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>
57 #include <sys/taskqueue.h>
58 #include <sys/vnode.h>
60 #include <sys/event.h>
61 #include <sys/mount.h>
62 #include <geom/geom.h>
64 #include <machine/atomic.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_extern.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_object.h>
76 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
77 * overflow. (XXX will be removed soon.)
79 static u_long jobrefid;
82 * Counter for aio_fsync.
84 static uint64_t jobseqno;
86 #ifndef MAX_AIO_PER_PROC
87 #define MAX_AIO_PER_PROC 32
90 #ifndef MAX_AIO_QUEUE_PER_PROC
91 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
95 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
99 #define MAX_BUF_AIO 16
102 FEATURE(aio, "Asynchronous I/O");
104 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
106 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
107 "Async IO management");
109 static int enable_aio_unsafe = 0;
110 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
111 "Permit asynchronous IO on all file types, not just known-safe types");
113 static int max_aio_procs = MAX_AIO_PROCS;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
115 "Maximum number of kernel processes to use for handling async IO ");
117 static int num_aio_procs = 0;
118 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
119 "Number of presently active kernel processes for async IO");
122 * The code will adjust the actual number of AIO processes towards this
123 * number when it gets a chance.
125 static int target_aio_procs = TARGET_AIO_PROCS;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
128 "Preferred number of ready kernel processes for async IO");
130 static int max_queue_count = MAX_AIO_QUEUE;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
132 "Maximum number of aio requests to queue, globally");
134 static int num_queue_count = 0;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
136 "Number of queued aio requests");
138 static int num_buf_aio = 0;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
140 "Number of aio requests presently handled by the buf subsystem");
142 /* Number of async I/O processes in the process of being started */
143 /* XXX This should be local to aio_aqueue() */
144 static int num_aio_resv_start = 0;
146 static int aiod_lifetime;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
148 "Maximum lifetime for idle aiod");
150 static int max_aio_per_proc = MAX_AIO_PER_PROC;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
153 "Maximum active aio requests per process (stored in the process)");
155 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
157 &max_aio_queue_per_proc, 0,
158 "Maximum queued aio requests per process (stored in the process)");
160 static int max_buf_aio = MAX_BUF_AIO;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
162 "Maximum buf aio requests per process (stored in the process)");
164 #ifdef COMPAT_FREEBSD6
165 typedef struct oaiocb {
166 int aio_fildes; /* File descriptor */
167 off_t aio_offset; /* File offset for I/O */
168 volatile void *aio_buf; /* I/O buffer in process space */
169 size_t aio_nbytes; /* Number of bytes for I/O */
170 struct osigevent aio_sigevent; /* Signal to deliver */
171 int aio_lio_opcode; /* LIO opcode */
172 int aio_reqprio; /* Request priority -- ignored */
173 struct __aiocb_private _aiocb_private;
178 * Below is a key of locks used to protect each member of struct kaiocb
179 * aioliojob and kaioinfo and any backends.
181 * * - need not protected
182 * a - locked by kaioinfo lock
183 * b - locked by backend lock, the backend lock can be null in some cases,
184 * for example, BIO belongs to this type, in this case, proc lock is
186 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
190 * If the routine that services an AIO request blocks while running in an
191 * AIO kernel process it can starve other I/O requests. BIO requests
192 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
193 * processes at all. Socket I/O requests use a separate pool of
194 * kprocs and also force non-blocking I/O. Other file I/O requests
195 * use the generic fo_read/fo_write operations which can block. The
196 * fsync and mlock operations can also block while executing. Ideally
197 * none of these requests would block while executing.
199 * Note that the service routines cannot toggle O_NONBLOCK in the file
200 * structure directly while handling a request due to races with
205 #define KAIOCB_QUEUEING 0x01
206 #define KAIOCB_CANCELLED 0x02
207 #define KAIOCB_CANCELLING 0x04
208 #define KAIOCB_CHECKSYNC 0x08
209 #define KAIOCB_CLEARED 0x10
210 #define KAIOCB_FINISHED 0x20
215 #define AIOP_FREE 0x1 /* proc on free queue */
218 int aioprocflags; /* (c) AIO proc flags */
219 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
220 struct proc *aioproc; /* (*) the AIO proc */
224 * data-structure for lio signal management
227 int lioj_flags; /* (a) listio flags */
228 int lioj_count; /* (a) listio flags */
229 int lioj_finished_count; /* (a) listio flags */
230 struct sigevent lioj_signal; /* (a) signal on all I/O done */
231 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
232 struct knlist klist; /* (a) list of knotes */
233 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
236 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
237 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
238 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
241 * per process aio data structure
244 struct mtx kaio_mtx; /* the lock to protect this struct */
245 int kaio_flags; /* (a) per process kaio flags */
246 int kaio_maxactive_count; /* (*) maximum number of AIOs */
247 int kaio_active_count; /* (c) number of currently used AIOs */
248 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
249 int kaio_count; /* (a) size of AIO queue */
250 int kaio_ballowed_count; /* (*) maximum number of buffers */
251 int kaio_buffer_count; /* (a) number of physio buffers */
252 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */
253 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */
254 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
255 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */
256 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */
257 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */
258 struct task kaio_task; /* (*) task to kick aio processes */
259 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */
262 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
263 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
264 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
265 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
267 #define KAIO_RUNDOWN 0x1 /* process is being run down */
268 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */
271 * Operations used to interact with userland aio control blocks.
272 * Different ABIs provide their own operations.
275 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
276 long (*fetch_status)(struct aiocb *ujob);
277 long (*fetch_error)(struct aiocb *ujob);
278 int (*store_status)(struct aiocb *ujob, long status);
279 int (*store_error)(struct aiocb *ujob, long error);
280 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
281 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
284 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */
285 static struct sema aio_newproc_sem;
286 static struct mtx aio_job_mtx;
287 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */
288 static struct unrhdr *aiod_unr;
290 void aio_init_aioinfo(struct proc *p);
291 static int aio_onceonly(void);
292 static int aio_free_entry(struct kaiocb *job);
293 static void aio_process_rw(struct kaiocb *job);
294 static void aio_process_sync(struct kaiocb *job);
295 static void aio_process_mlock(struct kaiocb *job);
296 static void aio_schedule_fsync(void *context, int pending);
297 static int aio_newproc(int *);
298 int aio_aqueue(struct thread *td, struct aiocb *ujob,
299 struct aioliojob *lio, int type, struct aiocb_ops *ops);
300 static int aio_queue_file(struct file *fp, struct kaiocb *job);
301 static void aio_physwakeup(struct bio *bp);
302 static void aio_proc_rundown(void *arg, struct proc *p);
303 static void aio_proc_rundown_exec(void *arg, struct proc *p,
304 struct image_params *imgp);
305 static int aio_qphysio(struct proc *p, struct kaiocb *job);
306 static void aio_daemon(void *param);
307 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
308 static int aio_kick(struct proc *userp);
309 static void aio_kick_nowait(struct proc *userp);
310 static void aio_kick_helper(void *context, int pending);
311 static int filt_aioattach(struct knote *kn);
312 static void filt_aiodetach(struct knote *kn);
313 static int filt_aio(struct knote *kn, long hint);
314 static int filt_lioattach(struct knote *kn);
315 static void filt_liodetach(struct knote *kn);
316 static int filt_lio(struct knote *kn, long hint);
320 * kaio Per process async io info
321 * aiop async io process data
322 * aiocb async io jobs
323 * aiol list io job pointer - internal to aio_suspend XXX
324 * aiolio list io jobs
326 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
328 /* kqueue filters for aio */
329 static struct filterops aio_filtops = {
331 .f_attach = filt_aioattach,
332 .f_detach = filt_aiodetach,
335 static struct filterops lio_filtops = {
337 .f_attach = filt_lioattach,
338 .f_detach = filt_liodetach,
342 static eventhandler_tag exit_tag, exec_tag;
344 TASKQUEUE_DEFINE_THREAD(aiod_kick);
347 * Main operations function for use as a kernel module.
