2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 1997 John S. Dyson. All rights reserved.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. John S. Dyson's name may not be used to endorse or promote products
12 * derived from this software without specific prior written permission.
14 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
15 * bad that happens because of using this software isn't the responsibility
16 * of the author. This software is distributed AS-IS.
20 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
23 #include <sys/cdefs.h>
24 __FBSDID("$FreeBSD$");
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
97 #define MAX_AIO_QUEUE 1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
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");
108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
111 "Async IO management");
113 static int enable_aio_unsafe = 0;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
115 "Permit asynchronous IO on all file types, not just known-safe types");
117 static unsigned int unsafe_warningcnt = 1;
118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
119 &unsafe_warningcnt, 0,
120 "Warnings that will be triggered upon failed IO requests on unsafe files");
122 static int max_aio_procs = MAX_AIO_PROCS;
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
124 "Maximum number of kernel processes to use for handling async IO ");
126 static int num_aio_procs = 0;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
128 "Number of presently active kernel processes for async IO");
131 * The code will adjust the actual number of AIO processes towards this
132 * number when it gets a chance.
134 static int target_aio_procs = TARGET_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
137 "Preferred number of ready kernel processes for async IO");
139 static int max_queue_count = MAX_AIO_QUEUE;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
141 "Maximum number of aio requests to queue, globally");
143 static int num_queue_count = 0;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
145 "Number of queued aio requests");
147 static int num_buf_aio = 0;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
149 "Number of aio requests presently handled by the buf subsystem");
151 static int num_unmapped_aio = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
154 "Number of aio requests presently handled by unmapped I/O buffers");
156 /* Number of async I/O processes in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
160 static int aiod_lifetime;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
162 "Maximum lifetime for idle aiod");
164 static int max_aio_per_proc = MAX_AIO_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
167 "Maximum active aio requests per process");
169 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
171 &max_aio_queue_per_proc, 0,
172 "Maximum queued aio requests per process");
174 static int max_buf_aio = MAX_BUF_AIO;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
176 "Maximum buf aio requests per process");
179 * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
180 * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
181 * vfs.aio.aio_listio_max.
183 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
184 CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
185 0, "Maximum aio requests for a single lio_listio call");
187 #ifdef COMPAT_FREEBSD6
188 typedef struct oaiocb {
189 int aio_fildes; /* File descriptor */
190 off_t aio_offset; /* File offset for I/O */
191 volatile void *aio_buf; /* I/O buffer in process space */
192 size_t aio_nbytes; /* Number of bytes for I/O */
193 struct osigevent aio_sigevent; /* Signal to deliver */
194 int aio_lio_opcode; /* LIO opcode */
195 int aio_reqprio; /* Request priority -- ignored */
196 struct __aiocb_private _aiocb_private;
201 * Below is a key of locks used to protect each member of struct kaiocb
202 * aioliojob and kaioinfo and any backends.
204 * * - need not protected
205 * a - locked by kaioinfo lock
206 * b - locked by backend lock, the backend lock can be null in some cases,
207 * for example, BIO belongs to this type, in this case, proc lock is
209 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
213 * If the routine that services an AIO request blocks while running in an
214 * AIO kernel process it can starve other I/O requests. BIO requests
215 * queued via aio_qbio() complete asynchronously and do not use AIO kernel
216 * processes at all. Socket I/O requests use a separate pool of
217 * kprocs and also force non-blocking I/O. Other file I/O requests
218 * use the generic fo_read/fo_write operations which can block. The
219 * fsync and mlock operations can also block while executing. Ideally
220 * none of these requests would block while executing.
222 * Note that the service routines cannot toggle O_NONBLOCK in the file
223 * structure directly while handling a request due to races with
228 #define KAIOCB_QUEUEING 0x01
229 #define KAIOCB_CANCELLED 0x02
230 #define KAIOCB_CANCELLING 0x04
231 #define KAIOCB_CHECKSYNC 0x08
232 #define KAIOCB_CLEARED 0x10
233 #define KAIOCB_FINISHED 0x20
238 #define AIOP_FREE 0x1 /* proc on free queue */
241 int aioprocflags; /* (c) AIO proc flags */
242 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
243 struct proc *aioproc; /* (*) the AIO proc */
247 * data-structure for lio signal management
250 int lioj_flags; /* (a) listio flags */
251 int lioj_count; /* (a) count of jobs */
252 int lioj_finished_count; /* (a) count of finished jobs */
253 struct sigevent lioj_signal; /* (a) signal on all I/O done */
254 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
255 struct knlist klist; /* (a) list of knotes */
256 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
259 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
260 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
261 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
264 * per process aio data structure
267 struct mtx kaio_mtx; /* the lock to protect this struct */
268 int kaio_flags; /* (a) per process kaio flags */
269 int kaio_active_count; /* (c) number of currently used AIOs */
270 int kaio_count; /* (a) size of AIO queue */
271 int kaio_buffer_count; /* (a) number of bio buffers */
272 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */
273 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */
274 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */
276 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */
277 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */
278 struct task kaio_task; /* (*) task to kick aio processes */
279 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */
282 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
287 #define KAIO_RUNDOWN 0x1 /* process is being run down */
288 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */
291 * Operations used to interact with userland aio control blocks.
292 * Different ABIs provide their own operations.
295 int (*aio_copyin)(struct aiocb *ujob, struct aiocb *kjob);
296 long (*fetch_status)(struct aiocb *ujob);
297 long (*fetch_error)(struct aiocb *ujob);
298 int (*store_status)(struct aiocb *ujob, long status);
299 int (*store_error)(struct aiocb *ujob, long error);
300 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
304 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */
308 static struct unrhdr *aiod_unr;
310 void aio_init_aioinfo(struct proc *p);
311 static int aio_onceonly(void);
312 static int aio_free_entry(struct kaiocb *job);
313 static void aio_process_rw(struct kaiocb *job);
314 static void aio_process_sync(struct kaiocb *job);
315 static void aio_process_mlock(struct kaiocb *job);
316 static void aio_schedule_fsync(void *context, int pending);
317 static int aio_newproc(int *);
318 int aio_aqueue(struct thread *td, struct aiocb *ujob,
319 struct aioliojob *lio, int type, struct aiocb_ops *ops);
320 static int aio_queue_file(struct file *fp, struct kaiocb *job);
321 static void aio_biowakeup(struct bio *bp);
322 static void aio_proc_rundown(void *arg, struct proc *p);
323 static void aio_proc_rundown_exec(void *arg, struct proc *p,
324 struct image_params *imgp);
325 static int aio_qbio(struct proc *p, struct kaiocb *job);
326 static void aio_daemon(void *param);
327 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
328 static bool aio_clear_cancel_function_locked(struct kaiocb *job);
329 static int aio_kick(struct proc *userp);
330 static void aio_kick_nowait(struct proc *userp);
331 static void aio_kick_helper(void *context, int pending);
332 static int filt_aioattach(struct knote *kn);
333 static void filt_aiodetach(struct knote *kn);
334 static int filt_aio(struct knote *kn, long hint);
335 static int filt_lioattach(struct knote *kn);
336 static void filt_liodetach(struct knote *kn);
337 static int filt_lio(struct knote *kn, long hint);
341 * kaio Per process async io info
342 * aiop async io process data
343 * aiocb async io jobs
344 * aiolio list io jobs
346 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
348 /* kqueue filters for aio */
349 static struct filterops aio_filtops = {
351 .f_attach = filt_aioattach,
352 .f_detach = filt_aiodetach,
355 static struct filterops lio_filtops = {
357 .f_attach = filt_lioattach,
358 .f_detach = filt_liodetach,
362 static eventhandler_tag exit_tag, exec_tag;
364 TASKQUEUE_DEFINE_THREAD(aiod_kick);
367 * Main operations function for use as a kernel module.
