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 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
124 static int max_aio_procs = MAX_AIO_PROCS;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
126 CTLFLAG_RW, &max_aio_procs, 0,
127 "Maximum number of kernel threads to use for handling async IO ");
129 static int num_aio_procs = 0;
130 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
131 CTLFLAG_RD, &num_aio_procs, 0,
132 "Number of presently active kernel threads for async IO");
135 * The code will adjust the actual number of AIO processes towards this
136 * number when it gets a chance.
138 static int target_aio_procs = TARGET_AIO_PROCS;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
140 0, "Preferred number of ready kernel threads for async IO");
142 static int max_queue_count = MAX_AIO_QUEUE;
143 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
144 "Maximum number of aio requests to queue, globally");
146 static int num_queue_count = 0;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
148 "Number of queued aio requests");
150 static int num_buf_aio = 0;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
152 "Number of aio requests presently handled by the buf subsystem");
154 /* Number of async I/O thread in the process of being started */
155 /* XXX This should be local to aio_aqueue() */
156 static int num_aio_resv_start = 0;
158 static int aiod_timeout;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
160 "Timeout value for synchronous aio operations");
162 static int aiod_lifetime;
163 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
164 "Maximum lifetime for idle aiod");
166 static int unloadable = 0;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
168 "Allow unload of aio (not recommended)");
171 static int max_aio_per_proc = MAX_AIO_PER_PROC;
172 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
173 0, "Maximum active aio requests per process (stored in the process)");
175 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
176 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
177 &max_aio_queue_per_proc, 0,
178 "Maximum queued aio requests per process (stored in the process)");
180 static int max_buf_aio = MAX_BUF_AIO;
181 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
182 "Maximum buf aio requests per process (stored in the process)");
184 typedef struct oaiocb {
185 int aio_fildes; /* File descriptor */
186 off_t aio_offset; /* File offset for I/O */
187 volatile void *aio_buf; /* I/O buffer in process space */
188 size_t aio_nbytes; /* Number of bytes for I/O */
189 struct osigevent aio_sigevent; /* Signal to deliver */
190 int aio_lio_opcode; /* LIO opcode */
191 int aio_reqprio; /* Request priority -- ignored */
192 struct __aiocb_private _aiocb_private;
196 * Below is a key of locks used to protect each member of struct aiocblist
197 * aioliojob and kaioinfo and any backends.
199 * * - need not protected
200 * a - locked by kaioinfo lock
201 * b - locked by backend lock, the backend lock can be null in some cases,
202 * for example, BIO belongs to this type, in this case, proc lock is
204 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
208 * Current, there is only two backends: BIO and generic file I/O.
209 * socket I/O is served by generic file I/O, this is not a good idea, since
210 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
211 * threads, if there is no thread to serve socket I/O, the socket I/O will be
212 * delayed too long or starved, we should create some threads dedicated to
213 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
214 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
215 * structure is not safe because there is race between userland and aio
220 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
221 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
222 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
223 int jobflags; /* (a) job flags */
224 int jobstate; /* (b) job state */
225 int inputcharge; /* (*) input blockes */
226 int outputcharge; /* (*) output blockes */
227 struct buf *bp; /* (*) private to BIO backend,
230 struct proc *userproc; /* (*) user process */
231 struct ucred *cred; /* (*) active credential when created */
232 struct file *fd_file; /* (*) pointer to file structure */
233 struct aioliojob *lio; /* (*) optional lio job */
234 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
235 struct knlist klist; /* (a) list of knotes */
236 struct aiocb uaiocb; /* (*) kernel I/O control block */
237 ksiginfo_t ksi; /* (a) realtime signal info */
238 struct task biotask; /* (*) private to BIO backend */
239 uint64_t seqno; /* (*) job number */
240 int pending; /* (a) number of pending I/O, aio_fsync only */
244 #define AIOCBLIST_DONE 0x01
245 #define AIOCBLIST_BUFDONE 0x02
246 #define AIOCBLIST_RUNDOWN 0x04
247 #define AIOCBLIST_CHECKSYNC 0x08
252 #define AIOP_FREE 0x1 /* proc on free queue */
254 struct aiothreadlist {
255 int aiothreadflags; /* (c) AIO proc flags */
256 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
257 struct thread *aiothread; /* (*) the AIO thread */
261 * data-structure for lio signal management
264 int lioj_flags; /* (a) listio flags */
265 int lioj_count; /* (a) listio flags */
266 int lioj_finished_count; /* (a) listio flags */
267 struct sigevent lioj_signal; /* (a) signal on all I/O done */
268 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
269 struct knlist klist; /* (a) list of knotes */
270 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
273 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
274 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
275 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
278 * per process aio data structure
281 struct mtx kaio_mtx; /* the lock to protect this struct */
282 int kaio_flags; /* (a) per process kaio flags */
283 int kaio_maxactive_count; /* (*) maximum number of AIOs */
284 int kaio_active_count; /* (c) number of currently used AIOs */
285 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
286 int kaio_count; /* (a) size of AIO queue */
287 int kaio_ballowed_count; /* (*) maximum number of buffers */
288 int kaio_buffer_count; /* (a) number of physio buffers */
289 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
290 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
291 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
292 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
293 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
294 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
297 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
298 struct task kaio_task; /* (*) task to kick aio threads */
301 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
302 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
303 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
304 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
306 #define KAIO_RUNDOWN 0x1 /* process is being run down */
307 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
309 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
310 static struct sema aio_newproc_sem;
311 static struct mtx aio_job_mtx;
312 static struct mtx aio_sock_mtx;
313 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
314 static struct unrhdr *aiod_unr;
316 void aio_init_aioinfo(struct proc *p);
317 static void aio_onceonly(void);
318 static int aio_free_entry(struct aiocblist *aiocbe);
319 static void aio_process(struct aiocblist *aiocbe);
320 static int aio_newproc(int *);
321 int aio_aqueue(struct thread *td, struct aiocb *job,
322 struct aioliojob *lio, int type, int osigev);
323 static void aio_physwakeup(struct buf *bp);
324 static void aio_proc_rundown(void *arg, struct proc *p);
325 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
326 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
327 static void biohelper(void *, int);
328 static void aio_daemon(void *param);
329 static void aio_swake_cb(struct socket *, struct sockbuf *);
330 static int aio_unload(void);
331 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
334 static int do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev);
335 static int aio_kick(struct proc *userp);
336 static void aio_kick_nowait(struct proc *userp);
337 static void aio_kick_helper(void *context, int pending);
338 static int filt_aioattach(struct knote *kn);
339 static void filt_aiodetach(struct knote *kn);
340 static int filt_aio(struct knote *kn, long hint);
341 static int filt_lioattach(struct knote *kn);
342 static void filt_liodetach(struct knote *kn);
343 static int filt_lio(struct knote *kn, long hint);
347 * kaio Per process async io info
348 * aiop async io thread data
349 * aiocb async io jobs
350 * aiol list io job pointer - internal to aio_suspend XXX
351 * aiolio list io jobs
353 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
355 /* kqueue filters for aio */
356 static struct filterops aio_filtops =
357 { 0, filt_aioattach, filt_aiodetach, filt_aio };
358 static struct filterops lio_filtops =
359 { 0, filt_lioattach, filt_liodetach, filt_lio };
361 static eventhandler_tag exit_tag, exec_tag;
363 TASKQUEUE_DEFINE_THREAD(aiod_bio);
366 * Main operations function for use as a kernel module.
