2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. John S. Dyson's name may not be used to endorse or promote products
10 * derived from this software without specific prior written permission.
12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
13 * bad that happens because of using this software isn't the responsibility
14 * of the author. This software is distributed AS-IS.
18 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
24 #include "opt_compat.h"
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
31 #include <sys/capability.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
40 #include <sys/limits.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/protosw.h>
49 #include <sys/rwlock.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/syscall.h>
54 #include <sys/sysent.h>
55 #include <sys/sysctl.h>
57 #include <sys/taskqueue.h>
58 #include <sys/vnode.h>
60 #include <sys/event.h>
61 #include <sys/mount.h>
63 #include <machine/atomic.h>
66 #include <vm/vm_extern.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
73 #include "opt_vfs_aio.h"
76 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
77 * overflow. (XXX will be removed soon.)
79 static u_long jobrefid;
82 * Counter for aio_fsync.
84 static uint64_t jobseqno;
87 #define JOBST_JOBQSOCK 1
88 #define JOBST_JOBQGLOBAL 2
89 #define JOBST_JOBRUNNING 3
90 #define JOBST_JOBFINISHED 4
91 #define JOBST_JOBQBUF 5
92 #define JOBST_JOBQSYNC 6
94 #ifndef MAX_AIO_PER_PROC
95 #define MAX_AIO_PER_PROC 32
98 #ifndef MAX_AIO_QUEUE_PER_PROC
99 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
102 #ifndef MAX_AIO_PROCS
103 #define MAX_AIO_PROCS 32
106 #ifndef MAX_AIO_QUEUE
107 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
110 #ifndef TARGET_AIO_PROCS
111 #define TARGET_AIO_PROCS 4
115 #define MAX_BUF_AIO 16
118 #ifndef AIOD_TIMEOUT_DEFAULT
119 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
122 #ifndef AIOD_LIFETIME_DEFAULT
123 #define AIOD_LIFETIME_DEFAULT (30 * hz)
126 FEATURE(aio, "Asynchronous I/O");
128 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
130 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
132 static int max_aio_procs = MAX_AIO_PROCS;
133 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
134 CTLFLAG_RW, &max_aio_procs, 0,
135 "Maximum number of kernel threads to use for handling async IO ");
137 static int num_aio_procs = 0;
138 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
139 CTLFLAG_RD, &num_aio_procs, 0,
140 "Number of presently active kernel threads for async IO");
143 * The code will adjust the actual number of AIO processes towards this
144 * number when it gets a chance.
146 static int target_aio_procs = TARGET_AIO_PROCS;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
148 0, "Preferred number of ready kernel threads for async IO");
150 static int max_queue_count = MAX_AIO_QUEUE;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
152 "Maximum number of aio requests to queue, globally");
154 static int num_queue_count = 0;
155 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
156 "Number of queued aio requests");
158 static int num_buf_aio = 0;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
160 "Number of aio requests presently handled by the buf subsystem");
162 /* Number of async I/O thread in the process of being started */
163 /* XXX This should be local to aio_aqueue() */
164 static int num_aio_resv_start = 0;
166 static int aiod_timeout;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
168 "Timeout value for synchronous aio operations");
170 static int aiod_lifetime;
171 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
172 "Maximum lifetime for idle aiod");
174 static int unloadable = 0;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
176 "Allow unload of aio (not recommended)");
179 static int max_aio_per_proc = MAX_AIO_PER_PROC;
180 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
181 0, "Maximum active aio requests per process (stored in the process)");
183 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
184 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
185 &max_aio_queue_per_proc, 0,
186 "Maximum queued aio requests per process (stored in the process)");
188 static int max_buf_aio = MAX_BUF_AIO;
189 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
190 "Maximum buf aio requests per process (stored in the process)");
192 typedef struct oaiocb {
193 int aio_fildes; /* File descriptor */
194 off_t aio_offset; /* File offset for I/O */
195 volatile void *aio_buf; /* I/O buffer in process space */
196 size_t aio_nbytes; /* Number of bytes for I/O */
197 struct osigevent aio_sigevent; /* Signal to deliver */
198 int aio_lio_opcode; /* LIO opcode */
199 int aio_reqprio; /* Request priority -- ignored */
200 struct __aiocb_private _aiocb_private;
204 * Below is a key of locks used to protect each member of struct aiocblist
205 * aioliojob and kaioinfo and any backends.