370 aio_modload(struct module *module, int cmd, void *arg)
387 static moduledata_t aio_mod = {
393 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
394 MODULE_VERSION(aio, 1);
397 * Startup initialization
403 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
404 EVENTHANDLER_PRI_ANY);
405 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
406 NULL, EVENTHANDLER_PRI_ANY);
407 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
408 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
409 TAILQ_INIT(&aio_freeproc);
410 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
411 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
412 TAILQ_INIT(&aio_jobs);
413 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
414 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
415 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
416 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
417 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
418 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
419 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
420 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
421 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
422 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
424 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
425 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
426 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
432 * Init the per-process aioinfo structure. The aioinfo limits are set
433 * per-process for user limit (resource) management.
436 aio_init_aioinfo(struct proc *p)
440 ki = uma_zalloc(kaio_zone, M_WAITOK);
441 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
443 ki->kaio_active_count = 0;
445 ki->kaio_buffer_count = 0;
446 TAILQ_INIT(&ki->kaio_all);
447 TAILQ_INIT(&ki->kaio_done);
448 TAILQ_INIT(&ki->kaio_jobqueue);
449 TAILQ_INIT(&ki->kaio_liojoblist);
450 TAILQ_INIT(&ki->kaio_syncqueue);
451 TAILQ_INIT(&ki->kaio_syncready);
452 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
453 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
455 if (p->p_aioinfo == NULL) {
460 mtx_destroy(&ki->kaio_mtx);
461 uma_zfree(kaio_zone, ki);
464 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
469 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
474 error = sigev_findtd(p, sigev, &td);
478 ksiginfo_set_sigev(ksi, sigev);
479 ksi->ksi_code = SI_ASYNCIO;
480 ksi->ksi_flags |= KSI_EXT | KSI_INS;
481 tdsendsignal(p, td, ksi->ksi_signo, ksi);
488 * Free a job entry. Wait for completion if it is currently active, but don't
489 * delay forever. If we delay, we return a flag that says that we have to
490 * restart the queue scan.
493 aio_free_entry(struct kaiocb *job)
496 struct aioliojob *lj;
504 AIO_LOCK_ASSERT(ki, MA_OWNED);
505 MPASS(job->jobflags & KAIOCB_FINISHED);
507 atomic_subtract_int(&num_queue_count, 1);
510 MPASS(ki->kaio_count >= 0);
512 TAILQ_REMOVE(&ki->kaio_done, job, plist);
513 TAILQ_REMOVE(&ki->kaio_all, job, allist);
518 lj->lioj_finished_count--;
520 if (lj->lioj_count == 0) {
521 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
522 /* lio is going away, we need to destroy any knotes */
523 knlist_delete(&lj->klist, curthread, 1);
525 sigqueue_take(&lj->lioj_ksi);
527 uma_zfree(aiolio_zone, lj);
531 /* job is going away, we need to destroy any knotes */
532 knlist_delete(&job->klist, curthread, 1);
534 sigqueue_take(&job->ksi);
540 * The thread argument here is used to find the owning process
541 * and is also passed to fo_close() which may pass it to various
542 * places such as devsw close() routines. Because of that, we
543 * need a thread pointer from the process owning the job that is
544 * persistent and won't disappear out from under us or move to
547 * Currently, all the callers of this function call it to remove
548 * a kaiocb from the current process' job list either via a
549 * syscall or due to the current process calling exit() or
550 * execve(). Thus, we know that p == curproc. We also know that
551 * curthread can't exit since we are curthread.
553 * Therefore, we use curthread as the thread to pass to
554 * knlist_delete(). This does mean that it is possible for the
555 * thread pointer at close time to differ from the thread pointer
556 * at open time, but this is already true of file descriptors in
557 * a multithreaded process.
560 fdrop(job->fd_file, curthread);
562 uma_zfree(aiocb_zone, job);
569 aio_proc_rundown_exec(void *arg, struct proc *p,
570 struct image_params *imgp __unused)
572 aio_proc_rundown(arg, p);
576 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
578 aio_cancel_fn_t *func;
581 AIO_LOCK_ASSERT(ki, MA_OWNED);
582 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
584 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
585 job->jobflags |= KAIOCB_CANCELLED;
587 func = job->cancel_fn;
590 * If there is no cancel routine, just leave the job marked as
591 * cancelled. The job should be in active use by a caller who
592 * should complete it normally or when it fails to install a
599 * Set the CANCELLING flag so that aio_complete() will defer
600 * completions of this job. This prevents the job from being
601 * freed out from under the cancel callback. After the
602 * callback any deferred completion (whether from the callback
603 * or any other source) will be completed.
605 job->jobflags |= KAIOCB_CANCELLING;
609 job->jobflags &= ~KAIOCB_CANCELLING;
610 if (job->jobflags & KAIOCB_FINISHED) {
611 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
612 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
613 aio_bio_done_notify(p, job);
616 * The cancel callback might have scheduled an
617 * operation to cancel this request, but it is
618 * only counted as cancelled if the request is
619 * cancelled when the callback returns.
627 * Rundown the jobs for a given process.
630 aio_proc_rundown(void *arg, struct proc *p)
633 struct aioliojob *lj;
634 struct kaiocb *job, *jobn;
636 KASSERT(curthread->td_proc == p,
637 ("%s: called on non-curproc", __func__));
643 ki->kaio_flags |= KAIO_RUNDOWN;
648 * Try to cancel all pending requests. This code simulates
649 * aio_cancel on all pending I/O requests.
651 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
652 aio_cancel_job(p, ki, job);
655 /* Wait for all running I/O to be finished */
656 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
657 ki->kaio_flags |= KAIO_WAKEUP;
658 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
662 /* Free all completed I/O requests. */
663 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
666 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
667 if (lj->lioj_count == 0) {
668 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
669 knlist_delete(&lj->klist, curthread, 1);
671 sigqueue_take(&lj->lioj_ksi);
673 uma_zfree(aiolio_zone, lj);
675 panic("LIO job not cleaned up: C:%d, FC:%d\n",
676 lj->lioj_count, lj->lioj_finished_count);
680 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
681 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
682 mtx_destroy(&ki->kaio_mtx);
683 uma_zfree(kaio_zone, ki);
688 * Select a job to run (called by an AIO daemon).
