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/capability.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>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/syscall.h>
53 #include <sys/sysent.h>
54 #include <sys/sysctl.h>
56 #include <sys/taskqueue.h>
57 #include <sys/vnode.h>
59 #include <sys/event.h>
60 #include <sys/mount.h>
62 #include <machine/atomic.h>
65 #include <vm/vm_extern.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_object.h>
72 #include "opt_vfs_aio.h"
75 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
76 * overflow. (XXX will be removed soon.)
78 static u_long jobrefid;
81 * Counter for aio_fsync.
83 static uint64_t jobseqno;
86 #define JOBST_JOBQSOCK 1
87 #define JOBST_JOBQGLOBAL 2
88 #define JOBST_JOBRUNNING 3
89 #define JOBST_JOBFINISHED 4
90 #define JOBST_JOBQBUF 5
91 #define JOBST_JOBQSYNC 6
93 #ifndef MAX_AIO_PER_PROC
94 #define MAX_AIO_PER_PROC 32
97 #ifndef MAX_AIO_QUEUE_PER_PROC
98 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
101 #ifndef MAX_AIO_PROCS
102 #define MAX_AIO_PROCS 32
105 #ifndef MAX_AIO_QUEUE
106 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
109 #ifndef TARGET_AIO_PROCS
110 #define TARGET_AIO_PROCS 4
114 #define MAX_BUF_AIO 16
117 #ifndef AIOD_TIMEOUT_DEFAULT
118 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
121 #ifndef AIOD_LIFETIME_DEFAULT
122 #define AIOD_LIFETIME_DEFAULT (30 * hz)
125 FEATURE(aio, "Asynchronous I/O");
127 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
129 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
131 static int max_aio_procs = MAX_AIO_PROCS;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
133 CTLFLAG_RW, &max_aio_procs, 0,
134 "Maximum number of kernel threads to use for handling async IO ");
136 static int num_aio_procs = 0;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
138 CTLFLAG_RD, &num_aio_procs, 0,
139 "Number of presently active kernel threads for async IO");
142 * The code will adjust the actual number of AIO processes towards this
143 * number when it gets a chance.
145 static int target_aio_procs = TARGET_AIO_PROCS;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
147 0, "Preferred number of ready kernel threads for async IO");
149 static int max_queue_count = MAX_AIO_QUEUE;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
151 "Maximum number of aio requests to queue, globally");
153 static int num_queue_count = 0;
154 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
155 "Number of queued aio requests");
157 static int num_buf_aio = 0;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
159 "Number of aio requests presently handled by the buf subsystem");
161 /* Number of async I/O thread in the process of being started */
162 /* XXX This should be local to aio_aqueue() */
163 static int num_aio_resv_start = 0;
165 static int aiod_timeout;
166 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
167 "Timeout value for synchronous aio operations");
169 static int aiod_lifetime;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
171 "Maximum lifetime for idle aiod");
173 static int unloadable = 0;
174 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
175 "Allow unload of aio (not recommended)");
178 static int max_aio_per_proc = MAX_AIO_PER_PROC;
179 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
180 0, "Maximum active aio requests per process (stored in the process)");
182 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
183 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
184 &max_aio_queue_per_proc, 0,
185 "Maximum queued aio requests per process (stored in the process)");
187 static int max_buf_aio = MAX_BUF_AIO;
188 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
189 "Maximum buf aio requests per process (stored in the process)");
191 typedef struct oaiocb {
192 int aio_fildes; /* File descriptor */
193 off_t aio_offset; /* File offset for I/O */
194 volatile void *aio_buf; /* I/O buffer in process space */
195 size_t aio_nbytes; /* Number of bytes for I/O */
196 struct osigevent aio_sigevent; /* Signal to deliver */
197 int aio_lio_opcode; /* LIO opcode */
198 int aio_reqprio; /* Request priority -- ignored */
199 struct __aiocb_private _aiocb_private;
203 * Below is a key of locks used to protect each member of struct aiocblist
204 * aioliojob and kaioinfo and any backends.
206 * * - need not protected
207 * a - locked by kaioinfo lock
208 * b - locked by backend lock, the backend lock can be null in some cases,
209 * for example, BIO belongs to this type, in this case, proc lock is
211 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
215 * Current, there is only two backends: BIO and generic file I/O.
216 * socket I/O is served by generic file I/O, this is not a good idea, since
217 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
218 * threads, if there is no thread to serve socket I/O, the socket I/O will be
219 * delayed too long or starved, we should create some threads dedicated to
220 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
221 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
222 * structure is not safe because there is race between userland and aio
227 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
228 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
229 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
230 int jobflags; /* (a) job flags */
231 int jobstate; /* (b) job state */
232 int inputcharge; /* (*) input blockes */
233 int outputcharge; /* (*) output blockes */
234 struct buf *bp; /* (*) private to BIO backend,
237 struct proc *userproc; /* (*) user process */
238 struct ucred *cred; /* (*) active credential when created */
239 struct file *fd_file; /* (*) pointer to file structure */
240 struct aioliojob *lio; /* (*) optional lio job */
241 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
242 struct knlist klist; /* (a) list of knotes */
243 struct aiocb uaiocb; /* (*) kernel I/O control block */
244 ksiginfo_t ksi; /* (a) realtime signal info */
245 struct task biotask; /* (*) private to BIO backend */
246 uint64_t seqno; /* (*) job number */
247 int pending; /* (a) number of pending I/O, aio_fsync only */
251 #define AIOCBLIST_DONE 0x01
252 #define AIOCBLIST_BUFDONE 0x02
253 #define AIOCBLIST_RUNDOWN 0x04
254 #define AIOCBLIST_CHECKSYNC 0x08
259 #define AIOP_FREE 0x1 /* proc on free queue */
261 struct aiothreadlist {
262 int aiothreadflags; /* (c) AIO proc flags */
263 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
264 struct thread *aiothread; /* (*) the AIO thread */
268 * data-structure for lio signal management
271 int lioj_flags; /* (a) listio flags */
272 int lioj_count; /* (a) listio flags */
273 int lioj_finished_count; /* (a) listio flags */
274 struct sigevent lioj_signal; /* (a) signal on all I/O done */
275 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
276 struct knlist klist; /* (a) list of knotes */
277 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
280 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
281 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
282 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
285 * per process aio data structure
288 struct mtx kaio_mtx; /* the lock to protect this struct */
289 int kaio_flags; /* (a) per process kaio flags */
290 int kaio_maxactive_count; /* (*) maximum number of AIOs */
291 int kaio_active_count; /* (c) number of currently used AIOs */
292 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
293 int kaio_count; /* (a) size of AIO queue */
294 int kaio_ballowed_count; /* (*) maximum number of buffers */
295 int kaio_buffer_count; /* (a) number of physio buffers */
296 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
297 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
298 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
299 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
300 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
301 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
304 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
305 struct task kaio_task; /* (*) task to kick aio threads */
308 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
309 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
310 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
311 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
313 #define KAIO_RUNDOWN 0x1 /* process is being run down */
314 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
317 * Operations used to interact with userland aio control blocks.
318 * Different ABIs provide their own operations.
