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 <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/malloc.h>
29 #include <sys/eventhandler.h>
30 #include <sys/sysproto.h>
31 #include <sys/filedesc.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/kthread.h>
35 #include <sys/fcntl.h>
37 #include <sys/limits.h>
39 #include <sys/mutex.h>
40 #include <sys/unistd.h>
41 #include <sys/posix4.h>
43 #include <sys/resourcevar.h>
44 #include <sys/signalvar.h>
45 #include <sys/protosw.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/syscall.h>
50 #include <sys/sysent.h>
51 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/vnode.h>
56 #include <sys/event.h>
57 #include <sys/mount.h>
59 #include <machine/atomic.h>
62 #include <vm/vm_extern.h>
64 #include <vm/vm_map.h>
65 #include <vm/vm_object.h>
69 #include "opt_vfs_aio.h"
72 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
73 * overflow. (XXX will be removed soon.)
75 static u_long jobrefid;
78 * Counter for aio_fsync.
80 static uint64_t jobseqno;
83 #define JOBST_JOBQSOCK 1
84 #define JOBST_JOBQGLOBAL 2
85 #define JOBST_JOBRUNNING 3
86 #define JOBST_JOBFINISHED 4
87 #define JOBST_JOBQBUF 5
88 #define JOBST_JOBQSYNC 6
90 #ifndef MAX_AIO_PER_PROC
91 #define MAX_AIO_PER_PROC 32
94 #ifndef MAX_AIO_QUEUE_PER_PROC
95 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
99 #define MAX_AIO_PROCS 32
102 #ifndef MAX_AIO_QUEUE
103 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
106 #ifndef TARGET_AIO_PROCS
107 #define TARGET_AIO_PROCS 4
111 #define MAX_BUF_AIO 16
114 #ifndef AIOD_TIMEOUT_DEFAULT
115 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
118 #ifndef AIOD_LIFETIME_DEFAULT
119 #define AIOD_LIFETIME_DEFAULT (30 * hz)
122 FEATURE(aio, "Asynchronous I/O");
124 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
126 static int max_aio_procs = MAX_AIO_PROCS;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
128 CTLFLAG_RW, &max_aio_procs, 0,
129 "Maximum number of kernel threads to use for handling async IO ");
131 static int num_aio_procs = 0;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
133 CTLFLAG_RD, &num_aio_procs, 0,
134 "Number of presently active kernel threads for async IO");
137 * The code will adjust the actual number of AIO processes towards this
138 * number when it gets a chance.
140 static int target_aio_procs = TARGET_AIO_PROCS;
141 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
142 0, "Preferred number of ready kernel threads for async IO");
144 static int max_queue_count = MAX_AIO_QUEUE;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
146 "Maximum number of aio requests to queue, globally");
148 static int num_queue_count = 0;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
150 "Number of queued aio requests");
152 static int num_buf_aio = 0;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
154 "Number of aio requests presently handled by the buf subsystem");
156 /* Number of async I/O thread in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
160 static int aiod_timeout;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
162 "Timeout value for synchronous aio operations");
164 static int aiod_lifetime;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
166 "Maximum lifetime for idle aiod");
168 static int unloadable = 0;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
170 "Allow unload of aio (not recommended)");
173 static int max_aio_per_proc = MAX_AIO_PER_PROC;
174 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
175 0, "Maximum active aio requests per process (stored in the process)");
177 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
178 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
179 &max_aio_queue_per_proc, 0,
180 "Maximum queued aio requests per process (stored in the process)");
182 static int max_buf_aio = MAX_BUF_AIO;
183 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
184 "Maximum buf aio requests per process (stored in the process)");
186 typedef struct oaiocb {
187 int aio_fildes; /* File descriptor */
188 off_t aio_offset; /* File offset for I/O */
189 volatile void *aio_buf; /* I/O buffer in process space */
190 size_t aio_nbytes; /* Number of bytes for I/O */
191 struct osigevent aio_sigevent; /* Signal to deliver */
192 int aio_lio_opcode; /* LIO opcode */
193 int aio_reqprio; /* Request priority -- ignored */
194 struct __aiocb_private _aiocb_private;
198 * Below is a key of locks used to protect each member of struct aiocblist
199 * aioliojob and kaioinfo and any backends.
201 * * - need not protected
202 * a - locked by kaioinfo lock
203 * b - locked by backend lock, the backend lock can be null in some cases,
204 * for example, BIO belongs to this type, in this case, proc lock is
206 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210 * Current, there is only two backends: BIO and generic file I/O.