690 static struct kaiocb *
691 aio_selectjob(struct aioproc *aiop)
697 mtx_assert(&aio_job_mtx, MA_OWNED);
699 TAILQ_FOREACH(job, &aio_jobs, list) {
700 userp = job->userproc;
701 ki = userp->p_aioinfo;
703 if (ki->kaio_active_count < max_aio_per_proc) {
704 TAILQ_REMOVE(&aio_jobs, job, list);
705 if (!aio_clear_cancel_function(job))
708 /* Account for currently active jobs. */
709 ki->kaio_active_count++;
717 * Move all data to a permanent storage device. This code
718 * simulates the fsync syscall.
721 aio_fsync_vnode(struct thread *td, struct vnode *vp)
726 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
728 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
729 if (vp->v_object != NULL) {
730 VM_OBJECT_WLOCK(vp->v_object);
731 vm_object_page_clean(vp->v_object, 0, 0, 0);
732 VM_OBJECT_WUNLOCK(vp->v_object);
734 error = VOP_FSYNC(vp, MNT_WAIT, td);
737 vn_finished_write(mp);
743 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
744 * does the I/O request for the non-bio version of the operations. The normal
745 * vn operations are used, and this code should work in all instances for every
746 * type of file, including pipes, sockets, fifos, and regular files.
748 * XXX I don't think it works well for socket, pipe, and fifo.
751 aio_process_rw(struct kaiocb *job)
753 struct ucred *td_savedcred;
760 long msgsnd_st, msgsnd_end;
761 long msgrcv_st, msgrcv_end;
762 long oublock_st, oublock_end;
763 long inblock_st, inblock_end;
766 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
767 job->uaiocb.aio_lio_opcode == LIO_WRITE,
768 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
770 aio_switch_vmspace(job);
772 td_savedcred = td->td_ucred;
773 td->td_ucred = job->cred;
777 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
778 aiov.iov_len = cb->aio_nbytes;
780 auio.uio_iov = &aiov;
782 auio.uio_offset = cb->aio_offset;
783 auio.uio_resid = cb->aio_nbytes;
784 cnt = cb->aio_nbytes;
785 auio.uio_segflg = UIO_USERSPACE;
788 msgrcv_st = td->td_ru.ru_msgrcv;
789 msgsnd_st = td->td_ru.ru_msgsnd;
790 inblock_st = td->td_ru.ru_inblock;
791 oublock_st = td->td_ru.ru_oublock;
794 * aio_aqueue() acquires a reference to the file that is
795 * released in aio_free_entry().
797 if (cb->aio_lio_opcode == LIO_READ) {
798 auio.uio_rw = UIO_READ;
799 if (auio.uio_resid == 0)
802 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
804 if (fp->f_type == DTYPE_VNODE)
806 auio.uio_rw = UIO_WRITE;
807 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
809 msgrcv_end = td->td_ru.ru_msgrcv;
810 msgsnd_end = td->td_ru.ru_msgsnd;
811 inblock_end = td->td_ru.ru_inblock;
812 oublock_end = td->td_ru.ru_oublock;
814 job->msgrcv = msgrcv_end - msgrcv_st;
815 job->msgsnd = msgsnd_end - msgsnd_st;
816 job->inblock = inblock_end - inblock_st;
817 job->outblock = oublock_end - oublock_st;
819 if ((error) && (auio.uio_resid != cnt)) {
820 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
822 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
823 PROC_LOCK(job->userproc);
824 kern_psignal(job->userproc, SIGPIPE);
825 PROC_UNLOCK(job->userproc);
829 cnt -= auio.uio_resid;
830 td->td_ucred = td_savedcred;
832 aio_complete(job, -1, error);
834 aio_complete(job, cnt, 0);
838 aio_process_sync(struct kaiocb *job)
840 struct thread *td = curthread;
841 struct ucred *td_savedcred = td->td_ucred;
842 struct file *fp = job->fd_file;
845 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
846 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
848 td->td_ucred = job->cred;
849 if (fp->f_vnode != NULL)
850 error = aio_fsync_vnode(td, fp->f_vnode);
851 td->td_ucred = td_savedcred;
853 aio_complete(job, -1, error);
855 aio_complete(job, 0, 0);
859 aio_process_mlock(struct kaiocb *job)
861 struct aiocb *cb = &job->uaiocb;
864 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
865 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
867 aio_switch_vmspace(job);
868 error = kern_mlock(job->userproc, job->cred,
869 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
870 aio_complete(job, error != 0 ? -1 : 0, error);
874 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
876 struct aioliojob *lj;
878 struct kaiocb *sjob, *sjobn;
882 ki = userp->p_aioinfo;
883 AIO_LOCK_ASSERT(ki, MA_OWNED);
887 lj->lioj_finished_count++;
888 if (lj->lioj_count == lj->lioj_finished_count)
891 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
892 MPASS(job->jobflags & KAIOCB_FINISHED);
894 if (ki->kaio_flags & KAIO_RUNDOWN)
895 goto notification_done;
897 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
898 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
899 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
901 KNOTE_LOCKED(&job->klist, 1);
904 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
905 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
906 KNOTE_LOCKED(&lj->klist, 1);
908 if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
910 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
911 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
912 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
913 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
918 if (job->jobflags & KAIOCB_CHECKSYNC) {
919 schedule_fsync = false;
920 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
921 if (job->fd_file != sjob->fd_file ||
922 job->seqno >= sjob->seqno)
924 if (--sjob->pending > 0)
926 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
927 if (!aio_clear_cancel_function_locked(sjob))
929 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
930 schedule_fsync = true;
933 taskqueue_enqueue(taskqueue_aiod_kick,
934 &ki->kaio_sync_task);
936 if (ki->kaio_flags & KAIO_WAKEUP) {
937 ki->kaio_flags &= ~KAIO_WAKEUP;
938 wakeup(&userp->p_aioinfo);
943 aio_schedule_fsync(void *context, int pending)
950 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
951 job = TAILQ_FIRST(&ki->kaio_syncready);
952 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
954 aio_schedule(job, aio_process_sync);
961 aio_cancel_cleared(struct kaiocb *job)
965 * The caller should hold the same queue lock held when
966 * aio_clear_cancel_function() was called and set this flag
967 * ensuring this check sees an up-to-date value. However,
968 * there is no way to assert that.
970 return ((job->jobflags & KAIOCB_CLEARED) != 0);
974 aio_clear_cancel_function_locked(struct kaiocb *job)
977 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
978 MPASS(job->cancel_fn != NULL);
979 if (job->jobflags & KAIOCB_CANCELLING) {
980 job->jobflags |= KAIOCB_CLEARED;
983 job->cancel_fn = NULL;
988 aio_clear_cancel_function(struct kaiocb *job)
993 ki = job->userproc->p_aioinfo;
995 ret = aio_clear_cancel_function_locked(job);
1001 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1004 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1005 if (job->jobflags & KAIOCB_CANCELLED)
1007 job->cancel_fn = func;
1012 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1014 struct kaioinfo *ki;
1017 ki = job->userproc->p_aioinfo;
1019 ret = aio_set_cancel_function_locked(job, func);
1025 aio_complete(struct kaiocb *job, long status, int error)
1027 struct kaioinfo *ki;
1030 job->uaiocb._aiocb_private.error = error;
1031 job->uaiocb._aiocb_private.status = status;
1033 userp = job->userproc;
1034 ki = userp->p_aioinfo;
1037 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1038 ("duplicate aio_complete"));
1039 job->jobflags |= KAIOCB_FINISHED;
1040 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1041 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1042 aio_bio_done_notify(userp, job);
1048 aio_cancel(struct kaiocb *job)
1051 aio_complete(job, -1, ECANCELED);
1055 aio_switch_vmspace(struct kaiocb *job)
1058 vmspace_switch_aio(job->userproc->p_vmspace);
1062 * The AIO daemon, most of the actual work is done in aio_process_*,
1063 * but the setup (and address space mgmt) is done in this routine.
