2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. John S. Dyson's name may not be used to endorse or promote products
10 * derived from this software without specific prior written permission.
12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
13 * bad that happens because of using this software isn't the responsibility
14 * of the author. This software is distributed AS-IS.
18 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
24 #include "opt_compat.h"
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
40 #include <sys/limits.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
66 #include <machine/atomic.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
78 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
79 * overflow. (XXX will be removed soon.)
81 static u_long jobrefid;
84 * Counter for aio_fsync.
86 static uint64_t jobseqno;
88 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC 32
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
97 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
101 #define MAX_BUF_AIO 16
104 FEATURE(aio, "Asynchronous I/O");
106 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
108 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
109 "Async IO management");
111 static int enable_aio_unsafe = 0;
112 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
113 "Permit asynchronous IO on all file types, not just known-safe types");
115 static unsigned int unsafe_warningcnt = 1;
116 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
117 &unsafe_warningcnt, 0,
118 "Warnings that will be triggered upon failed IO requests on unsafe files");
120 static int max_aio_procs = MAX_AIO_PROCS;
121 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
122 "Maximum number of kernel processes to use for handling async IO ");
124 static int num_aio_procs = 0;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
126 "Number of presently active kernel processes for async IO");
129 * The code will adjust the actual number of AIO processes towards this
130 * number when it gets a chance.
132 static int target_aio_procs = TARGET_AIO_PROCS;
133 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
135 "Preferred number of ready kernel processes for async IO");
137 static int max_queue_count = MAX_AIO_QUEUE;
138 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
139 "Maximum number of aio requests to queue, globally");
141 static int num_queue_count = 0;
142 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
143 "Number of queued aio requests");
145 static int num_buf_aio = 0;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
147 "Number of aio requests presently handled by the buf subsystem");
149 /* Number of async I/O processes in the process of being started */
150 /* XXX This should be local to aio_aqueue() */
151 static int num_aio_resv_start = 0;
153 static int aiod_lifetime;
154 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
155 "Maximum lifetime for idle aiod");
157 static int max_aio_per_proc = MAX_AIO_PER_PROC;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
160 "Maximum active aio requests per process (stored in the process)");
162 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
163 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
164 &max_aio_queue_per_proc, 0,
165 "Maximum queued aio requests per process (stored in the process)");
167 static int max_buf_aio = MAX_BUF_AIO;
168 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
169 "Maximum buf aio requests per process (stored in the process)");
171 #ifdef COMPAT_FREEBSD6
172 typedef struct oaiocb {
173 int aio_fildes; /* File descriptor */
174 off_t aio_offset; /* File offset for I/O */
175 volatile void *aio_buf; /* I/O buffer in process space */
176 size_t aio_nbytes; /* Number of bytes for I/O */
177 struct osigevent aio_sigevent; /* Signal to deliver */
178 int aio_lio_opcode; /* LIO opcode */
179 int aio_reqprio; /* Request priority -- ignored */
180 struct __aiocb_private _aiocb_private;
185 * Below is a key of locks used to protect each member of struct kaiocb
186 * aioliojob and kaioinfo and any backends.
188 * * - need not protected
189 * a - locked by kaioinfo lock
190 * b - locked by backend lock, the backend lock can be null in some cases,
191 * for example, BIO belongs to this type, in this case, proc lock is
193 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
197 * If the routine that services an AIO request blocks while running in an
198 * AIO kernel process it can starve other I/O requests. BIO requests
199 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
200 * processes at all. Socket I/O requests use a separate pool of
201 * kprocs and also force non-blocking I/O. Other file I/O requests
202 * use the generic fo_read/fo_write operations which can block. The
203 * fsync and mlock operations can also block while executing. Ideally
204 * none of these requests would block while executing.
206 * Note that the service routines cannot toggle O_NONBLOCK in the file
207 * structure directly while handling a request due to races with
212 #define KAIOCB_QUEUEING 0x01
213 #define KAIOCB_CANCELLED 0x02
214 #define KAIOCB_CANCELLING 0x04
215 #define KAIOCB_CHECKSYNC 0x08
216 #define KAIOCB_CLEARED 0x10
217 #define KAIOCB_FINISHED 0x20
222 #define AIOP_FREE 0x1 /* proc on free queue */
225 int aioprocflags; /* (c) AIO proc flags */
226 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
227 struct proc *aioproc; /* (*) the AIO proc */
231 * data-structure for lio signal management
234 int lioj_flags; /* (a) listio flags */
235 int lioj_count; /* (a) listio flags */
236 int lioj_finished_count; /* (a) listio flags */
237 struct sigevent lioj_signal; /* (a) signal on all I/O done */
238 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
239 struct knlist klist; /* (a) list of knotes */
240 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
243 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
244 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
245 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
248 * per process aio data structure
251 struct mtx kaio_mtx; /* the lock to protect this struct */
252 int kaio_flags; /* (a) per process kaio flags */
253 int kaio_maxactive_count; /* (*) maximum number of AIOs */
254 int kaio_active_count; /* (c) number of currently used AIOs */
255 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
256 int kaio_count; /* (a) size of AIO queue */
257 int kaio_ballowed_count; /* (*) maximum number of buffers */
258 int kaio_buffer_count; /* (a) number of physio buffers */
259 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */
260 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */
261 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
262 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */
263 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */
264 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */
265 struct task kaio_task; /* (*) task to kick aio processes */
266 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */
269 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
270 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
271 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
272 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
274 #define KAIO_RUNDOWN 0x1 /* process is being run down */
275 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */
278 * Operations used to interact with userland aio control blocks.
279 * Different ABIs provide their own operations.
282 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
283 long (*fetch_status)(struct aiocb *ujob);
284 long (*fetch_error)(struct aiocb *ujob);
285 int (*store_status)(struct aiocb *ujob, long status);
286 int (*store_error)(struct aiocb *ujob, long error);
287 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
288 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
291 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */
292 static struct sema aio_newproc_sem;
293 static struct mtx aio_job_mtx;
294 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */
295 static struct unrhdr *aiod_unr;
297 void aio_init_aioinfo(struct proc *p);
298 static int aio_onceonly(void);
299 static int aio_free_entry(struct kaiocb *job);
300 static void aio_process_rw(struct kaiocb *job);
301 static void aio_process_sync(struct kaiocb *job);
302 static void aio_process_mlock(struct kaiocb *job);
303 static void aio_schedule_fsync(void *context, int pending);
304 static int aio_newproc(int *);
305 int aio_aqueue(struct thread *td, struct aiocb *ujob,
306 struct aioliojob *lio, int type, struct aiocb_ops *ops);
307 static int aio_queue_file(struct file *fp, struct kaiocb *job);
308 static void aio_physwakeup(struct bio *bp);
309 static void aio_proc_rundown(void *arg, struct proc *p);
310 static void aio_proc_rundown_exec(void *arg, struct proc *p,
311 struct image_params *imgp);
312 static int aio_qphysio(struct proc *p, struct kaiocb *job);
313 static void aio_daemon(void *param);
314 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
315 static bool aio_clear_cancel_function_locked(struct kaiocb *job);
316 static int aio_kick(struct proc *userp);
317 static void aio_kick_nowait(struct proc *userp);
318 static void aio_kick_helper(void *context, int pending);
319 static int filt_aioattach(struct knote *kn);
320 static void filt_aiodetach(struct knote *kn);
321 static int filt_aio(struct knote *kn, long hint);
322 static int filt_lioattach(struct knote *kn);
323 static void filt_liodetach(struct knote *kn);
324 static int filt_lio(struct knote *kn, long hint);
328 * kaio Per process async io info
329 * aiop async io process data
330 * aiocb async io jobs
331 * aiol list io job pointer - internal to aio_suspend XXX
332 * aiolio list io jobs
334 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
336 /* kqueue filters for aio */
337 static struct filterops aio_filtops = {
339 .f_attach = filt_aioattach,
340 .f_detach = filt_aiodetach,
343 static struct filterops lio_filtops = {
345 .f_attach = filt_lioattach,
346 .f_detach = filt_liodetach,
350 static eventhandler_tag exit_tag, exec_tag;
352 TASKQUEUE_DEFINE_THREAD(aiod_kick);
355 * Main operations function for use as a kernel module.
