2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. John S. Dyson's name may not be used to endorse or promote products
10 * derived from this software without specific prior written permission.
12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
13 * bad that happens because of using this software isn't the responsibility
14 * of the author. This software is distributed AS-IS.
18 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
24 #include "opt_compat.h"
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
40 #include <sys/limits.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/protosw.h>
49 #include <sys/rwlock.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/syscall.h>
54 #include <sys/sysent.h>
55 #include <sys/sysctl.h>
56 #include <sys/syslog.h>
58 #include <sys/taskqueue.h>
59 #include <sys/vnode.h>
61 #include <sys/event.h>
62 #include <sys/mount.h>
63 #include <geom/geom.h>
65 #include <machine/atomic.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_extern.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_object.h>
77 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
78 * overflow. (XXX will be removed soon.)
80 static u_long jobrefid;
83 * Counter for aio_fsync.
85 static uint64_t jobseqno;
87 #ifndef MAX_AIO_PER_PROC
88 #define MAX_AIO_PER_PROC 32
91 #ifndef MAX_AIO_QUEUE_PER_PROC
92 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
96 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
100 #define MAX_BUF_AIO 16
103 FEATURE(aio, "Asynchronous I/O");
105 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
107 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
108 "Async IO management");
110 static int enable_aio_unsafe = 0;
111 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
112 "Permit asynchronous IO on all file types, not just known-safe types");
114 static unsigned int unsafe_warningcnt = 1;
115 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
116 &unsafe_warningcnt, 0,
117 "Warnings that will be triggered upon failed IO requests on unsafe files");
119 static int max_aio_procs = MAX_AIO_PROCS;
120 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
121 "Maximum number of kernel processes to use for handling async IO ");
123 static int num_aio_procs = 0;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
125 "Number of presently active kernel processes for async IO");
128 * The code will adjust the actual number of AIO processes towards this
129 * number when it gets a chance.
131 static int target_aio_procs = TARGET_AIO_PROCS;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
134 "Preferred number of ready kernel processes for async IO");
136 static int max_queue_count = MAX_AIO_QUEUE;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
138 "Maximum number of aio requests to queue, globally");
140 static int num_queue_count = 0;
141 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
142 "Number of queued aio requests");
144 static int num_buf_aio = 0;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
146 "Number of aio requests presently handled by the buf subsystem");
148 /* Number of async I/O processes in the process of being started */
149 /* XXX This should be local to aio_aqueue() */
150 static int num_aio_resv_start = 0;
152 static int aiod_lifetime;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
154 "Maximum lifetime for idle aiod");
156 static int max_aio_per_proc = MAX_AIO_PER_PROC;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
159 "Maximum active aio requests per process (stored in the process)");
161 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
163 &max_aio_queue_per_proc, 0,
164 "Maximum queued aio requests per process (stored in the process)");
166 static int max_buf_aio = MAX_BUF_AIO;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
168 "Maximum buf aio requests per process (stored in the process)");
170 #ifdef COMPAT_FREEBSD6
171 typedef struct oaiocb {
172 int aio_fildes; /* File descriptor */
173 off_t aio_offset; /* File offset for I/O */
174 volatile void *aio_buf; /* I/O buffer in process space */
175 size_t aio_nbytes; /* Number of bytes for I/O */
176 struct osigevent aio_sigevent; /* Signal to deliver */
177 int aio_lio_opcode; /* LIO opcode */
178 int aio_reqprio; /* Request priority -- ignored */
179 struct __aiocb_private _aiocb_private;
184 * Below is a key of locks used to protect each member of struct kaiocb
185 * aioliojob and kaioinfo and any backends.
187 * * - need not protected
188 * a - locked by kaioinfo lock
189 * b - locked by backend lock, the backend lock can be null in some cases,
190 * for example, BIO belongs to this type, in this case, proc lock is
192 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
196 * If the routine that services an AIO request blocks while running in an
197 * AIO kernel process it can starve other I/O requests. BIO requests
198 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
199 * processes at all. Socket I/O requests use a separate pool of
200 * kprocs and also force non-blocking I/O. Other file I/O requests
201 * use the generic fo_read/fo_write operations which can block. The
202 * fsync and mlock operations can also block while executing. Ideally
203 * none of these requests would block while executing.
205 * Note that the service routines cannot toggle O_NONBLOCK in the file
206 * structure directly while handling a request due to races with
211 #define KAIOCB_QUEUEING 0x01
212 #define KAIOCB_CANCELLED 0x02
213 #define KAIOCB_CANCELLING 0x04
214 #define KAIOCB_CHECKSYNC 0x08
215 #define KAIOCB_CLEARED 0x10
216 #define KAIOCB_FINISHED 0x20
221 #define AIOP_FREE 0x1 /* proc on free queue */
224 int aioprocflags; /* (c) AIO proc flags */
225 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
226 struct proc *aioproc; /* (*) the AIO proc */
230 * data-structure for lio signal management
233 int lioj_flags; /* (a) listio flags */
234 int lioj_count; /* (a) listio flags */
235 int lioj_finished_count; /* (a) listio flags */
236 struct sigevent lioj_signal; /* (a) signal on all I/O done */
237 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
238 struct knlist klist; /* (a) list of knotes */
239 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
242 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
243 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
244 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
247 * per process aio data structure
250 struct mtx kaio_mtx; /* the lock to protect this struct */
251 int kaio_flags; /* (a) per process kaio flags */
252 int kaio_maxactive_count; /* (*) maximum number of AIOs */
253 int kaio_active_count; /* (c) number of currently used AIOs */
254 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
255 int kaio_count; /* (a) size of AIO queue */
256 int kaio_ballowed_count; /* (*) maximum number of buffers */
257 int kaio_buffer_count; /* (a) number of physio buffers */
258 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */
259 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */
260 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
261 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */
262 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */
263 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */
264 struct task kaio_task; /* (*) task to kick aio processes */
265 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */
268 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
269 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
270 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
271 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
273 #define KAIO_RUNDOWN 0x1 /* process is being run down */
274 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */
277 * Operations used to interact with userland aio control blocks.
278 * Different ABIs provide their own operations.
