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 bool aio_clear_cancel_function_locked(struct kaiocb *job);
315 static int aio_kick(struct proc *userp);
316 static void aio_kick_nowait(struct proc *userp);
317 static void aio_kick_helper(void *context, int pending);
318 static int filt_aioattach(struct knote *kn);
319 static void filt_aiodetach(struct knote *kn);
320 static int filt_aio(struct knote *kn, long hint);
321 static int filt_lioattach(struct knote *kn);
322 static void filt_liodetach(struct knote *kn);
323 static int filt_lio(struct knote *kn, long hint);
327 * kaio Per process async io info
328 * aiop async io process data
329 * aiocb async io jobs
330 * aiol list io job pointer - internal to aio_suspend XXX
331 * aiolio list io jobs
333 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
335 /* kqueue filters for aio */
336 static struct filterops aio_filtops = {
338 .f_attach = filt_aioattach,
339 .f_detach = filt_aiodetach,
342 static struct filterops lio_filtops = {
344 .f_attach = filt_lioattach,
345 .f_detach = filt_liodetach,
349 static eventhandler_tag exit_tag, exec_tag;
351 TASKQUEUE_DEFINE_THREAD(aiod_kick);
354 * Main operations function for use as a kernel module.
357 aio_modload(struct module *module, int cmd, void *arg)
374 static moduledata_t aio_mod = {
380 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
381 MODULE_VERSION(aio, 1);
384 * Startup initialization
390 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
391 EVENTHANDLER_PRI_ANY);
392 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
393 NULL, EVENTHANDLER_PRI_ANY);
394 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
395 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
396 TAILQ_INIT(&aio_freeproc);
397 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
398 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
399 TAILQ_INIT(&aio_jobs);
400 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
401 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
402 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
403 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
404 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
405 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
406 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
407 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
408 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
409 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
410 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
411 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
413 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
414 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
415 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
416 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
422 * Init the per-process aioinfo structure. The aioinfo limits are set
423 * per-process for user limit (resource) management.
426 aio_init_aioinfo(struct proc *p)
430 ki = uma_zalloc(kaio_zone, M_WAITOK);
431 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
433 ki->kaio_maxactive_count = max_aio_per_proc;
434 ki->kaio_active_count = 0;
435 ki->kaio_qallowed_count = max_aio_queue_per_proc;
437 ki->kaio_ballowed_count = max_buf_aio;
438 ki->kaio_buffer_count = 0;
439 TAILQ_INIT(&ki->kaio_all);
440 TAILQ_INIT(&ki->kaio_done);
441 TAILQ_INIT(&ki->kaio_jobqueue);
442 TAILQ_INIT(&ki->kaio_liojoblist);
443 TAILQ_INIT(&ki->kaio_syncqueue);
444 TAILQ_INIT(&ki->kaio_syncready);
445 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
446 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
448 if (p->p_aioinfo == NULL) {
453 mtx_destroy(&ki->kaio_mtx);
454 uma_zfree(kaio_zone, ki);
457 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
462 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
467 error = sigev_findtd(p, sigev, &td);
471 ksiginfo_set_sigev(ksi, sigev);
472 ksi->ksi_code = SI_ASYNCIO;
473 ksi->ksi_flags |= KSI_EXT | KSI_INS;
474 tdsendsignal(p, td, ksi->ksi_signo, ksi);
481 * Free a job entry. Wait for completion if it is currently active, but don't
482 * delay forever. If we delay, we return a flag that says that we have to
483 * restart the queue scan.
486 aio_free_entry(struct kaiocb *job)
489 struct aioliojob *lj;
497 AIO_LOCK_ASSERT(ki, MA_OWNED);
498 MPASS(job->jobflags & KAIOCB_FINISHED);
500 atomic_subtract_int(&num_queue_count, 1);
503 MPASS(ki->kaio_count >= 0);
505 TAILQ_REMOVE(&ki->kaio_done, job, plist);
506 TAILQ_REMOVE(&ki->kaio_all, job, allist);
511 lj->lioj_finished_count--;
513 if (lj->lioj_count == 0) {
514 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
515 /* lio is going away, we need to destroy any knotes */
516 knlist_delete(&lj->klist, curthread, 1);
518 sigqueue_take(&lj->lioj_ksi);
520 uma_zfree(aiolio_zone, lj);
524 /* job is going away, we need to destroy any knotes */
525 knlist_delete(&job->klist, curthread, 1);
527 sigqueue_take(&job->ksi);
533 * The thread argument here is used to find the owning process
534 * and is also passed to fo_close() which may pass it to various
535 * places such as devsw close() routines. Because of that, we
536 * need a thread pointer from the process owning the job that is
537 * persistent and won't disappear out from under us or move to
540 * Currently, all the callers of this function call it to remove
541 * a kaiocb from the current process' job list either via a
542 * syscall or due to the current process calling exit() or
543 * execve(). Thus, we know that p == curproc. We also know that
544 * curthread can't exit since we are curthread.