369 aio_modload(struct module *module, int cmd, void *arg)
378 error = aio_unload();
389 static moduledata_t aio_mod = {
395 SYSCALL_MODULE_HELPER(aio_cancel);
396 SYSCALL_MODULE_HELPER(aio_error);
397 SYSCALL_MODULE_HELPER(aio_fsync);
398 SYSCALL_MODULE_HELPER(aio_read);
399 SYSCALL_MODULE_HELPER(aio_return);
400 SYSCALL_MODULE_HELPER(aio_suspend);
401 SYSCALL_MODULE_HELPER(aio_waitcomplete);
402 SYSCALL_MODULE_HELPER(aio_write);
403 SYSCALL_MODULE_HELPER(lio_listio);
404 SYSCALL_MODULE_HELPER(oaio_read);
405 SYSCALL_MODULE_HELPER(oaio_write);
406 SYSCALL_MODULE_HELPER(olio_listio);
408 DECLARE_MODULE(aio, aio_mod,
409 SI_SUB_VFS, SI_ORDER_ANY);
410 MODULE_VERSION(aio, 1);
413 * Startup initialization
419 /* XXX: should probably just use so->callback */
420 aio_swake = &aio_swake_cb;
421 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
422 EVENTHANDLER_PRI_ANY);
423 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
424 EVENTHANDLER_PRI_ANY);
425 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
426 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
427 TAILQ_INIT(&aio_freeproc);
428 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
429 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
430 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
431 TAILQ_INIT(&aio_jobs);
432 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
433 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
434 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
435 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
436 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
437 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
438 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
439 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
440 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
441 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
442 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
443 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
444 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
446 async_io_version = _POSIX_VERSION;
447 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
448 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
449 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
453 * Callback for unload of AIO when used as a module.
461 * XXX: no unloads by default, it's too dangerous.
462 * perhaps we could do it if locked out callers and then
463 * did an aio_proc_rundown() on each process.
465 * jhb: aio_proc_rundown() needs to run on curproc though,
466 * so I don't think that would fly.
471 error = kqueue_del_filteropts(EVFILT_AIO);
474 error = kqueue_del_filteropts(EVFILT_LIO);
477 async_io_version = 0;
479 taskqueue_free(taskqueue_aiod_bio);
480 delete_unrhdr(aiod_unr);
481 uma_zdestroy(kaio_zone);
482 uma_zdestroy(aiop_zone);
483 uma_zdestroy(aiocb_zone);
484 uma_zdestroy(aiol_zone);
485 uma_zdestroy(aiolio_zone);
486 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
487 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
488 mtx_destroy(&aio_job_mtx);
489 mtx_destroy(&aio_sock_mtx);
490 sema_destroy(&aio_newproc_sem);
491 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
492 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
493 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
498 * Init the per-process aioinfo structure. The aioinfo limits are set
499 * per-process for user limit (resource) management.
502 aio_init_aioinfo(struct proc *p)
506 ki = uma_zalloc(kaio_zone, M_WAITOK);
507 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
509 ki->kaio_maxactive_count = max_aio_per_proc;
510 ki->kaio_active_count = 0;
511 ki->kaio_qallowed_count = max_aio_queue_per_proc;
513 ki->kaio_ballowed_count = max_buf_aio;
514 ki->kaio_buffer_count = 0;
515 TAILQ_INIT(&ki->kaio_all);
516 TAILQ_INIT(&ki->kaio_done);
517 TAILQ_INIT(&ki->kaio_jobqueue);
518 TAILQ_INIT(&ki->kaio_bufqueue);
519 TAILQ_INIT(&ki->kaio_liojoblist);
520 TAILQ_INIT(&ki->kaio_sockqueue);
521 TAILQ_INIT(&ki->kaio_syncqueue);
522 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
524 if (p->p_aioinfo == NULL) {
529 mtx_destroy(&ki->kaio_mtx);
530 uma_zfree(kaio_zone, ki);
533 while (num_aio_procs < target_aio_procs)
538 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
544 ksi->ksi_code = SI_ASYNCIO;
545 ksi->ksi_flags |= KSI_EXT | KSI_INS;
546 ret = psignal_event(p, sigev, ksi);
553 * Free a job entry. Wait for completion if it is currently active, but don't
554 * delay forever. If we delay, we return a flag that says that we have to
555 * restart the queue scan.