358 aio_modload(struct module *module, int cmd, void *arg)
375 static moduledata_t aio_mod = {
381 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
382 MODULE_VERSION(aio, 1);
385 * Startup initialization
391 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
392 EVENTHANDLER_PRI_ANY);
393 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
394 NULL, EVENTHANDLER_PRI_ANY);
395 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
396 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
397 TAILQ_INIT(&aio_freeproc);
398 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
399 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
400 TAILQ_INIT(&aio_jobs);
401 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
402 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
403 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
404 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
405 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
406 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
407 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
408 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
409 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
410 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
411 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
412 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
414 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
415 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
416 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
417 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
423 * Init the per-process aioinfo structure. The aioinfo limits are set
424 * per-process for user limit (resource) management.
427 aio_init_aioinfo(struct proc *p)
431 ki = uma_zalloc(kaio_zone, M_WAITOK);
432 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
434 ki->kaio_maxactive_count = max_aio_per_proc;
435 ki->kaio_active_count = 0;
436 ki->kaio_qallowed_count = max_aio_queue_per_proc;
438 ki->kaio_ballowed_count = max_buf_aio;
439 ki->kaio_buffer_count = 0;
440 TAILQ_INIT(&ki->kaio_all);
441 TAILQ_INIT(&ki->kaio_done);
442 TAILQ_INIT(&ki->kaio_jobqueue);
443 TAILQ_INIT(&ki->kaio_liojoblist);
444 TAILQ_INIT(&ki->kaio_syncqueue);
445 TAILQ_INIT(&ki->kaio_syncready);
446 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
447 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
449 if (p->p_aioinfo == NULL) {
454 mtx_destroy(&ki->kaio_mtx);
455 uma_zfree(kaio_zone, ki);
458 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
463 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
468 error = sigev_findtd(p, sigev, &td);
472 ksiginfo_set_sigev(ksi, sigev);
473 ksi->ksi_code = SI_ASYNCIO;
474 ksi->ksi_flags |= KSI_EXT | KSI_INS;
475 tdsendsignal(p, td, ksi->ksi_signo, ksi);
482 * Free a job entry. Wait for completion if it is currently active, but don't
483 * delay forever. If we delay, we return a flag that says that we have to
484 * restart the queue scan.
487 aio_free_entry(struct kaiocb *job)
490 struct aioliojob *lj;
498 AIO_LOCK_ASSERT(ki, MA_OWNED);
499 MPASS(job->jobflags & KAIOCB_FINISHED);
501 atomic_subtract_int(&num_queue_count, 1);
504 MPASS(ki->kaio_count >= 0);
506 TAILQ_REMOVE(&ki->kaio_done, job, plist);
507 TAILQ_REMOVE(&ki->kaio_all, job, allist);
512 lj->lioj_finished_count--;
514 if (lj->lioj_count == 0) {
515 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
516 /* lio is going away, we need to destroy any knotes */
517 knlist_delete(&lj->klist, curthread, 1);
519 sigqueue_take(&lj->lioj_ksi);
521 uma_zfree(aiolio_zone, lj);
525 /* job is going away, we need to destroy any knotes */
526 knlist_delete(&job->klist, curthread, 1);
528 sigqueue_take(&job->ksi);
534 * The thread argument here is used to find the owning process
535 * and is also passed to fo_close() which may pass it to various
536 * places such as devsw close() routines. Because of that, we
537 * need a thread pointer from the process owning the job that is
538 * persistent and won't disappear out from under us or move to
541 * Currently, all the callers of this function call it to remove
542 * a kaiocb from the current process' job list either via a
543 * syscall or due to the current process calling exit() or
544 * execve(). Thus, we know that p == curproc. We also know that
545 * curthread can't exit since we are curthread.
547 * Therefore, we use curthread as the thread to pass to
548 * knlist_delete(). This does mean that it is possible for the
549 * thread pointer at close time to differ from the thread pointer
550 * at open time, but this is already true of file descriptors in
551 * a multithreaded process.
554 fdrop(job->fd_file, curthread);
556 uma_zfree(aiocb_zone, job);
563 aio_proc_rundown_exec(void *arg, struct proc *p,
564 struct image_params *imgp __unused)
566 aio_proc_rundown(arg, p);
570 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
572 aio_cancel_fn_t *func;
575 AIO_LOCK_ASSERT(ki, MA_OWNED);
576 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
578 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
579 job->jobflags |= KAIOCB_CANCELLED;
581 func = job->cancel_fn;
584 * If there is no cancel routine, just leave the job marked as
585 * cancelled. The job should be in active use by a caller who
586 * should complete it normally or when it fails to install a
593 * Set the CANCELLING flag so that aio_complete() will defer
594 * completions of this job. This prevents the job from being
595 * freed out from under the cancel callback. After the
596 * callback any deferred completion (whether from the callback
597 * or any other source) will be completed.
599 job->jobflags |= KAIOCB_CANCELLING;
603 job->jobflags &= ~KAIOCB_CANCELLING;
604 if (job->jobflags & KAIOCB_FINISHED) {
605 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
606 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
607 aio_bio_done_notify(p, job);
610 * The cancel callback might have scheduled an
611 * operation to cancel this request, but it is
612 * only counted as cancelled if the request is
613 * cancelled when the callback returns.
621 * Rundown the jobs for a given process.
624 aio_proc_rundown(void *arg, struct proc *p)
627 struct aioliojob *lj;
628 struct kaiocb *job, *jobn;
630 KASSERT(curthread->td_proc == p,
631 ("%s: called on non-curproc", __func__));
637 ki->kaio_flags |= KAIO_RUNDOWN;
642 * Try to cancel all pending requests. This code simulates
643 * aio_cancel on all pending I/O requests.
645 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
646 aio_cancel_job(p, ki, job);
649 /* Wait for all running I/O to be finished */
650 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
651 ki->kaio_flags |= KAIO_WAKEUP;
652 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
656 /* Free all completed I/O requests. */
657 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
660 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
661 if (lj->lioj_count == 0) {
662 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
663 knlist_delete(&lj->klist, curthread, 1);
665 sigqueue_take(&lj->lioj_ksi);
667 uma_zfree(aiolio_zone, lj);
669 panic("LIO job not cleaned up: C:%d, FC:%d\n",
670 lj->lioj_count, lj->lioj_finished_count);
674 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
675 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
676 mtx_destroy(&ki->kaio_mtx);
677 uma_zfree(kaio_zone, ki);
682 * Select a job to run (called by an AIO daemon).