546 * Therefore, we use curthread as the thread to pass to
547 * knlist_delete(). This does mean that it is possible for the
548 * thread pointer at close time to differ from the thread pointer
549 * at open time, but this is already true of file descriptors in
550 * a multithreaded process.
553 fdrop(job->fd_file, curthread);
555 uma_zfree(aiocb_zone, job);
562 aio_proc_rundown_exec(void *arg, struct proc *p,
563 struct image_params *imgp __unused)
565 aio_proc_rundown(arg, p);
569 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
571 aio_cancel_fn_t *func;
574 AIO_LOCK_ASSERT(ki, MA_OWNED);
575 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
577 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
578 job->jobflags |= KAIOCB_CANCELLED;
580 func = job->cancel_fn;
583 * If there is no cancel routine, just leave the job marked as
584 * cancelled. The job should be in active use by a caller who
585 * should complete it normally or when it fails to install a
592 * Set the CANCELLING flag so that aio_complete() will defer
593 * completions of this job. This prevents the job from being
594 * freed out from under the cancel callback. After the
595 * callback any deferred completion (whether from the callback
596 * or any other source) will be completed.
598 job->jobflags |= KAIOCB_CANCELLING;
602 job->jobflags &= ~KAIOCB_CANCELLING;
603 if (job->jobflags & KAIOCB_FINISHED) {
604 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
605 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
606 aio_bio_done_notify(p, job);
609 * The cancel callback might have scheduled an
610 * operation to cancel this request, but it is
611 * only counted as cancelled if the request is
612 * cancelled when the callback returns.
620 * Rundown the jobs for a given process.
623 aio_proc_rundown(void *arg, struct proc *p)
626 struct aioliojob *lj;
627 struct kaiocb *job, *jobn;
629 KASSERT(curthread->td_proc == p,
630 ("%s: called on non-curproc", __func__));
636 ki->kaio_flags |= KAIO_RUNDOWN;
641 * Try to cancel all pending requests. This code simulates
642 * aio_cancel on all pending I/O requests.
644 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
645 aio_cancel_job(p, ki, job);
648 /* Wait for all running I/O to be finished */
649 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
650 ki->kaio_flags |= KAIO_WAKEUP;
651 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
655 /* Free all completed I/O requests. */
656 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
659 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
660 if (lj->lioj_count == 0) {
661 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
662 knlist_delete(&lj->klist, curthread, 1);
664 sigqueue_take(&lj->lioj_ksi);
666 uma_zfree(aiolio_zone, lj);
668 panic("LIO job not cleaned up: C:%d, FC:%d\n",
669 lj->lioj_count, lj->lioj_finished_count);
673 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
674 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
675 mtx_destroy(&ki->kaio_mtx);
676 uma_zfree(kaio_zone, ki);
681 * Select a job to run (called by an AIO daemon).
683 static struct kaiocb *
684 aio_selectjob(struct aioproc *aiop)
690 mtx_assert(&aio_job_mtx, MA_OWNED);
692 TAILQ_FOREACH(job, &aio_jobs, list) {
693 userp = job->userproc;
694 ki = userp->p_aioinfo;
696 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
697 TAILQ_REMOVE(&aio_jobs, job, list);
698 if (!aio_clear_cancel_function(job))
701 /* Account for currently active jobs. */
702 ki->kaio_active_count++;
710 * Move all data to a permanent storage device. This code
711 * simulates the fsync syscall.
714 aio_fsync_vnode(struct thread *td, struct vnode *vp)
719 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
721 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
722 if (vp->v_object != NULL) {
723 VM_OBJECT_WLOCK(vp->v_object);
724 vm_object_page_clean(vp->v_object, 0, 0, 0);
725 VM_OBJECT_WUNLOCK(vp->v_object);
727 error = VOP_FSYNC(vp, MNT_WAIT, td);
730 vn_finished_write(mp);
736 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
737 * does the I/O request for the non-physio version of the operations. The
738 * normal vn operations are used, and this code should work in all instances
739 * for every type of file, including pipes, sockets, fifos, and regular files.
741 * XXX I don't think it works well for socket, pipe, and fifo.