207 * * - need not protected
208 * a - locked by kaioinfo lock
209 * b - locked by backend lock, the backend lock can be null in some cases,
210 * for example, BIO belongs to this type, in this case, proc lock is
212 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
216 * Current, there is only two backends: BIO and generic file I/O.
217 * socket I/O is served by generic file I/O, this is not a good idea, since
218 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
219 * threads, if there is no thread to serve socket I/O, the socket I/O will be
220 * delayed too long or starved, we should create some threads dedicated to
221 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
222 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
223 * structure is not safe because there is race between userland and aio
228 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
229 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
230 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
231 int jobflags; /* (a) job flags */
232 int jobstate; /* (b) job state */
233 int inputcharge; /* (*) input blockes */
234 int outputcharge; /* (*) output blockes */
235 struct buf *bp; /* (*) private to BIO backend,
238 struct proc *userproc; /* (*) user process */
239 struct ucred *cred; /* (*) active credential when created */
240 struct file *fd_file; /* (*) pointer to file structure */
241 struct aioliojob *lio; /* (*) optional lio job */
242 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
243 struct knlist klist; /* (a) list of knotes */
244 struct aiocb uaiocb; /* (*) kernel I/O control block */
245 ksiginfo_t ksi; /* (a) realtime signal info */
246 struct task biotask; /* (*) private to BIO backend */
247 uint64_t seqno; /* (*) job number */
248 int pending; /* (a) number of pending I/O, aio_fsync only */
252 #define AIOCBLIST_DONE 0x01
253 #define AIOCBLIST_BUFDONE 0x02
254 #define AIOCBLIST_RUNDOWN 0x04
255 #define AIOCBLIST_CHECKSYNC 0x08
260 #define AIOP_FREE 0x1 /* proc on free queue */
262 struct aiothreadlist {
263 int aiothreadflags; /* (c) AIO proc flags */
264 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
265 struct thread *aiothread; /* (*) the AIO thread */
269 * data-structure for lio signal management
272 int lioj_flags; /* (a) listio flags */
273 int lioj_count; /* (a) listio flags */
274 int lioj_finished_count; /* (a) listio flags */
275 struct sigevent lioj_signal; /* (a) signal on all I/O done */
276 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
277 struct knlist klist; /* (a) list of knotes */
278 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
281 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
282 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
283 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
286 * per process aio data structure
289 struct mtx kaio_mtx; /* the lock to protect this struct */
290 int kaio_flags; /* (a) per process kaio flags */
291 int kaio_maxactive_count; /* (*) maximum number of AIOs */
292 int kaio_active_count; /* (c) number of currently used AIOs */
293 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
294 int kaio_count; /* (a) size of AIO queue */
295 int kaio_ballowed_count; /* (*) maximum number of buffers */
296 int kaio_buffer_count; /* (a) number of physio buffers */
297 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
298 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
299 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
300 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
301 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
302 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
305 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
306 struct task kaio_task; /* (*) task to kick aio threads */
309 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
310 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
311 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
312 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
314 #define KAIO_RUNDOWN 0x1 /* process is being run down */
315 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
318 * Operations used to interact with userland aio control blocks.
319 * Different ABIs provide their own operations.
