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 "opt_compat.h"
28 #include <sys/param.h>
29 #include <sys/systm.h>
30 #include <sys/malloc.h>
33 #include <sys/capsicum.h>
34 #include <sys/eventhandler.h>
35 #include <sys/sysproto.h>
36 #include <sys/filedesc.h>
37 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/kthread.h>
40 #include <sys/fcntl.h>
42 #include <sys/limits.h>
44 #include <sys/mutex.h>
45 #include <sys/unistd.h>
46 #include <sys/posix4.h>
48 #include <sys/resourcevar.h>
49 #include <sys/signalvar.h>
50 #include <sys/syscallsubr.h>
51 #include <sys/protosw.h>
52 #include <sys/rwlock.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syscall.h>
57 #include <sys/sysent.h>
58 #include <sys/sysctl.h>
59 #include <sys/syslog.h>
61 #include <sys/taskqueue.h>
62 #include <sys/vnode.h>
64 #include <sys/event.h>
65 #include <sys/mount.h>
66 #include <geom/geom.h>
68 #include <machine/atomic.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_extern.h>
74 #include <vm/vm_map.h>
75 #include <vm/vm_object.h>
80 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
81 * overflow. (XXX will be removed soon.)
83 static u_long jobrefid;
86 * Counter for aio_fsync.
88 static uint64_t jobseqno;
90 #ifndef MAX_AIO_PER_PROC
91 #define MAX_AIO_PER_PROC 32
94 #ifndef MAX_AIO_QUEUE_PER_PROC
95 #define MAX_AIO_QUEUE_PER_PROC 256
99 #define MAX_AIO_QUEUE 1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
103 #define MAX_BUF_AIO 16
106 FEATURE(aio, "Asynchronous I/O");
107 SYSCTL_DECL(_p1003_1b);
109 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
110 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
112 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
113 "Async IO management");
115 static int enable_aio_unsafe = 0;
116 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
117 "Permit asynchronous IO on all file types, not just known-safe types");
119 static unsigned int unsafe_warningcnt = 1;
120 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
121 &unsafe_warningcnt, 0,
122 "Warnings that will be triggered upon failed IO requests on unsafe files");
124 static int max_aio_procs = MAX_AIO_PROCS;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
126 "Maximum number of kernel processes to use for handling async IO ");
128 static int num_aio_procs = 0;
129 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
130 "Number of presently active kernel processes for async IO");
133 * The code will adjust the actual number of AIO processes towards this
134 * number when it gets a chance.
136 static int target_aio_procs = TARGET_AIO_PROCS;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
139 "Preferred number of ready kernel processes for async IO");
141 static int max_queue_count = MAX_AIO_QUEUE;
142 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
143 "Maximum number of aio requests to queue, globally");
145 static int num_queue_count = 0;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
147 "Number of queued aio requests");
149 static int num_buf_aio = 0;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
151 "Number of aio requests presently handled by the buf subsystem");
153 /* Number of async I/O processes in the process of being started */
154 /* XXX This should be local to aio_aqueue() */
155 static int num_aio_resv_start = 0;
157 static int aiod_lifetime;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
159 "Maximum lifetime for idle aiod");
161 static int max_aio_per_proc = MAX_AIO_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
164 "Maximum active aio requests per process (stored in the process)");
166 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
168 &max_aio_queue_per_proc, 0,
169 "Maximum queued aio requests per process (stored in the process)");
171 static int max_buf_aio = MAX_BUF_AIO;
172 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
173 "Maximum buf aio requests per process (stored in the process)");
176 * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
177 * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
178 * vfs.aio.aio_listio_max.
180 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
181 CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
182 0, "Maximum aio requests for a single lio_listio call");
184 #ifdef COMPAT_FREEBSD6
185 typedef struct oaiocb {
186 int aio_fildes; /* File descriptor */
187 off_t aio_offset; /* File offset for I/O */
188 volatile void *aio_buf; /* I/O buffer in process space */
189 size_t aio_nbytes; /* Number of bytes for I/O */
190 struct osigevent aio_sigevent; /* Signal to deliver */
191 int aio_lio_opcode; /* LIO opcode */
192 int aio_reqprio; /* Request priority -- ignored */
193 struct __aiocb_private _aiocb_private;
198 * Below is a key of locks used to protect each member of struct kaiocb
199 * aioliojob and kaioinfo and any backends.
201 * * - need not protected
202 * a - locked by kaioinfo lock
203 * b - locked by backend lock, the backend lock can be null in some cases,
204 * for example, BIO belongs to this type, in this case, proc lock is
206 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210 * If the routine that services an AIO request blocks while running in an
211 * AIO kernel process it can starve other I/O requests. BIO requests
212 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
213 * processes at all. Socket I/O requests use a separate pool of
214 * kprocs and also force non-blocking I/O. Other file I/O requests
215 * use the generic fo_read/fo_write operations which can block. The
216 * fsync and mlock operations can also block while executing. Ideally
217 * none of these requests would block while executing.
219 * Note that the service routines cannot toggle O_NONBLOCK in the file
220 * structure directly while handling a request due to races with
225 #define KAIOCB_QUEUEING 0x01
226 #define KAIOCB_CANCELLED 0x02
227 #define KAIOCB_CANCELLING 0x04
228 #define KAIOCB_CHECKSYNC 0x08
229 #define KAIOCB_CLEARED 0x10
230 #define KAIOCB_FINISHED 0x20
235 #define AIOP_FREE 0x1 /* proc on free queue */
238 int aioprocflags; /* (c) AIO proc flags */
239 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
240 struct proc *aioproc; /* (*) the AIO proc */
244 * data-structure for lio signal management
247 int lioj_flags; /* (a) listio flags */
248 int lioj_count; /* (a) listio flags */
249 int lioj_finished_count; /* (a) listio flags */
250 struct sigevent lioj_signal; /* (a) signal on all I/O done */
251 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
252 struct knlist klist; /* (a) list of knotes */
253 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
256 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
257 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
258 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
261 * per process aio data structure
264 struct mtx kaio_mtx; /* the lock to protect this struct */
265 int kaio_flags; /* (a) per process kaio flags */
266 int kaio_maxactive_count; /* (*) maximum number of AIOs */
267 int kaio_active_count; /* (c) number of currently used AIOs */
268 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
269 int kaio_count; /* (a) size of AIO queue */
270 int kaio_ballowed_count; /* (*) maximum number of buffers */
271 int kaio_buffer_count; /* (a) number of physio 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 (*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_physwakeup(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_qphysio(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_maxactive_count = max_aio_per_proc;
444 ki->kaio_active_count = 0;
445 ki->kaio_qallowed_count = max_aio_queue_per_proc;
447 ki->kaio_ballowed_count = max_buf_aio;
448 ki->kaio_buffer_count = 0;
449 TAILQ_INIT(&ki->kaio_all);
450 TAILQ_INIT(&ki->kaio_done);
451 TAILQ_INIT(&ki->kaio_jobqueue);
452 TAILQ_INIT(&ki->kaio_liojoblist);
453 TAILQ_INIT(&ki->kaio_syncqueue);
454 TAILQ_INIT(&ki->kaio_syncready);
455 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
456 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
458 if (p->p_aioinfo == NULL) {
463 mtx_destroy(&ki->kaio_mtx);
464 uma_zfree(kaio_zone, ki);
467 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
472 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
477 error = sigev_findtd(p, sigev, &td);
481 ksiginfo_set_sigev(ksi, sigev);
482 ksi->ksi_code = SI_ASYNCIO;
483 ksi->ksi_flags |= KSI_EXT | KSI_INS;
484 tdsendsignal(p, td, ksi->ksi_signo, ksi);
491 * Free a job entry. Wait for completion if it is currently active, but don't
492 * delay forever. If we delay, we return a flag that says that we have to
493 * restart the queue scan.