684 static struct kaiocb *
685 aio_selectjob(struct aioproc *aiop)
691 mtx_assert(&aio_job_mtx, MA_OWNED);
693 TAILQ_FOREACH(job, &aio_jobs, list) {
694 userp = job->userproc;
695 ki = userp->p_aioinfo;
697 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
698 TAILQ_REMOVE(&aio_jobs, job, list);
699 if (!aio_clear_cancel_function(job))
702 /* Account for currently active jobs. */
703 ki->kaio_active_count++;
711 * Move all data to a permanent storage device. This code
712 * simulates the fsync syscall.
715 aio_fsync_vnode(struct thread *td, struct vnode *vp)
720 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
722 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
723 if (vp->v_object != NULL) {
724 VM_OBJECT_WLOCK(vp->v_object);
725 vm_object_page_clean(vp->v_object, 0, 0, 0);
726 VM_OBJECT_WUNLOCK(vp->v_object);
728 error = VOP_FSYNC(vp, MNT_WAIT, td);
731 vn_finished_write(mp);
737 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
738 * does the I/O request for the non-physio version of the operations. The
739 * normal vn operations are used, and this code should work in all instances
740 * for every type of file, including pipes, sockets, fifos, and regular files.
742 * XXX I don't think it works well for socket, pipe, and fifo.
745 aio_process_rw(struct kaiocb *job)
747 struct ucred *td_savedcred;
754 long msgsnd_st, msgsnd_end;
755 long msgrcv_st, msgrcv_end;
756 long oublock_st, oublock_end;
757 long inblock_st, inblock_end;
760 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
761 job->uaiocb.aio_lio_opcode == LIO_WRITE,
762 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
764 aio_switch_vmspace(job);
766 td_savedcred = td->td_ucred;
767 td->td_ucred = job->cred;
771 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
772 aiov.iov_len = cb->aio_nbytes;
774 auio.uio_iov = &aiov;
776 auio.uio_offset = cb->aio_offset;
777 auio.uio_resid = cb->aio_nbytes;
778 cnt = cb->aio_nbytes;
779 auio.uio_segflg = UIO_USERSPACE;
782 msgrcv_st = td->td_ru.ru_msgrcv;
783 msgsnd_st = td->td_ru.ru_msgsnd;
784 inblock_st = td->td_ru.ru_inblock;
785 oublock_st = td->td_ru.ru_oublock;
788 * aio_aqueue() acquires a reference to the file that is
789 * released in aio_free_entry().
791 if (cb->aio_lio_opcode == LIO_READ) {
792 auio.uio_rw = UIO_READ;
793 if (auio.uio_resid == 0)
796 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
798 if (fp->f_type == DTYPE_VNODE)
800 auio.uio_rw = UIO_WRITE;
801 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
803 msgrcv_end = td->td_ru.ru_msgrcv;
804 msgsnd_end = td->td_ru.ru_msgsnd;
805 inblock_end = td->td_ru.ru_inblock;
806 oublock_end = td->td_ru.ru_oublock;
808 job->msgrcv = msgrcv_end - msgrcv_st;
809 job->msgsnd = msgsnd_end - msgsnd_st;
810 job->inblock = inblock_end - inblock_st;
811 job->outblock = oublock_end - oublock_st;
813 if ((error) && (auio.uio_resid != cnt)) {
814 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
816 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
817 PROC_LOCK(job->userproc);
818 kern_psignal(job->userproc, SIGPIPE);
819 PROC_UNLOCK(job->userproc);
823 cnt -= auio.uio_resid;
824 td->td_ucred = td_savedcred;
826 aio_complete(job, -1, error);
828 aio_complete(job, cnt, 0);
832 aio_process_sync(struct kaiocb *job)
834 struct thread *td = curthread;
835 struct ucred *td_savedcred = td->td_ucred;
836 struct file *fp = job->fd_file;
839 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
840 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
842 td->td_ucred = job->cred;
843 if (fp->f_vnode != NULL)
844 error = aio_fsync_vnode(td, fp->f_vnode);
845 td->td_ucred = td_savedcred;
847 aio_complete(job, -1, error);
849 aio_complete(job, 0, 0);
853 aio_process_mlock(struct kaiocb *job)
855 struct aiocb *cb = &job->uaiocb;
858 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
859 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
861 aio_switch_vmspace(job);
862 error = kern_mlock(job->userproc, job->cred,
863 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
864 aio_complete(job, error != 0 ? -1 : 0, error);
868 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
870 struct aioliojob *lj;
872 struct kaiocb *sjob, *sjobn;
876 ki = userp->p_aioinfo;
877 AIO_LOCK_ASSERT(ki, MA_OWNED);
881 lj->lioj_finished_count++;
882 if (lj->lioj_count == lj->lioj_finished_count)
885 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
886 MPASS(job->jobflags & KAIOCB_FINISHED);
888 if (ki->kaio_flags & KAIO_RUNDOWN)
889 goto notification_done;
891 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
892 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
893 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
895 KNOTE_LOCKED(&job->klist, 1);
898 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
899 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
900 KNOTE_LOCKED(&lj->klist, 1);
902 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
904 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
905 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
906 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
907 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
912 if (job->jobflags & KAIOCB_CHECKSYNC) {
913 schedule_fsync = false;
914 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
915 if (job->fd_file != sjob->fd_file ||
916 job->seqno >= sjob->seqno)
918 if (--sjob->pending > 0)
920 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
921 if (!aio_clear_cancel_function_locked(sjob))
923 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
924 schedule_fsync = true;
927 taskqueue_enqueue(taskqueue_aiod_kick,
928 &ki->kaio_sync_task);
930 if (ki->kaio_flags & KAIO_WAKEUP) {
931 ki->kaio_flags &= ~KAIO_WAKEUP;
932 wakeup(&userp->p_aioinfo);
937 aio_schedule_fsync(void *context, int pending)
944 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
945 job = TAILQ_FIRST(&ki->kaio_syncready);
946 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
948 aio_schedule(job, aio_process_sync);
955 aio_cancel_cleared(struct kaiocb *job)
960 * The caller should hold the same queue lock held when
961 * aio_clear_cancel_function() was called and set this flag
962 * ensuring this check sees an up-to-date value. However,
963 * there is no way to assert that.
965 ki = job->userproc->p_aioinfo;
966 return ((job->jobflags & KAIOCB_CLEARED) != 0);
970 aio_clear_cancel_function_locked(struct kaiocb *job)
973 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
974 MPASS(job->cancel_fn != NULL);
975 if (job->jobflags & KAIOCB_CANCELLING) {
976 job->jobflags |= KAIOCB_CLEARED;
979 job->cancel_fn = NULL;
984 aio_clear_cancel_function(struct kaiocb *job)
989 ki = job->userproc->p_aioinfo;
991 ret = aio_clear_cancel_function_locked(job);
997 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1000 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1001 if (job->jobflags & KAIOCB_CANCELLED)
1003 job->cancel_fn = func;
1008 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1010 struct kaioinfo *ki;
1013 ki = job->userproc->p_aioinfo;
1015 ret = aio_set_cancel_function_locked(job, func);
1021 aio_complete(struct kaiocb *job, long status, int error)
1023 struct kaioinfo *ki;
1026 job->uaiocb._aiocb_private.error = error;
1027 job->uaiocb._aiocb_private.status = status;
1029 userp = job->userproc;
1030 ki = userp->p_aioinfo;
1033 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1034 ("duplicate aio_complete"));
1035 job->jobflags |= KAIOCB_FINISHED;
1036 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1037 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1038 aio_bio_done_notify(userp, job);
1044 aio_cancel(struct kaiocb *job)
1047 aio_complete(job, -1, ECANCELED);
1051 aio_switch_vmspace(struct kaiocb *job)
1054 vmspace_switch_aio(job->userproc->p_vmspace);
1058 * The AIO daemon, most of the actual work is done in aio_process_*,
1059 * but the setup (and address space mgmt) is done in this routine.