558 aio_free_entry(struct aiocblist *aiocbe)
561 struct aioliojob *lj;
564 p = aiocbe->userproc;
569 AIO_LOCK_ASSERT(ki, MA_OWNED);
570 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
572 atomic_subtract_int(&num_queue_count, 1);
575 MPASS(ki->kaio_count >= 0);
577 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
578 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
583 lj->lioj_finished_count--;
585 if (lj->lioj_count == 0) {
586 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
587 /* lio is going away, we need to destroy any knotes */
588 knlist_delete(&lj->klist, curthread, 1);
590 sigqueue_take(&lj->lioj_ksi);
592 uma_zfree(aiolio_zone, lj);
596 /* aiocbe is going away, we need to destroy any knotes */
597 knlist_delete(&aiocbe->klist, curthread, 1);
599 sigqueue_take(&aiocbe->ksi);
602 MPASS(aiocbe->bp == NULL);
603 aiocbe->jobstate = JOBST_NULL;
607 * The thread argument here is used to find the owning process
608 * and is also passed to fo_close() which may pass it to various
609 * places such as devsw close() routines. Because of that, we
610 * need a thread pointer from the process owning the job that is
611 * persistent and won't disappear out from under us or move to
614 * Currently, all the callers of this function call it to remove
615 * an aiocblist from the current process' job list either via a
616 * syscall or due to the current process calling exit() or
617 * execve(). Thus, we know that p == curproc. We also know that
618 * curthread can't exit since we are curthread.
620 * Therefore, we use curthread as the thread to pass to
621 * knlist_delete(). This does mean that it is possible for the
622 * thread pointer at close time to differ from the thread pointer
623 * at open time, but this is already true of file descriptors in
624 * a multithreaded process.
626 fdrop(aiocbe->fd_file, curthread);
627 crfree(aiocbe->cred);
628 uma_zfree(aiocb_zone, aiocbe);
635 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
637 aio_proc_rundown(arg, p);
641 * Rundown the jobs for a given process.
644 aio_proc_rundown(void *arg, struct proc *p)
647 struct aioliojob *lj;
648 struct aiocblist *cbe, *cbn;
653 KASSERT(curthread->td_proc == p,
654 ("%s: called on non-curproc", __func__));
660 ki->kaio_flags |= KAIO_RUNDOWN;
665 * Try to cancel all pending requests. This code simulates
666 * aio_cancel on all pending I/O requests.
668 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
670 mtx_lock(&aio_job_mtx);
671 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
672 TAILQ_REMOVE(&aio_jobs, cbe, list);
674 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
676 MPASS(fp->f_type == DTYPE_SOCKET);
678 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
680 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
681 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
684 mtx_unlock(&aio_job_mtx);
687 cbe->jobstate = JOBST_JOBFINISHED;
688 cbe->uaiocb._aiocb_private.status = -1;
689 cbe->uaiocb._aiocb_private.error = ECANCELED;
690 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
691 aio_bio_done_notify(p, cbe, DONE_QUEUE);
695 /* Wait for all running I/O to be finished */
696 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
697 TAILQ_FIRST(&ki->kaio_jobqueue)) {
698 ki->kaio_flags |= KAIO_WAKEUP;
699 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
703 /* Free all completed I/O requests. */
704 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
707 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
708 if (lj->lioj_count == 0) {
709 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
710 knlist_delete(&lj->klist, curthread, 1);
712 sigqueue_take(&lj->lioj_ksi);
714 uma_zfree(aiolio_zone, lj);
716 panic("LIO job not cleaned up: C:%d, FC:%d\n",
717 lj->lioj_count, lj->lioj_finished_count);
721 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
722 uma_zfree(kaio_zone, ki);
727 * Select a job to run (called by an AIO daemon).
729 static struct aiocblist *
730 aio_selectjob(struct aiothreadlist *aiop)
732 struct aiocblist *aiocbe;
736 mtx_assert(&aio_job_mtx, MA_OWNED);
737 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
738 userp = aiocbe->userproc;
739 ki = userp->p_aioinfo;
741 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
742 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
743 /* Account for currently active jobs. */
744 ki->kaio_active_count++;
745 aiocbe->jobstate = JOBST_JOBRUNNING;
753 * Move all data to a permanent storage device, this code
754 * simulates fsync syscall.
757 aio_fsync_vnode(struct thread *td, struct vnode *vp)
763 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
764 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
766 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
767 if (vp->v_object != NULL) {
768 VM_OBJECT_LOCK(vp->v_object);
769 vm_object_page_clean(vp->v_object, 0, 0, 0);
770 VM_OBJECT_UNLOCK(vp->v_object);
772 error = VOP_FSYNC(vp, MNT_WAIT, td);
774 VOP_UNLOCK(vp, 0, td);
775 vn_finished_write(mp);
777 VFS_UNLOCK_GIANT(vfslocked);
782 * The AIO processing activity. This is the code that does the I/O request for
783 * the non-physio version of the operations. The normal vn operations are used,
784 * and this code should work in all instances for every type of file, including
785 * pipes, sockets, fifos, and regular files.
787 * XXX I don't think it works well for socket, pipe, and fifo.
790 aio_process(struct aiocblist *aiocbe)
792 struct ucred *td_savedcred;
802 int oublock_st, oublock_end;
803 int inblock_st, inblock_end;
806 td_savedcred = td->td_ucred;
807 td->td_ucred = aiocbe->cred;
809 cb = &aiocbe->uaiocb;
810 fp = aiocbe->fd_file;
812 if (cb->aio_lio_opcode == LIO_SYNC) {
815 if (fp->f_vnode != NULL)
816 error = aio_fsync_vnode(td, fp->f_vnode);
817 cb->_aiocb_private.error = error;
818 cb->_aiocb_private.status = 0;
819 td->td_ucred = td_savedcred;
823 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
824 aiov.iov_len = cb->aio_nbytes;
826 auio.uio_iov = &aiov;
828 auio.uio_offset = cb->aio_offset;
829 auio.uio_resid = cb->aio_nbytes;
830 cnt = cb->aio_nbytes;
831 auio.uio_segflg = UIO_USERSPACE;
834 inblock_st = mycp->p_stats->p_ru.ru_inblock;
835 oublock_st = mycp->p_stats->p_ru.ru_oublock;
837 * aio_aqueue() acquires a reference to the file that is
838 * released in aio_free_entry().