496 aio_free_entry(struct kaiocb *job)
499 struct aioliojob *lj;
507 AIO_LOCK_ASSERT(ki, MA_OWNED);
508 MPASS(job->jobflags & KAIOCB_FINISHED);
510 atomic_subtract_int(&num_queue_count, 1);
513 MPASS(ki->kaio_count >= 0);
515 TAILQ_REMOVE(&ki->kaio_done, job, plist);
516 TAILQ_REMOVE(&ki->kaio_all, job, allist);
521 lj->lioj_finished_count--;
523 if (lj->lioj_count == 0) {
524 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
525 /* lio is going away, we need to destroy any knotes */
526 knlist_delete(&lj->klist, curthread, 1);
528 sigqueue_take(&lj->lioj_ksi);
530 uma_zfree(aiolio_zone, lj);
534 /* job is going away, we need to destroy any knotes */
535 knlist_delete(&job->klist, curthread, 1);
537 sigqueue_take(&job->ksi);
543 * The thread argument here is used to find the owning process
544 * and is also passed to fo_close() which may pass it to various
545 * places such as devsw close() routines. Because of that, we
546 * need a thread pointer from the process owning the job that is
547 * persistent and won't disappear out from under us or move to
550 * Currently, all the callers of this function call it to remove
551 * a kaiocb from the current process' job list either via a
552 * syscall or due to the current process calling exit() or
553 * execve(). Thus, we know that p == curproc. We also know that
554 * curthread can't exit since we are curthread.
556 * Therefore, we use curthread as the thread to pass to
557 * knlist_delete(). This does mean that it is possible for the
558 * thread pointer at close time to differ from the thread pointer
559 * at open time, but this is already true of file descriptors in
560 * a multithreaded process.
563 fdrop(job->fd_file, curthread);
565 uma_zfree(aiocb_zone, job);
572 aio_proc_rundown_exec(void *arg, struct proc *p,
573 struct image_params *imgp __unused)
575 aio_proc_rundown(arg, p);
579 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
581 aio_cancel_fn_t *func;
584 AIO_LOCK_ASSERT(ki, MA_OWNED);
585 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
587 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
588 job->jobflags |= KAIOCB_CANCELLED;
590 func = job->cancel_fn;
593 * If there is no cancel routine, just leave the job marked as
594 * cancelled. The job should be in active use by a caller who
595 * should complete it normally or when it fails to install a
602 * Set the CANCELLING flag so that aio_complete() will defer
603 * completions of this job. This prevents the job from being
604 * freed out from under the cancel callback. After the
605 * callback any deferred completion (whether from the callback
606 * or any other source) will be completed.
608 job->jobflags |= KAIOCB_CANCELLING;
612 job->jobflags &= ~KAIOCB_CANCELLING;
613 if (job->jobflags & KAIOCB_FINISHED) {
614 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
615 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
616 aio_bio_done_notify(p, job);
619 * The cancel callback might have scheduled an
620 * operation to cancel this request, but it is
621 * only counted as cancelled if the request is
622 * cancelled when the callback returns.
630 * Rundown the jobs for a given process.
633 aio_proc_rundown(void *arg, struct proc *p)
636 struct aioliojob *lj;
637 struct kaiocb *job, *jobn;
639 KASSERT(curthread->td_proc == p,
640 ("%s: called on non-curproc", __func__));
646 ki->kaio_flags |= KAIO_RUNDOWN;
651 * Try to cancel all pending requests. This code simulates
652 * aio_cancel on all pending I/O requests.
654 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
655 aio_cancel_job(p, ki, job);
658 /* Wait for all running I/O to be finished */
659 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
660 ki->kaio_flags |= KAIO_WAKEUP;
661 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
665 /* Free all completed I/O requests. */
666 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
669 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
670 if (lj->lioj_count == 0) {
671 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
672 knlist_delete(&lj->klist, curthread, 1);
674 sigqueue_take(&lj->lioj_ksi);
676 uma_zfree(aiolio_zone, lj);
678 panic("LIO job not cleaned up: C:%d, FC:%d\n",
679 lj->lioj_count, lj->lioj_finished_count);
683 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
684 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
685 mtx_destroy(&ki->kaio_mtx);
686 uma_zfree(kaio_zone, ki);
691 * Select a job to run (called by an AIO daemon).
693 static struct kaiocb *
694 aio_selectjob(struct aioproc *aiop)
700 mtx_assert(&aio_job_mtx, MA_OWNED);
702 TAILQ_FOREACH(job, &aio_jobs, list) {
703 userp = job->userproc;
704 ki = userp->p_aioinfo;
706 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
707 TAILQ_REMOVE(&aio_jobs, job, list);
708 if (!aio_clear_cancel_function(job))
711 /* Account for currently active jobs. */
712 ki->kaio_active_count++;
720 * Move all data to a permanent storage device. This code
721 * simulates the fsync syscall.
724 aio_fsync_vnode(struct thread *td, struct vnode *vp)
729 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
731 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
732 if (vp->v_object != NULL) {
733 VM_OBJECT_WLOCK(vp->v_object);
734 vm_object_page_clean(vp->v_object, 0, 0, 0);
735 VM_OBJECT_WUNLOCK(vp->v_object);
737 error = VOP_FSYNC(vp, MNT_WAIT, td);
740 vn_finished_write(mp);
746 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
747 * does the I/O request for the non-physio version of the operations. The
748 * normal vn operations are used, and this code should work in all instances
749 * for every type of file, including pipes, sockets, fifos, and regular files.
751 * XXX I don't think it works well for socket, pipe, and fifo.