1062 aio_daemon(void *_id)
1065 struct aioproc *aiop;
1066 struct kaioinfo *ki;
1068 struct vmspace *myvm;
1069 struct thread *td = curthread;
1070 int id = (intptr_t)_id;
1073 * Grab an extra reference on the daemon's vmspace so that it
1074 * doesn't get freed by jobs that switch to a different
1078 myvm = vmspace_acquire_ref(p);
1080 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1083 * Allocate and ready the aio control info. There is one aiop structure
1086 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1088 aiop->aioprocflags = 0;
1091 * Wakeup parent process. (Parent sleeps to keep from blasting away
1092 * and creating too many daemons.)
1094 sema_post(&aio_newproc_sem);
1096 mtx_lock(&aio_job_mtx);
1099 * Take daemon off of free queue
1101 if (aiop->aioprocflags & AIOP_FREE) {
1102 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1103 aiop->aioprocflags &= ~AIOP_FREE;
1109 while ((job = aio_selectjob(aiop)) != NULL) {
1110 mtx_unlock(&aio_job_mtx);
1112 ki = job->userproc->p_aioinfo;
1113 job->handle_fn(job);
1115 mtx_lock(&aio_job_mtx);
1116 /* Decrement the active job count. */
1117 ki->kaio_active_count--;
1121 * Disconnect from user address space.
1123 if (p->p_vmspace != myvm) {
1124 mtx_unlock(&aio_job_mtx);
1125 vmspace_switch_aio(myvm);
1126 mtx_lock(&aio_job_mtx);
1128 * We have to restart to avoid race, we only sleep if
1129 * no job can be selected.
1134 mtx_assert(&aio_job_mtx, MA_OWNED);
1136 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1137 aiop->aioprocflags |= AIOP_FREE;
1140 * If daemon is inactive for a long time, allow it to exit,
1141 * thereby freeing resources.
1143 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1144 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1145 (aiop->aioprocflags & AIOP_FREE) &&
1146 num_aio_procs > target_aio_procs)
1149 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1151 mtx_unlock(&aio_job_mtx);
1152 uma_zfree(aiop_zone, aiop);
1153 free_unr(aiod_unr, id);
1156 KASSERT(p->p_vmspace == myvm,
1157 ("AIOD: bad vmspace for exiting daemon"));
1158 KASSERT(myvm->vm_refcnt > 1,
1159 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1164 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1165 * AIO daemon modifies its environment itself.
1168 aio_newproc(int *start)
1174 id = alloc_unr(aiod_unr);
1175 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1176 RFNOWAIT, 0, "aiod%d", id);
1179 * Wait until daemon is started.
1181 sema_wait(&aio_newproc_sem);
1182 mtx_lock(&aio_job_mtx);
1186 mtx_unlock(&aio_job_mtx);
1188 free_unr(aiod_unr, id);
1194 * Try the high-performance, low-overhead physio method for eligible
1195 * VCHR devices. This method doesn't use an aio helper thread, and
1196 * thus has very low overhead.
1198 * Assumes that the caller, aio_aqueue(), has incremented the file
1199 * structure's reference count, preventing its deallocation for the
1200 * duration of this call.
1203 aio_qphysio(struct proc *p, struct kaiocb *job)
1212 struct kaioinfo *ki;
1213 int error, ref, poff;
1219 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1223 if (vp->v_type != VCHR)
1225 if (vp->v_bufobj.bo_bsize == 0)
1227 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1231 csw = devvn_refthread(vp, &dev, &ref);
1235 if ((csw->d_flags & D_DISK) == 0) {
1239 if (cb->aio_nbytes > dev->si_iosize_max) {
1245 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1246 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1247 if (cb->aio_nbytes > MAXPHYS) {
1254 if (cb->aio_nbytes > MAXPHYS - poff) {
1258 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1263 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1266 ki->kaio_buffer_count++;
1269 job->bp = bp = g_alloc_bio();
1271 bp->bio_length = cb->aio_nbytes;
1272 bp->bio_bcount = cb->aio_nbytes;
1273 bp->bio_done = aio_physwakeup;
1274 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1275 bp->bio_offset = cb->aio_offset;
1276 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1278 bp->bio_caller1 = (void *)job;
1280 prot = VM_PROT_READ;
1281 if (cb->aio_lio_opcode == LIO_READ)
1282 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1283 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1284 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1285 nitems(job->pages));
1286 if (job->npages < 0) {
1291 pmap_qenter((vm_offset_t)pbuf->b_data,
1292 job->pages, job->npages);
1293 bp->bio_data = pbuf->b_data + poff;
1294 atomic_add_int(&num_buf_aio, 1);
1296 bp->bio_ma = job->pages;
1297 bp->bio_ma_n = job->npages;
1298 bp->bio_ma_offset = poff;
1299 bp->bio_data = unmapped_buf;
1300 bp->bio_flags |= BIO_UNMAPPED;
1303 /* Perform transfer. */
1304 csw->d_strategy(bp);
1305 dev_relthread(dev, ref);
1311 ki->kaio_buffer_count--;
1313 relpbuf(pbuf, NULL);
1319 dev_relthread(dev, ref);
1323 #ifdef COMPAT_FREEBSD6
1325 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1329 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1330 * supported by AIO with the old sigevent structure.
1332 nsig->sigev_notify = osig->sigev_notify;
1333 switch (nsig->sigev_notify) {
1337 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1340 nsig->sigev_notify_kqueue =
1341 osig->__sigev_u.__sigev_notify_kqueue;
1342 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1351 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1353 struct oaiocb *ojob;
1356 bzero(kjob, sizeof(struct aiocb));
1357 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1360 ojob = (struct oaiocb *)kjob;
1361 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1366 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1369 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1373 aiocb_fetch_status(struct aiocb *ujob)
1376 return (fuword(&ujob->_aiocb_private.status));
1380 aiocb_fetch_error(struct aiocb *ujob)
1383 return (fuword(&ujob->_aiocb_private.error));
1387 aiocb_store_status(struct aiocb *ujob, long status)
1390 return (suword(&ujob->_aiocb_private.status, status));
1394 aiocb_store_error(struct aiocb *ujob, long error)
1397 return (suword(&ujob->_aiocb_private.error, error));
1401 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1404 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1408 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1411 return (suword(ujobp, (long)ujob));
1414 static struct aiocb_ops aiocb_ops = {
1415 .copyin = aiocb_copyin,
1416 .fetch_status = aiocb_fetch_status,
1417 .fetch_error = aiocb_fetch_error,
1418 .store_status = aiocb_store_status,
1419 .store_error = aiocb_store_error,
1420 .store_kernelinfo = aiocb_store_kernelinfo,
1421 .store_aiocb = aiocb_store_aiocb,
1424 #ifdef COMPAT_FREEBSD6
1425 static struct aiocb_ops aiocb_ops_osigevent = {
1426 .copyin = aiocb_copyin_old_sigevent,
1427 .fetch_status = aiocb_fetch_status,
1428 .fetch_error = aiocb_fetch_error,
1429 .store_status = aiocb_store_status,
1430 .store_error = aiocb_store_error,
1431 .store_kernelinfo = aiocb_store_kernelinfo,
1432 .store_aiocb = aiocb_store_aiocb,
1437 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1438 * technique is done in this code.