744 aio_process_rw(struct kaiocb *job)
746 struct ucred *td_savedcred;
753 long msgsnd_st, msgsnd_end;
754 long msgrcv_st, msgrcv_end;
755 long oublock_st, oublock_end;
756 long inblock_st, inblock_end;
759 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
760 job->uaiocb.aio_lio_opcode == LIO_WRITE,
761 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
763 aio_switch_vmspace(job);
765 td_savedcred = td->td_ucred;
766 td->td_ucred = job->cred;
770 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
771 aiov.iov_len = cb->aio_nbytes;
773 auio.uio_iov = &aiov;
775 auio.uio_offset = cb->aio_offset;
776 auio.uio_resid = cb->aio_nbytes;
777 cnt = cb->aio_nbytes;
778 auio.uio_segflg = UIO_USERSPACE;
781 msgrcv_st = td->td_ru.ru_msgrcv;
782 msgsnd_st = td->td_ru.ru_msgsnd;
783 inblock_st = td->td_ru.ru_inblock;
784 oublock_st = td->td_ru.ru_oublock;
787 * aio_aqueue() acquires a reference to the file that is
788 * released in aio_free_entry().
790 if (cb->aio_lio_opcode == LIO_READ) {
791 auio.uio_rw = UIO_READ;
792 if (auio.uio_resid == 0)
795 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
797 if (fp->f_type == DTYPE_VNODE)
799 auio.uio_rw = UIO_WRITE;
800 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
802 msgrcv_end = td->td_ru.ru_msgrcv;
803 msgsnd_end = td->td_ru.ru_msgsnd;
804 inblock_end = td->td_ru.ru_inblock;
805 oublock_end = td->td_ru.ru_oublock;
807 job->msgrcv = msgrcv_end - msgrcv_st;
808 job->msgsnd = msgsnd_end - msgsnd_st;
809 job->inblock = inblock_end - inblock_st;
810 job->outblock = oublock_end - oublock_st;
812 if ((error) && (auio.uio_resid != cnt)) {
813 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
815 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
816 PROC_LOCK(job->userproc);
817 kern_psignal(job->userproc, SIGPIPE);
818 PROC_UNLOCK(job->userproc);
822 cnt -= auio.uio_resid;
823 td->td_ucred = td_savedcred;
825 aio_complete(job, -1, error);
827 aio_complete(job, cnt, 0);
831 aio_process_sync(struct kaiocb *job)
833 struct thread *td = curthread;
834 struct ucred *td_savedcred = td->td_ucred;
835 struct file *fp = job->fd_file;
838 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
839 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
841 td->td_ucred = job->cred;
842 if (fp->f_vnode != NULL)
843 error = aio_fsync_vnode(td, fp->f_vnode);
844 td->td_ucred = td_savedcred;
846 aio_complete(job, -1, error);
848 aio_complete(job, 0, 0);
852 aio_process_mlock(struct kaiocb *job)
854 struct aiocb *cb = &job->uaiocb;
857 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
858 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
860 aio_switch_vmspace(job);
861 error = vm_mlock(job->userproc, job->cred,
862 __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
864 aio_complete(job, -1, error);
866 aio_complete(job, 0, 0);
870 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
872 struct aioliojob *lj;
874 struct kaiocb *sjob, *sjobn;
878 ki = userp->p_aioinfo;
879 AIO_LOCK_ASSERT(ki, MA_OWNED);
883 lj->lioj_finished_count++;
884 if (lj->lioj_count == lj->lioj_finished_count)
887 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
888 MPASS(job->jobflags & KAIOCB_FINISHED);
890 if (ki->kaio_flags & KAIO_RUNDOWN)
891 goto notification_done;
893 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
894 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
895 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
897 KNOTE_LOCKED(&job->klist, 1);
900 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
901 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
902 KNOTE_LOCKED(&lj->klist, 1);
904 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
906 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
907 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
908 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
909 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
914 if (job->jobflags & KAIOCB_CHECKSYNC) {
915 schedule_fsync = false;
916 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
917 if (job->fd_file != sjob->fd_file ||
918 job->seqno >= sjob->seqno)
920 if (--sjob->pending > 0)
922 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
923 if (!aio_clear_cancel_function_locked(sjob))
925 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
926 schedule_fsync = true;
929 taskqueue_enqueue(taskqueue_aiod_kick,
930 &ki->kaio_sync_task);
932 if (ki->kaio_flags & KAIO_WAKEUP) {
933 ki->kaio_flags &= ~KAIO_WAKEUP;
934 wakeup(&userp->p_aioinfo);
939 aio_schedule_fsync(void *context, int pending)
946 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
947 job = TAILQ_FIRST(&ki->kaio_syncready);
948 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
950 aio_schedule(job, aio_process_sync);
957 aio_cancel_cleared(struct kaiocb *job)
962 * The caller should hold the same queue lock held when
963 * aio_clear_cancel_function() was called and set this flag
964 * ensuring this check sees an up-to-date value. However,
965 * there is no way to assert that.