211 * socket I/O is served by generic file I/O, this is not a good idea, since
212 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
213 * threads, if there is no thread to serve socket I/O, the socket I/O will be
214 * delayed too long or starved, we should create some threads dedicated to
215 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
216 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
217 * structure is not safe because there is race between userland and aio
222 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
223 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
224 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
225 int jobflags; /* (a) job flags */
226 int jobstate; /* (b) job state */
227 int inputcharge; /* (*) input blockes */
228 int outputcharge; /* (*) output blockes */
229 struct buf *bp; /* (*) private to BIO backend,
232 struct proc *userproc; /* (*) user process */
233 struct ucred *cred; /* (*) active credential when created */
234 struct file *fd_file; /* (*) pointer to file structure */
235 struct aioliojob *lio; /* (*) optional lio job */
236 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
237 struct knlist klist; /* (a) list of knotes */
238 struct aiocb uaiocb; /* (*) kernel I/O control block */
239 ksiginfo_t ksi; /* (a) realtime signal info */
240 struct task biotask; /* (*) private to BIO backend */
241 uint64_t seqno; /* (*) job number */
242 int pending; /* (a) number of pending I/O, aio_fsync only */
246 #define AIOCBLIST_DONE 0x01
247 #define AIOCBLIST_BUFDONE 0x02
248 #define AIOCBLIST_RUNDOWN 0x04
249 #define AIOCBLIST_CHECKSYNC 0x08
254 #define AIOP_FREE 0x1 /* proc on free queue */
256 struct aiothreadlist {
257 int aiothreadflags; /* (c) AIO proc flags */
258 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
259 struct thread *aiothread; /* (*) the AIO thread */
263 * data-structure for lio signal management
266 int lioj_flags; /* (a) listio flags */
267 int lioj_count; /* (a) listio flags */
268 int lioj_finished_count; /* (a) listio flags */
269 struct sigevent lioj_signal; /* (a) signal on all I/O done */
270 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
271 struct knlist klist; /* (a) list of knotes */
272 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
275 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
276 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
277 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
280 * per process aio data structure
283 struct mtx kaio_mtx; /* the lock to protect this struct */
284 int kaio_flags; /* (a) per process kaio flags */
285 int kaio_maxactive_count; /* (*) maximum number of AIOs */
286 int kaio_active_count; /* (c) number of currently used AIOs */
287 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
288 int kaio_count; /* (a) size of AIO queue */
289 int kaio_ballowed_count; /* (*) maximum number of buffers */
290 int kaio_buffer_count; /* (a) number of physio buffers */
291 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
292 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
293 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
294 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
295 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
296 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
299 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
300 struct task kaio_task; /* (*) task to kick aio threads */
303 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
304 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
305 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
306 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
308 #define KAIO_RUNDOWN 0x1 /* process is being run down */
309 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
311 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
312 static struct sema aio_newproc_sem;
313 static struct mtx aio_job_mtx;
314 static struct mtx aio_sock_mtx;
315 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
316 static struct unrhdr *aiod_unr;
318 void aio_init_aioinfo(struct proc *p);
319 static void aio_onceonly(void);
320 static int aio_free_entry(struct aiocblist *aiocbe);
321 static void aio_process(struct aiocblist *aiocbe);
322 static int aio_newproc(int *);
323 int aio_aqueue(struct thread *td, struct aiocb *job,
324 struct aioliojob *lio, int type, int osigev);
325 static void aio_physwakeup(struct buf *bp);
326 static void aio_proc_rundown(void *arg, struct proc *p);
327 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
328 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
329 static void biohelper(void *, int);
330 static void aio_daemon(void *param);
331 static void aio_swake_cb(struct socket *, struct sockbuf *);
332 static int aio_unload(void);
333 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
336 static int do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev);
337 static int aio_kick(struct proc *userp);
338 static void aio_kick_nowait(struct proc *userp);
339 static void aio_kick_helper(void *context, int pending);
340 static int filt_aioattach(struct knote *kn);
341 static void filt_aiodetach(struct knote *kn);
342 static int filt_aio(struct knote *kn, long hint);
343 static int filt_lioattach(struct knote *kn);
344 static void filt_liodetach(struct knote *kn);
345 static int filt_lio(struct knote *kn, long hint);
349 * kaio Per process async io info
350 * aiop async io thread data
351 * aiocb async io jobs
352 * aiol list io job pointer - internal to aio_suspend XXX
353 * aiolio list io jobs
355 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
357 /* kqueue filters for aio */
358 static struct filterops aio_filtops =
359 { 0, filt_aioattach, filt_aiodetach, filt_aio };
360 static struct filterops lio_filtops =
361 { 0, filt_lioattach, filt_liodetach, filt_lio };
363 static eventhandler_tag exit_tag, exec_tag;
365 TASKQUEUE_DEFINE_THREAD(aiod_bio);
368 * Main operations function for use as a kernel module.
371 aio_modload(struct module *module, int cmd, void *arg)
380 error = aio_unload();
391 static moduledata_t aio_mod = {
397 SYSCALL_MODULE_HELPER(aio_cancel);
398 SYSCALL_MODULE_HELPER(aio_error);
399 SYSCALL_MODULE_HELPER(aio_fsync);
400 SYSCALL_MODULE_HELPER(aio_read);
401 SYSCALL_MODULE_HELPER(aio_return);
402 SYSCALL_MODULE_HELPER(aio_suspend);
403 SYSCALL_MODULE_HELPER(aio_waitcomplete);
404 SYSCALL_MODULE_HELPER(aio_write);
405 SYSCALL_MODULE_HELPER(lio_listio);
406 SYSCALL_MODULE_HELPER(oaio_read);
407 SYSCALL_MODULE_HELPER(oaio_write);
408 SYSCALL_MODULE_HELPER(olio_listio);
410 DECLARE_MODULE(aio, aio_mod,
411 SI_SUB_VFS, SI_ORDER_ANY);
412 MODULE_VERSION(aio, 1);
415 * Startup initialization
421 /* XXX: should probably just use so->callback */
422 aio_swake = &aio_swake_cb;
423 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
424 EVENTHANDLER_PRI_ANY);
425 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
426 EVENTHANDLER_PRI_ANY);
427 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
428 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
429 TAILQ_INIT(&aio_freeproc);
430 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
431 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
432 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
433 TAILQ_INIT(&aio_jobs);
434 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
435 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
436 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
437 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
438 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
439 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
440 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
441 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
442 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
443 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
444 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
445 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
446 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
448 async_io_version = _POSIX_VERSION;
449 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
450 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
451 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
455 * Callback for unload of AIO when used as a module.
463 * XXX: no unloads by default, it's too dangerous.
464 * perhaps we could do it if locked out callers and then
465 * did an aio_proc_rundown() on each process.
467 * jhb: aio_proc_rundown() needs to run on curproc though,
468 * so I don't think that would fly.
473 error = kqueue_del_filteropts(EVFILT_AIO);
476 error = kqueue_del_filteropts(EVFILT_LIO);
479 async_io_version = 0;
481 taskqueue_free(taskqueue_aiod_bio);
482 delete_unrhdr(aiod_unr);
483 uma_zdestroy(kaio_zone);
484 uma_zdestroy(aiop_zone);
485 uma_zdestroy(aiocb_zone);
486 uma_zdestroy(aiol_zone);
487 uma_zdestroy(aiolio_zone);
488 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
489 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
490 mtx_destroy(&aio_job_mtx);
491 mtx_destroy(&aio_sock_mtx);
492 sema_destroy(&aio_newproc_sem);
493 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
494 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
495 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
500 * Init the per-process aioinfo structure. The aioinfo limits are set
501 * per-process for user limit (resource) management.