350 aio_modload(struct module *module, int cmd, void *arg)
367 static moduledata_t aio_mod = {
373 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
374 MODULE_VERSION(aio, 1);
377 * Startup initialization
383 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
384 EVENTHANDLER_PRI_ANY);
385 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
386 NULL, EVENTHANDLER_PRI_ANY);
387 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
388 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
389 TAILQ_INIT(&aio_freeproc);
390 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
391 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
392 TAILQ_INIT(&aio_jobs);
393 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
394 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
395 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
396 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
397 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
398 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
399 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
400 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
401 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
402 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
403 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
404 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
406 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
407 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
408 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
409 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
415 * Init the per-process aioinfo structure. The aioinfo limits are set
416 * per-process for user limit (resource) management.
419 aio_init_aioinfo(struct proc *p)
423 ki = uma_zalloc(kaio_zone, M_WAITOK);
424 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
426 ki->kaio_maxactive_count = max_aio_per_proc;
427 ki->kaio_active_count = 0;
428 ki->kaio_qallowed_count = max_aio_queue_per_proc;
430 ki->kaio_ballowed_count = max_buf_aio;
431 ki->kaio_buffer_count = 0;
432 TAILQ_INIT(&ki->kaio_all);
433 TAILQ_INIT(&ki->kaio_done);
434 TAILQ_INIT(&ki->kaio_jobqueue);
435 TAILQ_INIT(&ki->kaio_liojoblist);
436 TAILQ_INIT(&ki->kaio_syncqueue);
437 TAILQ_INIT(&ki->kaio_syncready);
438 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
439 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
441 if (p->p_aioinfo == NULL) {
446 mtx_destroy(&ki->kaio_mtx);
447 uma_zfree(kaio_zone, ki);
450 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
455 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
460 error = sigev_findtd(p, sigev, &td);
464 ksiginfo_set_sigev(ksi, sigev);
465 ksi->ksi_code = SI_ASYNCIO;
466 ksi->ksi_flags |= KSI_EXT | KSI_INS;
467 tdsendsignal(p, td, ksi->ksi_signo, ksi);
474 * Free a job entry. Wait for completion if it is currently active, but don't
475 * delay forever. If we delay, we return a flag that says that we have to
476 * restart the queue scan.
479 aio_free_entry(struct kaiocb *job)
482 struct aioliojob *lj;
490 AIO_LOCK_ASSERT(ki, MA_OWNED);
491 MPASS(job->jobflags & KAIOCB_FINISHED);
493 atomic_subtract_int(&num_queue_count, 1);
496 MPASS(ki->kaio_count >= 0);
498 TAILQ_REMOVE(&ki->kaio_done, job, plist);
499 TAILQ_REMOVE(&ki->kaio_all, job, allist);
504 lj->lioj_finished_count--;
506 if (lj->lioj_count == 0) {
507 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
508 /* lio is going away, we need to destroy any knotes */
509 knlist_delete(&lj->klist, curthread, 1);
511 sigqueue_take(&lj->lioj_ksi);
513 uma_zfree(aiolio_zone, lj);
517 /* job is going away, we need to destroy any knotes */
518 knlist_delete(&job->klist, curthread, 1);
520 sigqueue_take(&job->ksi);
523 MPASS(job->bp == NULL);
527 * The thread argument here is used to find the owning process
528 * and is also passed to fo_close() which may pass it to various
529 * places such as devsw close() routines. Because of that, we
530 * need a thread pointer from the process owning the job that is
531 * persistent and won't disappear out from under us or move to
534 * Currently, all the callers of this function call it to remove
535 * a kaiocb from the current process' job list either via a
536 * syscall or due to the current process calling exit() or
537 * execve(). Thus, we know that p == curproc. We also know that
538 * curthread can't exit since we are curthread.
540 * Therefore, we use curthread as the thread to pass to
541 * knlist_delete(). This does mean that it is possible for the
542 * thread pointer at close time to differ from the thread pointer
543 * at open time, but this is already true of file descriptors in
544 * a multithreaded process.
547 fdrop(job->fd_file, curthread);
549 uma_zfree(aiocb_zone, job);
556 aio_proc_rundown_exec(void *arg, struct proc *p,
557 struct image_params *imgp __unused)
559 aio_proc_rundown(arg, p);
563 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
565 aio_cancel_fn_t *func;
568 AIO_LOCK_ASSERT(ki, MA_OWNED);
569 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
571 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
572 job->jobflags |= KAIOCB_CANCELLED;
574 func = job->cancel_fn;
577 * If there is no cancel routine, just leave the job marked as
578 * cancelled. The job should be in active use by a caller who
579 * should complete it normally or when it fails to install a
586 * Set the CANCELLING flag so that aio_complete() will defer
587 * completions of this job. This prevents the job from being
588 * freed out from under the cancel callback. After the
589 * callback any deferred completion (whether from the callback
590 * or any other source) will be completed.
592 job->jobflags |= KAIOCB_CANCELLING;
596 job->jobflags &= ~KAIOCB_CANCELLING;
597 if (job->jobflags & KAIOCB_FINISHED) {
598 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
599 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
600 aio_bio_done_notify(p, job);
603 * The cancel callback might have scheduled an
604 * operation to cancel this request, but it is
605 * only counted as cancelled if the request is
606 * cancelled when the callback returns.
614 * Rundown the jobs for a given process.
617 aio_proc_rundown(void *arg, struct proc *p)
620 struct aioliojob *lj;
621 struct kaiocb *job, *jobn;
623 KASSERT(curthread->td_proc == p,
624 ("%s: called on non-curproc", __func__));
630 ki->kaio_flags |= KAIO_RUNDOWN;
635 * Try to cancel all pending requests. This code simulates
636 * aio_cancel on all pending I/O requests.
638 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
639 aio_cancel_job(p, ki, job);
642 /* Wait for all running I/O to be finished */
643 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
644 ki->kaio_flags |= KAIO_WAKEUP;
645 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
649 /* Free all completed I/O requests. */
650 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
653 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
654 if (lj->lioj_count == 0) {
655 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
656 knlist_delete(&lj->klist, curthread, 1);
658 sigqueue_take(&lj->lioj_ksi);
660 uma_zfree(aiolio_zone, lj);
662 panic("LIO job not cleaned up: C:%d, FC:%d\n",
663 lj->lioj_count, lj->lioj_finished_count);
667 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
668 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
669 mtx_destroy(&ki->kaio_mtx);
670 uma_zfree(kaio_zone, ki);
675 * Select a job to run (called by an AIO daemon).