321 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
322 long (*fetch_status)(struct aiocb *ujob);
323 long (*fetch_error)(struct aiocb *ujob);
324 int (*store_status)(struct aiocb *ujob, long status);
325 int (*store_error)(struct aiocb *ujob, long error);
326 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
327 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
330 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
331 static struct sema aio_newproc_sem;
332 static struct mtx aio_job_mtx;
333 static struct mtx aio_sock_mtx;
334 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
335 static struct unrhdr *aiod_unr;
337 void aio_init_aioinfo(struct proc *p);
338 static int aio_onceonly(void);
339 static int aio_free_entry(struct aiocblist *aiocbe);
340 static void aio_process(struct aiocblist *aiocbe);
341 static int aio_newproc(int *);
342 int aio_aqueue(struct thread *td, struct aiocb *job,
343 struct aioliojob *lio, int type, struct aiocb_ops *ops);
344 static void aio_physwakeup(struct buf *bp);
345 static void aio_proc_rundown(void *arg, struct proc *p);
346 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
347 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
348 static void biohelper(void *, int);
349 static void aio_daemon(void *param);
350 static void aio_swake_cb(struct socket *, struct sockbuf *);
351 static int aio_unload(void);
352 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
355 static int aio_kick(struct proc *userp);
356 static void aio_kick_nowait(struct proc *userp);
357 static void aio_kick_helper(void *context, int pending);
358 static int filt_aioattach(struct knote *kn);
359 static void filt_aiodetach(struct knote *kn);
360 static int filt_aio(struct knote *kn, long hint);
361 static int filt_lioattach(struct knote *kn);
362 static void filt_liodetach(struct knote *kn);
363 static int filt_lio(struct knote *kn, long hint);
367 * kaio Per process async io info
368 * aiop async io thread data
369 * aiocb async io jobs
370 * aiol list io job pointer - internal to aio_suspend XXX
371 * aiolio list io jobs
373 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
375 /* kqueue filters for aio */
376 static struct filterops aio_filtops = {
378 .f_attach = filt_aioattach,
379 .f_detach = filt_aiodetach,
382 static struct filterops lio_filtops = {
384 .f_attach = filt_lioattach,
385 .f_detach = filt_liodetach,
389 static eventhandler_tag exit_tag, exec_tag;
391 TASKQUEUE_DEFINE_THREAD(aiod_bio);
394 * Main operations function for use as a kernel module.
397 aio_modload(struct module *module, int cmd, void *arg)
406 error = aio_unload();
417 static moduledata_t aio_mod = {
423 static struct syscall_helper_data aio_syscalls[] = {
424 SYSCALL_INIT_HELPER(aio_cancel),
425 SYSCALL_INIT_HELPER(aio_error),
426 SYSCALL_INIT_HELPER(aio_fsync),
427 SYSCALL_INIT_HELPER(aio_read),
428 SYSCALL_INIT_HELPER(aio_return),
429 SYSCALL_INIT_HELPER(aio_suspend),
430 SYSCALL_INIT_HELPER(aio_waitcomplete),
431 SYSCALL_INIT_HELPER(aio_write),
432 SYSCALL_INIT_HELPER(lio_listio),
433 SYSCALL_INIT_HELPER(oaio_read),
434 SYSCALL_INIT_HELPER(oaio_write),
435 SYSCALL_INIT_HELPER(olio_listio),
439 #ifdef COMPAT_FREEBSD32
440 #include <sys/mount.h>
441 #include <sys/socket.h>
442 #include <compat/freebsd32/freebsd32.h>
443 #include <compat/freebsd32/freebsd32_proto.h>
444 #include <compat/freebsd32/freebsd32_signal.h>
445 #include <compat/freebsd32/freebsd32_syscall.h>
446 #include <compat/freebsd32/freebsd32_util.h>
448 static struct syscall_helper_data aio32_syscalls[] = {
449 SYSCALL32_INIT_HELPER(freebsd32_aio_return),
450 SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
451 SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
452 SYSCALL32_INIT_HELPER(freebsd32_aio_error),
453 SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
454 SYSCALL32_INIT_HELPER(freebsd32_aio_read),
455 SYSCALL32_INIT_HELPER(freebsd32_aio_write),
456 SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
457 SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
458 SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
459 SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
460 SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
465 DECLARE_MODULE(aio, aio_mod,
466 SI_SUB_VFS, SI_ORDER_ANY);
467 MODULE_VERSION(aio, 1);
470 * Startup initialization
477 /* XXX: should probably just use so->callback */
478 aio_swake = &aio_swake_cb;
479 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
480 EVENTHANDLER_PRI_ANY);
481 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
482 EVENTHANDLER_PRI_ANY);
483 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
484 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
485 TAILQ_INIT(&aio_freeproc);
486 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
487 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
488 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
489 TAILQ_INIT(&aio_jobs);
490 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
491 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
492 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
493 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
494 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
495 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
496 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
497 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
498 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
499 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
500 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
501 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
502 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
504 async_io_version = _POSIX_VERSION;
505 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
506 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
507 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
509 error = syscall_helper_register(aio_syscalls);
512 #ifdef COMPAT_FREEBSD32
513 error = syscall32_helper_register(aio32_syscalls);
521 * Callback for unload of AIO when used as a module.
529 * XXX: no unloads by default, it's too dangerous.
530 * perhaps we could do it if locked out callers and then
531 * did an aio_proc_rundown() on each process.
533 * jhb: aio_proc_rundown() needs to run on curproc though,
534 * so I don't think that would fly.
539 #ifdef COMPAT_FREEBSD32
540 syscall32_helper_unregister(aio32_syscalls);
542 syscall_helper_unregister(aio_syscalls);
544 error = kqueue_del_filteropts(EVFILT_AIO);
547 error = kqueue_del_filteropts(EVFILT_LIO);
550 async_io_version = 0;
552 taskqueue_free(taskqueue_aiod_bio);
553 delete_unrhdr(aiod_unr);
554 uma_zdestroy(kaio_zone);
555 uma_zdestroy(aiop_zone);
556 uma_zdestroy(aiocb_zone);
557 uma_zdestroy(aiol_zone);
558 uma_zdestroy(aiolio_zone);
559 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
560 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
561 mtx_destroy(&aio_job_mtx);
562 mtx_destroy(&aio_sock_mtx);
563 sema_destroy(&aio_newproc_sem);
564 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
565 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
566 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
571 * Init the per-process aioinfo structure. The aioinfo limits are set
572 * per-process for user limit (resource) management.
575 aio_init_aioinfo(struct proc *p)
579 ki = uma_zalloc(kaio_zone, M_WAITOK);
580 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
582 ki->kaio_maxactive_count = max_aio_per_proc;
583 ki->kaio_active_count = 0;
584 ki->kaio_qallowed_count = max_aio_queue_per_proc;
586 ki->kaio_ballowed_count = max_buf_aio;
587 ki->kaio_buffer_count = 0;
588 TAILQ_INIT(&ki->kaio_all);
589 TAILQ_INIT(&ki->kaio_done);
590 TAILQ_INIT(&ki->kaio_jobqueue);
591 TAILQ_INIT(&ki->kaio_bufqueue);
592 TAILQ_INIT(&ki->kaio_liojoblist);
593 TAILQ_INIT(&ki->kaio_sockqueue);
594 TAILQ_INIT(&ki->kaio_syncqueue);
595 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
597 if (p->p_aioinfo == NULL) {
602 mtx_destroy(&ki->kaio_mtx);
603 uma_zfree(kaio_zone, ki);
606 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
611 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
616 error = sigev_findtd(p, sigev, &td);
620 ksiginfo_set_sigev(ksi, sigev);
621 ksi->ksi_code = SI_ASYNCIO;
622 ksi->ksi_flags |= KSI_EXT | KSI_INS;
623 tdsendsignal(p, td, ksi->ksi_signo, ksi);
630 * Free a job entry. Wait for completion if it is currently active, but don't
631 * delay forever. If we delay, we return a flag that says that we have to
632 * restart the queue scan.