281 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
282 long (*fetch_status)(struct aiocb *ujob);
283 long (*fetch_error)(struct aiocb *ujob);
284 int (*store_status)(struct aiocb *ujob, long status);
285 int (*store_error)(struct aiocb *ujob, long error);
286 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
287 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
290 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */
291 static struct sema aio_newproc_sem;
292 static struct mtx aio_job_mtx;
293 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */
294 static struct unrhdr *aiod_unr;
296 void aio_init_aioinfo(struct proc *p);
297 static int aio_onceonly(void);
298 static int aio_free_entry(struct kaiocb *job);
299 static void aio_process_rw(struct kaiocb *job);
300 static void aio_process_sync(struct kaiocb *job);
301 static void aio_process_mlock(struct kaiocb *job);
302 static void aio_schedule_fsync(void *context, int pending);
303 static int aio_newproc(int *);
304 int aio_aqueue(struct thread *td, struct aiocb *ujob,
305 struct aioliojob *lio, int type, struct aiocb_ops *ops);
306 static int aio_queue_file(struct file *fp, struct kaiocb *job);
307 static void aio_physwakeup(struct bio *bp);
308 static void aio_proc_rundown(void *arg, struct proc *p);
309 static void aio_proc_rundown_exec(void *arg, struct proc *p,
310 struct image_params *imgp);
311 static int aio_qphysio(struct proc *p, struct kaiocb *job);
312 static void aio_daemon(void *param);
313 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
314 static int aio_kick(struct proc *userp);
315 static void aio_kick_nowait(struct proc *userp);
316 static void aio_kick_helper(void *context, int pending);
317 static int filt_aioattach(struct knote *kn);
318 static void filt_aiodetach(struct knote *kn);
319 static int filt_aio(struct knote *kn, long hint);
320 static int filt_lioattach(struct knote *kn);
321 static void filt_liodetach(struct knote *kn);
322 static int filt_lio(struct knote *kn, long hint);
326 * kaio Per process async io info
327 * aiop async io process data
328 * aiocb async io jobs
329 * aiol list io job pointer - internal to aio_suspend XXX
330 * aiolio list io jobs
332 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
334 /* kqueue filters for aio */
335 static struct filterops aio_filtops = {
337 .f_attach = filt_aioattach,
338 .f_detach = filt_aiodetach,
341 static struct filterops lio_filtops = {
343 .f_attach = filt_lioattach,
344 .f_detach = filt_liodetach,
348 static eventhandler_tag exit_tag, exec_tag;
350 TASKQUEUE_DEFINE_THREAD(aiod_kick);
353 * Main operations function for use as a kernel module.
356 aio_modload(struct module *module, int cmd, void *arg)
373 static moduledata_t aio_mod = {
379 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
380 MODULE_VERSION(aio, 1);
383 * Startup initialization
389 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
390 EVENTHANDLER_PRI_ANY);
391 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
392 NULL, EVENTHANDLER_PRI_ANY);
393 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
394 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
395 TAILQ_INIT(&aio_freeproc);
396 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
397 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
398 TAILQ_INIT(&aio_jobs);
399 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
400 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
401 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
402 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
403 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
404 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
405 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
406 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
407 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
408 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
409 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
410 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
412 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
413 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
414 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
415 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
421 * Init the per-process aioinfo structure. The aioinfo limits are set
422 * per-process for user limit (resource) management.
425 aio_init_aioinfo(struct proc *p)
429 ki = uma_zalloc(kaio_zone, M_WAITOK);
430 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
432 ki->kaio_maxactive_count = max_aio_per_proc;
433 ki->kaio_active_count = 0;
434 ki->kaio_qallowed_count = max_aio_queue_per_proc;
436 ki->kaio_ballowed_count = max_buf_aio;
437 ki->kaio_buffer_count = 0;
438 TAILQ_INIT(&ki->kaio_all);
439 TAILQ_INIT(&ki->kaio_done);
440 TAILQ_INIT(&ki->kaio_jobqueue);
441 TAILQ_INIT(&ki->kaio_liojoblist);
442 TAILQ_INIT(&ki->kaio_syncqueue);
443 TAILQ_INIT(&ki->kaio_syncready);
444 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
445 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
447 if (p->p_aioinfo == NULL) {
452 mtx_destroy(&ki->kaio_mtx);
453 uma_zfree(kaio_zone, ki);
456 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
461 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
466 error = sigev_findtd(p, sigev, &td);
470 ksiginfo_set_sigev(ksi, sigev);
471 ksi->ksi_code = SI_ASYNCIO;
472 ksi->ksi_flags |= KSI_EXT | KSI_INS;
473 tdsendsignal(p, td, ksi->ksi_signo, ksi);
480 * Free a job entry. Wait for completion if it is currently active, but don't
481 * delay forever. If we delay, we return a flag that says that we have to
482 * restart the queue scan.
485 aio_free_entry(struct kaiocb *job)
488 struct aioliojob *lj;
496 AIO_LOCK_ASSERT(ki, MA_OWNED);
497 MPASS(job->jobflags & KAIOCB_FINISHED);
499 atomic_subtract_int(&num_queue_count, 1);
502 MPASS(ki->kaio_count >= 0);
504 TAILQ_REMOVE(&ki->kaio_done, job, plist);
505 TAILQ_REMOVE(&ki->kaio_all, job, allist);
510 lj->lioj_finished_count--;
512 if (lj->lioj_count == 0) {
513 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
514 /* lio is going away, we need to destroy any knotes */
515 knlist_delete(&lj->klist, curthread, 1);
517 sigqueue_take(&lj->lioj_ksi);
519 uma_zfree(aiolio_zone, lj);
523 /* job is going away, we need to destroy any knotes */
524 knlist_delete(&job->klist, curthread, 1);
526 sigqueue_take(&job->ksi);
532 * The thread argument here is used to find the owning process
533 * and is also passed to fo_close() which may pass it to various
534 * places such as devsw close() routines. Because of that, we
535 * need a thread pointer from the process owning the job that is
536 * persistent and won't disappear out from under us or move to
539 * Currently, all the callers of this function call it to remove
540 * a kaiocb from the current process' job list either via a
541 * syscall or due to the current process calling exit() or
542 * execve(). Thus, we know that p == curproc. We also know that
543 * curthread can't exit since we are curthread.
545 * Therefore, we use curthread as the thread to pass to
546 * knlist_delete(). This does mean that it is possible for the
547 * thread pointer at close time to differ from the thread pointer
548 * at open time, but this is already true of file descriptors in
549 * a multithreaded process.
552 fdrop(job->fd_file, curthread);
554 uma_zfree(aiocb_zone, job);
561 aio_proc_rundown_exec(void *arg, struct proc *p,
562 struct image_params *imgp __unused)
564 aio_proc_rundown(arg, p);
568 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
570 aio_cancel_fn_t *func;
573 AIO_LOCK_ASSERT(ki, MA_OWNED);
574 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
576 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
577 job->jobflags |= KAIOCB_CANCELLED;
579 func = job->cancel_fn;
582 * If there is no cancel routine, just leave the job marked as
583 * cancelled. The job should be in active use by a caller who
584 * should complete it normally or when it fails to install a
591 * Set the CANCELLING flag so that aio_complete() will defer
592 * completions of this job. This prevents the job from being
593 * freed out from under the cancel callback. After the
594 * callback any deferred completion (whether from the callback
595 * or any other source) will be completed.
597 job->jobflags |= KAIOCB_CANCELLING;
601 job->jobflags &= ~KAIOCB_CANCELLING;
602 if (job->jobflags & KAIOCB_FINISHED) {
603 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
604 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
605 aio_bio_done_notify(p, job);
608 * The cancel callback might have scheduled an
609 * operation to cancel this request, but it is
610 * only counted as cancelled if the request is
611 * cancelled when the callback returns.
619 * Rundown the jobs for a given process.
622 aio_proc_rundown(void *arg, struct proc *p)
625 struct aioliojob *lj;
626 struct kaiocb *job, *jobn;
628 KASSERT(curthread->td_proc == p,
629 ("%s: called on non-curproc", __func__));
635 ki->kaio_flags |= KAIO_RUNDOWN;
640 * Try to cancel all pending requests. This code simulates
641 * aio_cancel on all pending I/O requests.