1066 aio_daemon(void *_id)
1069 struct aioproc *aiop;
1070 struct kaioinfo *ki;
1072 struct vmspace *myvm;
1073 struct thread *td = curthread;
1074 int id = (intptr_t)_id;
1077 * Grab an extra reference on the daemon's vmspace so that it
1078 * doesn't get freed by jobs that switch to a different
1082 myvm = vmspace_acquire_ref(p);
1084 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1087 * Allocate and ready the aio control info. There is one aiop structure
1090 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1092 aiop->aioprocflags = 0;
1095 * Wakeup parent process. (Parent sleeps to keep from blasting away
1096 * and creating too many daemons.)
1098 sema_post(&aio_newproc_sem);
1100 mtx_lock(&aio_job_mtx);
1103 * Take daemon off of free queue
1105 if (aiop->aioprocflags & AIOP_FREE) {
1106 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1107 aiop->aioprocflags &= ~AIOP_FREE;
1113 while ((job = aio_selectjob(aiop)) != NULL) {
1114 mtx_unlock(&aio_job_mtx);
1116 ki = job->userproc->p_aioinfo;
1117 job->handle_fn(job);
1119 mtx_lock(&aio_job_mtx);
1120 /* Decrement the active job count. */
1121 ki->kaio_active_count--;
1125 * Disconnect from user address space.
1127 if (p->p_vmspace != myvm) {
1128 mtx_unlock(&aio_job_mtx);
1129 vmspace_switch_aio(myvm);
1130 mtx_lock(&aio_job_mtx);
1132 * We have to restart to avoid race, we only sleep if
1133 * no job can be selected.
1138 mtx_assert(&aio_job_mtx, MA_OWNED);
1140 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1141 aiop->aioprocflags |= AIOP_FREE;
1144 * If daemon is inactive for a long time, allow it to exit,
1145 * thereby freeing resources.
1147 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1148 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1149 (aiop->aioprocflags & AIOP_FREE) &&
1150 num_aio_procs > target_aio_procs)
1153 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1155 mtx_unlock(&aio_job_mtx);
1156 uma_zfree(aiop_zone, aiop);
1157 free_unr(aiod_unr, id);
1160 KASSERT(p->p_vmspace == myvm,
1161 ("AIOD: bad vmspace for exiting daemon"));
1162 KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1163 ("AIOD: bad vm refcnt for exiting daemon: %d",
1164 refcount_load(&myvm->vm_refcnt)));
1169 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1170 * AIO daemon modifies its environment itself.
1173 aio_newproc(int *start)
1179 id = alloc_unr(aiod_unr);
1180 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1181 RFNOWAIT, 0, "aiod%d", id);
1184 * Wait until daemon is started.
1186 sema_wait(&aio_newproc_sem);
1187 mtx_lock(&aio_job_mtx);
1191 mtx_unlock(&aio_job_mtx);
1193 free_unr(aiod_unr, id);
1199 * Try the high-performance, low-overhead bio method for eligible
1200 * VCHR devices. This method doesn't use an aio helper thread, and
1201 * thus has very low overhead.
1203 * Assumes that the caller, aio_aqueue(), has incremented the file
1204 * structure's reference count, preventing its deallocation for the
1205 * duration of this call.
1208 aio_qbio(struct proc *p, struct kaiocb *job)
1217 struct kaioinfo *ki;
1218 int error, ref, poff;
1224 if (!(cb->aio_lio_opcode == LIO_WRITE ||
1225 cb->aio_lio_opcode == LIO_READ))
1227 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1231 if (vp->v_type != VCHR)
1233 if (vp->v_bufobj.bo_bsize == 0)
1235 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1239 csw = devvn_refthread(vp, &dev, &ref);
1243 if ((csw->d_flags & D_DISK) == 0) {
1247 if (cb->aio_nbytes > dev->si_iosize_max) {
1253 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1254 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1255 if (cb->aio_nbytes > maxphys) {
1261 job->pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1262 cb->aio_nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1264 if (cb->aio_nbytes > maxphys) {
1268 if (ki->kaio_buffer_count >= max_buf_aio) {
1273 job->pbuf = pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1276 ki->kaio_buffer_count++;
1278 job->pages = pbuf->b_pages;
1280 job->bp = bp = g_alloc_bio();
1282 bp->bio_length = cb->aio_nbytes;
1283 bp->bio_bcount = cb->aio_nbytes;
1284 bp->bio_done = aio_biowakeup;
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)cb->aio_buf, bp->bio_length, prot, 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;
1311 atomic_add_int(&num_unmapped_aio, 1);
1314 /* Perform transfer. */
1315 csw->d_strategy(bp);
1316 dev_relthread(dev, ref);
1322 ki->kaio_buffer_count--;
1324 uma_zfree(pbuf_zone, pbuf);
1327 free(job->pages, M_TEMP);
1332 dev_relthread(dev, ref);
1336 #ifdef COMPAT_FREEBSD6
1338 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1342 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1343 * supported by AIO with the old sigevent structure.
1345 nsig->sigev_notify = osig->sigev_notify;
1346 switch (nsig->sigev_notify) {
1350 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1353 nsig->sigev_notify_kqueue =
1354 osig->__sigev_u.__sigev_notify_kqueue;
1355 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1364 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1366 struct oaiocb *ojob;
1369 bzero(kjob, sizeof(struct aiocb));
1370 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1373 ojob = (struct oaiocb *)kjob;
1374 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1379 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1382 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1386 aiocb_fetch_status(struct aiocb *ujob)
1389 return (fuword(&ujob->_aiocb_private.status));
1393 aiocb_fetch_error(struct aiocb *ujob)
1396 return (fuword(&ujob->_aiocb_private.error));
1400 aiocb_store_status(struct aiocb *ujob, long status)
1403 return (suword(&ujob->_aiocb_private.status, status));
1407 aiocb_store_error(struct aiocb *ujob, long error)
1410 return (suword(&ujob->_aiocb_private.error, error));
1414 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1417 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1421 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1424 return (suword(ujobp, (long)ujob));
1427 static struct aiocb_ops aiocb_ops = {
1428 .aio_copyin = aiocb_copyin,
1429 .fetch_status = aiocb_fetch_status,
1430 .fetch_error = aiocb_fetch_error,
1431 .store_status = aiocb_store_status,
1432 .store_error = aiocb_store_error,
1433 .store_kernelinfo = aiocb_store_kernelinfo,
1434 .store_aiocb = aiocb_store_aiocb,
1437 #ifdef COMPAT_FREEBSD6
1438 static struct aiocb_ops aiocb_ops_osigevent = {
1439 .aio_copyin = aiocb_copyin_old_sigevent,
1440 .fetch_status = aiocb_fetch_status,
1441 .fetch_error = aiocb_fetch_error,
1442 .store_status = aiocb_store_status,
1443 .store_error = aiocb_store_error,
1444 .store_kernelinfo = aiocb_store_kernelinfo,
1445 .store_aiocb = aiocb_store_aiocb,
1450 * Queue a new AIO request. Choosing either the threaded or direct bio VCHR
1451 * technique is done in this code.