967 ki = job->userproc->p_aioinfo;
968 return ((job->jobflags & KAIOCB_CLEARED) != 0);
972 aio_clear_cancel_function_locked(struct kaiocb *job)
975 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
976 MPASS(job->cancel_fn != NULL);
977 if (job->jobflags & KAIOCB_CANCELLING) {
978 job->jobflags |= KAIOCB_CLEARED;
981 job->cancel_fn = NULL;
986 aio_clear_cancel_function(struct kaiocb *job)
991 ki = job->userproc->p_aioinfo;
993 ret = aio_clear_cancel_function_locked(job);
999 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1002 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1003 if (job->jobflags & KAIOCB_CANCELLED)
1005 job->cancel_fn = func;
1010 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1012 struct kaioinfo *ki;
1015 ki = job->userproc->p_aioinfo;
1017 ret = aio_set_cancel_function_locked(job, func);
1023 aio_complete(struct kaiocb *job, long status, int error)
1025 struct kaioinfo *ki;
1028 job->uaiocb._aiocb_private.error = error;
1029 job->uaiocb._aiocb_private.status = status;
1031 userp = job->userproc;
1032 ki = userp->p_aioinfo;
1035 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1036 ("duplicate aio_complete"));
1037 job->jobflags |= KAIOCB_FINISHED;
1038 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1039 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1040 aio_bio_done_notify(userp, job);
1046 aio_cancel(struct kaiocb *job)
1049 aio_complete(job, -1, ECANCELED);
1053 aio_switch_vmspace(struct kaiocb *job)
1056 vmspace_switch_aio(job->userproc->p_vmspace);
1060 * The AIO daemon, most of the actual work is done in aio_process_*,
1061 * but the setup (and address space mgmt) is done in this routine.
1064 aio_daemon(void *_id)
1067 struct aioproc *aiop;
1068 struct kaioinfo *ki;
1070 struct vmspace *myvm;
1071 struct thread *td = curthread;
1072 int id = (intptr_t)_id;
1075 * Grab an extra reference on the daemon's vmspace so that it
1076 * doesn't get freed by jobs that switch to a different
1080 myvm = vmspace_acquire_ref(p);
1082 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1085 * Allocate and ready the aio control info. There is one aiop structure
1088 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1090 aiop->aioprocflags = 0;
1093 * Wakeup parent process. (Parent sleeps to keep from blasting away
1094 * and creating too many daemons.)
1096 sema_post(&aio_newproc_sem);
1098 mtx_lock(&aio_job_mtx);
1101 * Take daemon off of free queue
1103 if (aiop->aioprocflags & AIOP_FREE) {
1104 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1105 aiop->aioprocflags &= ~AIOP_FREE;
1111 while ((job = aio_selectjob(aiop)) != NULL) {
1112 mtx_unlock(&aio_job_mtx);
1114 ki = job->userproc->p_aioinfo;
1115 job->handle_fn(job);
1117 mtx_lock(&aio_job_mtx);
1118 /* Decrement the active job count. */
1119 ki->kaio_active_count--;
1123 * Disconnect from user address space.
1125 if (p->p_vmspace != myvm) {
1126 mtx_unlock(&aio_job_mtx);
1127 vmspace_switch_aio(myvm);
1128 mtx_lock(&aio_job_mtx);
1130 * We have to restart to avoid race, we only sleep if
1131 * no job can be selected.
1136 mtx_assert(&aio_job_mtx, MA_OWNED);
1138 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1139 aiop->aioprocflags |= AIOP_FREE;
1142 * If daemon is inactive for a long time, allow it to exit,
1143 * thereby freeing resources.
1145 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1146 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1147 (aiop->aioprocflags & AIOP_FREE) &&
1148 num_aio_procs > target_aio_procs)
1151 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1153 mtx_unlock(&aio_job_mtx);
1154 uma_zfree(aiop_zone, aiop);
1155 free_unr(aiod_unr, id);
1158 KASSERT(p->p_vmspace == myvm,
1159 ("AIOD: bad vmspace for exiting daemon"));
1160 KASSERT(myvm->vm_refcnt > 1,
1161 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1166 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1167 * AIO daemon modifies its environment itself.
