_old_sigevent,
2777 .fetch_status = aiocb32_fetch_status,
2778 .fetch_error = aiocb32_fetch_error,
2779 .store_status = aiocb32_store_status,
2780 .store_error = aiocb32_store_error,
2781 .store_kernelinfo = aiocb32_store_kernelinfo,
2782 .store_aiocb = aiocb32_store_aiocb,
2787 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2790 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2794 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2796 struct timespec32 ts32;
2797 struct timespec ts, *tsp;
2798 struct aiocb **ujoblist;
2799 uint32_t *ujoblist32;
2802 if (uap->nent < 0 || uap->nent > aio_listio_max)
2806 /* Get timespec struct. */
2807 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2809 CP(ts32, ts, tv_sec);
2810 CP(ts32, ts, tv_nsec);
2815 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2816 ujoblist32 = (uint32_t *)ujoblist;
2817 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2818 sizeof(ujoblist32[0]));
2820 for (i = uap->nent; i > 0; i--)
2821 ujoblist[i] = PTRIN(ujoblist32[i]);
2823 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2825 uma_zfree(aiol_zone, ujoblist);
2830 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2833 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2836 #ifdef COMPAT_FREEBSD6
2838 freebsd6_freebsd32_aio_read(struct thread *td,
2839 struct freebsd6_freebsd32_aio_read_args *uap)
2842 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2843 &aiocb32_ops_osigevent));
2848 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2851 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2855 #ifdef COMPAT_FREEBSD6
2857 freebsd6_freebsd32_aio_write(struct thread *td,
2858 struct freebsd6_freebsd32_aio_write_args *uap)
2861 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2862 &aiocb32_ops_osigevent));
2867 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2870 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2875 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2878 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2883 freebsd32_aio_waitcomplete(struct thread *td,
2884 struct freebsd32_aio_waitcomplete_args *uap)
2886 struct timespec32 ts32;
2887 struct timespec ts, *tsp;
2891 /* Get timespec struct. */
2892 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2895 CP(ts32, ts, tv_sec);
2896 CP(ts32, ts, tv_nsec);
2901 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2906 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2909 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2913 #ifdef COMPAT_FREEBSD6
2915 freebsd6_freebsd32_lio_listio(struct thread *td,
2916 struct freebsd6_freebsd32_lio_listio_args *uap)
2918 struct aiocb **acb_list;
2919 struct sigevent *sigp, sig;
2920 struct osigevent32 osig;
2921 uint32_t *acb_list32;
2924 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2928 if (nent < 0 || nent > aio_listio_max)
2931 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2932 error = copyin(uap->sig, &osig, sizeof(osig));
2935 error = convert_old_sigevent32(&osig, &sig);
2942 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2943 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2945 free(acb_list32, M_LIO);
2948 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2949 for (i = 0; i < nent; i++)
2950 acb_list[i] = PTRIN(acb_list32[i]);
2951 free(acb_list32, M_LIO);
2953 error = kern_lio_listio(td, uap->mode,
2954 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2955 &aiocb32_ops_osigevent);
2956 free(acb_list, M_LIO);
2962 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2964 struct aiocb **acb_list;
2965 struct sigevent *sigp, sig;
2966 struct sigevent32 sig32;
2967 uint32_t *acb_list32;
2970 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2974 if (nent < 0 || nent > aio_listio_max)
2977 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2978 error = copyin(uap->sig, &sig32, sizeof(sig32));
2981 error = convert_sigevent32(&sig32, &sig);
2988 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2989 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2991 free(acb_list32, M_LIO);
2994 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2995 for (i = 0; i < nent; i++)
2996 acb_list[i] = PTRIN(acb_list32[i]);
2997 free(acb_list32, M_LIO);
2999 error = kern_lio_listio(td, uap->mode,
3000 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3002 free(acb_list, M_LIO);