635 aio_free_entry(struct aiocblist *aiocbe)
638 struct aioliojob *lj;
641 p = aiocbe->userproc;
646 AIO_LOCK_ASSERT(ki, MA_OWNED);
647 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
649 atomic_subtract_int(&num_queue_count, 1);
652 MPASS(ki->kaio_count >= 0);
654 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
655 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
660 lj->lioj_finished_count--;
662 if (lj->lioj_count == 0) {
663 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
664 /* lio is going away, we need to destroy any knotes */
665 knlist_delete(&lj->klist, curthread, 1);
667 sigqueue_take(&lj->lioj_ksi);
669 uma_zfree(aiolio_zone, lj);
673 /* aiocbe is going away, we need to destroy any knotes */
674 knlist_delete(&aiocbe->klist, curthread, 1);
676 sigqueue_take(&aiocbe->ksi);
679 MPASS(aiocbe->bp == NULL);
680 aiocbe->jobstate = JOBST_NULL;
684 * The thread argument here is used to find the owning process
685 * and is also passed to fo_close() which may pass it to various
686 * places such as devsw close() routines. Because of that, we
687 * need a thread pointer from the process owning the job that is
688 * persistent and won't disappear out from under us or move to
691 * Currently, all the callers of this function call it to remove
692 * an aiocblist from the current process' job list either via a
693 * syscall or due to the current process calling exit() or
694 * execve(). Thus, we know that p == curproc. We also know that
695 * curthread can't exit since we are curthread.
697 * Therefore, we use curthread as the thread to pass to
698 * knlist_delete(). This does mean that it is possible for the
699 * thread pointer at close time to differ from the thread pointer
700 * at open time, but this is already true of file descriptors in
701 * a multithreaded process.
703 fdrop(aiocbe->fd_file, curthread);
704 crfree(aiocbe->cred);
705 uma_zfree(aiocb_zone, aiocbe);
712 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
714 aio_proc_rundown(arg, p);
718 * Rundown the jobs for a given process.
721 aio_proc_rundown(void *arg, struct proc *p)
724 struct aioliojob *lj;
725 struct aiocblist *cbe, *cbn;
730 KASSERT(curthread->td_proc == p,
731 ("%s: called on non-curproc", __func__));
737 ki->kaio_flags |= KAIO_RUNDOWN;
742 * Try to cancel all pending requests. This code simulates
743 * aio_cancel on all pending I/O requests.
745 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
747 mtx_lock(&aio_job_mtx);
748 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
749 TAILQ_REMOVE(&aio_jobs, cbe, list);
751 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
753 MPASS(fp->f_type == DTYPE_SOCKET);
755 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
757 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
758 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
761 mtx_unlock(&aio_job_mtx);
764 cbe->jobstate = JOBST_JOBFINISHED;
765 cbe->uaiocb._aiocb_private.status = -1;
766 cbe->uaiocb._aiocb_private.error = ECANCELED;
767 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
768 aio_bio_done_notify(p, cbe, DONE_QUEUE);
772 /* Wait for all running I/O to be finished */
773 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
774 TAILQ_FIRST(&ki->kaio_jobqueue)) {
775 ki->kaio_flags |= KAIO_WAKEUP;
776 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
780 /* Free all completed I/O requests. */
781 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
784 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
785 if (lj->lioj_count == 0) {
786 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
787 knlist_delete(&lj->klist, curthread, 1);
789 sigqueue_take(&lj->lioj_ksi);
791 uma_zfree(aiolio_zone, lj);
793 panic("LIO job not cleaned up: C:%d, FC:%d\n",
794 lj->lioj_count, lj->lioj_finished_count);
798 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
799 mtx_destroy(&ki->kaio_mtx);
800 uma_zfree(kaio_zone, ki);
805 * Select a job to run (called by an AIO daemon).
807 static struct aiocblist *
808 aio_selectjob(struct aiothreadlist *aiop)
810 struct aiocblist *aiocbe;
814 mtx_assert(&aio_job_mtx, MA_OWNED);
815 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
816 userp = aiocbe->userproc;
817 ki = userp->p_aioinfo;
819 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
820 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
821 /* Account for currently active jobs. */
822 ki->kaio_active_count++;
823 aiocbe->jobstate = JOBST_JOBRUNNING;
831 * Move all data to a permanent storage device, this code
832 * simulates fsync syscall.
835 aio_fsync_vnode(struct thread *td, struct vnode *vp)
840 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
842 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
843 if (vp->v_object != NULL) {
844 VM_OBJECT_LOCK(vp->v_object);
845 vm_object_page_clean(vp->v_object, 0, 0, 0);
846 VM_OBJECT_UNLOCK(vp->v_object);
848 error = VOP_FSYNC(vp, MNT_WAIT, td);
851 vn_finished_write(mp);
857 * The AIO processing activity. This is the code that does the I/O request for
858 * the non-physio version of the operations. The normal vn operations are used,
859 * and this code should work in all instances for every type of file, including
860 * pipes, sockets, fifos, and regular files.
862 * XXX I don't think it works well for socket, pipe, and fifo.
865 aio_process(struct aiocblist *aiocbe)
867 struct ucred *td_savedcred;
876 int oublock_st, oublock_end;
877 int inblock_st, inblock_end;
880 td_savedcred = td->td_ucred;
881 td->td_ucred = aiocbe->cred;
882 cb = &aiocbe->uaiocb;
883 fp = aiocbe->fd_file;
885 if (cb->aio_lio_opcode == LIO_SYNC) {
888 if (fp->f_vnode != NULL)
889 error = aio_fsync_vnode(td, fp->f_vnode);
890 cb->_aiocb_private.error = error;
891 cb->_aiocb_private.status = 0;
892 td->td_ucred = td_savedcred;
896 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
897 aiov.iov_len = cb->aio_nbytes;
899 auio.uio_iov = &aiov;
901 auio.uio_offset = cb->aio_offset;
902 auio.uio_resid = cb->aio_nbytes;
903 cnt = cb->aio_nbytes;
904 auio.uio_segflg = UIO_USERSPACE;
907 inblock_st = td->td_ru.ru_inblock;
908 oublock_st = td->td_ru.ru_oublock;
910 * aio_aqueue() acquires a reference to the file that is
911 * released in aio_free_entry().
913 if (cb->aio_lio_opcode == LIO_READ) {
914 auio.uio_rw = UIO_READ;
915 if (auio.uio_resid == 0)
918 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
920 if (fp->f_type == DTYPE_VNODE)
922 auio.uio_rw = UIO_WRITE;
923 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
925 inblock_end = td->td_ru.ru_inblock;
926 oublock_end = td->td_ru.ru_oublock;
928 aiocbe->inputcharge = inblock_end - inblock_st;
929 aiocbe->outputcharge = oublock_end - oublock_st;
931 if ((error) && (auio.uio_resid != cnt)) {
932 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
934 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
936 if (fp->f_type == DTYPE_SOCKET) {
938 if (so->so_options & SO_NOSIGPIPE)
942 PROC_LOCK(aiocbe->userproc);
943 kern_psignal(aiocbe->userproc, SIGPIPE);
944 PROC_UNLOCK(aiocbe->userproc);
949 cnt -= auio.uio_resid;
950 cb->_aiocb_private.error = error;
951 cb->_aiocb_private.status = cnt;
952 td->td_ucred = td_savedcred;
956 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
958 struct aioliojob *lj;
960 struct aiocblist *scb, *scbn;
963 ki = userp->p_aioinfo;
964 AIO_LOCK_ASSERT(ki, MA_OWNED);
968 lj->lioj_finished_count++;
969 if (lj->lioj_count == lj->lioj_finished_count)
972 if (type == DONE_QUEUE) {
973 aiocbe->jobflags |= AIOCBLIST_DONE;
975 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
977 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
978 aiocbe->jobstate = JOBST_JOBFINISHED;
980 if (ki->kaio_flags & KAIO_RUNDOWN)
981 goto notification_done;
983 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
984 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
985 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
987 KNOTE_LOCKED(&aiocbe->klist, 1);
990 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
991 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
992 KNOTE_LOCKED(&lj->klist, 1);
994 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
996 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
997 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
998 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
999 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1004 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1005 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1006 if (aiocbe->fd_file == scb->fd_file &&
1007 aiocbe->seqno < scb->seqno) {
1008 if (--scb->pending == 0) {
1009 mtx_lock(&aio_job_mtx);
1010 scb->jobstate = JOBST_JOBQGLOBAL;
1011 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1012 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1013 aio_kick_nowait(userp);
1014 mtx_unlock(&aio_job_mtx);
1019 if (ki->kaio_flags & KAIO_WAKEUP) {
1020 ki->kaio_flags &= ~KAIO_WAKEUP;
1021 wakeup(&userp->p_aioinfo);
1026 * The AIO daemon, most of the actual work is done in aio_process,
1027 * but the setup (and address space mgmt) is done in this routine.