677 static struct kaiocb *
678 aio_selectjob(struct aioproc *aiop)
684 mtx_assert(&aio_job_mtx, MA_OWNED);
686 TAILQ_FOREACH(job, &aio_jobs, list) {
687 userp = job->userproc;
688 ki = userp->p_aioinfo;
690 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
691 TAILQ_REMOVE(&aio_jobs, job, list);
692 if (!aio_clear_cancel_function(job))
695 /* Account for currently active jobs. */
696 ki->kaio_active_count++;
704 * Move all data to a permanent storage device. This code
705 * simulates the fsync syscall.
708 aio_fsync_vnode(struct thread *td, struct vnode *vp)
713 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
715 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
716 if (vp->v_object != NULL) {
717 VM_OBJECT_WLOCK(vp->v_object);
718 vm_object_page_clean(vp->v_object, 0, 0, 0);
719 VM_OBJECT_WUNLOCK(vp->v_object);
721 error = VOP_FSYNC(vp, MNT_WAIT, td);
724 vn_finished_write(mp);
730 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
731 * does the I/O request for the non-physio version of the operations. The
732 * normal vn operations are used, and this code should work in all instances
733 * for every type of file, including pipes, sockets, fifos, and regular files.
735 * XXX I don't think it works well for socket, pipe, and fifo.
738 aio_process_rw(struct kaiocb *job)
740 struct ucred *td_savedcred;
748 int oublock_st, oublock_end;
749 int inblock_st, inblock_end;
751 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
752 job->uaiocb.aio_lio_opcode == LIO_WRITE,
753 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
755 aio_switch_vmspace(job);
757 td_savedcred = td->td_ucred;
758 td->td_ucred = job->cred;
762 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
763 aiov.iov_len = cb->aio_nbytes;
765 auio.uio_iov = &aiov;
767 auio.uio_offset = cb->aio_offset;
768 auio.uio_resid = cb->aio_nbytes;
769 cnt = cb->aio_nbytes;
770 auio.uio_segflg = UIO_USERSPACE;
773 inblock_st = td->td_ru.ru_inblock;
774 oublock_st = td->td_ru.ru_oublock;
776 * aio_aqueue() acquires a reference to the file that is
777 * released in aio_free_entry().
779 if (cb->aio_lio_opcode == LIO_READ) {
780 auio.uio_rw = UIO_READ;
781 if (auio.uio_resid == 0)
784 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
786 if (fp->f_type == DTYPE_VNODE)
788 auio.uio_rw = UIO_WRITE;
789 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
791 inblock_end = td->td_ru.ru_inblock;
792 oublock_end = td->td_ru.ru_oublock;
794 job->inputcharge = inblock_end - inblock_st;
795 job->outputcharge = oublock_end - oublock_st;
797 if ((error) && (auio.uio_resid != cnt)) {
798 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
800 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
801 PROC_LOCK(job->userproc);
802 kern_psignal(job->userproc, SIGPIPE);
803 PROC_UNLOCK(job->userproc);
807 cnt -= auio.uio_resid;
808 td->td_ucred = td_savedcred;
809 aio_complete(job, cnt, error);
813 aio_process_sync(struct kaiocb *job)
815 struct thread *td = curthread;
816 struct ucred *td_savedcred = td->td_ucred;
817 struct file *fp = job->fd_file;
820 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
821 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
823 td->td_ucred = job->cred;
824 if (fp->f_vnode != NULL)
825 error = aio_fsync_vnode(td, fp->f_vnode);
826 td->td_ucred = td_savedcred;
827 aio_complete(job, 0, error);
831 aio_process_mlock(struct kaiocb *job)
833 struct aiocb *cb = &job->uaiocb;
836 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
837 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
839 aio_switch_vmspace(job);
840 error = vm_mlock(job->userproc, job->cred,
841 __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
842 aio_complete(job, 0, error);
846 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
848 struct aioliojob *lj;
850 struct kaiocb *sjob, *sjobn;
854 ki = userp->p_aioinfo;
855 AIO_LOCK_ASSERT(ki, MA_OWNED);
859 lj->lioj_finished_count++;
860 if (lj->lioj_count == lj->lioj_finished_count)
863 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
864 MPASS(job->jobflags & KAIOCB_FINISHED);
866 if (ki->kaio_flags & KAIO_RUNDOWN)
867 goto notification_done;
869 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
870 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
871 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
873 KNOTE_LOCKED(&job->klist, 1);
876 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
877 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
878 KNOTE_LOCKED(&lj->klist, 1);
880 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
882 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
883 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
884 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
885 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
890 if (job->jobflags & KAIOCB_CHECKSYNC) {
891 schedule_fsync = false;
892 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
893 if (job->fd_file == sjob->fd_file &&
894 job->seqno < sjob->seqno) {
895 if (--sjob->pending == 0) {
896 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob,
898 if (!aio_clear_cancel_function(sjob))
900 TAILQ_INSERT_TAIL(&ki->kaio_syncready,
902 schedule_fsync = true;
907 taskqueue_enqueue(taskqueue_aiod_kick,
908 &ki->kaio_sync_task);
910 if (ki->kaio_flags & KAIO_WAKEUP) {
911 ki->kaio_flags &= ~KAIO_WAKEUP;
912 wakeup(&userp->p_aioinfo);
917 aio_schedule_fsync(void *context, int pending)
924 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
925 job = TAILQ_FIRST(&ki->kaio_syncready);
926 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
928 aio_schedule(job, aio_process_sync);
935 aio_cancel_cleared(struct kaiocb *job)
940 * The caller should hold the same queue lock held when
941 * aio_clear_cancel_function() was called and set this flag
942 * ensuring this check sees an up-to-date value. However,
943 * there is no way to assert that.
945 ki = job->userproc->p_aioinfo;
946 return ((job->jobflags & KAIOCB_CLEARED) != 0);
950 aio_clear_cancel_function(struct kaiocb *job)
954 ki = job->userproc->p_aioinfo;
956 MPASS(job->cancel_fn != NULL);
957 if (job->jobflags & KAIOCB_CANCELLING) {
958 job->jobflags |= KAIOCB_CLEARED;
962 job->cancel_fn = NULL;
968 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
972 ki = job->userproc->p_aioinfo;
974 if (job->jobflags & KAIOCB_CANCELLED) {
978 job->cancel_fn = func;
984 aio_complete(struct kaiocb *job, long status, int error)
989 job->uaiocb._aiocb_private.error = error;
990 job->uaiocb._aiocb_private.status = status;
992 userp = job->userproc;
993 ki = userp->p_aioinfo;
996 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
997 ("duplicate aio_complete"));
998 job->jobflags |= KAIOCB_FINISHED;
999 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1000 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1001 aio_bio_done_notify(userp, job);
1007 aio_cancel(struct kaiocb *job)
1010 aio_complete(job, -1, ECANCELED);
1014 aio_switch_vmspace(struct kaiocb *job)
1017 vmspace_switch_aio(job->userproc->p_vmspace);
1021 * The AIO daemon, most of the actual work is done in aio_process_*,
1022 * but the setup (and address space mgmt) is done in this routine.