754 aio_process_rw(struct kaiocb *job)
756 struct ucred *td_savedcred;
763 long msgsnd_st, msgsnd_end;
764 long msgrcv_st, msgrcv_end;
765 long oublock_st, oublock_end;
766 long inblock_st, inblock_end;
769 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
770 job->uaiocb.aio_lio_opcode == LIO_WRITE,
771 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
773 aio_switch_vmspace(job);
775 td_savedcred = td->td_ucred;
776 td->td_ucred = job->cred;
780 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
781 aiov.iov_len = cb->aio_nbytes;
783 auio.uio_iov = &aiov;
785 auio.uio_offset = cb->aio_offset;
786 auio.uio_resid = cb->aio_nbytes;
787 cnt = cb->aio_nbytes;
788 auio.uio_segflg = UIO_USERSPACE;
791 msgrcv_st = td->td_ru.ru_msgrcv;
792 msgsnd_st = td->td_ru.ru_msgsnd;
793 inblock_st = td->td_ru.ru_inblock;
794 oublock_st = td->td_ru.ru_oublock;
797 * aio_aqueue() acquires a reference to the file that is
798 * released in aio_free_entry().
800 if (cb->aio_lio_opcode == LIO_READ) {
801 auio.uio_rw = UIO_READ;
802 if (auio.uio_resid == 0)
805 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
807 if (fp->f_type == DTYPE_VNODE)
809 auio.uio_rw = UIO_WRITE;
810 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
812 msgrcv_end = td->td_ru.ru_msgrcv;
813 msgsnd_end = td->td_ru.ru_msgsnd;
814 inblock_end = td->td_ru.ru_inblock;
815 oublock_end = td->td_ru.ru_oublock;
817 job->msgrcv = msgrcv_end - msgrcv_st;
818 job->msgsnd = msgsnd_end - msgsnd_st;
819 job->inblock = inblock_end - inblock_st;
820 job->outblock = oublock_end - oublock_st;
822 if ((error) && (auio.uio_resid != cnt)) {
823 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
825 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
826 PROC_LOCK(job->userproc);
827 kern_psignal(job->userproc, SIGPIPE);
828 PROC_UNLOCK(job->userproc);
832 cnt -= auio.uio_resid;
833 td->td_ucred = td_savedcred;
835 aio_complete(job, -1, error);
837 aio_complete(job, cnt, 0);
841 aio_process_sync(struct kaiocb *job)
843 struct thread *td = curthread;
844 struct ucred *td_savedcred = td->td_ucred;
845 struct file *fp = job->fd_file;
848 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
849 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
851 td->td_ucred = job->cred;
852 if (fp->f_vnode != NULL)
853 error = aio_fsync_vnode(td, fp->f_vnode);
854 td->td_ucred = td_savedcred;
856 aio_complete(job, -1, error);
858 aio_complete(job, 0, 0);
862 aio_process_mlock(struct kaiocb *job)
864 struct aiocb *cb = &job->uaiocb;
867 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
868 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
870 aio_switch_vmspace(job);
871 error = kern_mlock(job->userproc, job->cred,
872 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
873 aio_complete(job, error != 0 ? -1 : 0, error);
877 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
879 struct aioliojob *lj;
881 struct kaiocb *sjob, *sjobn;
885 ki = userp->p_aioinfo;
886 AIO_LOCK_ASSERT(ki, MA_OWNED);
890 lj->lioj_finished_count++;
891 if (lj->lioj_count == lj->lioj_finished_count)
894 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
895 MPASS(job->jobflags & KAIOCB_FINISHED);
897 if (ki->kaio_flags & KAIO_RUNDOWN)
898 goto notification_done;
900 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
901 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
902 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
904 KNOTE_LOCKED(&job->klist, 1);
907 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
908 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
909 KNOTE_LOCKED(&lj->klist, 1);
911 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
913 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
914 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
915 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
916 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
921 if (job->jobflags & KAIOCB_CHECKSYNC) {
922 schedule_fsync = false;
923 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
924 if (job->fd_file != sjob->fd_file ||
925 job->seqno >= sjob->seqno)
927 if (--sjob->pending > 0)
929 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
930 if (!aio_clear_cancel_function_locked(sjob))
932 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
933 schedule_fsync = true;
936 taskqueue_enqueue(taskqueue_aiod_kick,
937 &ki->kaio_sync_task);
939 if (ki->kaio_flags & KAIO_WAKEUP) {
940 ki->kaio_flags &= ~KAIO_WAKEUP;
941 wakeup(&userp->p_aioinfo);
946 aio_schedule_fsync(void *context, int pending)
953 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
954 job = TAILQ_FIRST(&ki->kaio_syncready);
955 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
957 aio_schedule(job, aio_process_sync);
964 aio_cancel_cleared(struct kaiocb *job)
968 * The caller should hold the same queue lock held when
969 * aio_clear_cancel_function() was called and set this flag
970 * ensuring this check sees an up-to-date value. However,
971 * there is no way to assert that.
973 return ((job->jobflags & KAIOCB_CLEARED) != 0);
977 aio_clear_cancel_function_locked(struct kaiocb *job)
980 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
981 MPASS(job->cancel_fn != NULL);
982 if (job->jobflags & KAIOCB_CANCELLING) {
983 job->jobflags |= KAIOCB_CLEARED;
986 job->cancel_fn = NULL;
991 aio_clear_cancel_function(struct kaiocb *job)
996 ki = job->userproc->p_aioinfo;
998 ret = aio_clear_cancel_function_locked(job);
1004 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1007 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1008 if (job->jobflags & KAIOCB_CANCELLED)
1010 job->cancel_fn = func;
1015 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1017 struct kaioinfo *ki;
1020 ki = job->userproc->p_aioinfo;
1022 ret = aio_set_cancel_function_locked(job, func);
1028 aio_complete(struct kaiocb *job, long status, int error)
1030 struct kaioinfo *ki;
1033 job->uaiocb._aiocb_private.error = error;
1034 job->uaiocb._aiocb_private.status = status;
1036 userp = job->userproc;
1037 ki = userp->p_aioinfo;
1040 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1041 ("duplicate aio_complete"));
1042 job->jobflags |= KAIOCB_FINISHED;
1043 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1044 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1045 aio_bio_done_notify(userp, job);
1051 aio_cancel(struct kaiocb *job)
1054 aio_complete(job, -1, ECANCELED);
1058 aio_switch_vmspace(struct kaiocb *job)
1061 vmspace_switch_aio(job->userproc->p_vmspace);
1065 * The AIO daemon, most of the actual work is done in aio_process_*,
1066 * but the setup (and address space mgmt) is done in this routine.
1069 aio_daemon(void *_id)
1072 struct aioproc *aiop;
1073 struct kaioinfo *ki;
1075 struct vmspace *myvm;
1076 struct thread *td = curthread;
1077 int id = (intptr_t)_id;
1080 * Grab an extra reference on the daemon's vmspace so that it
1081 * doesn't get freed by jobs that switch to a different
1085 myvm = vmspace_acquire_ref(p);
1087 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1090 * Allocate and ready the aio control info. There is one aiop structure
1093 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1095 aiop->aioprocflags = 0;
1098 * Wakeup parent process. (Parent sleeps to keep from blasting away
1099 * and creating too many daemons.)