1030 aio_daemon(void *_id)
1032 struct aiocblist *aiocbe;
1033 struct aiothreadlist *aiop;
1034 struct kaioinfo *ki;
1035 struct proc *curcp, *mycp, *userp;
1036 struct vmspace *myvm, *tmpvm;
1037 struct thread *td = curthread;
1038 int id = (intptr_t)_id;
1041 * Local copies of curproc (cp) and vmspace (myvm)
1044 myvm = mycp->p_vmspace;
1046 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1049 * Allocate and ready the aio control info. There is one aiop structure
1052 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1053 aiop->aiothread = td;
1054 aiop->aiothreadflags = 0;
1056 /* The daemon resides in its own pgrp. */
1057 sys_setsid(td, NULL);
1060 * Wakeup parent process. (Parent sleeps to keep from blasting away
1061 * and creating too many daemons.)
1063 sema_post(&aio_newproc_sem);
1065 mtx_lock(&aio_job_mtx);
1068 * curcp is the current daemon process context.
1069 * userp is the current user process context.
1074 * Take daemon off of free queue
1076 if (aiop->aiothreadflags & AIOP_FREE) {
1077 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1078 aiop->aiothreadflags &= ~AIOP_FREE;
1084 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1085 mtx_unlock(&aio_job_mtx);
1086 userp = aiocbe->userproc;
1089 * Connect to process address space for user program.
1091 if (userp != curcp) {
1093 * Save the current address space that we are
1096 tmpvm = mycp->p_vmspace;
1099 * Point to the new user address space, and
1102 mycp->p_vmspace = userp->p_vmspace;
1103 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1105 /* Activate the new mapping. */
1106 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1109 * If the old address space wasn't the daemons
1110 * own address space, then we need to remove the
1111 * daemon's reference from the other process
1112 * that it was acting on behalf of.
1114 if (tmpvm != myvm) {
1115 vmspace_free(tmpvm);
1120 ki = userp->p_aioinfo;
1122 /* Do the I/O function. */
1123 aio_process(aiocbe);
1125 mtx_lock(&aio_job_mtx);
1126 /* Decrement the active job count. */
1127 ki->kaio_active_count--;
1128 mtx_unlock(&aio_job_mtx);
1131 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1132 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1135 mtx_lock(&aio_job_mtx);
1139 * Disconnect from user address space.
1141 if (curcp != mycp) {
1143 mtx_unlock(&aio_job_mtx);
1145 /* Get the user address space to disconnect from. */
1146 tmpvm = mycp->p_vmspace;
1148 /* Get original address space for daemon. */
1149 mycp->p_vmspace = myvm;
1151 /* Activate the daemon's address space. */
1152 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1154 if (tmpvm == myvm) {
1155 printf("AIOD: vmspace problem -- %d\n",
1159 /* Remove our vmspace reference. */
1160 vmspace_free(tmpvm);
1164 mtx_lock(&aio_job_mtx);
1166 * We have to restart to avoid race, we only sleep if
1167 * no job can be selected, that should be
1173 mtx_assert(&aio_job_mtx, MA_OWNED);
1175 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1176 aiop->aiothreadflags |= AIOP_FREE;
1179 * If daemon is inactive for a long time, allow it to exit,
1180 * thereby freeing resources.
1182 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1184 if (TAILQ_EMPTY(&aio_jobs)) {
1185 if ((aiop->aiothreadflags & AIOP_FREE) &&
1186 (num_aio_procs > target_aio_procs)) {
1187 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1189 mtx_unlock(&aio_job_mtx);
1190 uma_zfree(aiop_zone, aiop);
1191 free_unr(aiod_unr, id);
1193 if (mycp->p_vmspace->vm_refcnt <= 1) {
1194 printf("AIOD: bad vm refcnt for"
1195 " exiting daemon: %d\n",
1196 mycp->p_vmspace->vm_refcnt);
1204 mtx_unlock(&aio_job_mtx);
1205 panic("shouldn't be here\n");
1209 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1210 * AIO daemon modifies its environment itself.
1213 aio_newproc(int *start)
1219 id = alloc_unr(aiod_unr);
1220 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1221 RFNOWAIT, 0, "aiod%d", id);
1224 * Wait until daemon is started.
1226 sema_wait(&aio_newproc_sem);
1227 mtx_lock(&aio_job_mtx);
1231 mtx_unlock(&aio_job_mtx);
1233 free_unr(aiod_unr, id);
1239 * Try the high-performance, low-overhead physio method for eligible
1240 * VCHR devices. This method doesn't use an aio helper thread, and
1241 * thus has very low overhead.
1243 * Assumes that the caller, aio_aqueue(), has incremented the file
1244 * structure's reference count, preventing its deallocation for the
1245 * duration of this call.
1248 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1254 struct kaioinfo *ki;
1255 struct aioliojob *lj;
1258 cb = &aiocbe->uaiocb;
1259 fp = aiocbe->fd_file;
1261 if (fp->f_type != DTYPE_VNODE)
1267 * If its not a disk, we don't want to return a positive error.
1268 * It causes the aio code to not fall through to try the thread
1269 * way when you're talking to a regular file.
1271 if (!vn_isdisk(vp, &error)) {
1272 if (error == ENOTBLK)
1278 if (vp->v_bufobj.bo_bsize == 0)
1281 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1284 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1287 if (cb->aio_nbytes >
1288 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1292 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1295 /* Create and build a buffer header for a transfer. */
1296 bp = (struct buf *)getpbuf(NULL);
1301 ki->kaio_buffer_count++;
1308 * Get a copy of the kva from the physical buffer.
1312 bp->b_bcount = cb->aio_nbytes;
1313 bp->b_bufsize = cb->aio_nbytes;
1314 bp->b_iodone = aio_physwakeup;
1315 bp->b_saveaddr = bp->b_data;
1316 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1317 bp->b_offset = cb->aio_offset;
1318 bp->b_iooffset = cb->aio_offset;
1319 bp->b_blkno = btodb(cb->aio_offset);
1320 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1323 * Bring buffer into kernel space.
1325 if (vmapbuf(bp) < 0) {
1332 bp->b_caller1 = (void *)aiocbe;
1333 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1334 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1335 aiocbe->jobstate = JOBST_JOBQBUF;
1336 cb->_aiocb_private.status = cb->aio_nbytes;
1339 atomic_add_int(&num_queue_count, 1);
1340 atomic_add_int(&num_buf_aio, 1);
1344 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1346 /* Perform transfer. */
1347 dev_strategy(vp->v_rdev, bp);
1353 ki->kaio_buffer_count--;
1363 * Wake up aio requests that may be serviceable now.