504 aio_init_aioinfo(struct proc *p)
508 ki = uma_zalloc(kaio_zone, M_WAITOK);
509 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
511 ki->kaio_maxactive_count = max_aio_per_proc;
512 ki->kaio_active_count = 0;
513 ki->kaio_qallowed_count = max_aio_queue_per_proc;
515 ki->kaio_ballowed_count = max_buf_aio;
516 ki->kaio_buffer_count = 0;
517 TAILQ_INIT(&ki->kaio_all);
518 TAILQ_INIT(&ki->kaio_done);
519 TAILQ_INIT(&ki->kaio_jobqueue);
520 TAILQ_INIT(&ki->kaio_bufqueue);
521 TAILQ_INIT(&ki->kaio_liojoblist);
522 TAILQ_INIT(&ki->kaio_sockqueue);
523 TAILQ_INIT(&ki->kaio_syncqueue);
524 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
526 if (p->p_aioinfo == NULL) {
531 mtx_destroy(&ki->kaio_mtx);
532 uma_zfree(kaio_zone, ki);
535 while (num_aio_procs < target_aio_procs)
540 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
546 ksi->ksi_code = SI_ASYNCIO;
547 ksi->ksi_flags |= KSI_EXT | KSI_INS;
548 ret = psignal_event(p, sigev, ksi);
555 * Free a job entry. Wait for completion if it is currently active, but don't
556 * delay forever. If we delay, we return a flag that says that we have to
557 * restart the queue scan.
560 aio_free_entry(struct aiocblist *aiocbe)
563 struct aioliojob *lj;
566 p = aiocbe->userproc;
571 AIO_LOCK_ASSERT(ki, MA_OWNED);
572 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
574 atomic_subtract_int(&num_queue_count, 1);
577 MPASS(ki->kaio_count >= 0);
579 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
580 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
585 lj->lioj_finished_count--;
587 if (lj->lioj_count == 0) {
588 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
589 /* lio is going away, we need to destroy any knotes */
590 knlist_delete(&lj->klist, curthread, 1);
592 sigqueue_take(&lj->lioj_ksi);
594 uma_zfree(aiolio_zone, lj);
598 /* aiocbe is going away, we need to destroy any knotes */
599 knlist_delete(&aiocbe->klist, curthread, 1);
601 sigqueue_take(&aiocbe->ksi);
604 MPASS(aiocbe->bp == NULL);
605 aiocbe->jobstate = JOBST_NULL;
609 * The thread argument here is used to find the owning process
610 * and is also passed to fo_close() which may pass it to various
611 * places such as devsw close() routines. Because of that, we
612 * need a thread pointer from the process owning the job that is
613 * persistent and won't disappear out from under us or move to
616 * Currently, all the callers of this function call it to remove
617 * an aiocblist from the current process' job list either via a
618 * syscall or due to the current process calling exit() or
619 * execve(). Thus, we know that p == curproc. We also know that
620 * curthread can't exit since we are curthread.
622 * Therefore, we use curthread as the thread to pass to
623 * knlist_delete(). This does mean that it is possible for the
624 * thread pointer at close time to differ from the thread pointer
625 * at open time, but this is already true of file descriptors in
626 * a multithreaded process.
628 fdrop(aiocbe->fd_file, curthread);
629 crfree(aiocbe->cred);
630 uma_zfree(aiocb_zone, aiocbe);
637 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
639 aio_proc_rundown(arg, p);
643 * Rundown the jobs for a given process.
646 aio_proc_rundown(void *arg, struct proc *p)
649 struct aioliojob *lj;
650 struct aiocblist *cbe, *cbn;
655 KASSERT(curthread->td_proc == p,
656 ("%s: called on non-curproc", __func__));
662 ki->kaio_flags |= KAIO_RUNDOWN;
667 * Try to cancel all pending requests. This code simulates
668 * aio_cancel on all pending I/O requests.
670 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
672 mtx_lock(&aio_job_mtx);
673 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
674 TAILQ_REMOVE(&aio_jobs, cbe, list);
676 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
678 MPASS(fp->f_type == DTYPE_SOCKET);
680 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
682 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
683 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
686 mtx_unlock(&aio_job_mtx);
689 cbe->jobstate = JOBST_JOBFINISHED;
690 cbe->uaiocb._aiocb_private.status = -1;
691 cbe->uaiocb._aiocb_private.error = ECANCELED;
692 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
693 aio_bio_done_notify(p, cbe, DONE_QUEUE);
697 /* Wait for all running I/O to be finished */
698 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
699 TAILQ_FIRST(&ki->kaio_jobqueue)) {
700 ki->kaio_flags |= KAIO_WAKEUP;
701 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
705 /* Free all completed I/O requests. */
706 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
709 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
710 if (lj->lioj_count == 0) {
711 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
712 knlist_delete(&lj->klist, curthread, 1);
714 sigqueue_take(&lj->lioj_ksi);
716 uma_zfree(aiolio_zone, lj);
718 panic("LIO job not cleaned up: C:%d, FC:%d\n",
719 lj->lioj_count, lj->lioj_finished_count);
723 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
724 mtx_destroy(&ki->kaio_mtx);
725 uma_zfree(kaio_zone, ki);
730 * Select a job to run (called by an AIO daemon).
732 static struct aiocblist *
733 aio_selectjob(struct aiothreadlist *aiop)
735 struct aiocblist *aiocbe;
739 mtx_assert(&aio_job_mtx, MA_OWNED);
740 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
741 userp = aiocbe->userproc;
742 ki = userp->p_aioinfo;
744 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
745 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
746 /* Account for currently active jobs. */
747 ki->kaio_active_count++;
748 aiocbe->jobstate = JOBST_JOBRUNNING;
756 * Move all data to a permanent storage device, this code
757 * simulates fsync syscall.
760 aio_fsync_vnode(struct thread *td, struct vnode *vp)
766 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
767 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
769 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
770 if (vp->v_object != NULL) {
771 VM_OBJECT_LOCK(vp->v_object);
772 vm_object_page_clean(vp->v_object, 0, 0, 0);
773 VM_OBJECT_UNLOCK(vp->v_object);
775 error = VOP_FSYNC(vp, MNT_WAIT, td);
778 vn_finished_write(mp);
780 VFS_UNLOCK_GIANT(vfslocked);
785 * The AIO processing activity. This is the code that does the I/O request for
786 * the non-physio version of the operations. The normal vn operations are used,
787 * and this code should work in all instances for every type of file, including
788 * pipes, sockets, fifos, and regular files.