840 if (cb->aio_lio_opcode == LIO_READ) {
841 auio.uio_rw = UIO_READ;
842 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
844 if (fp->f_type == DTYPE_VNODE)
846 auio.uio_rw = UIO_WRITE;
847 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
849 inblock_end = mycp->p_stats->p_ru.ru_inblock;
850 oublock_end = mycp->p_stats->p_ru.ru_oublock;
852 aiocbe->inputcharge = inblock_end - inblock_st;
853 aiocbe->outputcharge = oublock_end - oublock_st;
855 if ((error) && (auio.uio_resid != cnt)) {
856 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
858 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
860 if (fp->f_type == DTYPE_SOCKET) {
862 if (so->so_options & SO_NOSIGPIPE)
866 PROC_LOCK(aiocbe->userproc);
867 psignal(aiocbe->userproc, SIGPIPE);
868 PROC_UNLOCK(aiocbe->userproc);
873 cnt -= auio.uio_resid;
874 cb->_aiocb_private.error = error;
875 cb->_aiocb_private.status = cnt;
876 td->td_ucred = td_savedcred;
880 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
882 struct aioliojob *lj;
884 struct aiocblist *scb, *scbn;
887 ki = userp->p_aioinfo;
888 AIO_LOCK_ASSERT(ki, MA_OWNED);
892 lj->lioj_finished_count++;
893 if (lj->lioj_count == lj->lioj_finished_count)
896 if (type == DONE_QUEUE) {
897 aiocbe->jobflags |= AIOCBLIST_DONE;
899 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
901 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
902 aiocbe->jobstate = JOBST_JOBFINISHED;
904 if (ki->kaio_flags & KAIO_RUNDOWN)
905 goto notification_done;
907 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
908 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
909 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
911 KNOTE_LOCKED(&aiocbe->klist, 1);
914 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
915 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
916 KNOTE_LOCKED(&lj->klist, 1);
918 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
920 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
921 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
922 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
923 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
928 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
929 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
930 if (aiocbe->fd_file == scb->fd_file &&
931 aiocbe->seqno < scb->seqno) {
932 if (--scb->pending == 0) {
933 mtx_lock(&aio_job_mtx);
934 scb->jobstate = JOBST_JOBQGLOBAL;
935 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
936 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
937 aio_kick_nowait(userp);
938 mtx_unlock(&aio_job_mtx);
943 if (ki->kaio_flags & KAIO_WAKEUP) {
944 ki->kaio_flags &= ~KAIO_WAKEUP;
945 wakeup(&userp->p_aioinfo);
950 * The AIO daemon, most of the actual work is done in aio_process,
951 * but the setup (and address space mgmt) is done in this routine.
954 aio_daemon(void *_id)
956 struct aiocblist *aiocbe;
957 struct aiothreadlist *aiop;
959 struct proc *curcp, *mycp, *userp;
960 struct vmspace *myvm, *tmpvm;
961 struct thread *td = curthread;
962 int id = (intptr_t)_id;
965 * Local copies of curproc (cp) and vmspace (myvm)
968 myvm = mycp->p_vmspace;
970 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
973 * Allocate and ready the aio control info. There is one aiop structure
976 aiop = uma_zalloc(aiop_zone, M_WAITOK);
977 aiop->aiothread = td;
978 aiop->aiothreadflags = 0;
980 /* The daemon resides in its own pgrp. */
984 * Wakeup parent process. (Parent sleeps to keep from blasting away
985 * and creating too many daemons.)
987 sema_post(&aio_newproc_sem);
989 mtx_lock(&aio_job_mtx);
992 * curcp is the current daemon process context.
993 * userp is the current user process context.
998 * Take daemon off of free queue
1000 if (aiop->aiothreadflags & AIOP_FREE) {
1001 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1002 aiop->aiothreadflags &= ~AIOP_FREE;
1008 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1009 mtx_unlock(&aio_job_mtx);
1010 userp = aiocbe->userproc;
1013 * Connect to process address space for user program.
1015 if (userp != curcp) {
1017 * Save the current address space that we are
1020 tmpvm = mycp->p_vmspace;
1023 * Point to the new user address space, and
1026 mycp->p_vmspace = userp->p_vmspace;
1027 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1029 /* Activate the new mapping. */
1030 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1033 * If the old address space wasn't the daemons
1034 * own address space, then we need to remove the
1035 * daemon's reference from the other process
1036 * that it was acting on behalf of.
1038 if (tmpvm != myvm) {
1039 vmspace_free(tmpvm);
1044 ki = userp->p_aioinfo;
1046 /* Do the I/O function. */
1047 aio_process(aiocbe);
1049 mtx_lock(&aio_job_mtx);
1050 /* Decrement the active job count. */
1051 ki->kaio_active_count--;
1052 mtx_unlock(&aio_job_mtx);
1055 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1056 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1059 mtx_lock(&aio_job_mtx);
1063 * Disconnect from user address space.
1065 if (curcp != mycp) {
1067 mtx_unlock(&aio_job_mtx);
1069 /* Get the user address space to disconnect from. */
1070 tmpvm = mycp->p_vmspace;
1072 /* Get original address space for daemon. */
1073 mycp->p_vmspace = myvm;
1075 /* Activate the daemon's address space. */
1076 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1078 if (tmpvm == myvm) {
1079 printf("AIOD: vmspace problem -- %d\n",
1083 /* Remove our vmspace reference. */
1084 vmspace_free(tmpvm);
1088 mtx_lock(&aio_job_mtx);
1090 * We have to restart to avoid race, we only sleep if
1091 * no job can be selected, that should be
1097 mtx_assert(&aio_job_mtx, MA_OWNED);
1099 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1100 aiop->aiothreadflags |= AIOP_FREE;
1103 * If daemon is inactive for a long time, allow it to exit,
1104 * thereby freeing resources.