1025 aio_daemon(void *_id)
1028 struct aioproc *aiop;
1029 struct kaioinfo *ki;
1031 struct vmspace *myvm;
1032 struct thread *td = curthread;
1033 int id = (intptr_t)_id;
1036 * Grab an extra reference on the daemon's vmspace so that it
1037 * doesn't get freed by jobs that switch to a different
1041 myvm = vmspace_acquire_ref(p);
1043 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1046 * Allocate and ready the aio control info. There is one aiop structure
1049 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1051 aiop->aioprocflags = 0;
1054 * Wakeup parent process. (Parent sleeps to keep from blasting away
1055 * and creating too many daemons.)
1057 sema_post(&aio_newproc_sem);
1059 mtx_lock(&aio_job_mtx);
1062 * Take daemon off of free queue
1064 if (aiop->aioprocflags & AIOP_FREE) {
1065 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1066 aiop->aioprocflags &= ~AIOP_FREE;
1072 while ((job = aio_selectjob(aiop)) != NULL) {
1073 mtx_unlock(&aio_job_mtx);
1075 ki = job->userproc->p_aioinfo;
1076 job->handle_fn(job);
1078 mtx_lock(&aio_job_mtx);
1079 /* Decrement the active job count. */
1080 ki->kaio_active_count--;
1084 * Disconnect from user address space.
1086 if (p->p_vmspace != myvm) {
1087 mtx_unlock(&aio_job_mtx);
1088 vmspace_switch_aio(myvm);
1089 mtx_lock(&aio_job_mtx);
1091 * We have to restart to avoid race, we only sleep if
1092 * no job can be selected.
1097 mtx_assert(&aio_job_mtx, MA_OWNED);
1099 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1100 aiop->aioprocflags |= AIOP_FREE;
1103 * If daemon is inactive for a long time, allow it to exit,
1104 * thereby freeing resources.
1106 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1107 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1108 (aiop->aioprocflags & AIOP_FREE) &&
1109 num_aio_procs > target_aio_procs)
1112 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1114 mtx_unlock(&aio_job_mtx);
1115 uma_zfree(aiop_zone, aiop);
1116 free_unr(aiod_unr, id);
1119 KASSERT(p->p_vmspace == myvm,
1120 ("AIOD: bad vmspace for exiting daemon"));
1121 KASSERT(myvm->vm_refcnt > 1,
1122 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1127 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1128 * AIO daemon modifies its environment itself.
1131 aio_newproc(int *start)
1137 id = alloc_unr(aiod_unr);
1138 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1139 RFNOWAIT, 0, "aiod%d", id);
1142 * Wait until daemon is started.
1144 sema_wait(&aio_newproc_sem);
1145 mtx_lock(&aio_job_mtx);
1149 mtx_unlock(&aio_job_mtx);
1151 free_unr(aiod_unr, id);
1157 * Try the high-performance, low-overhead physio method for eligible
1158 * VCHR devices. This method doesn't use an aio helper thread, and
1159 * thus has very low overhead.
1161 * Assumes that the caller, aio_aqueue(), has incremented the file
1162 * structure's reference count, preventing its deallocation for the
1163 * duration of this call.
1166 aio_qphysio(struct proc *p, struct kaiocb *job)
1175 struct kaioinfo *ki;
1176 int error, ref, poff;
1182 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1186 if (vp->v_type != VCHR)
1188 if (vp->v_bufobj.bo_bsize == 0)
1190 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1194 csw = devvn_refthread(vp, &dev, &ref);
1198 if ((csw->d_flags & D_DISK) == 0) {
1202 if (cb->aio_nbytes > dev->si_iosize_max) {
1208 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1209 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1210 if (cb->aio_nbytes > MAXPHYS) {
1217 if (cb->aio_nbytes > MAXPHYS - poff) {
1221 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1226 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1229 ki->kaio_buffer_count++;
1232 job->bp = bp = g_alloc_bio();
1234 bp->bio_length = cb->aio_nbytes;
1235 bp->bio_bcount = cb->aio_nbytes;
1236 bp->bio_done = aio_physwakeup;
1237 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1238 bp->bio_offset = cb->aio_offset;
1239 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1241 bp->bio_caller1 = (void *)job;
1243 prot = VM_PROT_READ;
1244 if (cb->aio_lio_opcode == LIO_READ)
1245 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1246 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1247 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1248 nitems(job->pages));
1249 if (job->npages < 0) {
1254 pmap_qenter((vm_offset_t)pbuf->b_data,
1255 job->pages, job->npages);
1256 bp->bio_data = pbuf->b_data + poff;
1257 atomic_add_int(&num_buf_aio, 1);
1259 bp->bio_ma = job->pages;
1260 bp->bio_ma_n = job->npages;
1261 bp->bio_ma_offset = poff;
1262 bp->bio_data = unmapped_buf;
1263 bp->bio_flags |= BIO_UNMAPPED;
1266 /* Perform transfer. */
1267 csw->d_strategy(bp);
1268 dev_relthread(dev, ref);
1274 ki->kaio_buffer_count--;
1276 relpbuf(pbuf, NULL);
1282 dev_relthread(dev, ref);
1286 #ifdef COMPAT_FREEBSD6
1288 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1292 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1293 * supported by AIO with the old sigevent structure.
1295 nsig->sigev_notify = osig->sigev_notify;
1296 switch (nsig->sigev_notify) {
1300 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1303 nsig->sigev_notify_kqueue =
1304 osig->__sigev_u.__sigev_notify_kqueue;
1305 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1314 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1316 struct oaiocb *ojob;
1319 bzero(kjob, sizeof(struct aiocb));
1320 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1323 ojob = (struct oaiocb *)kjob;
1324 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1329 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1332 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1336 aiocb_fetch_status(struct aiocb *ujob)
1339 return (fuword(&ujob->_aiocb_private.status));
1343 aiocb_fetch_error(struct aiocb *ujob)
1346 return (fuword(&ujob->_aiocb_private.error));
1350 aiocb_store_status(struct aiocb *ujob, long status)
1353 return (suword(&ujob->_aiocb_private.status, status));
1357 aiocb_store_error(struct aiocb *ujob, long error)
1360 return (suword(&ujob->_aiocb_private.error, error));
1364 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1367 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1371 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1374 return (suword(ujobp, (long)ujob));
1377 static struct aiocb_ops aiocb_ops = {
1378 .copyin = aiocb_copyin,
1379 .fetch_status = aiocb_fetch_status,
1380 .fetch_error = aiocb_fetch_error,
1381 .store_status = aiocb_store_status,
1382 .store_error = aiocb_store_error,
1383 .store_kernelinfo = aiocb_store_kernelinfo,
1384 .store_aiocb = aiocb_store_aiocb,
1387 #ifdef COMPAT_FREEBSD6
1388 static struct aiocb_ops aiocb_ops_osigevent = {
1389 .copyin = aiocb_copyin_old_sigevent,
1390 .fetch_status = aiocb_fetch_status,
1391 .fetch_error = aiocb_fetch_error,
1392 .store_status = aiocb_store_status,
1393 .store_error = aiocb_store_error,
1394 .store_kernelinfo = aiocb_store_kernelinfo,
1395 .store_aiocb = aiocb_store_aiocb,
1400 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1401 * technique is done in this code.