1101 sema_post(&aio_newproc_sem);
1103 mtx_lock(&aio_job_mtx);
1106 * Take daemon off of free queue
1108 if (aiop->aioprocflags & AIOP_FREE) {
1109 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1110 aiop->aioprocflags &= ~AIOP_FREE;
1116 while ((job = aio_selectjob(aiop)) != NULL) {
1117 mtx_unlock(&aio_job_mtx);
1119 ki = job->userproc->p_aioinfo;
1120 job->handle_fn(job);
1122 mtx_lock(&aio_job_mtx);
1123 /* Decrement the active job count. */
1124 ki->kaio_active_count--;
1128 * Disconnect from user address space.
1130 if (p->p_vmspace != myvm) {
1131 mtx_unlock(&aio_job_mtx);
1132 vmspace_switch_aio(myvm);
1133 mtx_lock(&aio_job_mtx);
1135 * We have to restart to avoid race, we only sleep if
1136 * no job can be selected.
1141 mtx_assert(&aio_job_mtx, MA_OWNED);
1143 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1144 aiop->aioprocflags |= AIOP_FREE;
1147 * If daemon is inactive for a long time, allow it to exit,
1148 * thereby freeing resources.
1150 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1151 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1152 (aiop->aioprocflags & AIOP_FREE) &&
1153 num_aio_procs > target_aio_procs)
1156 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1158 mtx_unlock(&aio_job_mtx);
1159 uma_zfree(aiop_zone, aiop);
1160 free_unr(aiod_unr, id);
1163 KASSERT(p->p_vmspace == myvm,
1164 ("AIOD: bad vmspace for exiting daemon"));
1165 KASSERT(myvm->vm_refcnt > 1,
1166 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1171 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1172 * AIO daemon modifies its environment itself.
1175 aio_newproc(int *start)
1181 id = alloc_unr(aiod_unr);
1182 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1183 RFNOWAIT, 0, "aiod%d", id);
1186 * Wait until daemon is started.
1188 sema_wait(&aio_newproc_sem);
1189 mtx_lock(&aio_job_mtx);
1193 mtx_unlock(&aio_job_mtx);
1195 free_unr(aiod_unr, id);
1201 * Try the high-performance, low-overhead physio method for eligible
1202 * VCHR devices. This method doesn't use an aio helper thread, and
1203 * thus has very low overhead.
1205 * Assumes that the caller, aio_aqueue(), has incremented the file
1206 * structure's reference count, preventing its deallocation for the
1207 * duration of this call.
1210 aio_qphysio(struct proc *p, struct kaiocb *job)
1219 struct kaioinfo *ki;
1220 int error, ref, poff;
1226 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1230 if (vp->v_type != VCHR)
1232 if (vp->v_bufobj.bo_bsize == 0)
1234 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1238 csw = devvn_refthread(vp, &dev, &ref);
1242 if ((csw->d_flags & D_DISK) == 0) {
1246 if (cb->aio_nbytes > dev->si_iosize_max) {
1252 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1253 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1254 if (cb->aio_nbytes > MAXPHYS) {
1261 if (cb->aio_nbytes > MAXPHYS - poff) {
1265 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1270 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1273 ki->kaio_buffer_count++;
1276 job->bp = bp = g_alloc_bio();
1278 bp->bio_length = cb->aio_nbytes;
1279 bp->bio_bcount = cb->aio_nbytes;
1280 bp->bio_done = aio_physwakeup;
1281 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1282 bp->bio_offset = cb->aio_offset;
1283 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1285 bp->bio_caller1 = (void *)job;
1287 prot = VM_PROT_READ;
1288 if (cb->aio_lio_opcode == LIO_READ)
1289 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1290 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1291 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1292 nitems(job->pages));
1293 if (job->npages < 0) {
1298 pmap_qenter((vm_offset_t)pbuf->b_data,
1299 job->pages, job->npages);
1300 bp->bio_data = pbuf->b_data + poff;
1301 atomic_add_int(&num_buf_aio, 1);
1303 bp->bio_ma = job->pages;
1304 bp->bio_ma_n = job->npages;
1305 bp->bio_ma_offset = poff;
1306 bp->bio_data = unmapped_buf;
1307 bp->bio_flags |= BIO_UNMAPPED;
1310 /* Perform transfer. */
1311 csw->d_strategy(bp);
1312 dev_relthread(dev, ref);
1318 ki->kaio_buffer_count--;
1320 relpbuf(pbuf, NULL);
1326 dev_relthread(dev, ref);
1330 #ifdef COMPAT_FREEBSD6
1332 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1336 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1337 * supported by AIO with the old sigevent structure.
1339 nsig->sigev_notify = osig->sigev_notify;
1340 switch (nsig->sigev_notify) {
1344 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1347 nsig->sigev_notify_kqueue =
1348 osig->__sigev_u.__sigev_notify_kqueue;
1349 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1358 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1360 struct oaiocb *ojob;
1363 bzero(kjob, sizeof(struct aiocb));
1364 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1367 ojob = (struct oaiocb *)kjob;
1368 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1373 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1376 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1380 aiocb_fetch_status(struct aiocb *ujob)
1383 return (fuword(&ujob->_aiocb_private.status));
1387 aiocb_fetch_error(struct aiocb *ujob)
1390 return (fuword(&ujob->_aiocb_private.error));
1394 aiocb_store_status(struct aiocb *ujob, long status)
1397 return (suword(&ujob->_aiocb_private.status, status));
1401 aiocb_store_error(struct aiocb *ujob, long error)
1404 return (suword(&ujob->_aiocb_private.error, error));
1408 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1411 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1415 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1418 return (suword(ujobp, (long)ujob));
1421 static struct aiocb_ops aiocb_ops = {
1422 .copyin = aiocb_copyin,
1423 .fetch_status = aiocb_fetch_status,
1424 .fetch_error = aiocb_fetch_error,
1425 .store_status = aiocb_store_status,
1426 .store_error = aiocb_store_error,
1427 .store_kernelinfo = aiocb_store_kernelinfo,
1428 .store_aiocb = aiocb_store_aiocb,
1431 #ifdef COMPAT_FREEBSD6
1432 static struct aiocb_ops aiocb_ops_osigevent = {
1433 .copyin = aiocb_copyin_old_sigevent,
1434 .fetch_status = aiocb_fetch_status,
1435 .fetch_error = aiocb_fetch_error,
1436 .store_status = aiocb_store_status,
1437 .store_error = aiocb_store_error,
1438 .store_kernelinfo = aiocb_store_kernelinfo,
1439 .store_aiocb = aiocb_store_aiocb,
1444 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1445 * technique is done in this code.