1106 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1108 if (TAILQ_EMPTY(&aio_jobs)) {
1109 if ((aiop->aiothreadflags & AIOP_FREE) &&
1110 (num_aio_procs > target_aio_procs)) {
1111 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1113 mtx_unlock(&aio_job_mtx);
1114 uma_zfree(aiop_zone, aiop);
1115 free_unr(aiod_unr, id);
1117 if (mycp->p_vmspace->vm_refcnt <= 1) {
1118 printf("AIOD: bad vm refcnt for"
1119 " exiting daemon: %d\n",
1120 mycp->p_vmspace->vm_refcnt);
1128 mtx_unlock(&aio_job_mtx);
1129 panic("shouldn't be here\n");
1133 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1134 * AIO daemon modifies its environment itself.
1137 aio_newproc(int *start)
1143 id = alloc_unr(aiod_unr);
1144 error = kthread_create(aio_daemon, (void *)(intptr_t)id, &p,
1145 RFNOWAIT, 0, "aiod%d", id);
1148 * Wait until daemon is started.
1150 sema_wait(&aio_newproc_sem);
1151 mtx_lock(&aio_job_mtx);
1155 mtx_unlock(&aio_job_mtx);
1157 free_unr(aiod_unr, id);
1163 * Try the high-performance, low-overhead physio method for eligible
1164 * VCHR devices. This method doesn't use an aio helper thread, and
1165 * thus has very low overhead.
1167 * Assumes that the caller, aio_aqueue(), has incremented the file
1168 * structure's reference count, preventing its deallocation for the
1169 * duration of this call.
1172 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1178 struct kaioinfo *ki;
1179 struct aioliojob *lj;
1182 cb = &aiocbe->uaiocb;
1183 fp = aiocbe->fd_file;
1185 if (fp->f_type != DTYPE_VNODE)
1191 * If its not a disk, we don't want to return a positive error.
1192 * It causes the aio code to not fall through to try the thread
1193 * way when you're talking to a regular file.
1195 if (!vn_isdisk(vp, &error)) {
1196 if (error == ENOTBLK)
1202 if (vp->v_bufobj.bo_bsize == 0)
1205 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1208 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1211 if (cb->aio_nbytes >
1212 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1216 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1219 /* Create and build a buffer header for a transfer. */
1220 bp = (struct buf *)getpbuf(NULL);
1225 ki->kaio_buffer_count++;
1232 * Get a copy of the kva from the physical buffer.
1236 bp->b_bcount = cb->aio_nbytes;
1237 bp->b_bufsize = cb->aio_nbytes;
1238 bp->b_iodone = aio_physwakeup;
1239 bp->b_saveaddr = bp->b_data;
1240 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1241 bp->b_offset = cb->aio_offset;
1242 bp->b_iooffset = cb->aio_offset;
1243 bp->b_blkno = btodb(cb->aio_offset);
1244 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1247 * Bring buffer into kernel space.
1249 if (vmapbuf(bp) < 0) {
1256 bp->b_caller1 = (void *)aiocbe;
1257 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1258 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1259 aiocbe->jobstate = JOBST_JOBQBUF;
1260 cb->_aiocb_private.status = cb->aio_nbytes;
1263 atomic_add_int(&num_queue_count, 1);
1264 atomic_add_int(&num_buf_aio, 1);
1268 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1270 /* Perform transfer. */
1271 dev_strategy(vp->v_rdev, bp);
1277 ki->kaio_buffer_count--;
1287 * Wake up aio requests that may be serviceable now.
1290 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1292 struct aiocblist *cb, *cbn;
1295 if (sb == &so->so_snd)
1301 sb->sb_flags &= ~SB_AIO;
1302 mtx_lock(&aio_job_mtx);
1303 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1304 if (opcode == cb->uaiocb.aio_lio_opcode) {
1305 if (cb->jobstate != JOBST_JOBQSOCK)
1306 panic("invalid queue value");
1308 * We don't have actual sockets backend yet,
1309 * so we simply move the requests to the generic
1312 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1313 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1314 aio_kick_nowait(cb->userproc);
1317 mtx_unlock(&aio_job_mtx);
1322 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1323 * technique is done in this code.