790 * XXX I don't think it works well for socket, pipe, and fifo.
793 aio_process(struct aiocblist *aiocbe)
795 struct ucred *td_savedcred;
804 int oublock_st, oublock_end;
805 int inblock_st, inblock_end;
808 td_savedcred = td->td_ucred;
809 td->td_ucred = aiocbe->cred;
810 cb = &aiocbe->uaiocb;
811 fp = aiocbe->fd_file;
813 if (cb->aio_lio_opcode == LIO_SYNC) {
816 if (fp->f_vnode != NULL)
817 error = aio_fsync_vnode(td, fp->f_vnode);
818 cb->_aiocb_private.error = error;
819 cb->_aiocb_private.status = 0;
820 td->td_ucred = td_savedcred;
824 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
825 aiov.iov_len = cb->aio_nbytes;
827 auio.uio_iov = &aiov;
829 auio.uio_offset = cb->aio_offset;
830 auio.uio_resid = cb->aio_nbytes;
831 cnt = cb->aio_nbytes;
832 auio.uio_segflg = UIO_USERSPACE;
835 inblock_st = td->td_ru.ru_inblock;
836 oublock_st = td->td_ru.ru_oublock;
838 * aio_aqueue() acquires a reference to the file that is
839 * released in aio_free_entry().
841 if (cb->aio_lio_opcode == LIO_READ) {
842 auio.uio_rw = UIO_READ;
843 if (auio.uio_resid == 0)
846 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
848 if (fp->f_type == DTYPE_VNODE)
850 auio.uio_rw = UIO_WRITE;
851 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
853 inblock_end = td->td_ru.ru_inblock;
854 oublock_end = td->td_ru.ru_oublock;
856 aiocbe->inputcharge = inblock_end - inblock_st;
857 aiocbe->outputcharge = oublock_end - oublock_st;
859 if ((error) && (auio.uio_resid != cnt)) {
860 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
862 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
864 if (fp->f_type == DTYPE_SOCKET) {
866 if (so->so_options & SO_NOSIGPIPE)
870 PROC_LOCK(aiocbe->userproc);
871 psignal(aiocbe->userproc, SIGPIPE);
872 PROC_UNLOCK(aiocbe->userproc);
877 cnt -= auio.uio_resid;
878 cb->_aiocb_private.error = error;
879 cb->_aiocb_private.status = cnt;
880 td->td_ucred = td_savedcred;
884 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
886 struct aioliojob *lj;
888 struct aiocblist *scb, *scbn;
891 ki = userp->p_aioinfo;
892 AIO_LOCK_ASSERT(ki, MA_OWNED);
896 lj->lioj_finished_count++;
897 if (lj->lioj_count == lj->lioj_finished_count)
900 if (type == DONE_QUEUE) {
901 aiocbe->jobflags |= AIOCBLIST_DONE;
903 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
905 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
906 aiocbe->jobstate = JOBST_JOBFINISHED;
908 if (ki->kaio_flags & KAIO_RUNDOWN)
909 goto notification_done;
911 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
912 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
913 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
915 KNOTE_LOCKED(&aiocbe->klist, 1);
918 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
919 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
920 KNOTE_LOCKED(&lj->klist, 1);
922 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
924 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
925 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
926 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
927 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
932 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
933 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
934 if (aiocbe->fd_file == scb->fd_file &&
935 aiocbe->seqno < scb->seqno) {
936 if (--scb->pending == 0) {
937 mtx_lock(&aio_job_mtx);
938 scb->jobstate = JOBST_JOBQGLOBAL;
939 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
940 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
941 aio_kick_nowait(userp);
942 mtx_unlock(&aio_job_mtx);
947 if (ki->kaio_flags & KAIO_WAKEUP) {
948 ki->kaio_flags &= ~KAIO_WAKEUP;
949 wakeup(&userp->p_aioinfo);
954 * The AIO daemon, most of the actual work is done in aio_process,
955 * but the setup (and address space mgmt) is done in this routine.
958 aio_daemon(void *_id)
960 struct aiocblist *aiocbe;
961 struct aiothreadlist *aiop;
963 struct proc *curcp, *mycp, *userp;
964 struct vmspace *myvm, *tmpvm;
965 struct thread *td = curthread;
966 int id = (intptr_t)_id;
969 * Local copies of curproc (cp) and vmspace (myvm)
972 myvm = mycp->p_vmspace;
974 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
977 * Allocate and ready the aio control info. There is one aiop structure
980 aiop = uma_zalloc(aiop_zone, M_WAITOK);
981 aiop->aiothread = td;
982 aiop->aiothreadflags = 0;
984 /* The daemon resides in its own pgrp. */
988 * Wakeup parent process. (Parent sleeps to keep from blasting away
989 * and creating too many daemons.)
991 sema_post(&aio_newproc_sem);
993 mtx_lock(&aio_job_mtx);
996 * curcp is the current daemon process context.
997 * userp is the current user process context.
1002 * Take daemon off of free queue
1004 if (aiop->aiothreadflags & AIOP_FREE) {
1005 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1006 aiop->aiothreadflags &= ~AIOP_FREE;
1012 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1013 mtx_unlock(&aio_job_mtx);
1014 userp = aiocbe->userproc;
1017 * Connect to process address space for user program.
1019 if (userp != curcp) {
1021 * Save the current address space that we are
1024 tmpvm = mycp->p_vmspace;
1027 * Point to the new user address space, and
1030 mycp->p_vmspace = userp->p_vmspace;
1031 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1033 /* Activate the new mapping. */
1034 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1037 * If the old address space wasn't the daemons
1038 * own address space, then we need to remove the
1039 * daemon's reference from the other process
1040 * that it was acting on behalf of.
1042 if (tmpvm != myvm) {
1043 vmspace_free(tmpvm);
1048 ki = userp->p_aioinfo;
1050 /* Do the I/O function. */
1051 aio_process(aiocbe);
1053 mtx_lock(&aio_job_mtx);
1054 /* Decrement the active job count. */
1055 ki->kaio_active_count--;
1056 mtx_unlock(&aio_job_mtx);
1059 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1060 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1063 mtx_lock(&aio_job_mtx);
1067 * Disconnect from user address space.