1448 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1449 int type, struct aiocb_ops *ops)
1451 struct proc *p = td->td_proc;
1452 cap_rights_t rights;
1455 struct kaioinfo *ki;
1463 if (p->p_aioinfo == NULL)
1464 aio_init_aioinfo(p);
1468 ops->store_status(ujob, -1);
1469 ops->store_error(ujob, 0);
1470 ops->store_kernelinfo(ujob, -1);
1472 if (num_queue_count >= max_queue_count ||
1473 ki->kaio_count >= ki->kaio_qallowed_count) {
1474 ops->store_error(ujob, EAGAIN);
1478 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1479 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1481 error = ops->copyin(ujob, &job->uaiocb);
1483 ops->store_error(ujob, error);
1484 uma_zfree(aiocb_zone, job);
1488 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1489 uma_zfree(aiocb_zone, job);
1493 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1494 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1495 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1496 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1497 ops->store_error(ujob, EINVAL);
1498 uma_zfree(aiocb_zone, job);
1502 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1503 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1504 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1505 uma_zfree(aiocb_zone, job);
1509 ksiginfo_init(&job->ksi);
1511 /* Save userspace address of the job info. */
1514 /* Get the opcode. */
1515 if (type != LIO_NOP)
1516 job->uaiocb.aio_lio_opcode = type;
1517 opcode = job->uaiocb.aio_lio_opcode;
1520 * Validate the opcode and fetch the file object for the specified
1523 * XXXRW: Moved the opcode validation up here so that we don't
1524 * retrieve a file descriptor without knowing what the capabiltity
1527 fd = job->uaiocb.aio_fildes;
1530 error = fget_write(td, fd,
1531 cap_rights_init(&rights, CAP_PWRITE), &fp);
1534 error = fget_read(td, fd,
1535 cap_rights_init(&rights, CAP_PREAD), &fp);
1538 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1544 error = fget(td, fd, cap_rights_init(&rights), &fp);
1550 uma_zfree(aiocb_zone, job);
1551 ops->store_error(ujob, error);
1555 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1560 if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1561 job->uaiocb.aio_offset < 0 &&
1562 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1569 mtx_lock(&aio_job_mtx);
1571 job->seqno = jobseqno++;
1572 mtx_unlock(&aio_job_mtx);
1573 error = ops->store_kernelinfo(ujob, jid);
1578 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1580 if (opcode == LIO_NOP) {
1582 uma_zfree(aiocb_zone, job);
1586 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1588 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1589 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1593 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1594 kev.ident = (uintptr_t)job->ujob;
1595 kev.filter = EVFILT_AIO;
1596 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1597 kev.data = (intptr_t)job;
1598 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1599 error = kqfd_register(kqfd, &kev, td, 1);
1605 ops->store_error(ujob, EINPROGRESS);
1606 job->uaiocb._aiocb_private.error = EINPROGRESS;
1608 job->cred = crhold(td->td_ucred);
1609 job->jobflags = KAIOCB_QUEUEING;
1612 if (opcode == LIO_MLOCK) {
1613 aio_schedule(job, aio_process_mlock);
1615 } else if (fp->f_ops->fo_aio_queue == NULL)
1616 error = aio_queue_file(fp, job);
1618 error = fo_aio_queue(fp, job);
1623 job->jobflags &= ~KAIOCB_QUEUEING;
1624 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1628 atomic_add_int(&num_queue_count, 1);
1629 if (job->jobflags & KAIOCB_FINISHED) {
1631 * The queue callback completed the request synchronously.
1632 * The bulk of the completion is deferred in that case
1635 aio_bio_done_notify(p, job);
1637 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1642 knlist_delete(&job->klist, curthread, 0);
1645 uma_zfree(aiocb_zone, job);
1646 ops->store_error(ujob, error);
1651 aio_cancel_daemon_job(struct kaiocb *job)
1654 mtx_lock(&aio_job_mtx);
1655 if (!aio_cancel_cleared(job))
1656 TAILQ_REMOVE(&aio_jobs, job, list);
1657 mtx_unlock(&aio_job_mtx);
1662 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1665 mtx_lock(&aio_job_mtx);
1666 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1667 mtx_unlock(&aio_job_mtx);
1671 job->handle_fn = func;
1672 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1673 aio_kick_nowait(job->userproc);
1674 mtx_unlock(&aio_job_mtx);
1678 aio_cancel_sync(struct kaiocb *job)
1680 struct kaioinfo *ki;
1682 ki = job->userproc->p_aioinfo;
1684 if (!aio_cancel_cleared(job))
1685 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1691 aio_queue_file(struct file *fp, struct kaiocb *job)
1693 struct kaioinfo *ki;
1694 struct kaiocb *job2;
1700 ki = job->userproc->p_aioinfo;
1701 opcode = job->uaiocb.aio_lio_opcode;
1702 if (opcode == LIO_SYNC)
1705 if ((error = aio_qphysio(job->userproc, job)) == 0)
1709 * XXX: This means qphysio() failed with EFAULT. The current
1710 * behavior is to retry the operation via fo_read/fo_write.
1711 * Wouldn't it be better to just complete the request with an
1719 if (fp->f_type == DTYPE_VNODE) {
1721 if (vp->v_type == VREG || vp->v_type == VDIR) {
1722 mp = fp->f_vnode->v_mount;
1723 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1727 if (!(safe || enable_aio_unsafe)) {
1728 counted_warning(&unsafe_warningcnt,
1729 "is attempting to use unsafe AIO requests");
1730 return (EOPNOTSUPP);
1733 if (opcode == LIO_SYNC) {
1735 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1736 if (job2->fd_file == job->fd_file &&
1737 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1738 job2->seqno < job->seqno) {
1739 job2->jobflags |= KAIOCB_CHECKSYNC;
1743 if (job->pending != 0) {
1744 if (!aio_set_cancel_function_locked(job,
1750 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1760 aio_schedule(job, aio_process_rw);
1764 aio_schedule(job, aio_process_sync);
1775 aio_kick_nowait(struct proc *userp)
1777 struct kaioinfo *ki = userp->p_aioinfo;
1778 struct aioproc *aiop;
1780 mtx_assert(&aio_job_mtx, MA_OWNED);
1781 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1782 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1783 aiop->aioprocflags &= ~AIOP_FREE;
1784 wakeup(aiop->aioproc);
1785 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1786 ki->kaio_active_count + num_aio_resv_start <
1787 ki->kaio_maxactive_count) {
1788 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1793 aio_kick(struct proc *userp)
1795 struct kaioinfo *ki = userp->p_aioinfo;
1796 struct aioproc *aiop;
1799 mtx_assert(&aio_job_mtx, MA_OWNED);
1801 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1802 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1803 aiop->aioprocflags &= ~AIOP_FREE;
1804 wakeup(aiop->aioproc);
1805 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1806 ki->kaio_active_count + num_aio_resv_start <
1807 ki->kaio_maxactive_count) {
1808 num_aio_resv_start++;
1809 mtx_unlock(&aio_job_mtx);
1810 error = aio_newproc(&num_aio_resv_start);
1811 mtx_lock(&aio_job_mtx);
1813 num_aio_resv_start--;
1823 aio_kick_helper(void *context, int pending)
1825 struct proc *userp = context;
1827 mtx_lock(&aio_job_mtx);
1828 while (--pending >= 0) {
1829 if (aio_kick(userp))
1832 mtx_unlock(&aio_job_mtx);
1836 * Support the aio_return system call, as a side-effect, kernel resources are
1840 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1842 struct proc *p = td->td_proc;
1844 struct kaioinfo *ki;
1851 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1852 if (job->ujob == ujob)
1856 MPASS(job->jobflags & KAIOCB_FINISHED);
1857 status = job->uaiocb._aiocb_private.status;
1858 error = job->uaiocb._aiocb_private.error;
1859 td->td_retval[0] = status;
1860 td->td_ru.ru_oublock += job->outblock;
1861 td->td_ru.ru_inblock += job->inblock;
1862 td->td_ru.ru_msgsnd += job->msgsnd;
1863 td->td_ru.ru_msgrcv += job->msgrcv;
1864 aio_free_entry(job);
1866 ops->store_error(ujob, error);
1867 ops->store_status(ujob, status);
1876 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1879 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1883 * Allow a process to wakeup when any of the I/O requests are completed.