1454 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1455 int type, struct aiocb_ops *ops)
1457 struct proc *p = td->td_proc;
1460 struct kaioinfo *ki;
1468 if (p->p_aioinfo == NULL)
1469 aio_init_aioinfo(p);
1473 ops->store_status(ujob, -1);
1474 ops->store_error(ujob, 0);
1475 ops->store_kernelinfo(ujob, -1);
1477 if (num_queue_count >= max_queue_count ||
1478 ki->kaio_count >= max_aio_queue_per_proc) {
1479 ops->store_error(ujob, EAGAIN);
1483 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1484 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1486 error = ops->aio_copyin(ujob, &job->uaiocb);
1488 ops->store_error(ujob, error);
1489 uma_zfree(aiocb_zone, job);
1493 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1494 uma_zfree(aiocb_zone, job);
1498 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1499 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1500 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1501 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1502 ops->store_error(ujob, EINVAL);
1503 uma_zfree(aiocb_zone, job);
1507 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1508 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1509 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1510 uma_zfree(aiocb_zone, job);
1514 ksiginfo_init(&job->ksi);
1516 /* Save userspace address of the job info. */
1519 /* Get the opcode. */
1520 if (type != LIO_NOP)
1521 job->uaiocb.aio_lio_opcode = type;
1522 opcode = job->uaiocb.aio_lio_opcode;
1525 * Validate the opcode and fetch the file object for the specified
1528 * XXXRW: Moved the opcode validation up here so that we don't
1529 * retrieve a file descriptor without knowing what the capabiltity
1532 fd = job->uaiocb.aio_fildes;
1535 error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1538 error = fget_read(td, fd, &cap_pread_rights, &fp);
1541 error = fget(td, fd, &cap_fsync_rights, &fp);
1547 error = fget(td, fd, &cap_no_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 memset(&kev, 0, sizeof(kev));
1598 kev.ident = (uintptr_t)job->ujob;
1599 kev.filter = EVFILT_AIO;
1600 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1601 kev.data = (intptr_t)job;
1602 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1603 error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1609 ops->store_error(ujob, EINPROGRESS);
1610 job->uaiocb._aiocb_private.error = EINPROGRESS;
1612 job->cred = crhold(td->td_ucred);
1613 job->jobflags = KAIOCB_QUEUEING;
1616 if (opcode == LIO_MLOCK) {
1617 aio_schedule(job, aio_process_mlock);
1619 } else if (fp->f_ops->fo_aio_queue == NULL)
1620 error = aio_queue_file(fp, job);
1622 error = fo_aio_queue(fp, job);
1627 job->jobflags &= ~KAIOCB_QUEUEING;
1628 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1632 atomic_add_int(&num_queue_count, 1);
1633 if (job->jobflags & KAIOCB_FINISHED) {
1635 * The queue callback completed the request synchronously.
1636 * The bulk of the completion is deferred in that case
1639 aio_bio_done_notify(p, job);
1641 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1646 knlist_delete(&job->klist, curthread, 0);
1649 uma_zfree(aiocb_zone, job);
1650 ops->store_error(ujob, error);
1655 aio_cancel_daemon_job(struct kaiocb *job)
1658 mtx_lock(&aio_job_mtx);
1659 if (!aio_cancel_cleared(job))
1660 TAILQ_REMOVE(&aio_jobs, job, list);
1661 mtx_unlock(&aio_job_mtx);
1666 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1669 mtx_lock(&aio_job_mtx);
1670 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1671 mtx_unlock(&aio_job_mtx);
1675 job->handle_fn = func;
1676 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1677 aio_kick_nowait(job->userproc);
1678 mtx_unlock(&aio_job_mtx);
1682 aio_cancel_sync(struct kaiocb *job)
1684 struct kaioinfo *ki;
1686 ki = job->userproc->p_aioinfo;
1688 if (!aio_cancel_cleared(job))
1689 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1695 aio_queue_file(struct file *fp, struct kaiocb *job)
1697 struct kaioinfo *ki;
1698 struct kaiocb *job2;
1704 ki = job->userproc->p_aioinfo;
1705 error = aio_qbio(job->userproc, job);
1709 if (fp->f_type == DTYPE_VNODE) {
1711 if (vp->v_type == VREG || vp->v_type == VDIR) {
1712 mp = fp->f_vnode->v_mount;
1713 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1717 if (!(safe || enable_aio_unsafe)) {
1718 counted_warning(&unsafe_warningcnt,
1719 "is attempting to use unsafe AIO requests");
1720 return (EOPNOTSUPP);
1723 switch (job->uaiocb.aio_lio_opcode) {
1726 aio_schedule(job, aio_process_rw);
1731 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1732 if (job2->fd_file == job->fd_file &&
1733 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1734 job2->seqno < job->seqno) {
1735 job2->jobflags |= KAIOCB_CHECKSYNC;
1739 if (job->pending != 0) {
1740 if (!aio_set_cancel_function_locked(job,
1746 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1751 aio_schedule(job, aio_process_sync);
1761 aio_kick_nowait(struct proc *userp)
1763 struct kaioinfo *ki = userp->p_aioinfo;
1764 struct aioproc *aiop;
1766 mtx_assert(&aio_job_mtx, MA_OWNED);
1767 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1768 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1769 aiop->aioprocflags &= ~AIOP_FREE;
1770 wakeup(aiop->aioproc);
1771 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1772 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1773 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1778 aio_kick(struct proc *userp)
1780 struct kaioinfo *ki = userp->p_aioinfo;
1781 struct aioproc *aiop;
1784 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 < max_aio_per_proc) {
1792 num_aio_resv_start++;
1793 mtx_unlock(&aio_job_mtx);
1794 error = aio_newproc(&num_aio_resv_start);
1795 mtx_lock(&aio_job_mtx);
1797 num_aio_resv_start--;
1807 aio_kick_helper(void *context, int pending)
1809 struct proc *userp = context;
1811 mtx_lock(&aio_job_mtx);
1812 while (--pending >= 0) {
1813 if (aio_kick(userp))
1816 mtx_unlock(&aio_job_mtx);
1820 * Support the aio_return system call, as a side-effect, kernel resources are
1824 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1826 struct proc *p = td->td_proc;
1828 struct kaioinfo *ki;
1835 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1836 if (job->ujob == ujob)
1840 MPASS(job->jobflags & KAIOCB_FINISHED);
1841 status = job->uaiocb._aiocb_private.status;
1842 error = job->uaiocb._aiocb_private.error;
1843 td->td_retval[0] = status;
1844 td->td_ru.ru_oublock += job->outblock;
1845 td->td_ru.ru_inblock += job->inblock;
1846 td->td_ru.ru_msgsnd += job->msgsnd;
1847 td->td_ru.ru_msgrcv += job->msgrcv;
1848 aio_free_entry(job);
1850 ops->store_error(ujob, error);
1851 ops->store_status(ujob, status);
1860 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1863 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1867 * Allow a process to wakeup when any of the I/O requests are completed.