1366 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1368 struct aiocblist *cb, *cbn;
1371 SOCKBUF_LOCK_ASSERT(sb);
1372 if (sb == &so->so_snd)
1377 sb->sb_flags &= ~SB_AIO;
1378 mtx_lock(&aio_job_mtx);
1379 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1380 if (opcode == cb->uaiocb.aio_lio_opcode) {
1381 if (cb->jobstate != JOBST_JOBQSOCK)
1382 panic("invalid queue value");
1384 * We don't have actual sockets backend yet,
1385 * so we simply move the requests to the generic
1388 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1389 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1390 aio_kick_nowait(cb->userproc);
1393 mtx_unlock(&aio_job_mtx);
1397 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1401 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1402 * supported by AIO with the old sigevent structure.
1404 nsig->sigev_notify = osig->sigev_notify;
1405 switch (nsig->sigev_notify) {
1409 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1412 nsig->sigev_notify_kqueue =
1413 osig->__sigev_u.__sigev_notify_kqueue;
1414 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1423 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1425 struct oaiocb *ojob;
1428 bzero(kjob, sizeof(struct aiocb));
1429 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1432 ojob = (struct oaiocb *)kjob;
1433 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1437 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1440 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1444 aiocb_fetch_status(struct aiocb *ujob)
1447 return (fuword(&ujob->_aiocb_private.status));
1451 aiocb_fetch_error(struct aiocb *ujob)
1454 return (fuword(&ujob->_aiocb_private.error));
1458 aiocb_store_status(struct aiocb *ujob, long status)
1461 return (suword(&ujob->_aiocb_private.status, status));
1465 aiocb_store_error(struct aiocb *ujob, long error)
1468 return (suword(&ujob->_aiocb_private.error, error));
1472 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1475 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1479 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1482 return (suword(ujobp, (long)ujob));
1485 static struct aiocb_ops aiocb_ops = {
1486 .copyin = aiocb_copyin,
1487 .fetch_status = aiocb_fetch_status,
1488 .fetch_error = aiocb_fetch_error,
1489 .store_status = aiocb_store_status,
1490 .store_error = aiocb_store_error,
1491 .store_kernelinfo = aiocb_store_kernelinfo,
1492 .store_aiocb = aiocb_store_aiocb,
1495 static struct aiocb_ops aiocb_ops_osigevent = {
1496 .copyin = aiocb_copyin_old_sigevent,
1497 .fetch_status = aiocb_fetch_status,
1498 .fetch_error = aiocb_fetch_error,
1499 .store_status = aiocb_store_status,
1500 .store_error = aiocb_store_error,
1501 .store_kernelinfo = aiocb_store_kernelinfo,
1502 .store_aiocb = aiocb_store_aiocb,
1506 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1507 * technique is done in this code.
1510 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1511 int type, struct aiocb_ops *ops)
1513 struct proc *p = td->td_proc;
1516 struct aiocblist *aiocbe, *cb;
1517 struct kaioinfo *ki;
1526 if (p->p_aioinfo == NULL)
1527 aio_init_aioinfo(p);
1531 ops->store_status(job, -1);
1532 ops->store_error(job, 0);
1533 ops->store_kernelinfo(job, -1);
1535 if (num_queue_count >= max_queue_count ||
1536 ki->kaio_count >= ki->kaio_qallowed_count) {
1537 ops->store_error(job, EAGAIN);
1541 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1542 aiocbe->inputcharge = 0;
1543 aiocbe->outputcharge = 0;
1544 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1546 error = ops->copyin(job, &aiocbe->uaiocb);
1548 ops->store_error(job, error);
1549 uma_zfree(aiocb_zone, aiocbe);
1553 /* XXX: aio_nbytes is later casted to signed types. */
1554 if (aiocbe->uaiocb.aio_nbytes > INT_MAX) {
1555 uma_zfree(aiocb_zone, aiocbe);
1559 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1560 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1561 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1562 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1563 ops->store_error(job, EINVAL);
1564 uma_zfree(aiocb_zone, aiocbe);
1568 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1569 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1570 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1571 uma_zfree(aiocb_zone, aiocbe);
1575 ksiginfo_init(&aiocbe->ksi);
1577 /* Save userspace address of the job info. */
1578 aiocbe->uuaiocb = job;
1580 /* Get the opcode. */
1581 if (type != LIO_NOP)
1582 aiocbe->uaiocb.aio_lio_opcode = type;
1583 opcode = aiocbe->uaiocb.aio_lio_opcode;
1586 * Validate the opcode and fetch the file object for the specified
1589 * XXXRW: Moved the opcode validation up here so that we don't
1590 * retrieve a file descriptor without knowing what the capabiltity
1593 fd = aiocbe->uaiocb.aio_fildes;
1596 error = fget_write(td, fd, CAP_WRITE | CAP_SEEK, &fp);
1599 error = fget_read(td, fd, CAP_READ | CAP_SEEK, &fp);
1602 error = fget(td, fd, CAP_FSYNC, &fp);
1605 error = fget(td, fd, 0, &fp);
1611 uma_zfree(aiocb_zone, aiocbe);
1612 ops->store_error(job, error);
1616 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1621 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1626 aiocbe->fd_file = fp;
1628 mtx_lock(&aio_job_mtx);
1630 aiocbe->seqno = jobseqno++;
1631 mtx_unlock(&aio_job_mtx);
1632 error = ops->store_kernelinfo(job, jid);
1637 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1639 if (opcode == LIO_NOP) {
1641 uma_zfree(aiocb_zone, aiocbe);
1645 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1647 evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1648 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1652 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1653 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1654 kev.filter = EVFILT_AIO;
1655 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1656 kev.data = (intptr_t)aiocbe;
1657 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1658 error = kqfd_register(kqfd, &kev, td, 1);
1662 uma_zfree(aiocb_zone, aiocbe);
1663 ops->store_error(job, error);
1668 ops->store_error(job, EINPROGRESS);
1669 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1670 aiocbe->userproc = p;
1671 aiocbe->cred = crhold(td->td_ucred);
1672 aiocbe->jobflags = 0;
1675 if (opcode == LIO_SYNC)
1678 if (fp->f_type == DTYPE_SOCKET) {
1680 * Alternate queueing for socket ops: Reach down into the
1681 * descriptor to get the socket data. Then check to see if the
1682 * socket is ready to be read or written (based on the requested
1685 * If it is not ready for io, then queue the aiocbe on the
1686 * socket, and set the flags so we get a call when sbnotify()
1689 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1690 * and unlock the snd sockbuf for no reason.