1886 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1887 struct timespec *ts)
1889 struct proc *p = td->td_proc;
1891 struct kaioinfo *ki;
1892 struct kaiocb *firstjob, *job;
1897 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1900 TIMESPEC_TO_TIMEVAL(&atv, ts);
1901 if (itimerfix(&atv))
1903 timo = tvtohz(&atv);
1917 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1918 for (i = 0; i < njoblist; i++) {
1919 if (job->ujob == ujoblist[i]) {
1920 if (firstjob == NULL)
1922 if (job->jobflags & KAIOCB_FINISHED)
1927 /* All tasks were finished. */
1928 if (firstjob == NULL)
1931 ki->kaio_flags |= KAIO_WAKEUP;
1932 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1934 if (error == ERESTART)
1945 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1947 struct timespec ts, *tsp;
1948 struct aiocb **ujoblist;
1951 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1955 /* Get timespec struct. */
1956 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1962 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
1963 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1965 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1966 free(ujoblist, M_AIOS);
1971 * aio_cancel cancels any non-physio aio operations not currently in
1975 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1977 struct proc *p = td->td_proc;
1978 struct kaioinfo *ki;
1979 struct kaiocb *job, *jobn;
1981 cap_rights_t rights;
1984 int notcancelled = 0;
1987 /* Lookup file object. */
1988 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1996 if (fp->f_type == DTYPE_VNODE) {
1998 if (vn_isdisk(vp, &error)) {
2000 td->td_retval[0] = AIO_NOTCANCELED;
2006 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2007 if ((uap->fd == job->uaiocb.aio_fildes) &&
2008 ((uap->aiocbp == NULL) ||
2009 (uap->aiocbp == job->ujob))) {
2010 if (aio_cancel_job(p, ki, job)) {
2015 if (uap->aiocbp != NULL)
2024 if (uap->aiocbp != NULL) {
2026 td->td_retval[0] = AIO_CANCELED;
2032 td->td_retval[0] = AIO_NOTCANCELED;
2037 td->td_retval[0] = AIO_CANCELED;
2041 td->td_retval[0] = AIO_ALLDONE;
2047 * aio_error is implemented in the kernel level for compatibility purposes
2048 * only. For a user mode async implementation, it would be best to do it in
2049 * a userland subroutine.
2052 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2054 struct proc *p = td->td_proc;
2056 struct kaioinfo *ki;
2061 td->td_retval[0] = EINVAL;
2066 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2067 if (job->ujob == ujob) {
2068 if (job->jobflags & KAIOCB_FINISHED)
2070 job->uaiocb._aiocb_private.error;
2072 td->td_retval[0] = EINPROGRESS;
2080 * Hack for failure of aio_aqueue.
2082 status = ops->fetch_status(ujob);
2084 td->td_retval[0] = ops->fetch_error(ujob);
2088 td->td_retval[0] = EINVAL;
2093 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2096 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2099 /* syscall - asynchronous read from a file (REALTIME) */
2100 #ifdef COMPAT_FREEBSD6
2102 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2105 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2106 &aiocb_ops_osigevent));
2111 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2114 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2117 /* syscall - asynchronous write to a file (REALTIME) */
2118 #ifdef COMPAT_FREEBSD6
2120 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2123 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2124 &aiocb_ops_osigevent));
2129 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2132 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2136 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2139 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2143 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2144 struct aiocb **acb_list, int nent, struct sigevent *sig,
2145 struct aiocb_ops *ops)
2147 struct proc *p = td->td_proc;
2149 struct kaioinfo *ki;
2150 struct aioliojob *lj;
2156 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2159 if (nent < 0 || nent > max_aio_queue_per_proc)
2162 if (p->p_aioinfo == NULL)
2163 aio_init_aioinfo(p);
2167 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2170 lj->lioj_finished_count = 0;
2171 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2172 ksiginfo_init(&lj->lioj_ksi);
2177 if (sig && (mode == LIO_NOWAIT)) {
2178 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2179 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2180 /* Assume only new style KEVENT */
2181 kev.filter = EVFILT_LIO;
2182 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2183 kev.ident = (uintptr_t)uacb_list; /* something unique */
2184 kev.data = (intptr_t)lj;
2185 /* pass user defined sigval data */
2186 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2187 error = kqfd_register(
2188 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2190 uma_zfree(aiolio_zone, lj);
2193 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2195 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2196 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2197 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2198 uma_zfree(aiolio_zone, lj);
2201 lj->lioj_flags |= LIOJ_SIGNAL;
2203 uma_zfree(aiolio_zone, lj);
2209 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2211 * Add extra aiocb count to avoid the lio to be freed
2212 * by other threads doing aio_waitcomplete or aio_return,
2213 * and prevent event from being sent until we have queued
2220 * Get pointers to the list of I/O requests.