1170 aio_newproc(int *start)
1176 id = alloc_unr(aiod_unr);
1177 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1178 RFNOWAIT, 0, "aiod%d", id);
1181 * Wait until daemon is started.
1183 sema_wait(&aio_newproc_sem);
1184 mtx_lock(&aio_job_mtx);
1188 mtx_unlock(&aio_job_mtx);
1190 free_unr(aiod_unr, id);
1196 * Try the high-performance, low-overhead physio method for eligible
1197 * VCHR devices. This method doesn't use an aio helper thread, and
1198 * thus has very low overhead.
1200 * Assumes that the caller, aio_aqueue(), has incremented the file
1201 * structure's reference count, preventing its deallocation for the
1202 * duration of this call.
1205 aio_qphysio(struct proc *p, struct kaiocb *job)
1214 struct kaioinfo *ki;
1215 int error, ref, poff;
1221 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1225 if (vp->v_type != VCHR)
1227 if (vp->v_bufobj.bo_bsize == 0)
1229 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1233 csw = devvn_refthread(vp, &dev, &ref);
1237 if ((csw->d_flags & D_DISK) == 0) {
1241 if (cb->aio_nbytes > dev->si_iosize_max) {
1247 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1248 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1249 if (cb->aio_nbytes > MAXPHYS) {
1256 if (cb->aio_nbytes > MAXPHYS - poff) {
1260 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1265 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1268 ki->kaio_buffer_count++;
1271 job->bp = bp = g_alloc_bio();
1273 bp->bio_length = cb->aio_nbytes;
1274 bp->bio_bcount = cb->aio_nbytes;
1275 bp->bio_done = aio_physwakeup;
1276 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1277 bp->bio_offset = cb->aio_offset;
1278 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1280 bp->bio_caller1 = (void *)job;
1282 prot = VM_PROT_READ;
1283 if (cb->aio_lio_opcode == LIO_READ)
1284 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1285 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1286 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1287 nitems(job->pages));
1288 if (job->npages < 0) {
1293 pmap_qenter((vm_offset_t)pbuf->b_data,
1294 job->pages, job->npages);
1295 bp->bio_data = pbuf->b_data + poff;
1296 atomic_add_int(&num_buf_aio, 1);
1298 bp->bio_ma = job->pages;
1299 bp->bio_ma_n = job->npages;
1300 bp->bio_ma_offset = poff;
1301 bp->bio_data = unmapped_buf;
1302 bp->bio_flags |= BIO_UNMAPPED;
1305 /* Perform transfer. */
1306 csw->d_strategy(bp);
1307 dev_relthread(dev, ref);
1313 ki->kaio_buffer_count--;
1315 relpbuf(pbuf, NULL);
1321 dev_relthread(dev, ref);
1325 #ifdef COMPAT_FREEBSD6
1327 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1331 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1332 * supported by AIO with the old sigevent structure.
1334 nsig->sigev_notify = osig->sigev_notify;
1335 switch (nsig->sigev_notify) {
1339 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1342 nsig->sigev_notify_kqueue =
1343 osig->__sigev_u.__sigev_notify_kqueue;
1344 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1353 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1355 struct oaiocb *ojob;
1358 bzero(kjob, sizeof(struct aiocb));
1359 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1362 ojob = (struct oaiocb *)kjob;
1363 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1368 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1371 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1375 aiocb_fetch_status(struct aiocb *ujob)
1378 return (fuword(&ujob->_aiocb_private.status));
1382 aiocb_fetch_error(struct aiocb *ujob)
1385 return (fuword(&ujob->_aiocb_private.error));
1389 aiocb_store_status(struct aiocb *ujob, long status)
1392 return (suword(&ujob->_aiocb_private.status, status));
1396 aiocb_store_error(struct aiocb *ujob, long error)
1399 return (suword(&ujob->_aiocb_private.error, error));
1403 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1406 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1410 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1413 return (suword(ujobp, (long)ujob));
1416 static struct aiocb_ops aiocb_ops = {
1417 .copyin = aiocb_copyin,
1418 .fetch_status = aiocb_fetch_status,
1419 .fetch_error = aiocb_fetch_error,
1420 .store_status = aiocb_store_status,
1421 .store_error = aiocb_store_error,
1422 .store_kernelinfo = aiocb_store_kernelinfo,
1423 .store_aiocb = aiocb_store_aiocb,
1426 #ifdef COMPAT_FREEBSD6
1427 static struct aiocb_ops aiocb_ops_osigevent = {
1428 .copyin = aiocb_copyin_old_sigevent,
1429 .fetch_status = aiocb_fetch_status,
1430 .fetch_error = aiocb_fetch_error,
1431 .store_status = aiocb_store_status,
1432 .store_error = aiocb_store_error,
1433 .store_kernelinfo = aiocb_store_kernelinfo,
1434 .store_aiocb = aiocb_store_aiocb,
1439 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1440 * technique is done in this code.