1441 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1442 int type, struct aiocb_ops *ops)
1444 struct proc *p = td->td_proc;
1445 cap_rights_t rights;
1448 struct kaioinfo *ki;
1456 if (p->p_aioinfo == NULL)
1457 aio_init_aioinfo(p);
1461 ops->store_status(ujob, -1);
1462 ops->store_error(ujob, 0);
1463 ops->store_kernelinfo(ujob, -1);
1465 if (num_queue_count >= max_queue_count ||
1466 ki->kaio_count >= ki->kaio_qallowed_count) {
1467 ops->store_error(ujob, EAGAIN);
1471 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1472 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1474 error = ops->copyin(ujob, &job->uaiocb);
1476 ops->store_error(ujob, error);
1477 uma_zfree(aiocb_zone, job);
1481 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1482 uma_zfree(aiocb_zone, job);
1486 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1487 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1488 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1489 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1490 ops->store_error(ujob, EINVAL);
1491 uma_zfree(aiocb_zone, job);
1495 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1496 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1497 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1498 uma_zfree(aiocb_zone, job);
1502 ksiginfo_init(&job->ksi);
1504 /* Save userspace address of the job info. */
1507 /* Get the opcode. */
1508 if (type != LIO_NOP)
1509 job->uaiocb.aio_lio_opcode = type;
1510 opcode = job->uaiocb.aio_lio_opcode;
1513 * Validate the opcode and fetch the file object for the specified
1516 * XXXRW: Moved the opcode validation up here so that we don't
1517 * retrieve a file descriptor without knowing what the capabiltity
1520 fd = job->uaiocb.aio_fildes;
1523 error = fget_write(td, fd,
1524 cap_rights_init(&rights, CAP_PWRITE), &fp);
1527 error = fget_read(td, fd,
1528 cap_rights_init(&rights, CAP_PREAD), &fp);
1531 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1537 error = fget(td, fd, cap_rights_init(&rights), &fp);
1543 uma_zfree(aiocb_zone, job);
1544 ops->store_error(ujob, error);
1548 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1553 if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1554 job->uaiocb.aio_offset < 0 &&
1555 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1562 mtx_lock(&aio_job_mtx);
1564 job->seqno = jobseqno++;
1565 mtx_unlock(&aio_job_mtx);
1566 error = ops->store_kernelinfo(ujob, jid);
1571 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1573 if (opcode == LIO_NOP) {
1575 uma_zfree(aiocb_zone, job);
1579 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1581 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1582 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1586 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1587 kev.ident = (uintptr_t)job->ujob;
1588 kev.filter = EVFILT_AIO;
1589 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1590 kev.data = (intptr_t)job;
1591 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1592 error = kqfd_register(kqfd, &kev, td, 1);
1598 ops->store_error(ujob, EINPROGRESS);
1599 job->uaiocb._aiocb_private.error = EINPROGRESS;
1601 job->cred = crhold(td->td_ucred);
1602 job->jobflags = KAIOCB_QUEUEING;
1605 if (opcode == LIO_MLOCK) {
1606 aio_schedule(job, aio_process_mlock);
1608 } else if (fp->f_ops->fo_aio_queue == NULL)
1609 error = aio_queue_file(fp, job);
1611 error = fo_aio_queue(fp, job);
1616 job->jobflags &= ~KAIOCB_QUEUEING;
1617 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1621 atomic_add_int(&num_queue_count, 1);
1622 if (job->jobflags & KAIOCB_FINISHED) {
1624 * The queue callback completed the request synchronously.
1625 * The bulk of the completion is deferred in that case
1628 aio_bio_done_notify(p, job);
1630 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1635 knlist_delete(&job->klist, curthread, 0);
1638 uma_zfree(aiocb_zone, job);
1639 ops->store_error(ujob, error);
1644 aio_cancel_daemon_job(struct kaiocb *job)
1647 mtx_lock(&aio_job_mtx);
1648 if (!aio_cancel_cleared(job))
1649 TAILQ_REMOVE(&aio_jobs, job, list);
1650 mtx_unlock(&aio_job_mtx);
1655 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1658 mtx_lock(&aio_job_mtx);
1659 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1660 mtx_unlock(&aio_job_mtx);
1664 job->handle_fn = func;
1665 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1666 aio_kick_nowait(job->userproc);
1667 mtx_unlock(&aio_job_mtx);
1671 aio_cancel_sync(struct kaiocb *job)
1673 struct kaioinfo *ki;
1675 ki = job->userproc->p_aioinfo;
1677 if (!aio_cancel_cleared(job))
1678 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1684 aio_queue_file(struct file *fp, struct kaiocb *job)
1686 struct aioliojob *lj;
1687 struct kaioinfo *ki;
1688 struct kaiocb *job2;
1695 ki = job->userproc->p_aioinfo;
1696 opcode = job->uaiocb.aio_lio_opcode;
1697 if (opcode == LIO_SYNC)
1700 if ((error = aio_qphysio(job->userproc, job)) == 0)
1704 * XXX: This means qphysio() failed with EFAULT. The current
1705 * behavior is to retry the operation via fo_read/fo_write.