1069 if (curcp != mycp) {
1071 mtx_unlock(&aio_job_mtx);
1073 /* Get the user address space to disconnect from. */
1074 tmpvm = mycp->p_vmspace;
1076 /* Get original address space for daemon. */
1077 mycp->p_vmspace = myvm;
1079 /* Activate the daemon's address space. */
1080 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1082 if (tmpvm == myvm) {
1083 printf("AIOD: vmspace problem -- %d\n",
1087 /* Remove our vmspace reference. */
1088 vmspace_free(tmpvm);
1092 mtx_lock(&aio_job_mtx);
1094 * We have to restart to avoid race, we only sleep if
1095 * no job can be selected, that should be
1101 mtx_assert(&aio_job_mtx, MA_OWNED);
1103 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1104 aiop->aiothreadflags |= AIOP_FREE;
1107 * If daemon is inactive for a long time, allow it to exit,
1108 * thereby freeing resources.
1110 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1112 if (TAILQ_EMPTY(&aio_jobs)) {
1113 if ((aiop->aiothreadflags & AIOP_FREE) &&
1114 (num_aio_procs > target_aio_procs)) {
1115 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1117 mtx_unlock(&aio_job_mtx);
1118 uma_zfree(aiop_zone, aiop);
1119 free_unr(aiod_unr, id);
1121 if (mycp->p_vmspace->vm_refcnt <= 1) {
1122 printf("AIOD: bad vm refcnt for"
1123 " exiting daemon: %d\n",
1124 mycp->p_vmspace->vm_refcnt);
1132 mtx_unlock(&aio_job_mtx);
1133 panic("shouldn't be here\n");
1137 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1138 * AIO daemon modifies its environment itself.
1141 aio_newproc(int *start)
1147 id = alloc_unr(aiod_unr);
1148 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1149 RFNOWAIT, 0, "aiod%d", id);
1152 * Wait until daemon is started.
1154 sema_wait(&aio_newproc_sem);
1155 mtx_lock(&aio_job_mtx);
1159 mtx_unlock(&aio_job_mtx);
1161 free_unr(aiod_unr, id);
1167 * Try the high-performance, low-overhead physio method for eligible
1168 * VCHR devices. This method doesn't use an aio helper thread, and
1169 * thus has very low overhead.
1171 * Assumes that the caller, aio_aqueue(), has incremented the file
1172 * structure's reference count, preventing its deallocation for the
1173 * duration of this call.
1176 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1182 struct kaioinfo *ki;
1183 struct aioliojob *lj;
1186 cb = &aiocbe->uaiocb;
1187 fp = aiocbe->fd_file;
1189 if (fp->f_type != DTYPE_VNODE)
1195 * If its not a disk, we don't want to return a positive error.
1196 * It causes the aio code to not fall through to try the thread
1197 * way when you're talking to a regular file.
1199 if (!vn_isdisk(vp, &error)) {
1200 if (error == ENOTBLK)
1206 if (vp->v_bufobj.bo_bsize == 0)
1209 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1212 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1215 if (cb->aio_nbytes >
1216 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1220 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1223 /* Create and build a buffer header for a transfer. */
1224 bp = (struct buf *)getpbuf(NULL);
1229 ki->kaio_buffer_count++;
1236 * Get a copy of the kva from the physical buffer.
1240 bp->b_bcount = cb->aio_nbytes;
1241 bp->b_bufsize = cb->aio_nbytes;
1242 bp->b_iodone = aio_physwakeup;
1243 bp->b_saveaddr = bp->b_data;
1244 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1245 bp->b_offset = cb->aio_offset;
1246 bp->b_iooffset = cb->aio_offset;
1247 bp->b_blkno = btodb(cb->aio_offset);
1248 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1251 * Bring buffer into kernel space.
1253 if (vmapbuf(bp) < 0) {
1260 bp->b_caller1 = (void *)aiocbe;
1261 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1262 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1263 aiocbe->jobstate = JOBST_JOBQBUF;
1264 cb->_aiocb_private.status = cb->aio_nbytes;
1267 atomic_add_int(&num_queue_count, 1);
1268 atomic_add_int(&num_buf_aio, 1);
1272 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1274 /* Perform transfer. */
1275 dev_strategy(vp->v_rdev, bp);
1281 ki->kaio_buffer_count--;
1291 * Wake up aio requests that may be serviceable now.
1294 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1296 struct aiocblist *cb, *cbn;
1299 if (sb == &so->so_snd)
1305 sb->sb_flags &= ~SB_AIO;
1306 mtx_lock(&aio_job_mtx);
1307 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1308 if (opcode == cb->uaiocb.aio_lio_opcode) {
1309 if (cb->jobstate != JOBST_JOBQSOCK)
1310 panic("invalid queue value");
1312 * We don't have actual sockets backend yet,
1313 * so we simply move the requests to the generic
1316 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1317 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1318 aio_kick_nowait(cb->userproc);
1321 mtx_unlock(&aio_job_mtx);
1326 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1327 * technique is done in this code.