1443 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1444 int type, struct aiocb_ops *ops)
1446 struct proc *p = td->td_proc;
1447 cap_rights_t rights;
1450 struct kaioinfo *ki;
1458 if (p->p_aioinfo == NULL)
1459 aio_init_aioinfo(p);
1463 ops->store_status(ujob, -1);
1464 ops->store_error(ujob, 0);
1465 ops->store_kernelinfo(ujob, -1);
1467 if (num_queue_count >= max_queue_count ||
1468 ki->kaio_count >= ki->kaio_qallowed_count) {
1469 ops->store_error(ujob, EAGAIN);
1473 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1474 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1476 error = ops->copyin(ujob, &job->uaiocb);
1478 ops->store_error(ujob, error);
1479 uma_zfree(aiocb_zone, job);
1483 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1484 uma_zfree(aiocb_zone, job);
1488 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1489 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1490 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1491 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1492 ops->store_error(ujob, EINVAL);
1493 uma_zfree(aiocb_zone, job);
1497 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1498 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1499 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1500 uma_zfree(aiocb_zone, job);
1504 ksiginfo_init(&job->ksi);
1506 /* Save userspace address of the job info. */
1509 /* Get the opcode. */
1510 if (type != LIO_NOP)
1511 job->uaiocb.aio_lio_opcode = type;
1512 opcode = job->uaiocb.aio_lio_opcode;
1515 * Validate the opcode and fetch the file object for the specified
1518 * XXXRW: Moved the opcode validation up here so that we don't
1519 * retrieve a file descriptor without knowing what the capabiltity
1522 fd = job->uaiocb.aio_fildes;
1525 error = fget_write(td, fd,
1526 cap_rights_init(&rights, CAP_PWRITE), &fp);
1529 error = fget_read(td, fd,
1530 cap_rights_init(&rights, CAP_PREAD), &fp);
1533 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1539 error = fget(td, fd, cap_rights_init(&rights), &fp);
1545 uma_zfree(aiocb_zone, job);
1546 ops->store_error(ujob, error);
1550 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1555 if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
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)
2496 struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2499 * The job pointer must be validated before using it, so
2500 * registration is restricted to the kernel; the user cannot
2503 if ((kn->kn_flags & EV_FLAG1) == 0)
2505 kn->kn_ptr.p_aio = job;
2506 kn->kn_flags &= ~EV_FLAG1;
2508 knlist_add(&job->klist, kn, 0);
2513 /* kqueue detach function */
2515 filt_aiodetach(struct knote *kn)
2519 knl = &kn->kn_ptr.p_aio->klist;
2520 knl->kl_lock(knl->kl_lockarg);
2521 if (!knlist_empty(knl))
2522 knlist_remove(knl, kn, 1);
2523 knl->kl_unlock(knl->kl_lockarg);
2526 /* kqueue filter function */
2529 filt_aio(struct knote *kn, long hint)
2531 struct kaiocb *job = kn->kn_ptr.p_aio;
2533 kn->kn_data = job->uaiocb._aiocb_private.error;
2534 if (!(job->jobflags & KAIOCB_FINISHED))
2536 kn->kn_flags |= EV_EOF;
2540 /* kqueue attach function */
2542 filt_lioattach(struct knote *kn)
2544 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2547 * The aioliojob pointer must be validated before using it, so
2548 * registration is restricted to the kernel; the user cannot
2551 if ((kn->kn_flags & EV_FLAG1) == 0)
2553 kn->kn_ptr.p_lio = lj;
2554 kn->kn_flags &= ~EV_FLAG1;
2556 knlist_add(&lj->klist, kn, 0);
2561 /* kqueue detach function */
2563 filt_liodetach(struct knote *kn)
2567 knl = &kn->kn_ptr.p_lio->klist;
2568 knl->kl_lock(knl->kl_lockarg);
2569 if (!knlist_empty(knl))
2570 knlist_remove(knl, kn, 1);
2571 knl->kl_unlock(knl->kl_lockarg);
2574 /* kqueue filter function */
2577 filt_lio(struct knote *kn, long hint)
2579 struct aioliojob * lj = kn->kn_ptr.p_lio;
2581 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2584 #ifdef COMPAT_FREEBSD32
2585 #include <sys/mount.h>
2586 #include <sys/socket.h>
2587 #include <compat/freebsd32/freebsd32.h>
2588 #include <compat/freebsd32/freebsd32_proto.h>
2589 #include <compat/freebsd32/freebsd32_signal.h>
2590 #include <compat/freebsd32/freebsd32_syscall.h>
2591 #include <compat/freebsd32/freebsd32_util.h>