1693 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1695 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1696 LIO_WRITE) && (!sowriteable(so)))) {
1697 sb->sb_flags |= SB_AIO;
1699 mtx_lock(&aio_job_mtx);
1700 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1701 mtx_unlock(&aio_job_mtx);
1704 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1705 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1706 aiocbe->jobstate = JOBST_JOBQSOCK;
1712 atomic_add_int(&num_queue_count, 1);
1719 if ((error = aio_qphysio(p, aiocbe)) == 0)
1723 aiocbe->uaiocb._aiocb_private.error = error;
1724 ops->store_error(job, error);
1729 /* No buffer for daemon I/O. */
1731 atomic_add_int(&num_queue_count, 1);
1737 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1738 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1739 if (opcode == LIO_SYNC) {
1740 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1741 if (cb->fd_file == aiocbe->fd_file &&
1742 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1743 cb->seqno < aiocbe->seqno) {
1744 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1748 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1749 if (cb->fd_file == aiocbe->fd_file &&
1750 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1751 cb->seqno < aiocbe->seqno) {
1752 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1756 if (aiocbe->pending != 0) {
1757 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1758 aiocbe->jobstate = JOBST_JOBQSYNC;
1763 mtx_lock(&aio_job_mtx);
1764 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1765 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1767 mtx_unlock(&aio_job_mtx);
1775 aio_kick_nowait(struct proc *userp)
1777 struct kaioinfo *ki = userp->p_aioinfo;
1778 struct aiothreadlist *aiop;
1780 mtx_assert(&aio_job_mtx, MA_OWNED);
1781 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1782 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1783 aiop->aiothreadflags &= ~AIOP_FREE;
1784 wakeup(aiop->aiothread);
1785 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1786 ((ki->kaio_active_count + num_aio_resv_start) <
1787 ki->kaio_maxactive_count)) {
1788 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1793 aio_kick(struct proc *userp)
1795 struct kaioinfo *ki = userp->p_aioinfo;
1796 struct aiothreadlist *aiop;
1799 mtx_assert(&aio_job_mtx, MA_OWNED);
1801 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1802 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1803 aiop->aiothreadflags &= ~AIOP_FREE;
1804 wakeup(aiop->aiothread);
1805 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1806 ((ki->kaio_active_count + num_aio_resv_start) <
1807 ki->kaio_maxactive_count)) {
1808 num_aio_resv_start++;
1809 mtx_unlock(&aio_job_mtx);
1810 error = aio_newproc(&num_aio_resv_start);
1811 mtx_lock(&aio_job_mtx);
1813 num_aio_resv_start--;
1823 aio_kick_helper(void *context, int pending)
1825 struct proc *userp = context;
1827 mtx_lock(&aio_job_mtx);
1828 while (--pending >= 0) {
1829 if (aio_kick(userp))
1832 mtx_unlock(&aio_job_mtx);
1836 * Support the aio_return system call, as a side-effect, kernel resources are
1840 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1842 struct proc *p = td->td_proc;
1843 struct aiocblist *cb;
1844 struct kaioinfo *ki;
1851 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1852 if (cb->uuaiocb == uaiocb)
1856 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1857 status = cb->uaiocb._aiocb_private.status;
1858 error = cb->uaiocb._aiocb_private.error;
1859 td->td_retval[0] = status;
1860 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1861 td->td_ru.ru_oublock += cb->outputcharge;
1862 cb->outputcharge = 0;
1863 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1864 td->td_ru.ru_inblock += cb->inputcharge;
1865 cb->inputcharge = 0;
1869 ops->store_error(uaiocb, error);
1870 ops->store_status(uaiocb, status);
1879 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1882 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1886 * Allow a process to wakeup when any of the I/O requests are completed.
1889 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1890 struct timespec *ts)
1892 struct proc *p = td->td_proc;
1894 struct kaioinfo *ki;
1895 struct aiocblist *cb, *cbfirst;
1900 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1903 TIMESPEC_TO_TIMEVAL(&atv, ts);
1904 if (itimerfix(&atv))
1906 timo = tvtohz(&atv);
1920 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1921 for (i = 0; i < njoblist; i++) {
1922 if (cb->uuaiocb == ujoblist[i]) {
1923 if (cbfirst == NULL)
1925 if (cb->jobstate == JOBST_JOBFINISHED)
1930 /* All tasks were finished. */
1931 if (cbfirst == NULL)
1934 ki->kaio_flags |= KAIO_WAKEUP;
1935 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1937 if (error == ERESTART)
1948 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1950 struct timespec ts, *tsp;
1951 struct aiocb **ujoblist;
1954 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1958 /* Get timespec struct. */
1959 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1965 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1966 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1968 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1969 uma_zfree(aiol_zone, ujoblist);
1974 * aio_cancel cancels any non-physio aio operations not currently in
1978 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1980 struct proc *p = td->td_proc;
1981 struct kaioinfo *ki;
1982 struct aiocblist *cbe, *cbn;
1988 int notcancelled = 0;
1991 /* Lookup file object. */
1992 error = fget(td, uap->fd, 0, &fp);
2000 if (fp->f_type == DTYPE_VNODE) {
2002 if (vn_isdisk(vp, &error)) {
2004 td->td_retval[0] = AIO_NOTCANCELED;
2010 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
2011 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
2012 ((uap->aiocbp == NULL) ||
2013 (uap->aiocbp == cbe->uuaiocb))) {
2016 mtx_lock(&aio_job_mtx);
2017 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
2018 TAILQ_REMOVE(&aio_jobs, cbe, list);
2020 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
2021 MPASS(fp->f_type == DTYPE_SOCKET);
2023 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2025 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
2026 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2029 mtx_unlock(&aio_job_mtx);
2032 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2033 cbe->uaiocb._aiocb_private.status = -1;
2034 cbe->uaiocb._aiocb_private.error = ECANCELED;
2035 aio_bio_done_notify(p, cbe, DONE_QUEUE);
2040 if (uap->aiocbp != NULL)
2049 if (uap->aiocbp != NULL) {
2051 td->td_retval[0] = AIO_CANCELED;
2057 td->td_retval[0] = AIO_NOTCANCELED;
2062 td->td_retval[0] = AIO_CANCELED;
2066 td->td_retval[0] = AIO_ALLDONE;
2072 * aio_error is implemented in the kernel level for compatibility purposes
2073 * only. For a user mode async implementation, it would be best to do it in
2074 * a userland subroutine.
2077 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2079 struct proc *p = td->td_proc;
2080 struct aiocblist *cb;
2081 struct kaioinfo *ki;
2086 td->td_retval[0] = EINVAL;
2091 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2092 if (cb->uuaiocb == aiocbp) {
2093 if (cb->jobstate == JOBST_JOBFINISHED)
2095 cb->uaiocb._aiocb_private.error;
2097 td->td_retval[0] = EINPROGRESS;
2105 * Hack for failure of aio_aqueue.
2107 status = ops->fetch_status(aiocbp);
2109 td->td_retval[0] = ops->fetch_error(aiocbp);
2113 td->td_retval[0] = EINVAL;
2118 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2121 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2124 /* syscall - asynchronous read from a file (REALTIME) */
2126 sys_oaio_read(struct thread *td, struct oaio_read_args *uap)
2129 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2130 &aiocb_ops_osigevent));
2134 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2137 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2140 /* syscall - asynchronous write to a file (REALTIME) */
2142 sys_oaio_write(struct thread *td, struct oaio_write_args *uap)
2145 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2146 &aiocb_ops_osigevent));
2150 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2153 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2157 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2158 struct aiocb **acb_list, int nent, struct sigevent *sig,
2159 struct aiocb_ops *ops)
2161 struct proc *p = td->td_proc;
2163 struct kaioinfo *ki;
2164 struct aioliojob *lj;
2170 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2173 if (nent < 0 || nent > AIO_LISTIO_MAX)
2176 if (p->p_aioinfo == NULL)
2177 aio_init_aioinfo(p);
2181 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2184 lj->lioj_finished_count = 0;
2185 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2186 ksiginfo_init(&lj->lioj_ksi);
2191 if (sig && (mode == LIO_NOWAIT)) {
2192 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2193 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2194 /* Assume only new style KEVENT */
2195 kev.filter = EVFILT_LIO;
2196 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2197 kev.ident = (uintptr_t)uacb_list; /* something unique */
2198 kev.data = (intptr_t)lj;
2199 /* pass user defined sigval data */
2200 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2201 error = kqfd_register(
2202 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2204 uma_zfree(aiolio_zone, lj);
2207 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2209 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2210 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2211 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2212 uma_zfree(aiolio_zone, lj);
2215 lj->lioj_flags |= LIOJ_SIGNAL;
2217 uma_zfree(aiolio_zone, lj);
2223 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2225 * Add extra aiocb count to avoid the lio to be freed
2226 * by other threads doing aio_waitcomplete or aio_return,
2227 * and prevent event from being sent until we have queued
2234 * Get pointers to the list of I/O requests.