2223 for (i = 0; i < nent; i++) {
2226 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2234 if (mode == LIO_WAIT) {
2235 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2236 ki->kaio_flags |= KAIO_WAKEUP;
2237 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2238 PRIBIO | PCATCH, "aiospn", 0);
2239 if (error == ERESTART)
2245 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2246 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2247 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2248 KNOTE_LOCKED(&lj->klist, 1);
2250 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2252 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2253 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2254 aio_sendsig(p, &lj->lioj_signal,
2256 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2261 if (lj->lioj_count == 0) {
2262 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2263 knlist_delete(&lj->klist, curthread, 1);
2265 sigqueue_take(&lj->lioj_ksi);
2268 uma_zfree(aiolio_zone, lj);
2277 /* syscall - list directed I/O (REALTIME) */
2278 #ifdef COMPAT_FREEBSD6
2280 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2282 struct aiocb **acb_list;
2283 struct sigevent *sigp, sig;
2284 struct osigevent osig;
2287 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2291 if (nent < 0 || nent > max_aio_queue_per_proc)
2294 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2295 error = copyin(uap->sig, &osig, sizeof(osig));
2298 error = convert_old_sigevent(&osig, &sig);
2305 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2306 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2308 error = kern_lio_listio(td, uap->mode,
2309 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2310 &aiocb_ops_osigevent);
2311 free(acb_list, M_LIO);
2316 /* syscall - list directed I/O (REALTIME) */
2318 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2320 struct aiocb **acb_list;
2321 struct sigevent *sigp, sig;
2324 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2328 if (nent < 0 || nent > max_aio_queue_per_proc)
2331 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2332 error = copyin(uap->sig, &sig, sizeof(sig));
2339 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2340 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2342 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2343 nent, sigp, &aiocb_ops);
2344 free(acb_list, M_LIO);
2349 aio_physwakeup(struct bio *bp)
2351 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2353 struct kaioinfo *ki;
2357 /* Release mapping into kernel space. */
2358 userp = job->userproc;
2359 ki = userp->p_aioinfo;
2361 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2362 relpbuf(job->pbuf, NULL);
2364 atomic_subtract_int(&num_buf_aio, 1);
2366 ki->kaio_buffer_count--;
2369 vm_page_unhold_pages(job->pages, job->npages);
2373 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2375 if (bp->bio_flags & BIO_ERROR)
2376 error = bp->bio_error;
2377 nblks = btodb(nbytes);
2378 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2379 job->outblock += nblks;
2381 job->inblock += nblks;
2384 aio_complete(job, -1, error);
2386 aio_complete(job, nbytes, 0);
2391 /* syscall - wait for the next completion of an aio request */
2393 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2394 struct timespec *ts, struct aiocb_ops *ops)
2396 struct proc *p = td->td_proc;
2398 struct kaioinfo *ki;
2404 ops->store_aiocb(ujobp, NULL);
2408 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2411 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2414 TIMESPEC_TO_TIMEVAL(&atv, ts);
2415 if (itimerfix(&atv))
2417 timo = tvtohz(&atv);
2420 if (p->p_aioinfo == NULL)
2421 aio_init_aioinfo(p);
2427 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2429 error = EWOULDBLOCK;
2432 ki->kaio_flags |= KAIO_WAKEUP;
2433 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2435 if (timo && error == ERESTART)
2442 MPASS(job->jobflags & KAIOCB_FINISHED);
2444 status = job->uaiocb._aiocb_private.status;
2445 error = job->uaiocb._aiocb_private.error;
2446 td->td_retval[0] = status;
2447 td->td_ru.ru_oublock += job->outblock;
2448 td->td_ru.ru_inblock += job->inblock;
2449 td->td_ru.ru_msgsnd += job->msgsnd;
2450 td->td_ru.ru_msgrcv += job->msgrcv;
2451 aio_free_entry(job);
2453 ops->store_aiocb(ujobp, ujob);
2454 ops->store_error(ujob, error);
2455 ops->store_status(ujob, status);
2463 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2465 struct timespec ts, *tsp;
2469 /* Get timespec struct. */
2470 error = copyin(uap->timeout, &ts, sizeof(ts));
2477 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2481 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2482 struct aiocb_ops *ops)
2485 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2487 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2491 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2494 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2497 /* kqueue attach function */
2499 filt_aioattach(struct knote *kn)
2503 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2506 * The job pointer must be validated before using it, so
2507 * registration is restricted to the kernel; the user cannot
2510 if ((kn->kn_flags & EV_FLAG1) == 0)
2512 kn->kn_ptr.p_aio = job;
2513 kn->kn_flags &= ~EV_FLAG1;
2515 knlist_add(&job->klist, kn, 0);
2520 /* kqueue detach function */
2522 filt_aiodetach(struct knote *kn)
2526 knl = &kn->kn_ptr.p_aio->klist;
2527 knl->kl_lock(knl->kl_lockarg);
2528 if (!knlist_empty(knl))
2529 knlist_remove(knl, kn, 1);
2530 knl->kl_unlock(knl->kl_lockarg);
2533 /* kqueue filter function */
2536 filt_aio(struct knote *kn, long hint)
2538 struct kaiocb *job = kn->kn_ptr.p_aio;
2540 kn->kn_data = job->uaiocb._aiocb_private.error;
2541 if (!(job->jobflags & KAIOCB_FINISHED))
2543 kn->kn_flags |= EV_EOF;
2547 /* kqueue attach function */
2549 filt_lioattach(struct knote *kn)
2551 struct aioliojob *lj;
2553 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2556 * The aioliojob pointer must be validated before using it, so
2557 * registration is restricted to the kernel; the user cannot
2560 if ((kn->kn_flags & EV_FLAG1) == 0)
2562 kn->kn_ptr.p_lio = lj;
2563 kn->kn_flags &= ~EV_FLAG1;
2565 knlist_add(&lj->klist, kn, 0);
2570 /* kqueue detach function */
2572 filt_liodetach(struct knote *kn)
2576 knl = &kn->kn_ptr.p_lio->klist;
2577 knl->kl_lock(knl->kl_lockarg);
2578 if (!knlist_empty(knl))
2579 knlist_remove(knl, kn, 1);
2580 knl->kl_unlock(knl->kl_lockarg);
2583 /* kqueue filter function */
2586 filt_lio(struct knote *kn, long hint)
2588 struct aioliojob * lj = kn->kn_ptr.p_lio;
2590 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2593 #ifdef COMPAT_FREEBSD32
2594 #include <sys/mount.h>
2595 #include <sys/socket.h>
2596 #include <compat/freebsd32/freebsd32.h>
2597 #include <compat/freebsd32/freebsd32_proto.h>
2598 #include <compat/freebsd32/freebsd32_signal.h>
2599 #include <compat/freebsd32/freebsd32_syscall.h>
2600 #include <compat/freebsd32/freebsd32_util.h>
2602 struct __aiocb_private32 {
2605 uint32_t kernelinfo;
2608 #ifdef COMPAT_FREEBSD6
2609 typedef struct oaiocb32 {
2610 int aio_fildes; /* File descriptor */
2611 uint64_t aio_offset __packed; /* File offset for I/O */
2612 uint32_t aio_buf; /* I/O buffer in process space */
2613 uint32_t aio_nbytes; /* Number of bytes for I/O */
2614 struct osigevent32 aio_sigevent; /* Signal to deliver */
2615 int aio_lio_opcode; /* LIO opcode */
2616 int aio_reqprio; /* Request priority -- ignored */
2617 struct __aiocb_private32 _aiocb_private;
2621 typedef struct aiocb32 {
2622 int32_t aio_fildes; /* File descriptor */
2623 uint64_t aio_offset __packed; /* File offset for I/O */
2624 uint32_t aio_buf; /* I/O buffer in process space */
2625 uint32_t aio_nbytes; /* Number of bytes for I/O */
2627 uint32_t __spare2__;
2628 int aio_lio_opcode; /* LIO opcode */
2629 int aio_reqprio; /* Request priority -- ignored */
2630 struct __aiocb_private32 _aiocb_private;
2631 struct sigevent32 aio_sigevent; /* Signal to deliver */
2634 #ifdef COMPAT_FREEBSD6
2636 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2640 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2641 * supported by AIO with the old sigevent structure.