2593 struct __aiocb_private32 {
2596 uint32_t kernelinfo;
2599 #ifdef COMPAT_FREEBSD6
2600 typedef struct oaiocb32 {
2601 int aio_fildes; /* File descriptor */
2602 uint64_t aio_offset __packed; /* File offset for I/O */
2603 uint32_t aio_buf; /* I/O buffer in process space */
2604 uint32_t aio_nbytes; /* Number of bytes for I/O */
2605 struct osigevent32 aio_sigevent; /* Signal to deliver */
2606 int aio_lio_opcode; /* LIO opcode */
2607 int aio_reqprio; /* Request priority -- ignored */
2608 struct __aiocb_private32 _aiocb_private;
2612 typedef struct aiocb32 {
2613 int32_t aio_fildes; /* File descriptor */
2614 uint64_t aio_offset __packed; /* File offset for I/O */
2615 uint32_t aio_buf; /* I/O buffer in process space */
2616 uint32_t aio_nbytes; /* Number of bytes for I/O */
2618 uint32_t __spare2__;
2619 int aio_lio_opcode; /* LIO opcode */
2620 int aio_reqprio; /* Request priority -- ignored */
2621 struct __aiocb_private32 _aiocb_private;
2622 struct sigevent32 aio_sigevent; /* Signal to deliver */
2625 #ifdef COMPAT_FREEBSD6
2627 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2631 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2632 * supported by AIO with the old sigevent structure.
2634 CP(*osig, *nsig, sigev_notify);
2635 switch (nsig->sigev_notify) {
2639 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2642 nsig->sigev_notify_kqueue =
2643 osig->__sigev_u.__sigev_notify_kqueue;
2644 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2653 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2655 struct oaiocb32 job32;
2658 bzero(kjob, sizeof(struct aiocb));
2659 error = copyin(ujob, &job32, sizeof(job32));
2663 CP(job32, *kjob, aio_fildes);
2664 CP(job32, *kjob, aio_offset);
2665 PTRIN_CP(job32, *kjob, aio_buf);
2666 CP(job32, *kjob, aio_nbytes);
2667 CP(job32, *kjob, aio_lio_opcode);
2668 CP(job32, *kjob, aio_reqprio);
2669 CP(job32, *kjob, _aiocb_private.status);
2670 CP(job32, *kjob, _aiocb_private.error);
2671 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2672 return (convert_old_sigevent32(&job32.aio_sigevent,
2673 &kjob->aio_sigevent));
2678 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2680 struct aiocb32 job32;
2683 error = copyin(ujob, &job32, sizeof(job32));
2686 CP(job32, *kjob, aio_fildes);
2687 CP(job32, *kjob, aio_offset);
2688 PTRIN_CP(job32, *kjob, aio_buf);
2689 CP(job32, *kjob, aio_nbytes);
2690 CP(job32, *kjob, aio_lio_opcode);
2691 CP(job32, *kjob, aio_reqprio);
2692 CP(job32, *kjob, _aiocb_private.status);
2693 CP(job32, *kjob, _aiocb_private.error);
2694 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2695 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2699 aiocb32_fetch_status(struct aiocb *ujob)
2701 struct aiocb32 *ujob32;
2703 ujob32 = (struct aiocb32 *)ujob;
2704 return (fuword32(&ujob32->_aiocb_private.status));
2708 aiocb32_fetch_error(struct aiocb *ujob)
2710 struct aiocb32 *ujob32;
2712 ujob32 = (struct aiocb32 *)ujob;
2713 return (fuword32(&ujob32->_aiocb_private.error));
2717 aiocb32_store_status(struct aiocb *ujob, long status)
2719 struct aiocb32 *ujob32;
2721 ujob32 = (struct aiocb32 *)ujob;
2722 return (suword32(&ujob32->_aiocb_private.status, status));
2726 aiocb32_store_error(struct aiocb *ujob, long error)
2728 struct aiocb32 *ujob32;
2730 ujob32 = (struct aiocb32 *)ujob;
2731 return (suword32(&ujob32->_aiocb_private.error, error));
2735 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2737 struct aiocb32 *ujob32;
2739 ujob32 = (struct aiocb32 *)ujob;
2740 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2744 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2747 return (suword32(ujobp, (long)ujob));
2750 static struct aiocb_ops aiocb32_ops = {
2751 .copyin = aiocb32_copyin,
2752 .fetch_status = aiocb32_fetch_status,
2753 .fetch_error = aiocb32_fetch_error,
2754 .store_status = aiocb32_store_status,
2755 .store_error = aiocb32_store_error,
2756 .store_kernelinfo = aiocb32_store_kernelinfo,
2757 .store_aiocb = aiocb32_store_aiocb,
2760 #ifdef COMPAT_FREEBSD6
2761 static struct aiocb_ops aiocb32_ops_osigevent = {
2762 .copyin = aiocb32_copyin_old_sigevent,
2763 .fetch_status = aiocb32_fetch_status,
2764 .fetch_error = aiocb32_fetch_error,