1330 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1331 int type, int oldsigev)
1333 struct proc *p = td->td_proc;
1336 struct aiocblist *aiocbe, *cb;
1337 struct kaioinfo *ki;
1345 if (p->p_aioinfo == NULL)
1346 aio_init_aioinfo(p);
1350 suword(&job->_aiocb_private.status, -1);
1351 suword(&job->_aiocb_private.error, 0);
1352 suword(&job->_aiocb_private.kernelinfo, -1);
1354 if (num_queue_count >= max_queue_count ||
1355 ki->kaio_count >= ki->kaio_qallowed_count) {
1356 suword(&job->_aiocb_private.error, EAGAIN);
1360 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1361 aiocbe->inputcharge = 0;
1362 aiocbe->outputcharge = 0;
1363 knlist_init(&aiocbe->klist, AIO_MTX(ki), NULL, NULL, NULL);
1366 bzero(&aiocbe->uaiocb, sizeof(struct aiocb));
1367 error = copyin(job, &aiocbe->uaiocb, sizeof(struct oaiocb));
1368 bcopy(&aiocbe->uaiocb.__spare__, &aiocbe->uaiocb.aio_sigevent,
1369 sizeof(struct osigevent));
1371 error = copyin(job, &aiocbe->uaiocb, sizeof(struct aiocb));
1374 suword(&job->_aiocb_private.error, error);
1375 uma_zfree(aiocb_zone, aiocbe);
1379 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1380 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1381 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1382 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1383 suword(&job->_aiocb_private.error, EINVAL);
1384 uma_zfree(aiocb_zone, aiocbe);
1388 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1389 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1390 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1391 uma_zfree(aiocb_zone, aiocbe);
1395 ksiginfo_init(&aiocbe->ksi);
1397 /* Save userspace address of the job info. */
1398 aiocbe->uuaiocb = job;
1400 /* Get the opcode. */
1401 if (type != LIO_NOP)
1402 aiocbe->uaiocb.aio_lio_opcode = type;
1403 opcode = aiocbe->uaiocb.aio_lio_opcode;
1405 /* Fetch the file object for the specified file descriptor. */
1406 fd = aiocbe->uaiocb.aio_fildes;
1409 error = fget_write(td, fd, &fp);
1412 error = fget_read(td, fd, &fp);
1415 error = fget(td, fd, &fp);
1418 uma_zfree(aiocb_zone, aiocbe);
1419 suword(&job->_aiocb_private.error, error);
1423 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1428 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1433 aiocbe->fd_file = fp;
1435 mtx_lock(&aio_job_mtx);
1437 aiocbe->seqno = jobseqno++;
1438 mtx_unlock(&aio_job_mtx);
1439 error = suword(&job->_aiocb_private.kernelinfo, jid);
1444 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1446 if (opcode == LIO_NOP) {
1448 uma_zfree(aiocb_zone, aiocbe);
1451 if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1452 (opcode != LIO_SYNC)) {
1457 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1459 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1460 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1461 kev.filter = EVFILT_AIO;
1462 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1463 kev.data = (intptr_t)aiocbe;
1464 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1465 error = kqfd_register(kqfd, &kev, td, 1);
1469 uma_zfree(aiocb_zone, aiocbe);
1470 suword(&job->_aiocb_private.error, error);
1475 suword(&job->_aiocb_private.error, EINPROGRESS);
1476 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1477 aiocbe->userproc = p;
1478 aiocbe->cred = crhold(td->td_ucred);
1479 aiocbe->jobflags = 0;
1482 if (opcode == LIO_SYNC)
1485 if (fp->f_type == DTYPE_SOCKET) {
1487 * Alternate queueing for socket ops: Reach down into the
1488 * descriptor to get the socket data. Then check to see if the
1489 * socket is ready to be read or written (based on the requested
1492 * If it is not ready for io, then queue the aiocbe on the
1493 * socket, and set the flags so we get a call when sbnotify()
1496 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1497 * and unlock the snd sockbuf for no reason.
1500 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1502 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1503 LIO_WRITE) && (!sowriteable(so)))) {
1504 sb->sb_flags |= SB_AIO;
1506 mtx_lock(&aio_job_mtx);
1507 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1508 mtx_unlock(&aio_job_mtx);
1511 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1512 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1513 aiocbe->jobstate = JOBST_JOBQSOCK;
1519 atomic_add_int(&num_queue_count, 1);
1526 if ((error = aio_qphysio(p, aiocbe)) == 0)
1530 aiocbe->uaiocb._aiocb_private.error = error;
1531 suword(&job->_aiocb_private.error, error);
1536 /* No buffer for daemon I/O. */
1538 atomic_add_int(&num_queue_count, 1);
1544 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1545 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1546 if (opcode == LIO_SYNC) {
1547 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1548 if (cb->fd_file == aiocbe->fd_file &&
1549 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1550 cb->seqno < aiocbe->seqno) {
1551 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1555 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1556 if (cb->fd_file == aiocbe->fd_file &&
1557 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1558 cb->seqno < aiocbe->seqno) {
1559 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1563 if (aiocbe->pending != 0) {
1564 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1565 aiocbe->jobstate = JOBST_JOBQSYNC;
1570 mtx_lock(&aio_job_mtx);
1571 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1572 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1574 mtx_unlock(&aio_job_mtx);
1582 aio_kick_nowait(struct proc *userp)
1584 struct kaioinfo *ki = userp->p_aioinfo;
1585 struct aiothreadlist *aiop;
1587 mtx_assert(&aio_job_mtx, MA_OWNED);
1588 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1589 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1590 aiop->aiothreadflags &= ~AIOP_FREE;
1591 wakeup(aiop->aiothread);
1592 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1593 ((ki->kaio_active_count + num_aio_resv_start) <
1594 ki->kaio_maxactive_count)) {
1595 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1600 aio_kick(struct proc *userp)
1602 struct kaioinfo *ki = userp->p_aioinfo;
1603 struct aiothreadlist *aiop;
1606 mtx_assert(&aio_job_mtx, MA_OWNED);
1608 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1609 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1610 aiop->aiothreadflags &= ~AIOP_FREE;
1611 wakeup(aiop->aiothread);
1612 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1613 ((ki->kaio_active_count + num_aio_resv_start) <
1614 ki->kaio_maxactive_count)) {
1615 num_aio_resv_start++;
1616 mtx_unlock(&aio_job_mtx);
1617 error = aio_newproc(&num_aio_resv_start);
1618 mtx_lock(&aio_job_mtx);
1620 num_aio_resv_start--;
1630 aio_kick_helper(void *context, int pending)
1632 struct proc *userp = context;
1634 mtx_lock(&aio_job_mtx);
1635 while (--pending >= 0) {
1636 if (aio_kick(userp))
1639 mtx_unlock(&aio_job_mtx);
1643 * Support the aio_return system call, as a side-effect, kernel resources are
1647 aio_return(struct thread *td, struct aio_return_args *uap)
1649 struct proc *p = td->td_proc;
1650 struct aiocblist *cb;
1651 struct aiocb *uaiocb;
1652 struct kaioinfo *ki;
1658 uaiocb = uap->aiocbp;
1660 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1661 if (cb->uuaiocb == uaiocb)
1665 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1666 status = cb->uaiocb._aiocb_private.status;
1667 error = cb->uaiocb._aiocb_private.error;
1668 td->td_retval[0] = status;
1669 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1670 td->td_ru.ru_oublock += cb->outputcharge;
1671 cb->outputcharge = 0;
1672 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1673 td->td_ru.ru_inblock += cb->inputcharge;
1674 cb->inputcharge = 0;
1678 suword(&uaiocb->_aiocb_private.error, error);
1679 suword(&uaiocb->_aiocb_private.status, status);
1688 * Allow a process to wakeup when any of the I/O requests are completed.