1404 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1405 int type, struct aiocb_ops *ops)
1407 struct proc *p = td->td_proc;
1408 cap_rights_t rights;
1411 struct kaioinfo *ki;
1419 if (p->p_aioinfo == NULL)
1420 aio_init_aioinfo(p);
1424 ops->store_status(ujob, -1);
1425 ops->store_error(ujob, 0);
1426 ops->store_kernelinfo(ujob, -1);
1428 if (num_queue_count >= max_queue_count ||
1429 ki->kaio_count >= ki->kaio_qallowed_count) {
1430 ops->store_error(ujob, EAGAIN);
1434 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1435 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1437 error = ops->copyin(ujob, &job->uaiocb);
1439 ops->store_error(ujob, error);
1440 uma_zfree(aiocb_zone, job);
1444 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1445 uma_zfree(aiocb_zone, job);
1449 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1450 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1451 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1452 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1453 ops->store_error(ujob, EINVAL);
1454 uma_zfree(aiocb_zone, job);
1458 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1459 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1460 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1461 uma_zfree(aiocb_zone, job);
1465 ksiginfo_init(&job->ksi);
1467 /* Save userspace address of the job info. */
1470 /* Get the opcode. */
1471 if (type != LIO_NOP)
1472 job->uaiocb.aio_lio_opcode = type;
1473 opcode = job->uaiocb.aio_lio_opcode;
1476 * Validate the opcode and fetch the file object for the specified
1479 * XXXRW: Moved the opcode validation up here so that we don't
1480 * retrieve a file descriptor without knowing what the capabiltity
1483 fd = job->uaiocb.aio_fildes;
1486 error = fget_write(td, fd,
1487 cap_rights_init(&rights, CAP_PWRITE), &fp);
1490 error = fget_read(td, fd,
1491 cap_rights_init(&rights, CAP_PREAD), &fp);
1494 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1500 error = fget(td, fd, cap_rights_init(&rights), &fp);
1506 uma_zfree(aiocb_zone, job);
1507 ops->store_error(ujob, error);
1511 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1516 if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1523 mtx_lock(&aio_job_mtx);
1525 job->seqno = jobseqno++;
1526 mtx_unlock(&aio_job_mtx);
1527 error = ops->store_kernelinfo(ujob, jid);
1532 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1534 if (opcode == LIO_NOP) {
1536 uma_zfree(aiocb_zone, job);
1540 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1542 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1543 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1547 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1548 kev.ident = (uintptr_t)job->ujob;
1549 kev.filter = EVFILT_AIO;
1550 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1551 kev.data = (intptr_t)job;
1552 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1553 error = kqfd_register(kqfd, &kev, td, 1);
1559 ops->store_error(ujob, EINPROGRESS);
1560 job->uaiocb._aiocb_private.error = EINPROGRESS;
1562 job->cred = crhold(td->td_ucred);
1563 job->jobflags = KAIOCB_QUEUEING;
1566 if (opcode == LIO_MLOCK) {
1567 aio_schedule(job, aio_process_mlock);
1569 } else if (fp->f_ops->fo_aio_queue == NULL)
1570 error = aio_queue_file(fp, job);
1572 error = fo_aio_queue(fp, job);
1577 job->jobflags &= ~KAIOCB_QUEUEING;
1578 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1582 atomic_add_int(&num_queue_count, 1);
1583 if (job->jobflags & KAIOCB_FINISHED) {
1585 * The queue callback completed the request synchronously.
1586 * The bulk of the completion is deferred in that case
1589 aio_bio_done_notify(p, job);
1591 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1596 knlist_delete(&job->klist, curthread, 0);
1599 uma_zfree(aiocb_zone, job);
1600 ops->store_error(ujob, error);
1605 aio_cancel_daemon_job(struct kaiocb *job)
1608 mtx_lock(&aio_job_mtx);
1609 if (!aio_cancel_cleared(job))
1610 TAILQ_REMOVE(&aio_jobs, job, list);
1611 mtx_unlock(&aio_job_mtx);
1616 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1619 mtx_lock(&aio_job_mtx);
1620 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1621 mtx_unlock(&aio_job_mtx);
1625 job->handle_fn = func;
1626 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1627 aio_kick_nowait(job->userproc);
1628 mtx_unlock(&aio_job_mtx);
1632 aio_cancel_sync(struct kaiocb *job)
1634 struct kaioinfo *ki;
1636 ki = job->userproc->p_aioinfo;
1637 mtx_lock(&aio_job_mtx);
1638 if (!aio_cancel_cleared(job))
1639 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1640 mtx_unlock(&aio_job_mtx);
1645 aio_queue_file(struct file *fp, struct kaiocb *job)
1647 struct aioliojob *lj;
1648 struct kaioinfo *ki;
1649 struct kaiocb *job2;
1653 ki = job->userproc->p_aioinfo;
1654 opcode = job->uaiocb.aio_lio_opcode;
1655 if (opcode == LIO_SYNC)
1658 if ((error = aio_qphysio(job->userproc, job)) == 0)
1662 * XXX: This means qphysio() failed with EFAULT. The current
1663 * behavior is to retry the operation via fo_read/fo_write.
1664 * Wouldn't it be better to just complete the request with an
1671 if (!enable_aio_unsafe)
1672 return (EOPNOTSUPP);
1674 if (opcode == LIO_SYNC) {
1676 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1677 if (job2->fd_file == job->fd_file &&
1678 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1679 job2->seqno < job->seqno) {
1680 job2->jobflags |= KAIOCB_CHECKSYNC;
1684 if (job->pending != 0) {
1685 if (!aio_set_cancel_function(job, aio_cancel_sync)) {
1690 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1700 aio_schedule(job, aio_process_rw);
1704 aio_schedule(job, aio_process_sync);
1715 aio_kick_nowait(struct proc *userp)
1717 struct kaioinfo *ki = userp->p_aioinfo;
1718 struct aioproc *aiop;
1720 mtx_assert(&aio_job_mtx, MA_OWNED);
1721 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1722 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1723 aiop->aioprocflags &= ~AIOP_FREE;
1724 wakeup(aiop->aioproc);
1725 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1726 ki->kaio_active_count + num_aio_resv_start <
1727 ki->kaio_maxactive_count) {
1728 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1733 aio_kick(struct proc *userp)
1735 struct kaioinfo *ki = userp->p_aioinfo;
1736 struct aioproc *aiop;
1739 mtx_assert(&aio_job_mtx, MA_OWNED);
1741 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1742 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1743 aiop->aioprocflags &= ~AIOP_FREE;
1744 wakeup(aiop->aioproc);
1745 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1746 ki->kaio_active_count + num_aio_resv_start <
1747 ki->kaio_maxactive_count) {
1748 num_aio_resv_start++;
1749 mtx_unlock(&aio_job_mtx);
1750 error = aio_newproc(&num_aio_resv_start);
1751 mtx_lock(&aio_job_mtx);
1753 num_aio_resv_start--;
1763 aio_kick_helper(void *context, int pending)
1765 struct proc *userp = context;
1767 mtx_lock(&aio_job_mtx);
1768 while (--pending >= 0) {
1769 if (aio_kick(userp))
1772 mtx_unlock(&aio_job_mtx);
1776 * Support the aio_return system call, as a side-effect, kernel resources are
1780 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1782 struct proc *p = td->td_proc;
1784 struct kaioinfo *ki;
1791 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1792 if (job->ujob == ujob)
1796 MPASS(job->jobflags & KAIOCB_FINISHED);
1797 status = job->uaiocb._aiocb_private.status;
1798 error = job->uaiocb._aiocb_private.error;
1799 td->td_retval[0] = status;
1800 if (job->uaiocb.aio_lio_opcode == LIO_WRITE) {
1801 td->td_ru.ru_oublock += job->outputcharge;
1802 job->outputcharge = 0;
1803 } else if (job->uaiocb.aio_lio_opcode == LIO_READ) {
1804 td->td_ru.ru_inblock += job->inputcharge;
1805 job->inputcharge = 0;
1807 aio_free_entry(job);
1809 ops->store_error(ujob, error);
1810 ops->store_status(ujob, status);
1819 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1822 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1826 * Allow a process to wakeup when any of the I/O requests are completed.