1706 * Wouldn't it be better to just complete the request with an
1714 if (fp->f_type == DTYPE_VNODE) {
1716 if (vp->v_type == VREG || vp->v_type == VDIR) {
1717 mp = fp->f_vnode->v_mount;
1718 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1722 if (!(safe || enable_aio_unsafe)) {
1723 counted_warning(&unsafe_warningcnt,
1724 "is attempting to use unsafe AIO requests");
1725 return (EOPNOTSUPP);
1728 if (opcode == LIO_SYNC) {
1730 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1731 if (job2->fd_file == job->fd_file &&
1732 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1733 job2->seqno < job->seqno) {
1734 job2->jobflags |= KAIOCB_CHECKSYNC;
1738 if (job->pending != 0) {
1739 if (!aio_set_cancel_function_locked(job,
1745 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1755 aio_schedule(job, aio_process_rw);
1759 aio_schedule(job, aio_process_sync);
1770 aio_kick_nowait(struct proc *userp)
1772 struct kaioinfo *ki = userp->p_aioinfo;
1773 struct aioproc *aiop;
1775 mtx_assert(&aio_job_mtx, MA_OWNED);
1776 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1777 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1778 aiop->aioprocflags &= ~AIOP_FREE;
1779 wakeup(aiop->aioproc);
1780 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1781 ki->kaio_active_count + num_aio_resv_start <
1782 ki->kaio_maxactive_count) {
1783 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1788 aio_kick(struct proc *userp)
1790 struct kaioinfo *ki = userp->p_aioinfo;
1791 struct aioproc *aiop;
1794 mtx_assert(&aio_job_mtx, MA_OWNED);
1796 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1797 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1798 aiop->aioprocflags &= ~AIOP_FREE;
1799 wakeup(aiop->aioproc);
1800 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1801 ki->kaio_active_count + num_aio_resv_start <
1802 ki->kaio_maxactive_count) {
1803 num_aio_resv_start++;
1804 mtx_unlock(&aio_job_mtx);
1805 error = aio_newproc(&num_aio_resv_start);
1806 mtx_lock(&aio_job_mtx);
1808 num_aio_resv_start--;
1818 aio_kick_helper(void *context, int pending)
1820 struct proc *userp = context;
1822 mtx_lock(&aio_job_mtx);
1823 while (--pending >= 0) {
1824 if (aio_kick(userp))
1827 mtx_unlock(&aio_job_mtx);
1831 * Support the aio_return system call, as a side-effect, kernel resources are
1835 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1837 struct proc *p = td->td_proc;
1839 struct kaioinfo *ki;
1846 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1847 if (job->ujob == ujob)
1851 MPASS(job->jobflags & KAIOCB_FINISHED);
1852 status = job->uaiocb._aiocb_private.status;
1853 error = job->uaiocb._aiocb_private.error;
1854 td->td_retval[0] = status;
1855 td->td_ru.ru_oublock += job->outblock;
1856 td->td_ru.ru_inblock += job->inblock;
1857 td->td_ru.ru_msgsnd += job->msgsnd;
1858 td->td_ru.ru_msgrcv += job->msgrcv;
1859 aio_free_entry(job);
1861 ops->store_error(ujob, error);
1862 ops->store_status(ujob, status);
1871 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1874 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1878 * Allow a process to wakeup when any of the I/O requests are completed.
1881 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1882 struct timespec *ts)
1884 struct proc *p = td->td_proc;
1886 struct kaioinfo *ki;
1887 struct kaiocb *firstjob, *job;
1892 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1895 TIMESPEC_TO_TIMEVAL(&atv, ts);
1896 if (itimerfix(&atv))
1898 timo = tvtohz(&atv);
1912 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1913 for (i = 0; i < njoblist; i++) {
1914 if (job->ujob == ujoblist[i]) {
1915 if (firstjob == NULL)
1917 if (job->jobflags & KAIOCB_FINISHED)
1922 /* All tasks were finished. */
1923 if (firstjob == NULL)
1926 ki->kaio_flags |= KAIO_WAKEUP;
1927 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1929 if (error == ERESTART)
1940 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1942 struct timespec ts, *tsp;
1943 struct aiocb **ujoblist;
1946 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1950 /* Get timespec struct. */
1951 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1957 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1958 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1960 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1961 uma_zfree(aiol_zone, ujoblist);
1966 * aio_cancel cancels any non-physio aio operations not currently in
1970 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1972 struct proc *p = td->td_proc;
1973 struct kaioinfo *ki;
1974 struct kaiocb *job, *jobn;
1976 cap_rights_t rights;
1979 int notcancelled = 0;
1982 /* Lookup file object. */
1983 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1991 if (fp->f_type == DTYPE_VNODE) {
1993 if (vn_isdisk(vp, &error)) {
1995 td->td_retval[0] = AIO_NOTCANCELED;
2001 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2002 if ((uap->fd == job->uaiocb.aio_fildes) &&
2003 ((uap->aiocbp == NULL) ||
2004 (uap->aiocbp == job->ujob))) {
2005 if (aio_cancel_job(p, ki, job)) {
2010 if (uap->aiocbp != NULL)
2019 if (uap->aiocbp != NULL) {
2021 td->td_retval[0] = AIO_CANCELED;
2027 td->td_retval[0] = AIO_NOTCANCELED;
2032 td->td_retval[0] = AIO_CANCELED;
2036 td->td_retval[0] = AIO_ALLDONE;
2042 * aio_error is implemented in the kernel level for compatibility purposes
2043 * only. For a user mode async implementation, it would be best to do it in
2044 * a userland subroutine.
2047 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2049 struct proc *p = td->td_proc;
2051 struct kaioinfo *ki;
2056 td->td_retval[0] = EINVAL;
2061 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2062 if (job->ujob == ujob) {
2063 if (job->jobflags & KAIOCB_FINISHED)
2065 job->uaiocb._aiocb_private.error;
2067 td->td_retval[0] = EINPROGRESS;
2075 * Hack for failure of aio_aqueue.
2077 status = ops->fetch_status(ujob);
2079 td->td_retval[0] = ops->fetch_error(ujob);
2083 td->td_retval[0] = EINVAL;
2088 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2091 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2094 /* syscall - asynchronous read from a file (REALTIME) */
2095 #ifdef COMPAT_FREEBSD6
2097 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2100 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2101 &aiocb_ops_osigevent));
2106 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2109 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2112 /* syscall - asynchronous write to a file (REALTIME) */
2113 #ifdef COMPAT_FREEBSD6
2115 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2118 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2119 &aiocb_ops_osigevent));
2124 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2127 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2131 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2134 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2138 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2139 struct aiocb **acb_list, int nent, struct sigevent *sig,
2140 struct aiocb_ops *ops)
2142 struct proc *p = td->td_proc;
2144 struct kaioinfo *ki;
2145 struct aioliojob *lj;
2151 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2154 if (nent < 0 || nent > AIO_LISTIO_MAX)
2157 if (p->p_aioinfo == NULL)
2158 aio_init_aioinfo(p);
2162 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2165 lj->lioj_finished_count = 0;
2166 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2167 ksiginfo_init(&lj->lioj_ksi);
2172 if (sig && (mode == LIO_NOWAIT)) {
2173 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2174 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2175 /* Assume only new style KEVENT */
2176 kev.filter = EVFILT_LIO;
2177 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2178 kev.ident = (uintptr_t)uacb_list; /* something unique */
2179 kev.data = (intptr_t)lj;
2180 /* pass user defined sigval data */
2181 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2182 error = kqfd_register(
2183 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2185 uma_zfree(aiolio_zone, lj);
2188 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2190 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2191 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2192 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2193 uma_zfree(aiolio_zone, lj);
2196 lj->lioj_flags |= LIOJ_SIGNAL;
2198 uma_zfree(aiolio_zone, lj);
2204 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2206 * Add extra aiocb count to avoid the lio to be freed
2207 * by other threads doing aio_waitcomplete or aio_return,
2208 * and prevent event from being sent until we have queued
2215 * Get pointers to the list of I/O requests.