322 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
323 long (*fetch_status)(struct aiocb *ujob);
324 long (*fetch_error)(struct aiocb *ujob);
325 int (*store_status)(struct aiocb *ujob, long status);
326 int (*store_error)(struct aiocb *ujob, long error);
327 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
328 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
331 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
332 static struct sema aio_newproc_sem;
333 static struct mtx aio_job_mtx;
334 static struct mtx aio_sock_mtx;
335 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
336 static struct unrhdr *aiod_unr;
338 void aio_init_aioinfo(struct proc *p);
339 static int aio_onceonly(void);
340 static int aio_free_entry(struct aiocblist *aiocbe);
341 static void aio_process(struct aiocblist *aiocbe);
342 static int aio_newproc(int *);
343 int aio_aqueue(struct thread *td, struct aiocb *job,
344 struct aioliojob *lio, int type, struct aiocb_ops *ops);
345 static void aio_physwakeup(struct buf *bp);
346 static void aio_proc_rundown(void *arg, struct proc *p);
347 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
348 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
349 static void biohelper(void *, int);
350 static void aio_daemon(void *param);
351 static void aio_swake_cb(struct socket *, struct sockbuf *);
352 static int aio_unload(void);
353 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
356 static int aio_kick(struct proc *userp);
357 static void aio_kick_nowait(struct proc *userp);
358 static void aio_kick_helper(void *context, int pending);
359 static int filt_aioattach(struct knote *kn);
360 static void filt_aiodetach(struct knote *kn);
361 static int filt_aio(struct knote *kn, long hint);
362 static int filt_lioattach(struct knote *kn);
363 static void filt_liodetach(struct knote *kn);
364 static int filt_lio(struct knote *kn, long hint);
368 * kaio Per process async io info
369 * aiop async io thread data
370 * aiocb async io jobs
371 * aiol list io job pointer - internal to aio_suspend XXX
372 * aiolio list io jobs
374 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
376 /* kqueue filters for aio */
377 static struct filterops aio_filtops = {
379 .f_attach = filt_aioattach,
380 .f_detach = filt_aiodetach,
383 static struct filterops lio_filtops = {
385 .f_attach = filt_lioattach,
386 .f_detach = filt_liodetach,
390 static eventhandler_tag exit_tag, exec_tag;
392 TASKQUEUE_DEFINE_THREAD(aiod_bio);
395 * Main operations function for use as a kernel module.
398 aio_modload(struct module *module, int cmd, void *arg)
407 error = aio_unload();
418 static moduledata_t aio_mod = {
424 static struct syscall_helper_data aio_syscalls[] = {
425 SYSCALL_INIT_HELPER(aio_cancel),
426 SYSCALL_INIT_HELPER(aio_error),
427 SYSCALL_INIT_HELPER(aio_fsync),
428 SYSCALL_INIT_HELPER(aio_read),
429 SYSCALL_INIT_HELPER(aio_return),
430 SYSCALL_INIT_HELPER(aio_suspend),
431 SYSCALL_INIT_HELPER(aio_waitcomplete),
432 SYSCALL_INIT_HELPER(aio_write),
433 SYSCALL_INIT_HELPER(lio_listio),
434 SYSCALL_INIT_HELPER(oaio_read),
435 SYSCALL_INIT_HELPER(oaio_write),
436 SYSCALL_INIT_HELPER(olio_listio),
440 #ifdef COMPAT_FREEBSD32
441 #include <sys/mount.h>
442 #include <sys/socket.h>
443 #include <compat/freebsd32/freebsd32.h>
444 #include <compat/freebsd32/freebsd32_proto.h>
445 #include <compat/freebsd32/freebsd32_signal.h>
446 #include <compat/freebsd32/freebsd32_syscall.h>
447 #include <compat/freebsd32/freebsd32_util.h>
449 static struct syscall_helper_data aio32_syscalls[] = {
450 SYSCALL32_INIT_HELPER(freebsd32_aio_return),
451 SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
452 SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
453 SYSCALL32_INIT_HELPER(freebsd32_aio_error),
454 SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
455 SYSCALL32_INIT_HELPER(freebsd32_aio_read),
456 SYSCALL32_INIT_HELPER(freebsd32_aio_write),
457 SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
458 SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
459 SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
460 SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
461 SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
466 DECLARE_MODULE(aio, aio_mod,
467 SI_SUB_VFS, SI_ORDER_ANY);
468 MODULE_VERSION(aio, 1);
471 * Startup initialization
478 /* XXX: should probably just use so->callback */
479 aio_swake = &aio_swake_cb;
480 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
481 EVENTHANDLER_PRI_ANY);
482 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
483 EVENTHANDLER_PRI_ANY);
484 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
485 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
486 TAILQ_INIT(&aio_freeproc);
487 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
488 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
489 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
490 TAILQ_INIT(&aio_jobs);
491 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
492 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
493 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
494 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
495 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
496 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
497 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
498 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
499 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
500 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
501 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
502 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
503 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
505 async_io_version = _POSIX_VERSION;
506 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
507 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
508 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
510 error = syscall_helper_register(aio_syscalls);
513 #ifdef COMPAT_FREEBSD32
514 error = syscall32_helper_register(aio32_syscalls);
522 * Callback for unload of AIO when used as a module.
530 * XXX: no unloads by default, it's too dangerous.
531 * perhaps we could do it if locked out callers and then
532 * did an aio_proc_rundown() on each process.
534 * jhb: aio_proc_rundown() needs to run on curproc though,
535 * so I don't think that would fly.