1829 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1830 struct timespec *ts)
1832 struct proc *p = td->td_proc;
1834 struct kaioinfo *ki;
1835 struct kaiocb *firstjob, *job;
1840 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1843 TIMESPEC_TO_TIMEVAL(&atv, ts);
1844 if (itimerfix(&atv))
1846 timo = tvtohz(&atv);
1860 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1861 for (i = 0; i < njoblist; i++) {
1862 if (job->ujob == ujoblist[i]) {
1863 if (firstjob == NULL)
1865 if (job->jobflags & KAIOCB_FINISHED)
1870 /* All tasks were finished. */
1871 if (firstjob == NULL)
1874 ki->kaio_flags |= KAIO_WAKEUP;
1875 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1877 if (error == ERESTART)
1888 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1890 struct timespec ts, *tsp;
1891 struct aiocb **ujoblist;
1894 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1898 /* Get timespec struct. */
1899 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1905 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1906 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1908 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1909 uma_zfree(aiol_zone, ujoblist);
1914 * aio_cancel cancels any non-physio aio operations not currently in
1918 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1920 struct proc *p = td->td_proc;
1921 struct kaioinfo *ki;
1922 struct kaiocb *job, *jobn;
1924 cap_rights_t rights;
1927 int notcancelled = 0;
1930 /* Lookup file object. */
1931 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1939 if (fp->f_type == DTYPE_VNODE) {
1941 if (vn_isdisk(vp, &error)) {
1943 td->td_retval[0] = AIO_NOTCANCELED;
1949 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1950 if ((uap->fd == job->uaiocb.aio_fildes) &&
1951 ((uap->aiocbp == NULL) ||
1952 (uap->aiocbp == job->ujob))) {
1953 if (aio_cancel_job(p, ki, job)) {
1958 if (uap->aiocbp != NULL)
1967 if (uap->aiocbp != NULL) {
1969 td->td_retval[0] = AIO_CANCELED;
1975 td->td_retval[0] = AIO_NOTCANCELED;
1980 td->td_retval[0] = AIO_CANCELED;
1984 td->td_retval[0] = AIO_ALLDONE;
1990 * aio_error is implemented in the kernel level for compatibility purposes
1991 * only. For a user mode async implementation, it would be best to do it in
1992 * a userland subroutine.
1995 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1997 struct proc *p = td->td_proc;
1999 struct kaioinfo *ki;
2004 td->td_retval[0] = EINVAL;
2009 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2010 if (job->ujob == ujob) {
2011 if (job->jobflags & KAIOCB_FINISHED)
2013 job->uaiocb._aiocb_private.error;
2015 td->td_retval[0] = EINPROGRESS;
2023 * Hack for failure of aio_aqueue.
2025 status = ops->fetch_status(ujob);
2027 td->td_retval[0] = ops->fetch_error(ujob);
2031 td->td_retval[0] = EINVAL;
2036 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2039 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2042 /* syscall - asynchronous read from a file (REALTIME) */
2043 #ifdef COMPAT_FREEBSD6
2045 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2048 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2049 &aiocb_ops_osigevent));
2054 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2057 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2060 /* syscall - asynchronous write to a file (REALTIME) */
2061 #ifdef COMPAT_FREEBSD6
2063 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2066 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2067 &aiocb_ops_osigevent));
2072 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2075 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2079 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2082 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2086 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2087 struct aiocb **acb_list, int nent, struct sigevent *sig,
2088 struct aiocb_ops *ops)
2090 struct proc *p = td->td_proc;
2092 struct kaioinfo *ki;
2093 struct aioliojob *lj;
2099 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2102 if (nent < 0 || nent > AIO_LISTIO_MAX)
2105 if (p->p_aioinfo == NULL)
2106 aio_init_aioinfo(p);
2110 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2113 lj->lioj_finished_count = 0;
2114 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2115 ksiginfo_init(&lj->lioj_ksi);
2120 if (sig && (mode == LIO_NOWAIT)) {
2121 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2122 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2123 /* Assume only new style KEVENT */
2124 kev.filter = EVFILT_LIO;
2125 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2126 kev.ident = (uintptr_t)uacb_list; /* something unique */
2127 kev.data = (intptr_t)lj;
2128 /* pass user defined sigval data */
2129 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2130 error = kqfd_register(
2131 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2133 uma_zfree(aiolio_zone, lj);
2136 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2138 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2139 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2140 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2141 uma_zfree(aiolio_zone, lj);
2144 lj->lioj_flags |= LIOJ_SIGNAL;
2146 uma_zfree(aiolio_zone, lj);
2152 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2154 * Add extra aiocb count to avoid the lio to be freed
2155 * by other threads doing aio_waitcomplete or aio_return,
2156 * and prevent event from being sent until we have queued
2163 * Get pointers to the list of I/O requests.