643 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
644 aio_cancel_job(p, ki, job);
647 /* Wait for all running I/O to be finished */
648 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
649 ki->kaio_flags |= KAIO_WAKEUP;
650 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
654 /* Free all completed I/O requests. */
655 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
658 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
659 if (lj->lioj_count == 0) {
660 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
661 knlist_delete(&lj->klist, curthread, 1);
663 sigqueue_take(&lj->lioj_ksi);
665 uma_zfree(aiolio_zone, lj);
667 panic("LIO job not cleaned up: C:%d, FC:%d\n",
668 lj->lioj_count, lj->lioj_finished_count);
672 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
673 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
674 mtx_destroy(&ki->kaio_mtx);
675 uma_zfree(kaio_zone, ki);
680 * Select a job to run (called by an AIO daemon).
682 static struct kaiocb *
683 aio_selectjob(struct aioproc *aiop)
689 mtx_assert(&aio_job_mtx, MA_OWNED);
691 TAILQ_FOREACH(job, &aio_jobs, list) {
692 userp = job->userproc;
693 ki = userp->p_aioinfo;
695 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
696 TAILQ_REMOVE(&aio_jobs, job, list);
697 if (!aio_clear_cancel_function(job))
700 /* Account for currently active jobs. */
701 ki->kaio_active_count++;
709 * Move all data to a permanent storage device. This code
710 * simulates the fsync syscall.
713 aio_fsync_vnode(struct thread *td, struct vnode *vp)
718 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
720 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
721 if (vp->v_object != NULL) {
722 VM_OBJECT_WLOCK(vp->v_object);
723 vm_object_page_clean(vp->v_object, 0, 0, 0);
724 VM_OBJECT_WUNLOCK(vp->v_object);
726 error = VOP_FSYNC(vp, MNT_WAIT, td);
729 vn_finished_write(mp);
735 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
736 * does the I/O request for the non-physio version of the operations. The
737 * normal vn operations are used, and this code should work in all instances
738 * for every type of file, including pipes, sockets, fifos, and regular files.
740 * XXX I don't think it works well for socket, pipe, and fifo.
743 aio_process_rw(struct kaiocb *job)
745 struct ucred *td_savedcred;
752 long msgsnd_st, msgsnd_end;
753 long msgrcv_st, msgrcv_end;
754 long oublock_st, oublock_end;
755 long inblock_st, inblock_end;
758 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
759 job->uaiocb.aio_lio_opcode == LIO_WRITE,
760 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
762 aio_switch_vmspace(job);
764 td_savedcred = td->td_ucred;
765 td->td_ucred = job->cred;
769 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
770 aiov.iov_len = cb->aio_nbytes;
772 auio.uio_iov = &aiov;
774 auio.uio_offset = cb->aio_offset;
775 auio.uio_resid = cb->aio_nbytes;
776 cnt = cb->aio_nbytes;
777 auio.uio_segflg = UIO_USERSPACE;
780 msgrcv_st = td->td_ru.ru_msgrcv;
781 msgsnd_st = td->td_ru.ru_msgsnd;
782 inblock_st = td->td_ru.ru_inblock;
783 oublock_st = td->td_ru.ru_oublock;
786 * aio_aqueue() acquires a reference to the file that is
787 * released in aio_free_entry().
789 if (cb->aio_lio_opcode == LIO_READ) {
790 auio.uio_rw = UIO_READ;
791 if (auio.uio_resid == 0)
794 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
796 if (fp->f_type == DTYPE_VNODE)
798 auio.uio_rw = UIO_WRITE;
799 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
801 msgrcv_end = td->td_ru.ru_msgrcv;
802 msgsnd_end = td->td_ru.ru_msgsnd;
803 inblock_end = td->td_ru.ru_inblock;
804 oublock_end = td->td_ru.ru_oublock;
806 job->msgrcv = msgrcv_end - msgrcv_st;
807 job->msgsnd = msgsnd_end - msgsnd_st;
808 job->inblock = inblock_end - inblock_st;
809 job->outblock = oublock_end - oublock_st;
811 if ((error) && (auio.uio_resid != cnt)) {
812 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
814 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
815 PROC_LOCK(job->userproc);
816 kern_psignal(job->userproc, SIGPIPE);
817 PROC_UNLOCK(job->userproc);
821 cnt -= auio.uio_resid;
822 td->td_ucred = td_savedcred;
824 aio_complete(job, -1, error);
826 aio_complete(job, cnt, 0);
830 aio_process_sync(struct kaiocb *job)
832 struct thread *td = curthread;
833 struct ucred *td_savedcred = td->td_ucred;
834 struct file *fp = job->fd_file;
837 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
838 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
840 td->td_ucred = job->cred;
841 if (fp->f_vnode != NULL)
842 error = aio_fsync_vnode(td, fp->f_vnode);
843 td->td_ucred = td_savedcred;
845 aio_complete(job, -1, error);
847 aio_complete(job, 0, 0);
851 aio_process_mlock(struct kaiocb *job)
853 struct aiocb *cb = &job->uaiocb;
856 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
857 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
859 aio_switch_vmspace(job);
860 error = vm_mlock(job->userproc, job->cred,
861 __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
863 aio_complete(job, -1, error);
865 aio_complete(job, 0, 0);
869 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
871 struct aioliojob *lj;
873 struct kaiocb *sjob, *sjobn;
877 ki = userp->p_aioinfo;
878 AIO_LOCK_ASSERT(ki, MA_OWNED);
882 lj->lioj_finished_count++;
883 if (lj->lioj_count == lj->lioj_finished_count)
886 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
887 MPASS(job->jobflags & KAIOCB_FINISHED);
889 if (ki->kaio_flags & KAIO_RUNDOWN)
890 goto notification_done;
892 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
893 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
894 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
896 KNOTE_LOCKED(&job->klist, 1);
899 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
900 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
901 KNOTE_LOCKED(&lj->klist, 1);
903 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
905 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
906 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
907 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
908 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
913 if (job->jobflags & KAIOCB_CHECKSYNC) {
914 schedule_fsync = false;
915 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
916 if (job->fd_file == sjob->fd_file &&
917 job->seqno < sjob->seqno) {
918 if (--sjob->pending == 0) {
919 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob,
921 if (!aio_clear_cancel_function(sjob))
923 TAILQ_INSERT_TAIL(&ki->kaio_syncready,
925 schedule_fsync = true;
930 taskqueue_enqueue(taskqueue_aiod_kick,
931 &ki->kaio_sync_task);
933 if (ki->kaio_flags & KAIO_WAKEUP) {
934 ki->kaio_flags &= ~KAIO_WAKEUP;
935 wakeup(&userp->p_aioinfo);
940 aio_schedule_fsync(void *context, int pending)
947 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
948 job = TAILQ_FIRST(&ki->kaio_syncready);
949 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
951 aio_schedule(job, aio_process_sync);
958 aio_cancel_cleared(struct kaiocb *job)
963 * The caller should hold the same queue lock held when
964 * aio_clear_cancel_function() was called and set this flag
965 * ensuring this check sees an up-to-date value. However,
966 * there is no way to assert that.