1326 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1327 int type, int oldsigev)
1329 struct proc *p = td->td_proc;
1332 struct aiocblist *aiocbe, *cb;
1333 struct kaioinfo *ki;
1341 if (p->p_aioinfo == NULL)
1342 aio_init_aioinfo(p);
1346 suword(&job->_aiocb_private.status, -1);
1347 suword(&job->_aiocb_private.error, 0);
1348 suword(&job->_aiocb_private.kernelinfo, -1);
1350 if (num_queue_count >= max_queue_count ||
1351 ki->kaio_count >= ki->kaio_qallowed_count) {
1352 suword(&job->_aiocb_private.error, EAGAIN);
1356 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1357 aiocbe->inputcharge = 0;
1358 aiocbe->outputcharge = 0;
1359 knlist_init(&aiocbe->klist, AIO_MTX(ki), NULL, NULL, NULL);
1362 bzero(&aiocbe->uaiocb, sizeof(struct aiocb));
1363 error = copyin(job, &aiocbe->uaiocb, sizeof(struct oaiocb));
1364 bcopy(&aiocbe->uaiocb.__spare__, &aiocbe->uaiocb.aio_sigevent,
1365 sizeof(struct osigevent));
1367 error = copyin(job, &aiocbe->uaiocb, sizeof(struct aiocb));
1370 suword(&job->_aiocb_private.error, error);
1371 uma_zfree(aiocb_zone, aiocbe);
1375 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1376 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1377 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1378 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1379 suword(&job->_aiocb_private.error, EINVAL);
1380 uma_zfree(aiocb_zone, aiocbe);
1384 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1385 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1386 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1387 uma_zfree(aiocb_zone, aiocbe);
1391 ksiginfo_init(&aiocbe->ksi);
1393 /* Save userspace address of the job info. */
1394 aiocbe->uuaiocb = job;
1396 /* Get the opcode. */
1397 if (type != LIO_NOP)
1398 aiocbe->uaiocb.aio_lio_opcode = type;
1399 opcode = aiocbe->uaiocb.aio_lio_opcode;
1401 /* Fetch the file object for the specified file descriptor. */
1402 fd = aiocbe->uaiocb.aio_fildes;
1405 error = fget_write(td, fd, &fp);
1408 error = fget_read(td, fd, &fp);
1411 error = fget(td, fd, &fp);
1414 uma_zfree(aiocb_zone, aiocbe);
1415 suword(&job->_aiocb_private.error, error);
1419 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1424 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1429 aiocbe->fd_file = fp;
1431 mtx_lock(&aio_job_mtx);
1433 aiocbe->seqno = jobseqno++;
1434 mtx_unlock(&aio_job_mtx);
1435 error = suword(&job->_aiocb_private.kernelinfo, jid);
1440 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1442 if (opcode == LIO_NOP) {
1444 uma_zfree(aiocb_zone, aiocbe);
1447 if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1448 (opcode != LIO_SYNC)) {
1453 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1455 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1456 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1457 kev.filter = EVFILT_AIO;
1458 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1459 kev.data = (intptr_t)aiocbe;
1460 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1461 error = kqfd_register(kqfd, &kev, td, 1);
1465 uma_zfree(aiocb_zone, aiocbe);
1466 suword(&job->_aiocb_private.error, error);
1471 suword(&job->_aiocb_private.error, EINPROGRESS);
1472 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1473 aiocbe->userproc = p;
1474 aiocbe->cred = crhold(td->td_ucred);
1475 aiocbe->jobflags = 0;
1478 if (opcode == LIO_SYNC)
1481 if (fp->f_type == DTYPE_SOCKET) {
1483 * Alternate queueing for socket ops: Reach down into the
1484 * descriptor to get the socket data. Then check to see if the
1485 * socket is ready to be read or written (based on the requested
1488 * If it is not ready for io, then queue the aiocbe on the
1489 * socket, and set the flags so we get a call when sbnotify()
1492 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1493 * and unlock the snd sockbuf for no reason.
1496 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1498 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1499 LIO_WRITE) && (!sowriteable(so)))) {
1500 sb->sb_flags |= SB_AIO;
1502 mtx_lock(&aio_job_mtx);
1503 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1504 mtx_unlock(&aio_job_mtx);
1507 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1508 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1509 aiocbe->jobstate = JOBST_JOBQSOCK;
1515 atomic_add_int(&num_queue_count, 1);
1522 if ((error = aio_qphysio(p, aiocbe)) == 0)
1526 aiocbe->uaiocb._aiocb_private.error = error;
1527 suword(&job->_aiocb_private.error, error);
1532 /* No buffer for daemon I/O. */
1534 atomic_add_int(&num_queue_count, 1);
1540 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1541 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1542 if (opcode == LIO_SYNC) {
1543 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1544 if (cb->fd_file == aiocbe->fd_file &&
1545 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1546 cb->seqno < aiocbe->seqno) {
1547 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1551 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1552 if (cb->fd_file == aiocbe->fd_file &&
1553 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1554 cb->seqno < aiocbe->seqno) {
1555 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1559 if (aiocbe->pending != 0) {
1560 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1561 aiocbe->jobstate = JOBST_JOBQSYNC;
1566 mtx_lock(&aio_job_mtx);
1567 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1568 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1570 mtx_unlock(&aio_job_mtx);
1578 aio_kick_nowait(struct proc *userp)
1580 struct kaioinfo *ki = userp->p_aioinfo;
1581 struct aiothreadlist *aiop;
1583 mtx_assert(&aio_job_mtx, MA_OWNED);
1584 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1585 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1586 aiop->aiothreadflags &= ~AIOP_FREE;
1587 wakeup(aiop->aiothread);
1588 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1589 ((ki->kaio_active_count + num_aio_resv_start) <
1590 ki->kaio_maxactive_count)) {
1591 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1596 aio_kick(struct proc *userp)
1598 struct kaioinfo *ki = userp->p_aioinfo;
1599 struct aiothreadlist *aiop;
1602 mtx_assert(&aio_job_mtx, MA_OWNED);
1604 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1605 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1606 aiop->aiothreadflags &= ~AIOP_FREE;
1607 wakeup(aiop->aiothread);
1608 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1609 ((ki->kaio_active_count + num_aio_resv_start) <
1610 ki->kaio_maxactive_count)) {
1611 num_aio_resv_start++;
1612 mtx_unlock(&aio_job_mtx);
1613 error = aio_newproc(&num_aio_resv_start);
1614 mtx_lock(&aio_job_mtx);
1616 num_aio_resv_start--;
1626 aio_kick_helper(void *context, int pending)
1628 struct proc *userp = context;
1630 mtx_lock(&aio_job_mtx);
1631 while (--pending >= 0) {
1632 if (aio_kick(userp))
1635 mtx_unlock(&aio_job_mtx);
1639 * Support the aio_return system call, as a side-effect, kernel resources are
1643 aio_return(struct thread *td, struct aio_return_args *uap)
1645 struct proc *p = td->td_proc;
1646 struct aiocblist *cb;
1647 struct aiocb *uaiocb;
1648 struct kaioinfo *ki;
1654 uaiocb = uap->aiocbp;
1656 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1657 if (cb->uuaiocb == uaiocb)
1661 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1662 status = cb->uaiocb._aiocb_private.status;
1663 error = cb->uaiocb._aiocb_private.error;
1664 td->td_retval[0] = status;
1665 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1666 p->p_stats->p_ru.ru_oublock +=
1668 cb->outputcharge = 0;
1669 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1670 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1671 cb->inputcharge = 0;
1675 suword(&uaiocb->_aiocb_private.error, error);
1676 suword(&uaiocb->_aiocb_private.status, status);
1685 * Allow a process to wakeup when any of the I/O requests are completed.