1870 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1871 struct timespec *ts)
1873 struct proc *p = td->td_proc;
1875 struct kaioinfo *ki;
1876 struct kaiocb *firstjob, *job;
1881 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1884 TIMESPEC_TO_TIMEVAL(&atv, ts);
1885 if (itimerfix(&atv))
1887 timo = tvtohz(&atv);
1901 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1902 for (i = 0; i < njoblist; i++) {
1903 if (job->ujob == ujoblist[i]) {
1904 if (firstjob == NULL)
1906 if (job->jobflags & KAIOCB_FINISHED)
1911 /* All tasks were finished. */
1912 if (firstjob == NULL)
1915 ki->kaio_flags |= KAIO_WAKEUP;
1916 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1918 if (error == ERESTART)
1929 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1931 struct timespec ts, *tsp;
1932 struct aiocb **ujoblist;
1935 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1939 /* Get timespec struct. */
1940 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1946 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
1947 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1949 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1950 free(ujoblist, M_AIOS);
1955 * aio_cancel cancels any non-bio aio operations not currently in progress.
1958 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1960 struct proc *p = td->td_proc;
1961 struct kaioinfo *ki;
1962 struct kaiocb *job, *jobn;
1966 int notcancelled = 0;
1969 /* Lookup file object. */
1970 error = fget(td, uap->fd, &cap_no_rights, &fp);
1978 if (fp->f_type == DTYPE_VNODE) {
1980 if (vn_isdisk(vp)) {
1982 td->td_retval[0] = AIO_NOTCANCELED;
1988 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1989 if ((uap->fd == job->uaiocb.aio_fildes) &&
1990 ((uap->aiocbp == NULL) ||
1991 (uap->aiocbp == job->ujob))) {
1992 if (aio_cancel_job(p, ki, job)) {
1997 if (uap->aiocbp != NULL)
2006 if (uap->aiocbp != NULL) {
2008 td->td_retval[0] = AIO_CANCELED;
2014 td->td_retval[0] = AIO_NOTCANCELED;
2019 td->td_retval[0] = AIO_CANCELED;
2023 td->td_retval[0] = AIO_ALLDONE;
2029 * aio_error is implemented in the kernel level for compatibility purposes
2030 * only. For a user mode async implementation, it would be best to do it in
2031 * a userland subroutine.
2034 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2036 struct proc *p = td->td_proc;
2038 struct kaioinfo *ki;
2043 td->td_retval[0] = EINVAL;
2048 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2049 if (job->ujob == ujob) {
2050 if (job->jobflags & KAIOCB_FINISHED)
2052 job->uaiocb._aiocb_private.error;
2054 td->td_retval[0] = EINPROGRESS;
2062 * Hack for failure of aio_aqueue.
2064 status = ops->fetch_status(ujob);
2066 td->td_retval[0] = ops->fetch_error(ujob);
2070 td->td_retval[0] = EINVAL;
2075 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2078 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2081 /* syscall - asynchronous read from a file (REALTIME) */
2082 #ifdef COMPAT_FREEBSD6
2084 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2087 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2088 &aiocb_ops_osigevent));
2093 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2096 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2099 /* syscall - asynchronous write to a file (REALTIME) */
2100 #ifdef COMPAT_FREEBSD6
2102 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2105 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2106 &aiocb_ops_osigevent));
2111 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2114 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2118 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2121 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2125 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2126 struct aiocb **acb_list, int nent, struct sigevent *sig,
2127 struct aiocb_ops *ops)
2129 struct proc *p = td->td_proc;
2131 struct kaioinfo *ki;
2132 struct aioliojob *lj;
2138 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2141 if (nent < 0 || nent > max_aio_queue_per_proc)
2144 if (p->p_aioinfo == NULL)
2145 aio_init_aioinfo(p);
2149 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2152 lj->lioj_finished_count = 0;
2153 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2154 ksiginfo_init(&lj->lioj_ksi);
2159 if (sig && (mode == LIO_NOWAIT)) {
2160 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2161 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2162 /* Assume only new style KEVENT */
2163 memset(&kev, 0, sizeof(kev));
2164 kev.filter = EVFILT_LIO;
2165 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2166 kev.ident = (uintptr_t)uacb_list; /* something unique */
2167 kev.data = (intptr_t)lj;
2168 /* pass user defined sigval data */
2169 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2170 error = kqfd_register(
2171 lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2174 uma_zfree(aiolio_zone, lj);
2177 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2179 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2180 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2181 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2182 uma_zfree(aiolio_zone, lj);
2185 lj->lioj_flags |= LIOJ_SIGNAL;
2187 uma_zfree(aiolio_zone, lj);
2193 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2195 * Add extra aiocb count to avoid the lio to be freed
2196 * by other threads doing aio_waitcomplete or aio_return,
2197 * and prevent event from being sent until we have queued
2204 * Get pointers to the list of I/O requests.