2166 for (i = 0; i < nent; i++) {
2169 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2177 if (mode == LIO_WAIT) {
2178 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2179 ki->kaio_flags |= KAIO_WAKEUP;
2180 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2181 PRIBIO | PCATCH, "aiospn", 0);
2182 if (error == ERESTART)
2188 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2189 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2190 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2191 KNOTE_LOCKED(&lj->klist, 1);
2193 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2195 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2196 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2197 aio_sendsig(p, &lj->lioj_signal,
2199 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2204 if (lj->lioj_count == 0) {
2205 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2206 knlist_delete(&lj->klist, curthread, 1);
2208 sigqueue_take(&lj->lioj_ksi);
2211 uma_zfree(aiolio_zone, lj);
2220 /* syscall - list directed I/O (REALTIME) */
2221 #ifdef COMPAT_FREEBSD6
2223 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2225 struct aiocb **acb_list;
2226 struct sigevent *sigp, sig;
2227 struct osigevent osig;
2230 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2234 if (nent < 0 || nent > AIO_LISTIO_MAX)
2237 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2238 error = copyin(uap->sig, &osig, sizeof(osig));
2241 error = convert_old_sigevent(&osig, &sig);
2248 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2249 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2251 error = kern_lio_listio(td, uap->mode,
2252 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2253 &aiocb_ops_osigevent);
2254 free(acb_list, M_LIO);
2259 /* syscall - list directed I/O (REALTIME) */
2261 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2263 struct aiocb **acb_list;
2264 struct sigevent *sigp, sig;
2267 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2271 if (nent < 0 || nent > AIO_LISTIO_MAX)
2274 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2275 error = copyin(uap->sig, &sig, sizeof(sig));
2282 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2283 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2285 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2286 nent, sigp, &aiocb_ops);
2287 free(acb_list, M_LIO);
2292 aio_physwakeup(struct bio *bp)
2294 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2296 struct kaioinfo *ki;
2300 /* Release mapping into kernel space. */
2301 userp = job->userproc;
2302 ki = userp->p_aioinfo;
2304 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2305 relpbuf(job->pbuf, NULL);
2307 atomic_subtract_int(&num_buf_aio, 1);
2309 ki->kaio_buffer_count--;
2312 vm_page_unhold_pages(job->pages, job->npages);
2316 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2318 if (bp->bio_flags & BIO_ERROR)
2319 error = bp->bio_error;
2320 nblks = btodb(nbytes);
2321 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2322 job->outputcharge += nblks;
2324 job->inputcharge += nblks;
2326 aio_complete(job, nbytes, error);
2331 /* syscall - wait for the next completion of an aio request */
2333 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2334 struct timespec *ts, struct aiocb_ops *ops)
2336 struct proc *p = td->td_proc;
2338 struct kaioinfo *ki;
2344 ops->store_aiocb(ujobp, NULL);
2348 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2351 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2354 TIMESPEC_TO_TIMEVAL(&atv, ts);
2355 if (itimerfix(&atv))
2357 timo = tvtohz(&atv);
2360 if (p->p_aioinfo == NULL)
2361 aio_init_aioinfo(p);
2367 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2369 error = EWOULDBLOCK;
2372 ki->kaio_flags |= KAIO_WAKEUP;
2373 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2375 if (timo && error == ERESTART)
2382 MPASS(job->jobflags & KAIOCB_FINISHED);
2384 status = job->uaiocb._aiocb_private.status;
2385 error = job->uaiocb._aiocb_private.error;
2386 td->td_retval[0] = status;
2387 if (job->uaiocb.aio_lio_opcode == LIO_WRITE) {
2388 td->td_ru.ru_oublock += job->outputcharge;
2389 job->outputcharge = 0;
2390 } else if (job->uaiocb.aio_lio_opcode == LIO_READ) {
2391 td->td_ru.ru_inblock += job->inputcharge;
2392 job->inputcharge = 0;
2394 aio_free_entry(job);
2396 ops->store_aiocb(ujobp, ujob);
2397 ops->store_error(ujob, error);
2398 ops->store_status(ujob, status);
2406 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2408 struct timespec ts, *tsp;
2412 /* Get timespec struct. */
2413 error = copyin(uap->timeout, &ts, sizeof(ts));
2420 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2424 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2425 struct aiocb_ops *ops)
2427 struct proc *p = td->td_proc;
2428 struct kaioinfo *ki;
2430 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2434 aio_init_aioinfo(p);
2435 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2439 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2442 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2445 /* kqueue attach function */
2447 filt_aioattach(struct knote *kn)
2449 struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2452 * The job pointer must be validated before using it, so
2453 * registration is restricted to the kernel; the user cannot
2456 if ((kn->kn_flags & EV_FLAG1) == 0)
2458 kn->kn_ptr.p_aio = job;
2459 kn->kn_flags &= ~EV_FLAG1;
2461 knlist_add(&job->klist, kn, 0);
2466 /* kqueue detach function */
2468 filt_aiodetach(struct knote *kn)
2472 knl = &kn->kn_ptr.p_aio->klist;
2473 knl->kl_lock(knl->kl_lockarg);
2474 if (!knlist_empty(knl))
2475 knlist_remove(knl, kn, 1);
2476 knl->kl_unlock(knl->kl_lockarg);
2479 /* kqueue filter function */
2482 filt_aio(struct knote *kn, long hint)
2484 struct kaiocb *job = kn->kn_ptr.p_aio;
2486 kn->kn_data = job->uaiocb._aiocb_private.error;
2487 if (!(job->jobflags & KAIOCB_FINISHED))
2489 kn->kn_flags |= EV_EOF;
2493 /* kqueue attach function */
2495 filt_lioattach(struct knote *kn)
2497 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2500 * The aioliojob pointer must be validated before using it, so
2501 * registration is restricted to the kernel; the user cannot
2504 if ((kn->kn_flags & EV_FLAG1) == 0)
2506 kn->kn_ptr.p_lio = lj;
2507 kn->kn_flags &= ~EV_FLAG1;
2509 knlist_add(&lj->klist, kn, 0);
2514 /* kqueue detach function */
2516 filt_liodetach(struct knote *kn)
2520 knl = &kn->kn_ptr.p_lio->klist;
2521 knl->kl_lock(knl->kl_lockarg);
2522 if (!knlist_empty(knl))
2523 knlist_remove(knl, kn, 1);
2524 knl->kl_unlock(knl->kl_lockarg);
2527 /* kqueue filter function */
2530 filt_lio(struct knote *kn, long hint)
2532 struct aioliojob * lj = kn->kn_ptr.p_lio;
2534 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2537 #ifdef COMPAT_FREEBSD32
2538 #include <sys/mount.h>
2539 #include <sys/socket.h>
2540 #include <compat/freebsd32/freebsd32.h>
2541 #include <compat/freebsd32/freebsd32_proto.h>
2542 #include <compat/freebsd32/freebsd32_signal.h>
2543 #include <compat/freebsd32/freebsd32_syscall.h>
2544 #include <compat/freebsd32/freebsd32_util.h>
2546 struct __aiocb_private32 {
2549 uint32_t kernelinfo;
2552 #ifdef COMPAT_FREEBSD6
2553 typedef struct oaiocb32 {
2554 int aio_fildes; /* File descriptor */
2555 uint64_t aio_offset __packed; /* File offset for I/O */
2556 uint32_t aio_buf; /* I/O buffer in process space */
2557 uint32_t aio_nbytes; /* Number of bytes for I/O */
2558 struct osigevent32 aio_sigevent; /* Signal to deliver */
2559 int aio_lio_opcode; /* LIO opcode */
2560 int aio_reqprio; /* Request priority -- ignored */
2561 struct __aiocb_private32 _aiocb_private;
2565 typedef struct aiocb32 {
2566 int32_t aio_fildes; /* File descriptor */
2567 uint64_t aio_offset __packed; /* File offset for I/O */
2568 uint32_t aio_buf; /* I/O buffer in process space */
2569 uint32_t aio_nbytes; /* Number of bytes for I/O */
2571 uint32_t __spare2__;
2572 int aio_lio_opcode; /* LIO opcode */
2573 int aio_reqprio; /* Request priority -- ignored */
2574 struct __aiocb_private32 _aiocb_private;
2575 struct sigevent32 aio_sigevent; /* Signal to deliver */
2578 #ifdef COMPAT_FREEBSD6
2580 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2584 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2585 * supported by AIO with the old sigevent structure.