1691 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1693 struct proc *p = td->td_proc;
1696 struct aiocb *const *cbptr, *cbp;
1697 struct kaioinfo *ki;
1698 struct aiocblist *cb, *cbfirst;
1699 struct aiocb **ujoblist;
1705 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1710 /* Get timespec struct. */
1711 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1714 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1717 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1718 if (itimerfix(&atv))
1720 timo = tvtohz(&atv);
1728 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1729 cbptr = uap->aiocbp;
1731 for (i = 0; i < uap->nent; i++) {
1732 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1735 ujoblist[njoblist] = cbp;
1739 if (njoblist == 0) {
1740 uma_zfree(aiol_zone, ujoblist);
1748 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1749 for (i = 0; i < njoblist; i++) {
1750 if (cb->uuaiocb == ujoblist[i]) {
1751 if (cbfirst == NULL)
1753 if (cb->jobstate == JOBST_JOBFINISHED)
1758 /* All tasks were finished. */
1759 if (cbfirst == NULL)
1762 ki->kaio_flags |= KAIO_WAKEUP;
1763 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1765 if (error == ERESTART)
1772 uma_zfree(aiol_zone, ujoblist);
1777 * aio_cancel cancels any non-physio aio operations not currently in
1781 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1783 struct proc *p = td->td_proc;
1784 struct kaioinfo *ki;
1785 struct aiocblist *cbe, *cbn;
1791 int notcancelled = 0;
1794 /* Lookup file object. */
1795 error = fget(td, uap->fd, &fp);
1803 if (fp->f_type == DTYPE_VNODE) {
1805 if (vn_isdisk(vp, &error)) {
1807 td->td_retval[0] = AIO_NOTCANCELED;
1813 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1814 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1815 ((uap->aiocbp == NULL) ||
1816 (uap->aiocbp == cbe->uuaiocb))) {
1819 mtx_lock(&aio_job_mtx);
1820 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1821 TAILQ_REMOVE(&aio_jobs, cbe, list);
1823 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
1824 MPASS(fp->f_type == DTYPE_SOCKET);
1826 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1828 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
1829 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
1832 mtx_unlock(&aio_job_mtx);
1835 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1836 cbe->uaiocb._aiocb_private.status = -1;
1837 cbe->uaiocb._aiocb_private.error = ECANCELED;
1838 aio_bio_done_notify(p, cbe, DONE_QUEUE);
1843 if (uap->aiocbp != NULL)
1852 if (uap->aiocbp != NULL) {
1854 td->td_retval[0] = AIO_CANCELED;
1860 td->td_retval[0] = AIO_NOTCANCELED;
1865 td->td_retval[0] = AIO_CANCELED;
1869 td->td_retval[0] = AIO_ALLDONE;
1875 * aio_error is implemented in the kernel level for compatibility purposes
1876 * only. For a user mode async implementation, it would be best to do it in
1877 * a userland subroutine.
1880 aio_error(struct thread *td, struct aio_error_args *uap)
1882 struct proc *p = td->td_proc;
1883 struct aiocblist *cb;
1884 struct kaioinfo *ki;
1889 td->td_retval[0] = EINVAL;
1894 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1895 if (cb->uuaiocb == uap->aiocbp) {
1896 if (cb->jobstate == JOBST_JOBFINISHED)
1898 cb->uaiocb._aiocb_private.error;
1900 td->td_retval[0] = EINPROGRESS;
1908 * Hack for failure of aio_aqueue.
1910 status = fuword(&uap->aiocbp->_aiocb_private.status);
1912 td->td_retval[0] = fuword(&uap->aiocbp->_aiocb_private.error);
1916 td->td_retval[0] = EINVAL;
1920 /* syscall - asynchronous read from a file (REALTIME) */
1922 oaio_read(struct thread *td, struct oaio_read_args *uap)
1925 return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 1);
1929 aio_read(struct thread *td, struct aio_read_args *uap)
1932 return aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, 0);
1935 /* syscall - asynchronous write to a file (REALTIME) */
1937 oaio_write(struct thread *td, struct oaio_write_args *uap)
1940 return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 1);
1944 aio_write(struct thread *td, struct aio_write_args *uap)
1947 return aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, 0);
1950 /* syscall - list directed I/O (REALTIME) */
1952 olio_listio(struct thread *td, struct olio_listio_args *uap)
1954 return do_lio_listio(td, (struct lio_listio_args *)uap, 1);
1957 /* syscall - list directed I/O (REALTIME) */
1959 lio_listio(struct thread *td, struct lio_listio_args *uap)
1961 return do_lio_listio(td, uap, 0);
1965 do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev)
1967 struct proc *p = td->td_proc;
1968 struct aiocb *iocb, * const *cbptr;
1969 struct kaioinfo *ki;
1970 struct aioliojob *lj;
1977 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1981 if (nent < 0 || nent > AIO_LISTIO_MAX)
1984 if (p->p_aioinfo == NULL)
1985 aio_init_aioinfo(p);
1989 lj = uma_zalloc(aiolio_zone, M_WAITOK);
1992 lj->lioj_finished_count = 0;
1993 knlist_init(&lj->klist, AIO_MTX(ki), NULL, NULL, NULL);
1994 ksiginfo_init(&lj->lioj_ksi);
1999 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2000 bzero(&lj->lioj_signal, sizeof(&lj->lioj_signal));
2001 error = copyin(uap->sig, &lj->lioj_signal,
2002 oldsigev ? sizeof(struct osigevent) :
2003 sizeof(struct sigevent));
2005 uma_zfree(aiolio_zone, lj);
2009 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2010 /* Assume only new style KEVENT */
2011 kev.filter = EVFILT_LIO;
2012 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2013 kev.ident = (uintptr_t)uap->acb_list; /* something unique */
2014 kev.data = (intptr_t)lj;
2015 /* pass user defined sigval data */
2016 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2017 error = kqfd_register(
2018 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2020 uma_zfree(aiolio_zone, lj);
2023 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2025 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2026 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2027 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2028 uma_zfree(aiolio_zone, lj);
2031 lj->lioj_flags |= LIOJ_SIGNAL;
2033 uma_zfree(aiolio_zone, lj);
2039 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2041 * Add extra aiocb count to avoid the lio to be freed
2042 * by other threads doing aio_waitcomplete or aio_return,
2043 * and prevent event from being sent until we have queued
2050 * Get pointers to the list of I/O requests.