2208 for (i = 0; i < nent; i++) {
2211 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2212 if (error == EAGAIN)
2214 else if (error != 0)
2221 if (mode == LIO_WAIT) {
2222 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2223 ki->kaio_flags |= KAIO_WAKEUP;
2224 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2225 PRIBIO | PCATCH, "aiospn", 0);
2226 if (error == ERESTART)
2232 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2233 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2234 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2235 KNOTE_LOCKED(&lj->klist, 1);
2237 if ((lj->lioj_flags & (LIOJ_SIGNAL |
2238 LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2239 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2240 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2241 aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi);
2242 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2247 if (lj->lioj_count == 0) {
2248 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2249 knlist_delete(&lj->klist, curthread, 1);
2251 sigqueue_take(&lj->lioj_ksi);
2254 uma_zfree(aiolio_zone, lj);
2266 /* syscall - list directed I/O (REALTIME) */
2267 #ifdef COMPAT_FREEBSD6
2269 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2271 struct aiocb **acb_list;
2272 struct sigevent *sigp, sig;
2273 struct osigevent osig;
2276 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2280 if (nent < 0 || nent > max_aio_queue_per_proc)
2283 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2284 error = copyin(uap->sig, &osig, sizeof(osig));
2287 error = convert_old_sigevent(&osig, &sig);
2294 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2295 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2297 error = kern_lio_listio(td, uap->mode,
2298 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2299 &aiocb_ops_osigevent);
2300 free(acb_list, M_LIO);
2305 /* syscall - list directed I/O (REALTIME) */
2307 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2309 struct aiocb **acb_list;
2310 struct sigevent *sigp, sig;
2313 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2317 if (nent < 0 || nent > max_aio_queue_per_proc)
2320 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2321 error = copyin(uap->sig, &sig, sizeof(sig));
2328 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2329 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2331 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2332 nent, sigp, &aiocb_ops);
2333 free(acb_list, M_LIO);
2338 aio_biowakeup(struct bio *bp)
2340 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2341 struct kaioinfo *ki;
2345 /* Release mapping into kernel space. */
2346 if (job->pbuf != NULL) {
2347 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2348 vm_page_unhold_pages(job->pages, job->npages);
2349 uma_zfree(pbuf_zone, job->pbuf);
2351 atomic_subtract_int(&num_buf_aio, 1);
2352 ki = job->userproc->p_aioinfo;
2354 ki->kaio_buffer_count--;
2357 vm_page_unhold_pages(job->pages, job->npages);
2358 free(job->pages, M_TEMP);
2359 atomic_subtract_int(&num_unmapped_aio, 1);
2364 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2366 if (bp->bio_flags & BIO_ERROR)
2367 error = bp->bio_error;
2368 nblks = btodb(nbytes);
2369 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2370 job->outblock += nblks;
2372 job->inblock += nblks;
2375 aio_complete(job, -1, error);
2377 aio_complete(job, nbytes, 0);
2382 /* syscall - wait for the next completion of an aio request */
2384 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2385 struct timespec *ts, struct aiocb_ops *ops)
2387 struct proc *p = td->td_proc;
2389 struct kaioinfo *ki;
2395 ops->store_aiocb(ujobp, NULL);
2399 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2402 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2405 TIMESPEC_TO_TIMEVAL(&atv, ts);
2406 if (itimerfix(&atv))
2408 timo = tvtohz(&atv);
2411 if (p->p_aioinfo == NULL)
2412 aio_init_aioinfo(p);
2418 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2420 error = EWOULDBLOCK;
2423 ki->kaio_flags |= KAIO_WAKEUP;
2424 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2426 if (timo && error == ERESTART)
2433 MPASS(job->jobflags & KAIOCB_FINISHED);
2435 status = job->uaiocb._aiocb_private.status;
2436 error = job->uaiocb._aiocb_private.error;
2437 td->td_retval[0] = status;
2438 td->td_ru.ru_oublock += job->outblock;
2439 td->td_ru.ru_inblock += job->inblock;
2440 td->td_ru.ru_msgsnd += job->msgsnd;
2441 td->td_ru.ru_msgrcv += job->msgrcv;
2442 aio_free_entry(job);
2444 ops->store_aiocb(ujobp, ujob);
2445 ops->store_error(ujob, error);
2446 ops->store_status(ujob, status);
2454 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2456 struct timespec ts, *tsp;
2460 /* Get timespec struct. */
2461 error = copyin(uap->timeout, &ts, sizeof(ts));
2468 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2472 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2473 struct aiocb_ops *ops)
2476 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2478 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2482 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2485 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2488 /* kqueue attach function */
2490 filt_aioattach(struct knote *kn)
2494 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2497 * The job pointer must be validated before using it, so
2498 * registration is restricted to the kernel; the user cannot
2501 if ((kn->kn_flags & EV_FLAG1) == 0)
2503 kn->kn_ptr.p_aio = job;
2504 kn->kn_flags &= ~EV_FLAG1;
2506 knlist_add(&job->klist, kn, 0);
2511 /* kqueue detach function */
2513 filt_aiodetach(struct knote *kn)
2517 knl = &kn->kn_ptr.p_aio->klist;
2518 knl->kl_lock(knl->kl_lockarg);
2519 if (!knlist_empty(knl))
2520 knlist_remove(knl, kn, 1);
2521 knl->kl_unlock(knl->kl_lockarg);
2524 /* kqueue filter function */
2527 filt_aio(struct knote *kn, long hint)
2529 struct kaiocb *job = kn->kn_ptr.p_aio;
2531 kn->kn_data = job->uaiocb._aiocb_private.error;
2532 if (!(job->jobflags & KAIOCB_FINISHED))
2534 kn->kn_flags |= EV_EOF;
2538 /* kqueue attach function */
2540 filt_lioattach(struct knote *kn)
2542 struct aioliojob *lj;
2544 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2547 * The aioliojob pointer must be validated before using it, so
2548 * registration is restricted to the kernel; the user cannot
2551 if ((kn->kn_flags & EV_FLAG1) == 0)
2553 kn->kn_ptr.p_lio = lj;
2554 kn->kn_flags &= ~EV_FLAG1;
2556 knlist_add(&lj->klist, kn, 0);
2561 /* kqueue detach function */
2563 filt_liodetach(struct knote *kn)
2567 knl = &kn->kn_ptr.p_lio->klist;
2568 knl->kl_lock(knl->kl_lockarg);
2569 if (!knlist_empty(knl))
2570 knlist_remove(knl, kn, 1);
2571 knl->kl_unlock(knl->kl_lockarg);
2574 /* kqueue filter function */
2577 filt_lio(struct knote *kn, long hint)
2579 struct aioliojob * lj = kn->kn_ptr.p_lio;
2581 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2584 #ifdef COMPAT_FREEBSD32
2585 #include <sys/mount.h>
2586 #include <sys/socket.h>
2587 #include <compat/freebsd32/freebsd32.h>
2588 #include <compat/freebsd32/freebsd32_proto.h>
2589 #include <compat/freebsd32/freebsd32_signal.h>
2590 #include <compat/freebsd32/freebsd32_syscall.h>
2591 #include <compat/freebsd32/freebsd32_util.h>
2593 struct __aiocb_private32 {
2596 uint32_t kernelinfo;
2599 #ifdef COMPAT_FREEBSD6
2600 typedef struct oaiocb32 {
2601 int aio_fildes; /* File descriptor */
2602 uint64_t aio_offset __packed; /* File offset for I/O */
2603 uint32_t aio_buf; /* I/O buffer in process space */
2604 uint32_t aio_nbytes; /* Number of bytes for I/O */
2605 struct osigevent32 aio_sigevent; /* Signal to deliver */
2606 int aio_lio_opcode; /* LIO opcode */
2607 int aio_reqprio; /* Request priority -- ignored */
2608 struct __aiocb_private32 _aiocb_private;
2612 typedef struct aiocb32 {
2613 int32_t aio_fildes; /* File descriptor */
2614 uint64_t aio_offset __packed; /* File offset for I/O */
2615 uint32_t aio_buf; /* I/O buffer in process space */
2616 uint32_t aio_nbytes; /* Number of bytes for I/O */
2618 uint32_t __spare2__;
2619 int aio_lio_opcode; /* LIO opcode */
2620 int aio_reqprio; /* Request priority -- ignored */
2621 struct __aiocb_private32 _aiocb_private;
2622 struct sigevent32 aio_sigevent; /* Signal to deliver */
2625 #ifdef COMPAT_FREEBSD6
2627 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2631 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2632 * supported by AIO with the old sigevent structure.