2587 CP(*osig, *nsig, sigev_notify);
2588 switch (nsig->sigev_notify) {
2592 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2595 nsig->sigev_notify_kqueue =
2596 osig->__sigev_u.__sigev_notify_kqueue;
2597 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2606 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2608 struct oaiocb32 job32;
2611 bzero(kjob, sizeof(struct aiocb));
2612 error = copyin(ujob, &job32, sizeof(job32));
2616 CP(job32, *kjob, aio_fildes);
2617 CP(job32, *kjob, aio_offset);
2618 PTRIN_CP(job32, *kjob, aio_buf);
2619 CP(job32, *kjob, aio_nbytes);
2620 CP(job32, *kjob, aio_lio_opcode);
2621 CP(job32, *kjob, aio_reqprio);
2622 CP(job32, *kjob, _aiocb_private.status);
2623 CP(job32, *kjob, _aiocb_private.error);
2624 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2625 return (convert_old_sigevent32(&job32.aio_sigevent,
2626 &kjob->aio_sigevent));
2631 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2633 struct aiocb32 job32;
2636 error = copyin(ujob, &job32, sizeof(job32));
2639 CP(job32, *kjob, aio_fildes);
2640 CP(job32, *kjob, aio_offset);
2641 PTRIN_CP(job32, *kjob, aio_buf);
2642 CP(job32, *kjob, aio_nbytes);
2643 CP(job32, *kjob, aio_lio_opcode);
2644 CP(job32, *kjob, aio_reqprio);
2645 CP(job32, *kjob, _aiocb_private.status);
2646 CP(job32, *kjob, _aiocb_private.error);
2647 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2648 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2652 aiocb32_fetch_status(struct aiocb *ujob)
2654 struct aiocb32 *ujob32;
2656 ujob32 = (struct aiocb32 *)ujob;
2657 return (fuword32(&ujob32->_aiocb_private.status));
2661 aiocb32_fetch_error(struct aiocb *ujob)
2663 struct aiocb32 *ujob32;
2665 ujob32 = (struct aiocb32 *)ujob;
2666 return (fuword32(&ujob32->_aiocb_private.error));
2670 aiocb32_store_status(struct aiocb *ujob, long status)
2672 struct aiocb32 *ujob32;
2674 ujob32 = (struct aiocb32 *)ujob;
2675 return (suword32(&ujob32->_aiocb_private.status, status));
2679 aiocb32_store_error(struct aiocb *ujob, long error)
2681 struct aiocb32 *ujob32;
2683 ujob32 = (struct aiocb32 *)ujob;
2684 return (suword32(&ujob32->_aiocb_private.error, error));
2688 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2690 struct aiocb32 *ujob32;
2692 ujob32 = (struct aiocb32 *)ujob;
2693 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2697 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2700 return (suword32(ujobp, (long)ujob));
2703 static struct aiocb_ops aiocb32_ops = {
2704 .copyin = aiocb32_copyin,
2705 .fetch_status = aiocb32_fetch_status,
2706 .fetch_error = aiocb32_fetch_error,
2707 .store_status = aiocb32_store_status,
2708 .store_error = aiocb32_store_error,
2709 .store_kernelinfo = aiocb32_store_kernelinfo,
2710 .store_aiocb = aiocb32_store_aiocb,
2713 #ifdef COMPAT_FREEBSD6
2714 static struct aiocb_ops aiocb32_ops_osigevent = {
2715 .copyin = aiocb32_copyin_old_sigevent,
2716 .fetch_status = aiocb32_fetch_status,
2717 .fetch_error = aiocb32_fetch_error,
2718 .store_status = aiocb32_store_status,
2719 .store_error = aiocb32_store_error,
2720 .store_kernelinfo = aiocb32_store_kernelinfo,
2721 .store_aiocb = aiocb32_store_aiocb,
2726 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2729 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2733 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2735 struct timespec32 ts32;
2736 struct timespec ts, *tsp;
2737 struct aiocb **ujoblist;
2738 uint32_t *ujoblist32;
2741 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2745 /* Get timespec struct. */
2746 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2748 CP(ts32, ts, tv_sec);
2749 CP(ts32, ts, tv_nsec);
2754 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2755 ujoblist32 = (uint32_t *)ujoblist;
2756 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2757 sizeof(ujoblist32[0]));
2759 for (i = uap->nent; i > 0; i--)
2760 ujoblist[i] = PTRIN(ujoblist32[i]);
2762 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2764 uma_zfree(aiol_zone, ujoblist);
2769 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2772 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2775 #ifdef COMPAT_FREEBSD6
2777 freebsd6_freebsd32_aio_read(struct thread *td,
2778 struct freebsd6_freebsd32_aio_read_args *uap)
2781 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2782 &aiocb32_ops_osigevent));
2787 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2790 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2794 #ifdef COMPAT_FREEBSD6
2796 freebsd6_freebsd32_aio_write(struct thread *td,
2797 struct freebsd6_freebsd32_aio_write_args *uap)
2800 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2801 &aiocb32_ops_osigevent));
2806 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2809 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2814 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2817 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2822 freebsd32_aio_waitcomplete(struct thread *td,
2823 struct freebsd32_aio_waitcomplete_args *uap)
2825 struct timespec32 ts32;
2826 struct timespec ts, *tsp;
2830 /* Get timespec struct. */
2831 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2834 CP(ts32, ts, tv_sec);
2835 CP(ts32, ts, tv_nsec);
2840 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2845 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2848 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2852 #ifdef COMPAT_FREEBSD6
2854 freebsd6_freebsd32_lio_listio(struct thread *td,
2855 struct freebsd6_freebsd32_lio_listio_args *uap)
2857 struct aiocb **acb_list;
2858 struct sigevent *sigp, sig;
2859 struct osigevent32 osig;
2860 uint32_t *acb_list32;
2863 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2867 if (nent < 0 || nent > AIO_LISTIO_MAX)
2870 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2871 error = copyin(uap->sig, &osig, sizeof(osig));
2874 error = convert_old_sigevent32(&osig, &sig);
2881 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2882 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2884 free(acb_list32, M_LIO);
2887 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2888 for (i = 0; i < nent; i++)
2889 acb_list[i] = PTRIN(acb_list32[i]);
2890 free(acb_list32, M_LIO);
2892 error = kern_lio_listio(td, uap->mode,
2893 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2894 &aiocb32_ops_osigevent);
2895 free(acb_list, M_LIO);
2901 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2903 struct aiocb **acb_list;
2904 struct sigevent *sigp, sig;
2905 struct sigevent32 sig32;
2906 uint32_t *acb_list32;
2909 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2913 if (nent < 0 || nent > AIO_LISTIO_MAX)
2916 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2917 error = copyin(uap->sig, &sig32, sizeof(sig32));
2920 error = convert_sigevent32(&sig32, &sig);
2927 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2928 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2930 free(acb_list32, M_LIO);
2933 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2934 for (i = 0; i < nent; i++)
2935 acb_list[i] = PTRIN(acb_list32[i]);
2936 free(acb_list32, M_LIO);
2938 error = kern_lio_listio(td, uap->mode,
2939 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2941 free(acb_list, M_LIO);