2218 for (i = 0; i < nent; i++) {
2221 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2229 if (mode == LIO_WAIT) {
2230 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2231 ki->kaio_flags |= KAIO_WAKEUP;
2232 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2233 PRIBIO | PCATCH, "aiospn", 0);
2234 if (error == ERESTART)
2240 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2241 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2242 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2243 KNOTE_LOCKED(&lj->klist, 1);
2245 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2247 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2248 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2249 aio_sendsig(p, &lj->lioj_signal,
2251 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2256 if (lj->lioj_count == 0) {
2257 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2258 knlist_delete(&lj->klist, curthread, 1);
2260 sigqueue_take(&lj->lioj_ksi);
2263 uma_zfree(aiolio_zone, lj);
2272 /* syscall - list directed I/O (REALTIME) */
2273 #ifdef COMPAT_FREEBSD6
2275 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2277 struct aiocb **acb_list;
2278 struct sigevent *sigp, sig;
2279 struct osigevent osig;
2282 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2286 if (nent < 0 || nent > AIO_LISTIO_MAX)
2289 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2290 error = copyin(uap->sig, &osig, sizeof(osig));
2293 error = convert_old_sigevent(&osig, &sig);
2300 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2301 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2303 error = kern_lio_listio(td, uap->mode,
2304 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2305 &aiocb_ops_osigevent);
2306 free(acb_list, M_LIO);
2311 /* syscall - list directed I/O (REALTIME) */
2313 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2315 struct aiocb **acb_list;
2316 struct sigevent *sigp, sig;
2319 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2323 if (nent < 0 || nent > AIO_LISTIO_MAX)
2326 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2327 error = copyin(uap->sig, &sig, sizeof(sig));
2334 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2335 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2337 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2338 nent, sigp, &aiocb_ops);
2339 free(acb_list, M_LIO);
2344 aio_physwakeup(struct bio *bp)
2346 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2348 struct kaioinfo *ki;
2352 /* Release mapping into kernel space. */
2353 userp = job->userproc;
2354 ki = userp->p_aioinfo;
2356 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2357 relpbuf(job->pbuf, NULL);
2359 atomic_subtract_int(&num_buf_aio, 1);
2361 ki->kaio_buffer_count--;
2364 vm_page_unhold_pages(job->pages, job->npages);
2368 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2370 if (bp->bio_flags & BIO_ERROR)
2371 error = bp->bio_error;
2372 nblks = btodb(nbytes);
2373 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2374 job->outblock += nblks;
2376 job->inblock += nblks;
2379 aio_complete(job, -1, error);
2381 aio_complete(job, nbytes, 0);
2386 /* syscall - wait for the next completion of an aio request */
2388 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2389 struct timespec *ts, struct aiocb_ops *ops)
2391 struct proc *p = td->td_proc;
2393 struct kaioinfo *ki;
2399 ops->store_aiocb(ujobp, NULL);
2403 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2406 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2409 TIMESPEC_TO_TIMEVAL(&atv, ts);
2410 if (itimerfix(&atv))
2412 timo = tvtohz(&atv);
2415 if (p->p_aioinfo == NULL)
2416 aio_init_aioinfo(p);
2422 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2424 error = EWOULDBLOCK;
2427 ki->kaio_flags |= KAIO_WAKEUP;
2428 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2430 if (timo && error == ERESTART)
2437 MPASS(job->jobflags & KAIOCB_FINISHED);
2439 status = job->uaiocb._aiocb_private.status;
2440 error = job->uaiocb._aiocb_private.error;
2441 td->td_retval[0] = status;
2442 td->td_ru.ru_oublock += job->outblock;
2443 td->td_ru.ru_inblock += job->inblock;
2444 td->td_ru.ru_msgsnd += job->msgsnd;
2445 td->td_ru.ru_msgrcv += job->msgrcv;
2446 aio_free_entry(job);
2448 ops->store_aiocb(ujobp, ujob);
2449 ops->store_error(ujob, error);
2450 ops->store_status(ujob, status);
2458 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2460 struct timespec ts, *tsp;
2464 /* Get timespec struct. */
2465 error = copyin(uap->timeout, &ts, sizeof(ts));
2472 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2476 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2477 struct aiocb_ops *ops)
2480 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2482 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2486 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2489 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2492 /* kqueue attach function */
2494 filt_aioattach(struct knote *kn)
2498 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2501 * The job pointer must be validated before using it, so
2502 * registration is restricted to the kernel; the user cannot
2505 if ((kn->kn_flags & EV_FLAG1) == 0)
2507 kn->kn_ptr.p_aio = job;
2508 kn->kn_flags &= ~EV_FLAG1;
2510 knlist_add(&job->klist, kn, 0);
2515 /* kqueue detach function */
2517 filt_aiodetach(struct knote *kn)
2521 knl = &kn->kn_ptr.p_aio->klist;
2522 knl->kl_lock(knl->kl_lockarg);
2523 if (!knlist_empty(knl))
2524 knlist_remove(knl, kn, 1);
2525 knl->kl_unlock(knl->kl_lockarg);
2528 /* kqueue filter function */
2531 filt_aio(struct knote *kn, long hint)
2533 struct kaiocb *job = kn->kn_ptr.p_aio;
2535 kn->kn_data = job->uaiocb._aiocb_private.error;
2536 if (!(job->jobflags & KAIOCB_FINISHED))
2538 kn->kn_flags |= EV_EOF;
2542 /* kqueue attach function */
2544 filt_lioattach(struct knote *kn)
2546 struct aioliojob *lj;
2548 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2551 * The aioliojob pointer must be validated before using it, so
2552 * registration is restricted to the kernel; the user cannot
2555 if ((kn->kn_flags & EV_FLAG1) == 0)
2557 kn->kn_ptr.p_lio = lj;
2558 kn->kn_flags &= ~EV_FLAG1;
2560 knlist_add(&lj->klist, kn, 0);
2565 /* kqueue detach function */
2567 filt_liodetach(struct knote *kn)
2571 knl = &kn->kn_ptr.p_lio->klist;
2572 knl->kl_lock(knl->kl_lockarg);
2573 if (!knlist_empty(knl))
2574 knlist_remove(knl, kn, 1);
2575 knl->kl_unlock(knl->kl_lockarg);
2578 /* kqueue filter function */
2581 filt_lio(struct knote *kn, long hint)
2583 struct aioliojob * lj = kn->kn_ptr.p_lio;
2585 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2588 #ifdef COMPAT_FREEBSD32
2589 #include <sys/mount.h>
2590 #include <sys/socket.h>
2591 #include <compat/freebsd32/freebsd32.h>
2592 #include <compat/freebsd32/freebsd32_proto.h>
2593 #include <compat/freebsd32/freebsd32_signal.h>
2594 #include <compat/freebsd32/freebsd32_syscall.h>
2595 #include <compat/freebsd32/freebsd32_util.h>
2597 struct __aiocb_private32 {
2600 uint32_t kernelinfo;
2603 #ifdef COMPAT_FREEBSD6
2604 typedef struct oaiocb32 {
2605 int aio_fildes; /* File descriptor */
2606 uint64_t aio_offset __packed; /* File offset for I/O */
2607 uint32_t aio_buf; /* I/O buffer in process space */
2608 uint32_t aio_nbytes; /* Number of bytes for I/O */
2609 struct osigevent32 aio_sigevent; /* Signal to deliver */
2610 int aio_lio_opcode; /* LIO opcode */
2611 int aio_reqprio; /* Request priority -- ignored */
2612 struct __aiocb_private32 _aiocb_private;
2616 typedef struct aiocb32 {
2617 int32_t aio_fildes; /* File descriptor */
2618 uint64_t aio_offset __packed; /* File offset for I/O */
2619 uint32_t aio_buf; /* I/O buffer in process space */
2620 uint32_t aio_nbytes; /* Number of bytes for I/O */
2622 uint32_t __spare2__;
2623 int aio_lio_opcode; /* LIO opcode */
2624 int aio_reqprio; /* Request priority -- ignored */
2625 struct __aiocb_private32 _aiocb_private;
2626 struct sigevent32 aio_sigevent; /* Signal to deliver */
2629 #ifdef COMPAT_FREEBSD6
2631 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2635 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2636 * supported by AIO with the old sigevent structure.