2053 cbptr = uap->acb_list;
2054 for (i = 0; i < uap->nent; i++) {
2055 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
2056 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
2057 error = aio_aqueue(td, iocb, lj, LIO_NOP, oldsigev);
2065 if (uap->mode == LIO_WAIT) {
2066 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2067 ki->kaio_flags |= KAIO_WAKEUP;
2068 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2069 PRIBIO | PCATCH, "aiospn", 0);
2070 if (error == ERESTART)
2076 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2077 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2078 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2079 KNOTE_LOCKED(&lj->klist, 1);
2081 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2083 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2084 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2085 aio_sendsig(p, &lj->lioj_signal,
2087 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2092 if (lj->lioj_count == 0) {
2093 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2094 knlist_delete(&lj->klist, curthread, 1);
2096 sigqueue_take(&lj->lioj_ksi);
2099 uma_zfree(aiolio_zone, lj);
2109 * Called from interrupt thread for physio, we should return as fast
2110 * as possible, so we schedule a biohelper task.
2113 aio_physwakeup(struct buf *bp)
2115 struct aiocblist *aiocbe;
2117 aiocbe = (struct aiocblist *)bp->b_caller1;
2118 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2122 * Task routine to perform heavy tasks, process wakeup, and signals.
2125 biohelper(void *context, int pending)
2127 struct aiocblist *aiocbe = context;
2130 struct kaioinfo *ki;
2134 userp = aiocbe->userproc;
2135 ki = userp->p_aioinfo;
2137 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2138 aiocbe->uaiocb._aiocb_private.error = 0;
2139 if (bp->b_ioflags & BIO_ERROR)
2140 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2141 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2142 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2143 aiocbe->outputcharge += nblks;
2145 aiocbe->inputcharge += nblks;
2147 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2148 ki->kaio_buffer_count--;
2149 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2152 /* Release mapping into kernel space. */
2155 atomic_subtract_int(&num_buf_aio, 1);
2158 /* syscall - wait for the next completion of an aio request */
2160 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2162 struct proc *p = td->td_proc;
2165 struct kaioinfo *ki;
2166 struct aiocblist *cb;
2167 struct aiocb *uuaiocb;
2168 int error, status, timo;
2170 suword(uap->aiocbp, (long)NULL);
2174 /* Get timespec struct. */
2175 error = copyin(uap->timeout, &ts, sizeof(ts));
2179 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2182 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2183 if (itimerfix(&atv))
2185 timo = tvtohz(&atv);
2188 if (p->p_aioinfo == NULL)
2189 aio_init_aioinfo(p);
2195 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2196 ki->kaio_flags |= KAIO_WAKEUP;
2197 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2199 if (timo && error == ERESTART)
2206 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2207 uuaiocb = cb->uuaiocb;
2208 status = cb->uaiocb._aiocb_private.status;
2209 error = cb->uaiocb._aiocb_private.error;
2210 td->td_retval[0] = status;
2211 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2212 td->td_ru.ru_oublock += cb->outputcharge;
2213 cb->outputcharge = 0;
2214 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2215 td->td_ru.ru_inblock += cb->inputcharge;
2216 cb->inputcharge = 0;
2220 suword(uap->aiocbp, (long)uuaiocb);
2221 suword(&uuaiocb->_aiocb_private.error, error);
2222 suword(&uuaiocb->_aiocb_private.status, status);
2230 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2232 struct proc *p = td->td_proc;
2233 struct kaioinfo *ki;
2235 if (uap->op != O_SYNC) /* XXX lack of O_DSYNC */
2239 aio_init_aioinfo(p);
2240 return aio_aqueue(td, uap->aiocbp, NULL, LIO_SYNC, 0);
2243 /* kqueue attach function */
2245 filt_aioattach(struct knote *kn)
2247 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2250 * The aiocbe pointer must be validated before using it, so
2251 * registration is restricted to the kernel; the user cannot
2254 if ((kn->kn_flags & EV_FLAG1) == 0)
2256 kn->kn_ptr.p_aio = aiocbe;
2257 kn->kn_flags &= ~EV_FLAG1;
2259 knlist_add(&aiocbe->klist, kn, 0);
2264 /* kqueue detach function */
2266 filt_aiodetach(struct knote *kn)
2268 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2270 if (!knlist_empty(&aiocbe->klist))
2271 knlist_remove(&aiocbe->klist, kn, 0);
2274 /* kqueue filter function */
2277 filt_aio(struct knote *kn, long hint)
2279 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2281 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2282 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2284 kn->kn_flags |= EV_EOF;
2288 /* kqueue attach function */
2290 filt_lioattach(struct knote *kn)
2292 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2295 * The aioliojob pointer must be validated before using it, so
2296 * registration is restricted to the kernel; the user cannot
2299 if ((kn->kn_flags & EV_FLAG1) == 0)
2301 kn->kn_ptr.p_lio = lj;
2302 kn->kn_flags &= ~EV_FLAG1;
2304 knlist_add(&lj->klist, kn, 0);
2309 /* kqueue detach function */
2311 filt_liodetach(struct knote *kn)
2313 struct aioliojob * lj = kn->kn_ptr.p_lio;
2315 if (!knlist_empty(&lj->klist))
2316 knlist_remove(&lj->klist, kn, 0);
2319 /* kqueue filter function */
2322 filt_lio(struct knote *kn, long hint)
2324 struct aioliojob * lj = kn->kn_ptr.p_lio;
2326 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);