2634 CP(*osig, *nsig, sigev_notify);
2635 switch (nsig->sigev_notify) {
2639 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2642 nsig->sigev_notify_kqueue =
2643 osig->__sigev_u.__sigev_notify_kqueue;
2644 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2653 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2655 struct oaiocb32 job32;
2658 bzero(kjob, sizeof(struct aiocb));
2659 error = copyin(ujob, &job32, sizeof(job32));
2663 CP(job32, *kjob, aio_fildes);
2664 CP(job32, *kjob, aio_offset);
2665 PTRIN_CP(job32, *kjob, aio_buf);
2666 CP(job32, *kjob, aio_nbytes);
2667 CP(job32, *kjob, aio_lio_opcode);
2668 CP(job32, *kjob, aio_reqprio);
2669 CP(job32, *kjob, _aiocb_private.status);
2670 CP(job32, *kjob, _aiocb_private.error);
2671 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2672 return (convert_old_sigevent32(&job32.aio_sigevent,
2673 &kjob->aio_sigevent));
2678 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2680 struct aiocb32 job32;
2683 error = copyin(ujob, &job32, sizeof(job32));
2686 CP(job32, *kjob, aio_fildes);
2687 CP(job32, *kjob, aio_offset);
2688 PTRIN_CP(job32, *kjob, aio_buf);
2689 CP(job32, *kjob, aio_nbytes);
2690 CP(job32, *kjob, aio_lio_opcode);
2691 CP(job32, *kjob, aio_reqprio);
2692 CP(job32, *kjob, _aiocb_private.status);
2693 CP(job32, *kjob, _aiocb_private.error);
2694 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2695 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2699 aiocb32_fetch_status(struct aiocb *ujob)
2701 struct aiocb32 *ujob32;
2703 ujob32 = (struct aiocb32 *)ujob;
2704 return (fuword32(&ujob32->_aiocb_private.status));
2708 aiocb32_fetch_error(struct aiocb *ujob)
2710 struct aiocb32 *ujob32;
2712 ujob32 = (struct aiocb32 *)ujob;
2713 return (fuword32(&ujob32->_aiocb_private.error));
2717 aiocb32_store_status(struct aiocb *ujob, long status)
2719 struct aiocb32 *ujob32;
2721 ujob32 = (struct aiocb32 *)ujob;
2722 return (suword32(&ujob32->_aiocb_private.status, status));
2726 aiocb32_store_error(struct aiocb *ujob, long error)
2728 struct aiocb32 *ujob32;
2730 ujob32 = (struct aiocb32 *)ujob;
2731 return (suword32(&ujob32->_aiocb_private.error, error));
2735 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2737 struct aiocb32 *ujob32;
2739 ujob32 = (struct aiocb32 *)ujob;
2740 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2744 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2747 return (suword32(ujobp, (long)ujob));
2750 static struct aiocb_ops aiocb32_ops = {
2751 .aio_copyin = aiocb32_copyin,
2752 .fetch_status = aiocb32_fetch_status,
2753 .fetch_error = aiocb32_fetch_error,
2754 .store_status = aiocb32_store_status,
2755 .store_error = aiocb32_store_error,
2756 .store_kernelinfo = aiocb32_store_kernelinfo,
2757 .store_aiocb = aiocb32_store_aiocb,
2760 #ifdef COMPAT_FREEBSD6
2761 static struct aiocb_ops aiocb32_ops_osigevent = {
2762 .aio_copyin = aiocb32_copyin_old_sigevent,
2763 .fetch_status = aiocb32_fetch_status,
2764 .fetch_error = aiocb32_fetch_error,
2765 .store_status = aiocb32_store_status,
2766 .store_error = aiocb32_store_error,
2767 .store_kernelinfo = aiocb32_store_kernelinfo,
2768 .store_aiocb = aiocb32_store_aiocb,
2773 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2776 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2780 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2782 struct timespec32 ts32;
2783 struct timespec ts, *tsp;
2784 struct aiocb **ujoblist;
2785 uint32_t *ujoblist32;
2788 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2792 /* Get timespec struct. */
2793 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2795 CP(ts32, ts, tv_sec);
2796 CP(ts32, ts, tv_nsec);
2801 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2802 ujoblist32 = (uint32_t *)ujoblist;
2803 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2804 sizeof(ujoblist32[0]));
2806 for (i = uap->nent - 1; i >= 0; i--)
2807 ujoblist[i] = PTRIN(ujoblist32[i]);
2809 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2811 free(ujoblist, M_AIOS);
2816 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2819 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2822 #ifdef COMPAT_FREEBSD6
2824 freebsd6_freebsd32_aio_read(struct thread *td,
2825 struct freebsd6_freebsd32_aio_read_args *uap)
2828 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2829 &aiocb32_ops_osigevent));
2834 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2837 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2841 #ifdef COMPAT_FREEBSD6
2843 freebsd6_freebsd32_aio_write(struct thread *td,
2844 struct freebsd6_freebsd32_aio_write_args *uap)
2847 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2848 &aiocb32_ops_osigevent));
2853 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2856 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2861 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2864 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2869 freebsd32_aio_waitcomplete(struct thread *td,
2870 struct freebsd32_aio_waitcomplete_args *uap)
2872 struct timespec32 ts32;
2873 struct timespec ts, *tsp;
2877 /* Get timespec struct. */
2878 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2881 CP(ts32, ts, tv_sec);
2882 CP(ts32, ts, tv_nsec);
2887 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2892 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2895 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2899 #ifdef COMPAT_FREEBSD6
2901 freebsd6_freebsd32_lio_listio(struct thread *td,
2902 struct freebsd6_freebsd32_lio_listio_args *uap)
2904 struct aiocb **acb_list;
2905 struct sigevent *sigp, sig;
2906 struct osigevent32 osig;
2907 uint32_t *acb_list32;
2910 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2914 if (nent < 0 || nent > max_aio_queue_per_proc)
2917 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2918 error = copyin(uap->sig, &osig, sizeof(osig));
2921 error = convert_old_sigevent32(&osig, &sig);
2928 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2929 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2931 free(acb_list32, M_LIO);
2934 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2935 for (i = 0; i < nent; i++)
2936 acb_list[i] = PTRIN(acb_list32[i]);
2937 free(acb_list32, M_LIO);
2939 error = kern_lio_listio(td, uap->mode,
2940 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2941 &aiocb32_ops_osigevent);
2942 free(acb_list, M_LIO);
2948 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2950 struct aiocb **acb_list;
2951 struct sigevent *sigp, sig;
2952 struct sigevent32 sig32;
2953 uint32_t *acb_list32;
2956 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2960 if (nent < 0 || nent > max_aio_queue_per_proc)
2963 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2964 error = copyin(uap->sig, &sig32, sizeof(sig32));
2967 error = convert_sigevent32(&sig32, &sig);
2974 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2975 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2977 free(acb_list32, M_LIO);
2980 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2981 for (i = 0; i < nent; i++)
2982 acb_list[i] = PTRIN(acb_list32[i]);
2983 free(acb_list32, M_LIO);
2985 error = kern_lio_listio(td, uap->mode,
2986 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2988 free(acb_list, M_LIO);