2643 CP(*osig, *nsig, sigev_notify);
2644 switch (nsig->sigev_notify) {
2648 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2651 nsig->sigev_notify_kqueue =
2652 osig->__sigev_u.__sigev_notify_kqueue;
2653 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2662 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2664 struct oaiocb32 job32;
2667 bzero(kjob, sizeof(struct aiocb));
2668 error = copyin(ujob, &job32, sizeof(job32));
2672 CP(job32, *kjob, aio_fildes);
2673 CP(job32, *kjob, aio_offset);
2674 PTRIN_CP(job32, *kjob, aio_buf);
2675 CP(job32, *kjob, aio_nbytes);
2676 CP(job32, *kjob, aio_lio_opcode);
2677 CP(job32, *kjob, aio_reqprio);
2678 CP(job32, *kjob, _aiocb_private.status);
2679 CP(job32, *kjob, _aiocb_private.error);
2680 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2681 return (convert_old_sigevent32(&job32.aio_sigevent,
2682 &kjob->aio_sigevent));
2687 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2689 struct aiocb32 job32;
2692 error = copyin(ujob, &job32, sizeof(job32));
2695 CP(job32, *kjob, aio_fildes);
2696 CP(job32, *kjob, aio_offset);
2697 PTRIN_CP(job32, *kjob, aio_buf);
2698 CP(job32, *kjob, aio_nbytes);
2699 CP(job32, *kjob, aio_lio_opcode);
2700 CP(job32, *kjob, aio_reqprio);
2701 CP(job32, *kjob, _aiocb_private.status);
2702 CP(job32, *kjob, _aiocb_private.error);
2703 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2704 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2708 aiocb32_fetch_status(struct aiocb *ujob)
2710 struct aiocb32 *ujob32;
2712 ujob32 = (struct aiocb32 *)ujob;
2713 return (fuword32(&ujob32->_aiocb_private.status));
2717 aiocb32_fetch_error(struct aiocb *ujob)
2719 struct aiocb32 *ujob32;
2721 ujob32 = (struct aiocb32 *)ujob;
2722 return (fuword32(&ujob32->_aiocb_private.error));
2726 aiocb32_store_status(struct aiocb *ujob, long status)
2728 struct aiocb32 *ujob32;
2730 ujob32 = (struct aiocb32 *)ujob;
2731 return (suword32(&ujob32->_aiocb_private.status, status));
2735 aiocb32_store_error(struct aiocb *ujob, long error)
2737 struct aiocb32 *ujob32;
2739 ujob32 = (struct aiocb32 *)ujob;
2740 return (suword32(&ujob32->_aiocb_private.error, error));
2744 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2746 struct aiocb32 *ujob32;
2748 ujob32 = (struct aiocb32 *)ujob;
2749 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2753 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2756 return (suword32(ujobp, (long)ujob));
2759 static struct aiocb_ops aiocb32_ops = {
2760 .copyin = aiocb32_copyin,
2761 .fetch_status = aiocb32_fetch_status,
2762 .fetch_error = aiocb32_fetch_error,
2763 .store_status = aiocb32_store_status,
2764 .store_error = aiocb32_store_error,
2765 .store_kernelinfo = aiocb32_store_kernelinfo,
2766 .store_aiocb = aiocb32_store_aiocb,
2769 #ifdef COMPAT_FREEBSD6
2770 static struct aiocb_ops aiocb32_ops_osigevent = {
2771 .copyin = aiocb32_copyin_old_sigevent,
2772 .fetch_status = aiocb32_fetch_status,
2773 .fetch_error = aiocb32_fetch_error,
2774 .store_status = aiocb32_store_status,
2775 .store_error = aiocb32_store_error,
2776 .store_kernelinfo = aiocb32_store_kernelinfo,
2777 .store_aiocb = aiocb32_store_aiocb,
2782 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2785 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2789 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2791 struct timespec32 ts32;
2792 struct timespec ts, *tsp;
2793 struct aiocb **ujoblist;
2794 uint32_t *ujoblist32;
2797 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2801 /* Get timespec struct. */
2802 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2804 CP(ts32, ts, tv_sec);
2805 CP(ts32, ts, tv_nsec);
2810 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2811 ujoblist32 = (uint32_t *)ujoblist;
2812 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2813 sizeof(ujoblist32[0]));
2815 for (i = uap->nent - 1; i >= 0; i--)
2816 ujoblist[i] = PTRIN(ujoblist32[i]);
2818 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2820 free(ujoblist, M_AIOS);
2825 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2828 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2831 #ifdef COMPAT_FREEBSD6
2833 freebsd6_freebsd32_aio_read(struct thread *td,
2834 struct freebsd6_freebsd32_aio_read_args *uap)
2837 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2838 &aiocb32_ops_osigevent));
2843 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2846 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2850 #ifdef COMPAT_FREEBSD6
2852 freebsd6_freebsd32_aio_write(struct thread *td,
2853 struct freebsd6_freebsd32_aio_write_args *uap)
2856 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2857 &aiocb32_ops_osigevent));
2862 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2865 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2870 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2873 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2878 freebsd32_aio_waitcomplete(struct thread *td,
2879 struct freebsd32_aio_waitcomplete_args *uap)
2881 struct timespec32 ts32;
2882 struct timespec ts, *tsp;
2886 /* Get timespec struct. */
2887 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2890 CP(ts32, ts, tv_sec);
2891 CP(ts32, ts, tv_nsec);
2896 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2901 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2904 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2908 #ifdef COMPAT_FREEBSD6
2910 freebsd6_freebsd32_lio_listio(struct thread *td,
2911 struct freebsd6_freebsd32_lio_listio_args *uap)
2913 struct aiocb **acb_list;
2914 struct sigevent *sigp, sig;
2915 struct osigevent32 osig;
2916 uint32_t *acb_list32;
2919 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2923 if (nent < 0 || nent > max_aio_queue_per_proc)
2926 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2927 error = copyin(uap->sig, &osig, sizeof(osig));
2930 error = convert_old_sigevent32(&osig, &sig);
2937 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2938 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2940 free(acb_list32, M_LIO);
2943 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2944 for (i = 0; i < nent; i++)
2945 acb_list[i] = PTRIN(acb_list32[i]);
2946 free(acb_list32, M_LIO);
2948 error = kern_lio_listio(td, uap->mode,
2949 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2950 &aiocb32_ops_osigevent);
2951 free(acb_list, M_LIO);
2957 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2959 struct aiocb **acb_list;
2960 struct sigevent *sigp, sig;
2961 struct sigevent32 sig32;
2962 uint32_t *acb_list32;
2965 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2969 if (nent < 0 || nent > max_aio_queue_per_proc)
2972 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2973 error = copyin(uap->sig, &sig32, sizeof(sig32));
2976 error = convert_sigevent32(&sig32, &sig);
2983 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2984 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2986 free(acb_list32, M_LIO);
2989 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2990 for (i = 0; i < nent; i++)
2991 acb_list[i] = PTRIN(acb_list32[i]);
2992 free(acb_list32, M_LIO);
2994 error = kern_lio_listio(td, uap->mode,
2995 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2997 free(acb_list, M_LIO);