2765 .store_status = aiocb32_store_status,
2766 .store_error = aiocb32_store_error,
2767 .store_kernelinfo = aiocb32_store_kernelinfo,
2768 .store_aiocb = aiocb32_store_aiocb,
2773 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2776 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2780 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2782 struct timespec32 ts32;
2783 struct timespec ts, *tsp;
2784 struct aiocb **ujoblist;
2785 uint32_t *ujoblist32;
2788 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2792 /* Get timespec struct. */
2793 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2795 CP(ts32, ts, tv_sec);
2796 CP(ts32, ts, tv_nsec);
2801 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2802 ujoblist32 = (uint32_t *)ujoblist;
2803 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2804 sizeof(ujoblist32[0]));
2806 for (i = uap->nent; i > 0; i--)
2807 ujoblist[i] = PTRIN(ujoblist32[i]);
2809 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2811 uma_zfree(aiol_zone, ujoblist);
2816 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2819 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2822 #ifdef COMPAT_FREEBSD6
2824 freebsd6_freebsd32_aio_read(struct thread *td,
2825 struct freebsd6_freebsd32_aio_read_args *uap)
2828 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2829 &aiocb32_ops_osigevent));
2834 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2837 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2841 #ifdef COMPAT_FREEBSD6
2843 freebsd6_freebsd32_aio_write(struct thread *td,
2844 struct freebsd6_freebsd32_aio_write_args *uap)
2847 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2848 &aiocb32_ops_osigevent));
2853 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2856 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2861 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2864 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2869 freebsd32_aio_waitcomplete(struct thread *td,
2870 struct freebsd32_aio_waitcomplete_args *uap)
2872 struct timespec32 ts32;
2873 struct timespec ts, *tsp;
2877 /* Get timespec struct. */
2878 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2881 CP(ts32, ts, tv_sec);
2882 CP(ts32, ts, tv_nsec);
2887 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2892 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2895 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2899 #ifdef COMPAT_FREEBSD6
2901 freebsd6_freebsd32_lio_listio(struct thread *td,
2902 struct freebsd6_freebsd32_lio_listio_args *uap)
2904 struct aiocb **acb_list;
2905 struct sigevent *sigp, sig;
2906 struct osigevent32 osig;
2907 uint32_t *acb_list32;
2910 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2914 if (nent < 0 || nent > AIO_LISTIO_MAX)
2917 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2918 error = copyin(uap->sig, &osig, sizeof(osig));
2921 error = convert_old_sigevent32(&osig, &sig);
2928 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2929 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2931 free(acb_list32, M_LIO);
2934 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2935 for (i = 0; i < nent; i++)
2936 acb_list[i] = PTRIN(acb_list32[i]);
2937 free(acb_list32, M_LIO);
2939 error = kern_lio_listio(td, uap->mode,
2940 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2941 &aiocb32_ops_osigevent);
2942 free(acb_list, M_LIO);
2948 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2950 struct aiocb **acb_list;
2951 struct sigevent *sigp, sig;
2952 struct sigevent32 sig32;
2953 uint32_t *acb_list32;
2956 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2960 if (nent < 0 || nent > AIO_LISTIO_MAX)
2963 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2964 error = copyin(uap->sig, &sig32, sizeof(sig32));
2967 error = convert_sigevent32(&sig32, &sig);
2974 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2975 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2977 free(acb_list32, M_LIO);
2980 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2981 for (i = 0; i < nent; i++)
2982 acb_list[i] = PTRIN(acb_list32[i]);
2983 free(acb_list32, M_LIO);
2985 error = kern_lio_listio(td, uap->mode,
2986 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2988 free(acb_list, M_LIO);