1688 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1690 struct proc *p = td->td_proc;
1693 struct aiocb *const *cbptr, *cbp;
1694 struct kaioinfo *ki;
1695 struct aiocblist *cb, *cbfirst;
1696 struct aiocb **ujoblist;
1702 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1707 /* Get timespec struct. */
1708 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1711 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1714 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1715 if (itimerfix(&atv))
1717 timo = tvtohz(&atv);
1725 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1726 cbptr = uap->aiocbp;
1728 for (i = 0; i < uap->nent; i++) {
1729 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1732 ujoblist[njoblist] = cbp;
1736 if (njoblist == 0) {
1737 uma_zfree(aiol_zone, ujoblist);
1745 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1746 for (i = 0; i < njoblist; i++) {
1747 if (cb->uuaiocb == ujoblist[i]) {
1748 if (cbfirst == NULL)
1750 if (cb->jobstate == JOBST_JOBFINISHED)
1755 /* All tasks were finished. */
1756 if (cbfirst == NULL)
1759 ki->kaio_flags |= KAIO_WAKEUP;
1760 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1762 if (error == ERESTART)
1769 uma_zfree(aiol_zone, ujoblist);
1774 * aio_cancel cancels any non-physio aio operations not currently in
1778 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1780 struct proc *p = td->td_proc;
1781 struct kaioinfo *ki;
1782 struct aiocblist *cbe, *cbn;
1788 int notcancelled = 0;
1791 /* Lookup file object. */
1792 error = fget(td, uap->fd, &fp);
1800 if (fp->f_type == DTYPE_VNODE) {
1802 if (vn_isdisk(vp, &error)) {
1804 td->td_retval[0] = AIO_NOTCANCELED;
1810 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1811 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1812 ((uap->aiocbp == NULL) ||
1813 (uap->aiocbp == cbe->uuaiocb))) {
1816 mtx_lock(&aio_job_mtx);
1817 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1818 TAILQ_REMOVE(&aio_jobs, cbe, list);
1820 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
1821 MPASS(fp->f_type == DTYPE_SOCKET);
1823 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1825 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
1826 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
1829 mtx_unlock(&aio_job_mtx);
1832 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1833 cbe->uaiocb._aiocb_private.status = -1;
1834 cbe->uaiocb._aiocb_private.error = ECANCELED;
1835 aio_bio_done_notify(p, cbe, DONE_QUEUE);
1840 if (uap->aiocbp != NULL)
1849 if (uap->aiocbp != NULL) {
1851 td->td_retval[0] = AIO_CANCELED;
1857 td->td_retval[0] = AIO_NOTCANCELED;
1862 td->td_retval[0] = AIO_CANCELED;
1866 td->td_retval[0] = AIO_ALLDONE;
1872 * aio_error is implemented in the kernel level for compatibility purposes
1873 * only. For a user mode async implementation, it would be best to do it in
1874 * a userland subroutine.
1877 aio_error(struct thread *td, struct aio_error_args *uap)
1879 struct proc *p = td->td_proc;
1880 struct aiocblist *cb;
1881 struct kaioinfo *ki;
1886 td->td_retval[0] = EINVAL;
1891 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1892 if (cb->uuaiocb == uap->aiocbp) {
1893 if (cb->jobstate == JOBST_JOBFINISHED)
1895 cb->uaiocb._aiocb_private.error;
1897 td->td_retval[0] = EINPROGRESS;
1905 * Hack for failure of aio_aqueue.
1907 status = fuword(&uap->aiocbp->_aiocb_private.status);
1909 td->td_retval[0] = fuword(&uap->aiocbp->_aiocb_private.error);
1913 td->td_retval[0] = EINVAL;
1917 /* syscall - asynchronous read from a file (REALTIME) */
1919 oaio_read(struct thread *td, struct oaio_read_args *uap)
1922 return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 1);
1926 aio_read(struct thread *td, struct aio_read_args *uap)
1929 return aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, 0);
1932 /* syscall - asynchronous write to a file (REALTIME) */
1934 oaio_write(struct thread *td, struct oaio_write_args *uap)
1937 return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 1);
1941 aio_write(struct thread *td, struct aio_write_args *uap)
1944 return aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, 0);
1947 /* syscall - list directed I/O (REALTIME) */
1949 olio_listio(struct thread *td, struct olio_listio_args *uap)
1951 return do_lio_listio(td, (struct lio_listio_args *)uap, 1);
1954 /* syscall - list directed I/O (REALTIME) */
1956 lio_listio(struct thread *td, struct lio_listio_args *uap)
1958 return do_lio_listio(td, uap, 0);
1962 do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev)
1964 struct proc *p = td->td_proc;
1965 struct aiocb *iocb, * const *cbptr;
1966 struct kaioinfo *ki;
1967 struct aioliojob *lj;
1974 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1978 if (nent < 0 || nent > AIO_LISTIO_MAX)
1981 if (p->p_aioinfo == NULL)
1982 aio_init_aioinfo(p);
1986 lj = uma_zalloc(aiolio_zone, M_WAITOK);
1989 lj->lioj_finished_count = 0;
1990 knlist_init(&lj->klist, AIO_MTX(ki), NULL, NULL, NULL);
1991 ksiginfo_init(&lj->lioj_ksi);
1996 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1997 bzero(&lj->lioj_signal, sizeof(&lj->lioj_signal));
1998 error = copyin(uap->sig, &lj->lioj_signal,
1999 oldsigev ? sizeof(struct osigevent) :
2000 sizeof(struct sigevent));
2002 uma_zfree(aiolio_zone, lj);
2006 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2007 /* Assume only new style KEVENT */
2008 kev.filter = EVFILT_LIO;
2009 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2010 kev.ident = (uintptr_t)uap->acb_list; /* something unique */
2011 kev.data = (intptr_t)lj;
2012 /* pass user defined sigval data */
2013 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2014 error = kqfd_register(
2015 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2017 uma_zfree(aiolio_zone, lj);
2020 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2022 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2023 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2024 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2025 uma_zfree(aiolio_zone, lj);
2028 lj->lioj_flags |= LIOJ_SIGNAL;
2030 uma_zfree(aiolio_zone, lj);
2036 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2038 * Add extra aiocb count to avoid the lio to be freed
2039 * by other threads doing aio_waitcomplete or aio_return,
2040 * and prevent event from being sent until we have queued
2047 * Get pointers to the list of I/O requests.