968 ki = job->userproc->p_aioinfo;
969 return ((job->jobflags & KAIOCB_CLEARED) != 0);
973 aio_clear_cancel_function(struct kaiocb *job)
977 ki = job->userproc->p_aioinfo;
979 MPASS(job->cancel_fn != NULL);
980 if (job->jobflags & KAIOCB_CANCELLING) {
981 job->jobflags |= KAIOCB_CLEARED;
985 job->cancel_fn = NULL;
991 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
995 ki = job->userproc->p_aioinfo;
997 if (job->jobflags & KAIOCB_CANCELLED) {
1001 job->cancel_fn = func;
1007 aio_complete(struct kaiocb *job, long status, int error)
1009 struct kaioinfo *ki;
1012 job->uaiocb._aiocb_private.error = error;
1013 job->uaiocb._aiocb_private.status = status;
1015 userp = job->userproc;
1016 ki = userp->p_aioinfo;
1019 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1020 ("duplicate aio_complete"));
1021 job->jobflags |= KAIOCB_FINISHED;
1022 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1023 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1024 aio_bio_done_notify(userp, job);
1030 aio_cancel(struct kaiocb *job)
1033 aio_complete(job, -1, ECANCELED);
1037 aio_switch_vmspace(struct kaiocb *job)
1040 vmspace_switch_aio(job->userproc->p_vmspace);
1044 * The AIO daemon, most of the actual work is done in aio_process_*,
1045 * but the setup (and address space mgmt) is done in this routine.
1048 aio_daemon(void *_id)
1051 struct aioproc *aiop;
1052 struct kaioinfo *ki;
1054 struct vmspace *myvm;
1055 struct thread *td = curthread;
1056 int id = (intptr_t)_id;
1059 * Grab an extra reference on the daemon's vmspace so that it
1060 * doesn't get freed by jobs that switch to a different
1064 myvm = vmspace_acquire_ref(p);
1066 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1069 * Allocate and ready the aio control info. There is one aiop structure
1072 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1074 aiop->aioprocflags = 0;
1077 * Wakeup parent process. (Parent sleeps to keep from blasting away
1078 * and creating too many daemons.)
1080 sema_post(&aio_newproc_sem);
1082 mtx_lock(&aio_job_mtx);
1085 * Take daemon off of free queue
1087 if (aiop->aioprocflags & AIOP_FREE) {
1088 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1089 aiop->aioprocflags &= ~AIOP_FREE;
1095 while ((job = aio_selectjob(aiop)) != NULL) {
1096 mtx_unlock(&aio_job_mtx);
1098 ki = job->userproc->p_aioinfo;
1099 job->handle_fn(job);
1101 mtx_lock(&aio_job_mtx);
1102 /* Decrement the active job count. */
1103 ki->kaio_active_count--;
1107 * Disconnect from user address space.
1109 if (p->p_vmspace != myvm) {
1110 mtx_unlock(&aio_job_mtx);
1111 vmspace_switch_aio(myvm);
1112 mtx_lock(&aio_job_mtx);
1114 * We have to restart to avoid race, we only sleep if
1115 * no job can be selected.
1120 mtx_assert(&aio_job_mtx, MA_OWNED);
1122 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1123 aiop->aioprocflags |= AIOP_FREE;
1126 * If daemon is inactive for a long time, allow it to exit,
1127 * thereby freeing resources.
1129 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1130 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1131 (aiop->aioprocflags & AIOP_FREE) &&
1132 num_aio_procs > target_aio_procs)
1135 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1137 mtx_unlock(&aio_job_mtx);
1138 uma_zfree(aiop_zone, aiop);
1139 free_unr(aiod_unr, id);
1142 KASSERT(p->p_vmspace == myvm,
1143 ("AIOD: bad vmspace for exiting daemon"));
1144 KASSERT(myvm->vm_refcnt > 1,
1145 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1150 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1151 * AIO daemon modifies its environment itself.
1154 aio_newproc(int *start)
1160 id = alloc_unr(aiod_unr);
1161 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1162 RFNOWAIT, 0, "aiod%d", id);
1165 * Wait until daemon is started.
1167 sema_wait(&aio_newproc_sem);
1168 mtx_lock(&aio_job_mtx);
1172 mtx_unlock(&aio_job_mtx);
1174 free_unr(aiod_unr, id);
1180 * Try the high-performance, low-overhead physio method for eligible
1181 * VCHR devices. This method doesn't use an aio helper thread, and
1182 * thus has very low overhead.
1184 * Assumes that the caller, aio_aqueue(), has incremented the file
1185 * structure's reference count, preventing its deallocation for the
1186 * duration of this call.
1189 aio_qphysio(struct proc *p, struct kaiocb *job)
1198 struct kaioinfo *ki;
1199 int error, ref, poff;
1205 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1209 if (vp->v_type != VCHR)
1211 if (vp->v_bufobj.bo_bsize == 0)
1213 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1217 csw = devvn_refthread(vp, &dev, &ref);
1221 if ((csw->d_flags & D_DISK) == 0) {
1225 if (cb->aio_nbytes > dev->si_iosize_max) {
1231 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1232 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1233 if (cb->aio_nbytes > MAXPHYS) {
1240 if (cb->aio_nbytes > MAXPHYS - poff) {
1244 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1249 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1252 ki->kaio_buffer_count++;
1255 job->bp = bp = g_alloc_bio();
1257 bp->bio_length = cb->aio_nbytes;
1258 bp->bio_bcount = cb->aio_nbytes;
1259 bp->bio_done = aio_physwakeup;
1260 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1261 bp->bio_offset = cb->aio_offset;
1262 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1264 bp->bio_caller1 = (void *)job;
1266 prot = VM_PROT_READ;
1267 if (cb->aio_lio_opcode == LIO_READ)
1268 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1269 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1270 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1271 nitems(job->pages));
1272 if (job->npages < 0) {
1277 pmap_qenter((vm_offset_t)pbuf->b_data,
1278 job->pages, job->npages);
1279 bp->bio_data = pbuf->b_data + poff;
1280 atomic_add_int(&num_buf_aio, 1);
1282 bp->bio_ma = job->pages;
1283 bp->bio_ma_n = job->npages;
1284 bp->bio_ma_offset = poff;
1285 bp->bio_data = unmapped_buf;
1286 bp->bio_flags |= BIO_UNMAPPED;
1289 /* Perform transfer. */
1290 csw->d_strategy(bp);
1291 dev_relthread(dev, ref);
1297 ki->kaio_buffer_count--;
1299 relpbuf(pbuf, NULL);
1305 dev_relthread(dev, ref);
1309 #ifdef COMPAT_FREEBSD6
1311 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1315 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1316 * supported by AIO with the old sigevent structure.
1318 nsig->sigev_notify = osig->sigev_notify;
1319 switch (nsig->sigev_notify) {
1323 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1326 nsig->sigev_notify_kqueue =
1327 osig->__sigev_u.__sigev_notify_kqueue;
1328 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1337 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1339 struct oaiocb *ojob;
1342 bzero(kjob, sizeof(struct aiocb));
1343 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1346 ojob = (struct oaiocb *)kjob;
1347 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1352 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1355 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1359 aiocb_fetch_status(struct aiocb *ujob)
1362 return (fuword(&ujob->_aiocb_private.status));
1366 aiocb_fetch_error(struct aiocb *ujob)
1369 return (fuword(&ujob->_aiocb_private.error));
1373 aiocb_store_status(struct aiocb *ujob, long status)
1376 return (suword(&ujob->_aiocb_private.status, status));
1380 aiocb_store_error(struct aiocb *ujob, long error)
1383 return (suword(&ujob->_aiocb_private.error, error));
1387 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1390 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1394 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1397 return (suword(ujobp, (long)ujob));
1400 static struct aiocb_ops aiocb_ops = {
1401 .copyin = aiocb_copyin,
1402 .fetch_status = aiocb_fetch_status,
1403 .fetch_error = aiocb_fetch_error,
1404 .store_status = aiocb_store_status,
1405 .store_error = aiocb_store_error,
1406 .store_kernelinfo = aiocb_store_kernelinfo,
1407 .store_aiocb = aiocb_store_aiocb,
1410 #ifdef COMPAT_FREEBSD6
1411 static struct aiocb_ops aiocb_ops_osigevent = {
1412 .copyin = aiocb_copyin_old_sigevent,
1413 .fetch_status = aiocb_fetch_status,
1414 .fetch_error = aiocb_fetch_error,
1415 .store_status = aiocb_store_status,
1416 .store_error = aiocb_store_error,
1417 .store_kernelinfo = aiocb_store_kernelinfo,
1418 .store_aiocb = aiocb_store_aiocb,
1423 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1424 * technique is done in this code.