2237 for (i = 0; i < nent; i++) {
2240 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2248 if (mode == LIO_WAIT) {
2249 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2250 ki->kaio_flags |= KAIO_WAKEUP;
2251 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2252 PRIBIO | PCATCH, "aiospn", 0);
2253 if (error == ERESTART)
2259 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2260 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2261 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2262 KNOTE_LOCKED(&lj->klist, 1);
2264 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2266 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2267 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2268 aio_sendsig(p, &lj->lioj_signal,
2270 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2275 if (lj->lioj_count == 0) {
2276 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2277 knlist_delete(&lj->klist, curthread, 1);
2279 sigqueue_take(&lj->lioj_ksi);
2282 uma_zfree(aiolio_zone, lj);
2291 /* syscall - list directed I/O (REALTIME) */
2293 sys_olio_listio(struct thread *td, struct olio_listio_args *uap)
2295 struct aiocb **acb_list;
2296 struct sigevent *sigp, sig;
2297 struct osigevent osig;
2300 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2304 if (nent < 0 || nent > AIO_LISTIO_MAX)
2307 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2308 error = copyin(uap->sig, &osig, sizeof(osig));
2311 error = convert_old_sigevent(&osig, &sig);
2318 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2319 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2321 error = kern_lio_listio(td, uap->mode,
2322 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2323 &aiocb_ops_osigevent);
2324 free(acb_list, M_LIO);
2328 /* syscall - list directed I/O (REALTIME) */
2330 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2332 struct aiocb **acb_list;
2333 struct sigevent *sigp, sig;
2336 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2340 if (nent < 0 || nent > AIO_LISTIO_MAX)
2343 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2344 error = copyin(uap->sig, &sig, sizeof(sig));
2351 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2352 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2354 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2355 nent, sigp, &aiocb_ops);
2356 free(acb_list, M_LIO);
2361 * Called from interrupt thread for physio, we should return as fast
2362 * as possible, so we schedule a biohelper task.
2365 aio_physwakeup(struct buf *bp)
2367 struct aiocblist *aiocbe;
2369 aiocbe = (struct aiocblist *)bp->b_caller1;
2370 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2374 * Task routine to perform heavy tasks, process wakeup, and signals.
2377 biohelper(void *context, int pending)
2379 struct aiocblist *aiocbe = context;
2382 struct kaioinfo *ki;
2386 userp = aiocbe->userproc;
2387 ki = userp->p_aioinfo;
2389 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2390 aiocbe->uaiocb._aiocb_private.error = 0;
2391 if (bp->b_ioflags & BIO_ERROR)
2392 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2393 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2394 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2395 aiocbe->outputcharge += nblks;
2397 aiocbe->inputcharge += nblks;
2399 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2400 ki->kaio_buffer_count--;
2401 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2404 /* Release mapping into kernel space. */
2407 atomic_subtract_int(&num_buf_aio, 1);
2410 /* syscall - wait for the next completion of an aio request */
2412 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2413 struct timespec *ts, struct aiocb_ops *ops)
2415 struct proc *p = td->td_proc;
2417 struct kaioinfo *ki;
2418 struct aiocblist *cb;
2419 struct aiocb *uuaiocb;
2420 int error, status, timo;
2422 ops->store_aiocb(aiocbp, NULL);
2426 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2429 TIMESPEC_TO_TIMEVAL(&atv, ts);
2430 if (itimerfix(&atv))
2432 timo = tvtohz(&atv);
2435 if (p->p_aioinfo == NULL)
2436 aio_init_aioinfo(p);
2442 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2443 ki->kaio_flags |= KAIO_WAKEUP;
2444 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2446 if (timo && error == ERESTART)
2453 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2454 uuaiocb = cb->uuaiocb;
2455 status = cb->uaiocb._aiocb_private.status;
2456 error = cb->uaiocb._aiocb_private.error;
2457 td->td_retval[0] = status;
2458 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2459 td->td_ru.ru_oublock += cb->outputcharge;
2460 cb->outputcharge = 0;
2461 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2462 td->td_ru.ru_inblock += cb->inputcharge;
2463 cb->inputcharge = 0;
2467 ops->store_aiocb(aiocbp, uuaiocb);
2468 ops->store_error(uuaiocb, error);
2469 ops->store_status(uuaiocb, status);
2477 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2479 struct timespec ts, *tsp;
2483 /* Get timespec struct. */
2484 error = copyin(uap->timeout, &ts, sizeof(ts));
2491 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2495 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2496 struct aiocb_ops *ops)
2498 struct proc *p = td->td_proc;
2499 struct kaioinfo *ki;
2501 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2505 aio_init_aioinfo(p);
2506 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2510 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2513 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2516 /* kqueue attach function */
2518 filt_aioattach(struct knote *kn)
2520 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2523 * The aiocbe pointer must be validated before using it, so
2524 * registration is restricted to the kernel; the user cannot
2527 if ((kn->kn_flags & EV_FLAG1) == 0)
2529 kn->kn_ptr.p_aio = aiocbe;
2530 kn->kn_flags &= ~EV_FLAG1;
2532 knlist_add(&aiocbe->klist, kn, 0);
2537 /* kqueue detach function */
2539 filt_aiodetach(struct knote *kn)
2543 knl = &kn->kn_ptr.p_aio->klist;
2544 knl->kl_lock(knl->kl_lockarg);
2545 if (!knlist_empty(knl))
2546 knlist_remove(knl, kn, 1);
2547 knl->kl_unlock(knl->kl_lockarg);
2550 /* kqueue filter function */
2553 filt_aio(struct knote *kn, long hint)
2555 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2557 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2558 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2560 kn->kn_flags |= EV_EOF;
2564 /* kqueue attach function */
2566 filt_lioattach(struct knote *kn)
2568 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2571 * The aioliojob pointer must be validated before using it, so
2572 * registration is restricted to the kernel; the user cannot
2575 if ((kn->kn_flags & EV_FLAG1) == 0)
2577 kn->kn_ptr.p_lio = lj;
2578 kn->kn_flags &= ~EV_FLAG1;
2580 knlist_add(&lj->klist, kn, 0);
2585 /* kqueue detach function */
2587 filt_liodetach(struct knote *kn)
2591 knl = &kn->kn_ptr.p_lio->klist;
2592 knl->kl_lock(knl->kl_lockarg);
2593 if (!knlist_empty(knl))
2594 knlist_remove(knl, kn, 1);
2595 knl->kl_unlock(knl->kl_lockarg);
2598 /* kqueue filter function */
2601 filt_lio(struct knote *kn, long hint)
2603 struct aioliojob * lj = kn->kn_ptr.p_lio;
2605 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2608 #ifdef COMPAT_FREEBSD32
2610 struct __aiocb_private32 {
2613 uint32_t kernelinfo;
2616 typedef struct oaiocb32 {
2617 int aio_fildes; /* File descriptor */
2618 uint64_t aio_offset __packed; /* File offset for I/O */
2619 uint32_t aio_buf; /* I/O buffer in process space */
2620 uint32_t aio_nbytes; /* Number of bytes for I/O */
2621 struct osigevent32 aio_sigevent; /* Signal to deliver */
2622 int aio_lio_opcode; /* LIO opcode */
2623 int aio_reqprio; /* Request priority -- ignored */
2624 struct __aiocb_private32 _aiocb_private;
2627 typedef struct aiocb32 {
2628 int32_t aio_fildes; /* File descriptor */
2629 uint64_t aio_offset __packed; /* File offset for I/O */
2630 uint32_t aio_buf; /* I/O buffer in process space */
2631 uint32_t aio_nbytes; /* Number of bytes for I/O */
2633 uint32_t __spare2__;
2634 int aio_lio_opcode; /* LIO opcode */
2635 int aio_reqprio; /* Request priority -- ignored */
2636 struct __aiocb_private32 _aiocb_private;
2637 struct sigevent32 aio_sigevent; /* Signal to deliver */
2641 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2645 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2646 * supported by AIO with the old sigevent structure.