2638 CP(*osig, *nsig, sigev_notify);
2639 switch (nsig->sigev_notify) {
2643 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2646 nsig->sigev_notify_kqueue =
2647 osig->__sigev_u.__sigev_notify_kqueue;
2648 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2657 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2659 struct oaiocb32 job32;
2662 bzero(kjob, sizeof(struct aiocb));
2663 error = copyin(ujob, &job32, sizeof(job32));
2667 CP(job32, *kjob, aio_fildes);
2668 CP(job32, *kjob, aio_offset);
2669 PTRIN_CP(job32, *kjob, aio_buf);
2670 CP(job32, *kjob, aio_nbytes);
2671 CP(job32, *kjob, aio_lio_opcode);
2672 CP(job32, *kjob, aio_reqprio);
2673 CP(job32, *kjob, _aiocb_private.status);
2674 CP(job32, *kjob, _aiocb_private.error);
2675 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2676 return (convert_old_sigevent32(&job32.aio_sigevent,
2677 &kjob->aio_sigevent));
2682 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2684 struct aiocb32 job32;
2687 error = copyin(ujob, &job32, sizeof(job32));
2690 CP(job32, *kjob, aio_fildes);
2691 CP(job32, *kjob, aio_offset);
2692 PTRIN_CP(job32, *kjob, aio_buf);
2693 CP(job32, *kjob, aio_nbytes);
2694 CP(job32, *kjob, aio_lio_opcode);
2695 CP(job32, *kjob, aio_reqprio);
2696 CP(job32, *kjob, _aiocb_private.status);
2697 CP(job32, *kjob, _aiocb_private.error);
2698 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2699 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2703 aiocb32_fetch_status(struct aiocb *ujob)
2705 struct aiocb32 *ujob32;
2707 ujob32 = (struct aiocb32 *)ujob;
2708 return (fuword32(&ujob32->_aiocb_private.status));
2712 aiocb32_fetch_error(struct aiocb *ujob)
2714 struct aiocb32 *ujob32;
2716 ujob32 = (struct aiocb32 *)ujob;
2717 return (fuword32(&ujob32->_aiocb_private.error));
2721 aiocb32_store_status(struct aiocb *ujob, long status)
2723 struct aiocb32 *ujob32;
2725 ujob32 = (struct aiocb32 *)ujob;
2726 return (suword32(&ujob32->_aiocb_private.status, status));
2730 aiocb32_store_error(struct aiocb *ujob, long error)
2732 struct aiocb32 *ujob32;
2734 ujob32 = (struct aiocb32 *)ujob;
2735 return (suword32(&ujob32->_aiocb_private.error, error));
2739 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2741 struct aiocb32 *ujob32;
2743 ujob32 = (struct aiocb32 *)ujob;
2744 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2748 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2751 return (suword32(ujobp, (long)ujob));
2754 static struct aiocb_ops aiocb32_ops = {
2755 .copyin = aiocb32_copyin,
2756 .fetch_status = aiocb32_fetch_status,
2757 .fetch_error = aiocb32_fetch_error,
2758 .store_status = aiocb32_store_status,
2759 .store_error = aiocb32_store_error,
2760 .store_kernelinfo = aiocb32_store_kernelinfo,
2761 .store_aiocb = aiocb32_store_aiocb,
2764 #ifdef COMPAT_FREEBSD6
2765 static struct aiocb_ops aiocb32_ops_osigevent = {
2766 .copyin = aiocb32_copyin_old_sigevent,
2767 .fetch_status = aiocb32_fetch_status,
2768 .fetch_error = aiocb32_fetch_error,
2769 .store_status = aiocb32_store_status,
2770 .store_error = aiocb32_store_error,
2771 .store_kernelinfo = aiocb32_store_kernelinfo,
2772 .store_aiocb = aiocb32_store_aiocb,
2777 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2780 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2784 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2786 struct timespec32 ts32;
2787 struct timespec ts, *tsp;
2788 struct aiocb **ujoblist;
2789 uint32_t *ujoblist32;
2792 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2796 /* Get timespec struct. */
2797 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2799 CP(ts32, ts, tv_sec);
2800 CP(ts32, ts, tv_nsec);
2805 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2806 ujoblist32 = (uint32_t *)ujoblist;
2807 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2808 sizeof(ujoblist32[0]));
2810 for (i = uap->nent; i > 0; i--)
2811 ujoblist[i] = PTRIN(ujoblist32[i]);
2813 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2815 uma_zfree(aiol_zone, ujoblist);
2820 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2823 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2826 #ifdef COMPAT_FREEBSD6
2828 freebsd6_freebsd32_aio_read(struct thread *td,
2829 struct freebsd6_freebsd32_aio_read_args *uap)
2832 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2833 &aiocb32_ops_osigevent));
2838 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2841 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2845 #ifdef COMPAT_FREEBSD6
2847 freebsd6_freebsd32_aio_write(struct thread *td,
2848 struct freebsd6_freebsd32_aio_write_args *uap)
2851 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2852 &aiocb32_ops_osigevent));
2857 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2860 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2865 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2868 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2873 freebsd32_aio_waitcomplete(struct thread *td,
2874 struct freebsd32_aio_waitcomplete_args *uap)
2876 struct timespec32 ts32;
2877 struct timespec ts, *tsp;
2881 /* Get timespec struct. */
2882 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2885 CP(ts32, ts, tv_sec);
2886 CP(ts32, ts, tv_nsec);
2891 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2896 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2899 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2903 #ifdef COMPAT_FREEBSD6
2905 freebsd6_freebsd32_lio_listio(struct thread *td,
2906 struct freebsd6_freebsd32_lio_listio_args *uap)
2908 struct aiocb **acb_list;
2909 struct sigevent *sigp, sig;
2910 struct osigevent32 osig;
2911 uint32_t *acb_list32;
2914 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2918 if (nent < 0 || nent > AIO_LISTIO_MAX)
2921 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2922 error = copyin(uap->sig, &osig, sizeof(osig));
2925 error = convert_old_sigevent32(&osig, &sig);
2932 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2933 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2935 free(acb_list32, M_LIO);
2938 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2939 for (i = 0; i < nent; i++)
2940 acb_list[i] = PTRIN(acb_list32[i]);
2941 free(acb_list32, M_LIO);
2943 error = kern_lio_listio(td, uap->mode,
2944 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2945 &aiocb32_ops_osigevent);
2946 free(acb_list, M_LIO);
2952 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2954 struct aiocb **acb_list;
2955 struct sigevent *sigp, sig;
2956 struct sigevent32 sig32;
2957 uint32_t *acb_list32;
2960 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2964 if (nent < 0 || nent > AIO_LISTIO_MAX)
2967 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2968 error = copyin(uap->sig, &sig32, sizeof(sig32));
2971 error = convert_sigevent32(&sig32, &sig);
2978 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2979 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2981 free(acb_list32, M_LIO);
2984 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2985 for (i = 0; i < nent; i++)
2986 acb_list[i] = PTRIN(acb_list32[i]);
2987 free(acb_list32, M_LIO);
2989 error = kern_lio_listio(td, uap->mode,
2990 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2992 free(acb_list, M_LIO);