1427 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1428 int type, struct aiocb_ops *ops)
1430 struct proc *p = td->td_proc;
1431 cap_rights_t rights;
1434 struct kaioinfo *ki;
1442 if (p->p_aioinfo == NULL)
1443 aio_init_aioinfo(p);
1447 ops->store_status(ujob, -1);
1448 ops->store_error(ujob, 0);
1449 ops->store_kernelinfo(ujob, -1);
1451 if (num_queue_count >= max_queue_count ||
1452 ki->kaio_count >= ki->kaio_qallowed_count) {
1453 ops->store_error(ujob, EAGAIN);
1457 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1458 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1460 error = ops->copyin(ujob, &job->uaiocb);
1462 ops->store_error(ujob, error);
1463 uma_zfree(aiocb_zone, job);
1467 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1468 uma_zfree(aiocb_zone, job);
1472 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1473 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1474 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1475 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1476 ops->store_error(ujob, EINVAL);
1477 uma_zfree(aiocb_zone, job);
1481 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1482 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1483 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1484 uma_zfree(aiocb_zone, job);
1488 ksiginfo_init(&job->ksi);
1490 /* Save userspace address of the job info. */
1493 /* Get the opcode. */
1494 if (type != LIO_NOP)
1495 job->uaiocb.aio_lio_opcode = type;
1496 opcode = job->uaiocb.aio_lio_opcode;
1499 * Validate the opcode and fetch the file object for the specified
1502 * XXXRW: Moved the opcode validation up here so that we don't
1503 * retrieve a file descriptor without knowing what the capabiltity
1506 fd = job->uaiocb.aio_fildes;
1509 error = fget_write(td, fd,
1510 cap_rights_init(&rights, CAP_PWRITE), &fp);
1513 error = fget_read(td, fd,
1514 cap_rights_init(&rights, CAP_PREAD), &fp);
1517 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1523 error = fget(td, fd, cap_rights_init(&rights), &fp);
1529 uma_zfree(aiocb_zone, job);
1530 ops->store_error(ujob, error);
1534 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1539 if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1546 mtx_lock(&aio_job_mtx);
1548 job->seqno = jobseqno++;
1549 mtx_unlock(&aio_job_mtx);
1550 error = ops->store_kernelinfo(ujob, jid);
1555 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1557 if (opcode == LIO_NOP) {
1559 uma_zfree(aiocb_zone, job);
1563 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1565 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1566 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1570 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1571 kev.ident = (uintptr_t)job->ujob;
1572 kev.filter = EVFILT_AIO;
1573 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1574 kev.data = (intptr_t)job;
1575 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1576 error = kqfd_register(kqfd, &kev, td, 1);
1582 ops->store_error(ujob, EINPROGRESS);
1583 job->uaiocb._aiocb_private.error = EINPROGRESS;
1585 job->cred = crhold(td->td_ucred);
1586 job->jobflags = KAIOCB_QUEUEING;
1589 if (opcode == LIO_MLOCK) {
1590 aio_schedule(job, aio_process_mlock);
1592 } else if (fp->f_ops->fo_aio_queue == NULL)
1593 error = aio_queue_file(fp, job);
1595 error = fo_aio_queue(fp, job);
1600 job->jobflags &= ~KAIOCB_QUEUEING;
1601 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1605 atomic_add_int(&num_queue_count, 1);
1606 if (job->jobflags & KAIOCB_FINISHED) {
1608 * The queue callback completed the request synchronously.
1609 * The bulk of the completion is deferred in that case
1612 aio_bio_done_notify(p, job);
1614 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1619 knlist_delete(&job->klist, curthread, 0);
1622 uma_zfree(aiocb_zone, job);
1623 ops->store_error(ujob, error);
1628 aio_cancel_daemon_job(struct kaiocb *job)
1631 mtx_lock(&aio_job_mtx);
1632 if (!aio_cancel_cleared(job))
1633 TAILQ_REMOVE(&aio_jobs, job, list);
1634 mtx_unlock(&aio_job_mtx);
1639 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1642 mtx_lock(&aio_job_mtx);
1643 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1644 mtx_unlock(&aio_job_mtx);
1648 job->handle_fn = func;
1649 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1650 aio_kick_nowait(job->userproc);
1651 mtx_unlock(&aio_job_mtx);
1655 aio_cancel_sync(struct kaiocb *job)
1657 struct kaioinfo *ki;
1659 ki = job->userproc->p_aioinfo;
1660 mtx_lock(&aio_job_mtx);
1661 if (!aio_cancel_cleared(job))
1662 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1663 mtx_unlock(&aio_job_mtx);
1668 aio_queue_file(struct file *fp, struct kaiocb *job)
1670 struct aioliojob *lj;
1671 struct kaioinfo *ki;
1672 struct kaiocb *job2;
1679 ki = job->userproc->p_aioinfo;
1680 opcode = job->uaiocb.aio_lio_opcode;
1681 if (opcode == LIO_SYNC)
1684 if ((error = aio_qphysio(job->userproc, job)) == 0)
1688 * XXX: This means qphysio() failed with EFAULT. The current
1689 * behavior is to retry the operation via fo_read/fo_write.