2648 CP(*osig, *nsig, sigev_notify);
2649 switch (nsig->sigev_notify) {
2653 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2656 nsig->sigev_notify_kqueue =
2657 osig->__sigev_u.__sigev_notify_kqueue;
2658 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2667 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2669 struct oaiocb32 job32;
2672 bzero(kjob, sizeof(struct aiocb));
2673 error = copyin(ujob, &job32, sizeof(job32));
2677 CP(job32, *kjob, aio_fildes);
2678 CP(job32, *kjob, aio_offset);
2679 PTRIN_CP(job32, *kjob, aio_buf);
2680 CP(job32, *kjob, aio_nbytes);
2681 CP(job32, *kjob, aio_lio_opcode);
2682 CP(job32, *kjob, aio_reqprio);
2683 CP(job32, *kjob, _aiocb_private.status);
2684 CP(job32, *kjob, _aiocb_private.error);
2685 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2686 return (convert_old_sigevent32(&job32.aio_sigevent,
2687 &kjob->aio_sigevent));
2691 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig)
2694 CP(*sig32, *sig, sigev_notify);
2695 switch (sig->sigev_notify) {
2698 case SIGEV_THREAD_ID:
2699 CP(*sig32, *sig, sigev_notify_thread_id);
2702 CP(*sig32, *sig, sigev_signo);
2705 CP(*sig32, *sig, sigev_notify_kqueue);
2706 CP(*sig32, *sig, sigev_notify_kevent_flags);
2707 PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr);
2716 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2718 struct aiocb32 job32;
2721 error = copyin(ujob, &job32, sizeof(job32));
2724 CP(job32, *kjob, aio_fildes);
2725 CP(job32, *kjob, aio_offset);
2726 PTRIN_CP(job32, *kjob, aio_buf);
2727 CP(job32, *kjob, aio_nbytes);
2728 CP(job32, *kjob, aio_lio_opcode);
2729 CP(job32, *kjob, aio_reqprio);
2730 CP(job32, *kjob, _aiocb_private.status);
2731 CP(job32, *kjob, _aiocb_private.error);
2732 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2733 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2737 aiocb32_fetch_status(struct aiocb *ujob)
2739 struct aiocb32 *ujob32;
2741 ujob32 = (struct aiocb32 *)ujob;
2742 return (fuword32(&ujob32->_aiocb_private.status));
2746 aiocb32_fetch_error(struct aiocb *ujob)
2748 struct aiocb32 *ujob32;
2750 ujob32 = (struct aiocb32 *)ujob;
2751 return (fuword32(&ujob32->_aiocb_private.error));
2755 aiocb32_store_status(struct aiocb *ujob, long status)
2757 struct aiocb32 *ujob32;
2759 ujob32 = (struct aiocb32 *)ujob;
2760 return (suword32(&ujob32->_aiocb_private.status, status));
2764 aiocb32_store_error(struct aiocb *ujob, long error)
2766 struct aiocb32 *ujob32;
2768 ujob32 = (struct aiocb32 *)ujob;
2769 return (suword32(&ujob32->_aiocb_private.error, error));
2773 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2775 struct aiocb32 *ujob32;
2777 ujob32 = (struct aiocb32 *)ujob;
2778 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2782 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2785 return (suword32(ujobp, (long)ujob));
2788 static struct aiocb_ops aiocb32_ops = {
2789 .copyin = aiocb32_copyin,
2790 .fetch_status = aiocb32_fetch_status,
2791 .fetch_error = aiocb32_fetch_error,
2792 .store_status = aiocb32_store_status,
2793 .store_error = aiocb32_store_error,
2794 .store_kernelinfo = aiocb32_store_kernelinfo,
2795 .store_aiocb = aiocb32_store_aiocb,
2798 static struct aiocb_ops aiocb32_ops_osigevent = {
2799 .copyin = aiocb32_copyin_old_sigevent,
2800 .fetch_status = aiocb32_fetch_status,
2801 .fetch_error = aiocb32_fetch_error,
2802 .store_status = aiocb32_store_status,
2803 .store_error = aiocb32_store_error,
2804 .store_kernelinfo = aiocb32_store_kernelinfo,
2805 .store_aiocb = aiocb32_store_aiocb,
2809 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2812 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2816 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2818 struct timespec32 ts32;
2819 struct timespec ts, *tsp;
2820 struct aiocb **ujoblist;
2821 uint32_t *ujoblist32;
2824 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2828 /* Get timespec struct. */
2829 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2831 CP(ts32, ts, tv_sec);
2832 CP(ts32, ts, tv_nsec);
2837 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2838 ujoblist32 = (uint32_t *)ujoblist;
2839 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2840 sizeof(ujoblist32[0]));
2842 for (i = uap->nent; i > 0; i--)
2843 ujoblist[i] = PTRIN(ujoblist32[i]);
2845 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2847 uma_zfree(aiol_zone, ujoblist);
2852 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2855 return (sys_aio_cancel(td, (struct aio_cancel_args *)uap));
2859 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2862 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2866 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2869 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2870 &aiocb32_ops_osigevent));
2874 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2877 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2882 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2885 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2886 &aiocb32_ops_osigevent));
2890 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2893 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2898 freebsd32_aio_waitcomplete(struct thread *td,
2899 struct freebsd32_aio_waitcomplete_args *uap)
2901 struct timespec32 ts32;
2902 struct timespec ts, *tsp;
2906 /* Get timespec struct. */
2907 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2910 CP(ts32, ts, tv_sec);
2911 CP(ts32, ts, tv_nsec);
2916 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2921 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2924 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2929 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2931 struct aiocb **acb_list;
2932 struct sigevent *sigp, sig;
2933 struct osigevent32 osig;
2934 uint32_t *acb_list32;
2937 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2941 if (nent < 0 || nent > AIO_LISTIO_MAX)
2944 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2945 error = copyin(uap->sig, &osig, sizeof(osig));
2948 error = convert_old_sigevent32(&osig, &sig);
2955 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2956 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2958 free(acb_list32, M_LIO);
2961 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2962 for (i = 0; i < nent; i++)
2963 acb_list[i] = PTRIN(acb_list32[i]);
2964 free(acb_list32, M_LIO);
2966 error = kern_lio_listio(td, uap->mode,
2967 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2968 &aiocb32_ops_osigevent);
2969 free(acb_list, M_LIO);
2974 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2976 struct aiocb **acb_list;
2977 struct sigevent *sigp, sig;
2978 struct sigevent32 sig32;
2979 uint32_t *acb_list32;
2982 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2986 if (nent < 0 || nent > AIO_LISTIO_MAX)
2989 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2990 error = copyin(uap->sig, &sig32, sizeof(sig32));
2993 error = convert_sigevent32(&sig32, &sig);
3000 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3001 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3003 free(acb_list32, M_LIO);
3006 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3007 for (i = 0; i < nent; i++)
3008 acb_list[i] = PTRIN(acb_list32[i]);
3009 free(acb_list32, M_LIO);
3011 error = kern_lio_listio(td, uap->mode,
3012 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3014 free(acb_list, M_LIO);