2050 cbptr = uap->acb_list;
2051 for (i = 0; i < uap->nent; i++) {
2052 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
2053 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
2054 error = aio_aqueue(td, iocb, lj, LIO_NOP, oldsigev);
2062 if (uap->mode == LIO_WAIT) {
2063 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2064 ki->kaio_flags |= KAIO_WAKEUP;
2065 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2066 PRIBIO | PCATCH, "aiospn", 0);
2067 if (error == ERESTART)
2073 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2074 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2075 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2076 KNOTE_LOCKED(&lj->klist, 1);
2078 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2080 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2081 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2082 aio_sendsig(p, &lj->lioj_signal,
2084 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2089 if (lj->lioj_count == 0) {
2090 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2091 knlist_delete(&lj->klist, curthread, 1);
2093 sigqueue_take(&lj->lioj_ksi);
2096 uma_zfree(aiolio_zone, lj);
2106 * Called from interrupt thread for physio, we should return as fast
2107 * as possible, so we schedule a biohelper task.
2110 aio_physwakeup(struct buf *bp)
2112 struct aiocblist *aiocbe;
2114 aiocbe = (struct aiocblist *)bp->b_caller1;
2115 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2119 * Task routine to perform heavy tasks, process wakeup, and signals.
2122 biohelper(void *context, int pending)
2124 struct aiocblist *aiocbe = context;
2127 struct kaioinfo *ki;
2131 userp = aiocbe->userproc;
2132 ki = userp->p_aioinfo;
2134 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2135 aiocbe->uaiocb._aiocb_private.error = 0;
2136 if (bp->b_ioflags & BIO_ERROR)
2137 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2138 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2139 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2140 aiocbe->outputcharge += nblks;
2142 aiocbe->inputcharge += nblks;
2144 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2145 ki->kaio_buffer_count--;
2146 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2149 /* Release mapping into kernel space. */
2152 atomic_subtract_int(&num_buf_aio, 1);
2155 /* syscall - wait for the next completion of an aio request */
2157 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2159 struct proc *p = td->td_proc;
2162 struct kaioinfo *ki;
2163 struct aiocblist *cb;
2164 struct aiocb *uuaiocb;
2165 int error, status, timo;
2167 suword(uap->aiocbp, (long)NULL);
2171 /* Get timespec struct. */
2172 error = copyin(uap->timeout, &ts, sizeof(ts));
2176 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2179 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2180 if (itimerfix(&atv))
2182 timo = tvtohz(&atv);
2185 if (p->p_aioinfo == NULL)
2186 aio_init_aioinfo(p);
2192 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2193 ki->kaio_flags |= KAIO_WAKEUP;
2194 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2196 if (timo && error == ERESTART)
2203 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2204 uuaiocb = cb->uuaiocb;
2205 status = cb->uaiocb._aiocb_private.status;
2206 error = cb->uaiocb._aiocb_private.error;
2207 td->td_retval[0] = status;
2208 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2209 p->p_stats->p_ru.ru_oublock += cb->outputcharge;
2210 cb->outputcharge = 0;
2211 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2212 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2213 cb->inputcharge = 0;
2217 suword(uap->aiocbp, (long)uuaiocb);
2218 suword(&uuaiocb->_aiocb_private.error, error);
2219 suword(&uuaiocb->_aiocb_private.status, status);
2227 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2229 struct proc *p = td->td_proc;
2230 struct kaioinfo *ki;
2232 if (uap->op != O_SYNC) /* XXX lack of O_DSYNC */
2236 aio_init_aioinfo(p);
2237 return aio_aqueue(td, uap->aiocbp, NULL, LIO_SYNC, 0);
2240 /* kqueue attach function */
2242 filt_aioattach(struct knote *kn)
2244 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2247 * The aiocbe pointer must be validated before using it, so
2248 * registration is restricted to the kernel; the user cannot
2251 if ((kn->kn_flags & EV_FLAG1) == 0)
2253 kn->kn_flags &= ~EV_FLAG1;
2255 knlist_add(&aiocbe->klist, kn, 0);
2260 /* kqueue detach function */
2262 filt_aiodetach(struct knote *kn)
2264 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2266 if (!knlist_empty(&aiocbe->klist))
2267 knlist_remove(&aiocbe->klist, kn, 0);
2270 /* kqueue filter function */
2273 filt_aio(struct knote *kn, long hint)
2275 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2277 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2278 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2280 kn->kn_flags |= EV_EOF;
2284 /* kqueue attach function */
2286 filt_lioattach(struct knote *kn)
2288 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2291 * The aioliojob pointer must be validated before using it, so
2292 * registration is restricted to the kernel; the user cannot
2295 if ((kn->kn_flags & EV_FLAG1) == 0)
2297 kn->kn_flags &= ~EV_FLAG1;
2299 knlist_add(&lj->klist, kn, 0);
2304 /* kqueue detach function */
2306 filt_liodetach(struct knote *kn)
2308 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2310 if (!knlist_empty(&lj->klist))
2311 knlist_remove(&lj->klist, kn, 0);
2314 /* kqueue filter function */
2317 filt_lio(struct knote *kn, long hint)
2319 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2321 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);