1690 * Wouldn't it be better to just complete the request with an
1698 if (fp->f_type == DTYPE_VNODE) {
1700 if (vp->v_type == VREG || vp->v_type == VDIR) {
1701 mp = fp->f_vnode->v_mount;
1702 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1706 if (!(safe || enable_aio_unsafe)) {
1707 counted_warning(&unsafe_warningcnt,
1708 "is attempting to use unsafe AIO requests");
1709 return (EOPNOTSUPP);
1712 if (opcode == LIO_SYNC) {
1714 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1715 if (job2->fd_file == job->fd_file &&
1716 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1717 job2->seqno < job->seqno) {
1718 job2->jobflags |= KAIOCB_CHECKSYNC;
1722 if (job->pending != 0) {
1723 if (!aio_set_cancel_function(job, aio_cancel_sync)) {
1728 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1738 aio_schedule(job, aio_process_rw);
1742 aio_schedule(job, aio_process_sync);
1753 aio_kick_nowait(struct proc *userp)
1755 struct kaioinfo *ki = userp->p_aioinfo;
1756 struct aioproc *aiop;
1758 mtx_assert(&aio_job_mtx, MA_OWNED);
1759 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1760 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1761 aiop->aioprocflags &= ~AIOP_FREE;
1762 wakeup(aiop->aioproc);
1763 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1764 ki->kaio_active_count + num_aio_resv_start <
1765 ki->kaio_maxactive_count) {
1766 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1771 aio_kick(struct proc *userp)
1773 struct kaioinfo *ki = userp->p_aioinfo;
1774 struct aioproc *aiop;
1777 mtx_assert(&aio_job_mtx, MA_OWNED);
1779 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1780 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1781 aiop->aioprocflags &= ~AIOP_FREE;
1782 wakeup(aiop->aioproc);
1783 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1784 ki->kaio_active_count + num_aio_resv_start <
1785 ki->kaio_maxactive_count) {
1786 num_aio_resv_start++;
1787 mtx_unlock(&aio_job_mtx);
1788 error = aio_newproc(&num_aio_resv_start);
1789 mtx_lock(&aio_job_mtx);
1791 num_aio_resv_start--;
1801 aio_kick_helper(void *context, int pending)
1803 struct proc *userp = context;
1805 mtx_lock(&aio_job_mtx);
1806 while (--pending >= 0) {
1807 if (aio_kick(userp))
1810 mtx_unlock(&aio_job_mtx);
1814 * Support the aio_return system call, as a side-effect, kernel resources are
1818 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1820 struct proc *p = td->td_proc;
1822 struct kaioinfo *ki;
1829 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1830 if (job->ujob == ujob)
1834 MPASS(job->jobflags & KAIOCB_FINISHED);
1835 status = job->uaiocb._aiocb_private.status;
1836 error = job->uaiocb._aiocb_private.error;
1837 td->td_retval[0] = status;
1838 td->td_ru.ru_oublock += job->outblock;
1839 td->td_ru.ru_inblock += job->inblock;
1840 td->td_ru.ru_msgsnd += job->msgsnd;
1841 td->td_ru.ru_msgrcv += job->msgrcv;
1842 aio_free_entry(job);
1844 ops->store_error(ujob, error);
1845 ops->store_status(ujob, status);
1854 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1857 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1861 * Allow a process to wakeup when any of the I/O requests are completed.
1864 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1865 struct timespec *ts)
1867 struct proc *p = td->td_proc;
1869 struct kaioinfo *ki;
1870 struct kaiocb *firstjob, *job;
1875 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1878 TIMESPEC_TO_TIMEVAL(&atv, ts);
1879 if (itimerfix(&atv))
1881 timo = tvtohz(&atv);
1895 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1896 for (i = 0; i < njoblist; i++) {
1897 if (job->ujob == ujoblist[i]) {
1898 if (firstjob == NULL)
1900 if (job->jobflags & KAIOCB_FINISHED)
1905 /* All tasks were finished. */
1906 if (firstjob == NULL)
1909 ki->kaio_flags |= KAIO_WAKEUP;
1910 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1912 if (error == ERESTART)
1923 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1925 struct timespec ts, *tsp;
1926 struct aiocb **ujoblist;
1929 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1933 /* Get timespec struct. */
1934 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1940 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1941 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1943 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1944 uma_zfree(aiol_zone, ujoblist);
1949 * aio_cancel cancels any non-physio aio operations not currently in
1953 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1955 struct proc *p = td->td_proc;
1956 struct kaioinfo *ki;
1957 struct kaiocb *job, *jobn;
1959 cap_rights_t rights;
1962 int notcancelled = 0;
1965 /* Lookup file object. */
1966 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1974 if (fp->f_type == DTYPE_VNODE) {
1976 if (vn_isdisk(vp, &error)) {
1978 td->td_retval[0] = AIO_NOTCANCELED;
1984 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1985 if ((uap->fd == job->uaiocb.aio_fildes) &&
1986 ((uap->aiocbp == NULL) ||
1987 (uap->aiocbp == job->ujob))) {
1988 if (aio_cancel_job(p, ki, job)) {
1993 if (uap->aiocbp != NULL)
2002 if (uap->aiocbp != NULL) {
2004 td->td_retval[0] = AIO_CANCELED;
2010 td->td_retval[0] = AIO_NOTCANCELED;
2015 td->td_retval[0] = AIO_CANCELED;
2019 td->td_retval[0] = AIO_ALLDONE;
2025 * aio_error is implemented in the kernel level for compatibility purposes
2026 * only. For a user mode async implementation, it would be best to do it in
2027 * a userland subroutine.
2030 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2032 struct proc *p = td->td_proc;
2034 struct kaioinfo *ki;
2039 td->td_retval[0] = EINVAL;
2044 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2045 if (job->ujob == ujob) {
2046 if (job->jobflags & KAIOCB_FINISHED)
2048 job->uaiocb._aiocb_private.error;
2050 td->td_retval[0] = EINPROGRESS;
2058 * Hack for failure of aio_aqueue.
2060 status = ops->fetch_status(ujob);
2062 td->td_retval[0] = ops->fetch_error(ujob);
2066 td->td_retval[0] = EINVAL;
2071 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2074 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2077 /* syscall - asynchronous read from a file (REALTIME) */
2078 #ifdef COMPAT_FREEBSD6
2080 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2083 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2084 &aiocb_ops_osigevent));
2089 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2092 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2095 /* syscall - asynchronous write to a file (REALTIME) */
2096 #ifdef COMPAT_FREEBSD6
2098 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2101 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2102 &aiocb_ops_osigevent));
2107 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2110 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2114 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2117 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2121 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2122 struct aiocb **acb_list, int nent, struct sigevent *sig,
2123 struct aiocb_ops *ops)
2125 struct proc *p = td->td_proc;
2127 struct kaioinfo *ki;
2128 struct aioliojob *lj;
2134 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2137 if (nent < 0 || nent > AIO_LISTIO_MAX)
2140 if (p->p_aioinfo == NULL)
2141 aio_init_aioinfo(p);
2145 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2148 lj->lioj_finished_count = 0;
2149 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2150 ksiginfo_init(&lj->lioj_ksi);
2155 if (sig && (mode == LIO_NOWAIT)) {
2156 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2157 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2158 /* Assume only new style KEVENT */
2159 kev.filter = EVFILT_LIO;
2160 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2161 kev.ident = (uintptr_t)uacb_list; /* something unique */
2162 kev.data = (intptr_t)lj;
2163 /* pass user defined sigval data */
2164 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2165 error = kqfd_register(
2166 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2168 uma_zfree(aiolio_zone, lj);
2171 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2173 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2174 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2175 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2176 uma_zfree(aiolio_zone, lj);
2179 lj->lioj_flags |= LIOJ_SIGNAL;
2181 uma_zfree(aiolio_zone, lj);
2187 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2189 * Add extra aiocb count to avoid the lio to be freed
2190 * by other threads doing aio_waitcomplete or aio_return,
2191 * and prevent event from being sent until we have queued
2198 * Get pointers to the list of I/O requests.
2201 for (i = 0; i < nent; i++) {
2204 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2212 if (mode == LIO_WAIT) {
2213 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2214 ki->kaio_flags |= KAIO_WAKEUP;
2215 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2216 PRIBIO | PCATCH, "aiospn", 0);
2217 if (error == ERESTART)
2223 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2224 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2225 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2226 KNOTE_LOCKED(&lj->klist, 1);
2228 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2230 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2231 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2232 aio_sendsig(p, &lj->lioj_signal,
2234 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2239 if (lj->lioj_count == 0) {
2240 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2241 knlist_delete(&lj->klist, curthread, 1);
2243 sigqueue_take(&lj->lioj_ksi);
2246 uma_zfree(aiolio_zone, lj);
2255 /* syscall - list directed I/O (REALTIME) */
2256 #ifdef COMPAT_FREEBSD6
2258 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2260 struct aiocb **acb_list;
2261 struct sigevent *sigp, sig;
2262 struct osigevent osig;
2265 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2269 if (nent < 0 || nent > AIO_LISTIO_MAX)
2272 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2273 error = copyin(uap->sig, &osig, sizeof(osig));
2276 error = convert_old_sigevent(&osig, &sig);
2283 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2284 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2286 error = kern_lio_listio(td, uap->mode,
2287 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2288 &aiocb_ops_osigevent);
2289 free(acb_list, M_LIO);
2294 /* syscall - list directed I/O (REALTIME) */
2296 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2298 struct aiocb **acb_list;
2299 struct sigevent *sigp, sig;
2302 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2306 if (nent < 0 || nent > AIO_LISTIO_MAX)
2309 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2310 error = copyin(uap->sig, &sig, sizeof(sig));
2317 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2318 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2320 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2321 nent, sigp, &aiocb_ops);
2322 free(acb_list, M_LIO);
2327 aio_physwakeup(struct bio *bp)
2329 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2331 struct kaioinfo *ki;
2335 /* Release mapping into kernel space. */
2336 userp = job->userproc;
2337 ki = userp->p_aioinfo;
2339 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2340 relpbuf(job->pbuf, NULL);
2342 atomic_subtract_int(&num_buf_aio, 1);
2344 ki->kaio_buffer_count--;
2347 vm_page_unhold_pages(job->pages, job->npages);
2351 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2353 if (bp->bio_flags & BIO_ERROR)
2354 error = bp->bio_error;
2355 nblks = btodb(nbytes);
2356 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2357 job->outblock += nblks;
2359 job->inblock += nblks;
2362 aio_complete(job, -1, error);
2364 aio_complete(job, nbytes, 0);
2369 /* syscall - wait for the next completion of an aio request */
2371 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2372 struct timespec *ts, struct aiocb_ops *ops)
2374 struct proc *p = td->td_proc;
2376 struct kaioinfo *ki;
2382 ops->store_aiocb(ujobp, NULL);
2386 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2389 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2392 TIMESPEC_TO_TIMEVAL(&atv, ts);
2393 if (itimerfix(&atv))
2395 timo = tvtohz(&atv);
2398 if (p->p_aioinfo == NULL)
2399 aio_init_aioinfo(p);
2405 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2407 error = EWOULDBLOCK;
2410 ki->kaio_flags |= KAIO_WAKEUP;
2411 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2413 if (timo && error == ERESTART)
2420 MPASS(job->jobflags & KAIOCB_FINISHED);
2422 status = job->uaiocb._aiocb_private.status;
2423 error = job->uaiocb._aiocb_private.error;
2424 td->td_retval[0] = status;
2425 td->td_ru.ru_oublock += job->outblock;
2426 td->td_ru.ru_inblock += job->inblock;
2427 td->td_ru.ru_msgsnd += job->msgsnd;
2428 td->td_ru.ru_msgrcv += job->msgrcv;
2429 aio_free_entry(job);
2431 ops->store_aiocb(ujobp, ujob);
2432 ops->store_error(ujob, error);
2433 ops->store_status(ujob, status);
2441 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2443 struct timespec ts, *tsp;
2447 /* Get timespec struct. */
2448 error = copyin(uap->timeout, &ts, sizeof(ts));
2455 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2459 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2460 struct aiocb_ops *ops)
2462 struct proc *p = td->td_proc;
2463 struct kaioinfo *ki;
2465 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2469 aio_init_aioinfo(p);
2470 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2474 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2477 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2480 /* kqueue attach function */
2482 filt_aioattach(struct knote *kn)
2484 struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2487 * The job pointer must be validated before using it, so
2488 * registration is restricted to the kernel; the user cannot
2491 if ((kn->kn_flags & EV_FLAG1) == 0)
2493 kn->kn_ptr.p_aio = job;
2494 kn->kn_flags &= ~EV_FLAG1;
2496 knlist_add(&job->klist, kn, 0);
2501 /* kqueue detach function */
2503 filt_aiodetach(struct knote *kn)
2507 knl = &kn->kn_ptr.p_aio->klist;
2508 knl->kl_lock(knl->kl_lockarg);
2509 if (!knlist_empty(knl))
2510 knlist_remove(knl, kn, 1);
2511 knl->kl_unlock(knl->kl_lockarg);
2514 /* kqueue filter function */
2517 filt_aio(struct knote *kn, long hint)
2519 struct kaiocb *job = kn->kn_ptr.p_aio;
2521 kn->kn_data = job->uaiocb._aiocb_private.error;
2522 if (!(job->jobflags & KAIOCB_FINISHED))
2524 kn->kn_flags |= EV_EOF;
2528 /* kqueue attach function */
2530 filt_lioattach(struct knote *kn)
2532 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2535 * The aioliojob pointer must be validated before using it, so
2536 * registration is restricted to the kernel; the user cannot
2539 if ((kn->kn_flags & EV_FLAG1) == 0)
2541 kn->kn_ptr.p_lio = lj;
2542 kn->kn_flags &= ~EV_FLAG1;
2544 knlist_add(&lj->klist, kn, 0);
2549 /* kqueue detach function */
2551 filt_liodetach(struct knote *kn)
2555 knl = &kn->kn_ptr.p_lio->klist;
2556 knl->kl_lock(knl->kl_lockarg);
2557 if (!knlist_empty(knl))
2558 knlist_remove(knl, kn, 1);
2559 knl->kl_unlock(knl->kl_lockarg);
2562 /* kqueue filter function */
2565 filt_lio(struct knote *kn, long hint)
2567 struct aioliojob * lj = kn->kn_ptr.p_lio;
2569 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2572 #ifdef COMPAT_FREEBSD32
2573 #include <sys/mount.h>
2574 #include <sys/socket.h>
2575 #include <compat/freebsd32/freebsd32.h>
2576 #include <compat/freebsd32/freebsd32_proto.h>
2577 #include <compat/freebsd32/freebsd32_signal.h>
2578 #include <compat/freebsd32/freebsd32_syscall.h>
2579 #include <compat/freebsd32/freebsd32_util.h>
2581 struct __aiocb_private32 {
2584 uint32_t kernelinfo;
2587 #ifdef COMPAT_FREEBSD6
2588 typedef struct oaiocb32 {
2589 int aio_fildes; /* File descriptor */
2590 uint64_t aio_offset __packed; /* File offset for I/O */
2591 uint32_t aio_buf; /* I/O buffer in process space */
2592 uint32_t aio_nbytes; /* Number of bytes for I/O */
2593 struct osigevent32 aio_sigevent; /* Signal to deliver */
2594 int aio_lio_opcode; /* LIO opcode */
2595 int aio_reqprio; /* Request priority -- ignored */
2596 struct __aiocb_private32 _aiocb_private;
2600 typedef struct aiocb32 {
2601 int32_t 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 */
2606 uint32_t __spare2__;
2607 int aio_lio_opcode; /* LIO opcode */
2608 int aio_reqprio; /* Request priority -- ignored */
2609 struct __aiocb_private32 _aiocb_private;
2610 struct sigevent32 aio_sigevent; /* Signal to deliver */
2613 #ifdef COMPAT_FREEBSD6
2615 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2619 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2620 * supported by AIO with the old sigevent structure.
2622 CP(*osig, *nsig, sigev_notify);
2623 switch (nsig->sigev_notify) {
2627 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2630 nsig->sigev_notify_kqueue =
2631 osig->__sigev_u.__sigev_notify_kqueue;
2632 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2641 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2643 struct oaiocb32 job32;
2646 bzero(kjob, sizeof(struct aiocb));
2647 error = copyin(ujob, &job32, sizeof(job32));
2651 CP(job32, *kjob, aio_fildes);
2652 CP(job32, *kjob, aio_offset);
2653 PTRIN_CP(job32, *kjob, aio_buf);
2654 CP(job32, *kjob, aio_nbytes);
2655 CP(job32, *kjob, aio_lio_opcode);
2656 CP(job32, *kjob, aio_reqprio);
2657 CP(job32, *kjob, _aiocb_private.status);
2658 CP(job32, *kjob, _aiocb_private.error);
2659 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2660 return (convert_old_sigevent32(&job32.aio_sigevent,
2661 &kjob->aio_sigevent));
2666 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2668 struct aiocb32 job32;
2671 error = copyin(ujob, &job32, sizeof(job32));
2674 CP(job32, *kjob, aio_fildes);
2675 CP(job32, *kjob, aio_offset);
2676 PTRIN_CP(job32, *kjob, aio_buf);
2677 CP(job32, *kjob, aio_nbytes);
2678 CP(job32, *kjob, aio_lio_opcode);
2679 CP(job32, *kjob, aio_reqprio);
2680 CP(job32, *kjob, _aiocb_private.status);
2681 CP(job32, *kjob, _aiocb_private.error);
2682 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2683 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2687 aiocb32_fetch_status(struct aiocb *ujob)
2689 struct aiocb32 *ujob32;
2691 ujob32 = (struct aiocb32 *)ujob;
2692 return (fuword32(&ujob32->_aiocb_private.status));
2696 aiocb32_fetch_error(struct aiocb *ujob)
2698 struct aiocb32 *ujob32;
2700 ujob32 = (struct aiocb32 *)ujob;
2701 return (fuword32(&ujob32->_aiocb_private.error));
2705 aiocb32_store_status(struct aiocb *ujob, long status)
2707 struct aiocb32 *ujob32;
2709 ujob32 = (struct aiocb32 *)ujob;
2710 return (suword32(&ujob32->_aiocb_private.status, status));
2714 aiocb32_store_error(struct aiocb *ujob, long error)
2716 struct aiocb32 *ujob32;
2718 ujob32 = (struct aiocb32 *)ujob;
2719 return (suword32(&ujob32->_aiocb_private.error, error));
2723 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2725 struct aiocb32 *ujob32;
2727 ujob32 = (struct aiocb32 *)ujob;
2728 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2732 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2735 return (suword32(ujobp, (long)ujob));
2738 static struct aiocb_ops aiocb32_ops = {
2739 .copyin = aiocb32_copyin,
2740 .fetch_status = aiocb32_fetch_status,
2741 .fetch_error = aiocb32_fetch_error,
2742 .store_status = aiocb32_store_status,
2743 .store_error = aiocb32_store_error,
2744 .store_kernelinfo = aiocb32_store_kernelinfo,
2745 .store_aiocb = aiocb32_store_aiocb,
2748 #ifdef COMPAT_FREEBSD6
2749 static struct aiocb_ops aiocb32_ops_osigevent = {
2750 .copyin = aiocb32_copyin_old_sigevent,
2751 .fetch_status = aiocb32_fetch_status,
2752 .fetch_error = aiocb32_fetch_error,
2753 .store_status = aiocb32_store_status,
2754 .store_error = aiocb32_store_error,
2755 .store_kernelinfo = aiocb32_store_kernelinfo,
2756 .store_aiocb = aiocb32_store_aiocb,
2761 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2764 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2768 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2770 struct timespec32 ts32;
2771 struct timespec ts, *tsp;
2772 struct aiocb **ujoblist;
2773 uint32_t *ujoblist32;
2776 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2780 /* Get timespec struct. */
2781 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2783 CP(ts32, ts, tv_sec);
2784 CP(ts32, ts, tv_nsec);
2789 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2790 ujoblist32 = (uint32_t *)ujoblist;
2791 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2792 sizeof(ujoblist32[0]));
2794 for (i = uap->nent; i > 0; i--)
2795 ujoblist[i] = PTRIN(ujoblist32[i]);
2797 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2799 uma_zfree(aiol_zone, ujoblist);
2804 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2807 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2810 #ifdef COMPAT_FREEBSD6
2812 freebsd6_freebsd32_aio_read(struct thread *td,
2813 struct freebsd6_freebsd32_aio_read_args *uap)
2816 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2817 &aiocb32_ops_osigevent));
2822 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2825 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2829 #ifdef COMPAT_FREEBSD6
2831 freebsd6_freebsd32_aio_write(struct thread *td,
2832 struct freebsd6_freebsd32_aio_write_args *uap)
2835 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2836 &aiocb32_ops_osigevent));
2841 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2844 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2849 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2852 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2857 freebsd32_aio_waitcomplete(struct thread *td,
2858 struct freebsd32_aio_waitcomplete_args *uap)
2860 struct timespec32 ts32;
2861 struct timespec ts, *tsp;
2865 /* Get timespec struct. */
2866 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2869 CP(ts32, ts, tv_sec);
2870 CP(ts32, ts, tv_nsec);
2875 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2880 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2883 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2887 #ifdef COMPAT_FREEBSD6
2889 freebsd6_freebsd32_lio_listio(struct thread *td,
2890 struct freebsd6_freebsd32_lio_listio_args *uap)
2892 struct aiocb **acb_list;
2893 struct sigevent *sigp, sig;
2894 struct osigevent32 osig;
2895 uint32_t *acb_list32;
2898 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2902 if (nent < 0 || nent > AIO_LISTIO_MAX)
2905 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2906 error = copyin(uap->sig, &osig, sizeof(osig));
2909 error = convert_old_sigevent32(&osig, &sig);
2916 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2917 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2919 free(acb_list32, M_LIO);
2922 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2923 for (i = 0; i < nent; i++)
2924 acb_list[i] = PTRIN(acb_list32[i]);
2925 free(acb_list32, M_LIO);
2927 error = kern_lio_listio(td, uap->mode,
2928 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2929 &aiocb32_ops_osigevent);
2930 free(acb_list, M_LIO);
2936 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2938 struct aiocb **acb_list;
2939 struct sigevent *sigp, sig;
2940 struct sigevent32 sig32;
2941 uint32_t *acb_list32;
2944 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2948 if (nent < 0 || nent > AIO_LISTIO_MAX)
2951 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2952 error = copyin(uap->sig, &sig32, sizeof(sig32));
2955 error = convert_sigevent32(&sig32, &sig);
2962 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2963 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2965 free(acb_list32, M_LIO);
2968 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2969 for (i = 0; i < nent; i++)
2970 acb_list[i] = PTRIN(acb_list32[i]);
2971 free(acb_list32, M_LIO);
2973 error = kern_lio_listio(td, uap->mode,
2974 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2976 free(acb_list, M_LIO);