540 #ifdef COMPAT_FREEBSD32
541 syscall32_helper_unregister(aio32_syscalls);
543 syscall_helper_unregister(aio_syscalls);
545 error = kqueue_del_filteropts(EVFILT_AIO);
548 error = kqueue_del_filteropts(EVFILT_LIO);
551 async_io_version = 0;
553 taskqueue_free(taskqueue_aiod_bio);
554 delete_unrhdr(aiod_unr);
555 uma_zdestroy(kaio_zone);
556 uma_zdestroy(aiop_zone);
557 uma_zdestroy(aiocb_zone);
558 uma_zdestroy(aiol_zone);
559 uma_zdestroy(aiolio_zone);
560 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
561 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
562 mtx_destroy(&aio_job_mtx);
563 mtx_destroy(&aio_sock_mtx);
564 sema_destroy(&aio_newproc_sem);
565 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
566 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
567 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
572 * Init the per-process aioinfo structure. The aioinfo limits are set
573 * per-process for user limit (resource) management.
576 aio_init_aioinfo(struct proc *p)
580 ki = uma_zalloc(kaio_zone, M_WAITOK);
581 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
583 ki->kaio_maxactive_count = max_aio_per_proc;
584 ki->kaio_active_count = 0;
585 ki->kaio_qallowed_count = max_aio_queue_per_proc;
587 ki->kaio_ballowed_count = max_buf_aio;
588 ki->kaio_buffer_count = 0;
589 TAILQ_INIT(&ki->kaio_all);
590 TAILQ_INIT(&ki->kaio_done);
591 TAILQ_INIT(&ki->kaio_jobqueue);
592 TAILQ_INIT(&ki->kaio_bufqueue);
593 TAILQ_INIT(&ki->kaio_liojoblist);
594 TAILQ_INIT(&ki->kaio_sockqueue);
595 TAILQ_INIT(&ki->kaio_syncqueue);
596 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
598 if (p->p_aioinfo == NULL) {
603 mtx_destroy(&ki->kaio_mtx);
604 uma_zfree(kaio_zone, ki);
607 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
612 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
617 error = sigev_findtd(p, sigev, &td);
621 ksiginfo_set_sigev(ksi, sigev);
622 ksi->ksi_code = SI_ASYNCIO;
623 ksi->ksi_flags |= KSI_EXT | KSI_INS;
624 tdsendsignal(p, td, ksi->ksi_signo, ksi);
631 * Free a job entry. Wait for completion if it is currently active, but don't
632 * delay forever. If we delay, we return a flag that says that we have to
633 * restart the queue scan.
636 aio_free_entry(struct aiocblist *aiocbe)
639 struct aioliojob *lj;
642 p = aiocbe->userproc;
647 AIO_LOCK_ASSERT(ki, MA_OWNED);
648 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
650 atomic_subtract_int(&num_queue_count, 1);
653 MPASS(ki->kaio_count >= 0);
655 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
656 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
661 lj->lioj_finished_count--;
663 if (lj->lioj_count == 0) {
664 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
665 /* lio is going away, we need to destroy any knotes */
666 knlist_delete(&lj->klist, curthread, 1);
668 sigqueue_take(&lj->lioj_ksi);
670 uma_zfree(aiolio_zone, lj);
674 /* aiocbe is going away, we need to destroy any knotes */
675 knlist_delete(&aiocbe->klist, curthread, 1);
677 sigqueue_take(&aiocbe->ksi);
680 MPASS(aiocbe->bp == NULL);
681 aiocbe->jobstate = JOBST_NULL;
685 * The thread argument here is used to find the owning process
686 * and is also passed to fo_close() which may pass it to various
687 * places such as devsw close() routines. Because of that, we
688 * need a thread pointer from the process owning the job that is
689 * persistent and won't disappear out from under us or move to
692 * Currently, all the callers of this function call it to remove
693 * an aiocblist from the current process' job list either via a
694 * syscall or due to the current process calling exit() or
695 * execve(). Thus, we know that p == curproc. We also know that
696 * curthread can't exit since we are curthread.
698 * Therefore, we use curthread as the thread to pass to
699 * knlist_delete(). This does mean that it is possible for the
700 * thread pointer at close time to differ from the thread pointer
701 * at open time, but this is already true of file descriptors in
702 * a multithreaded process.
704 fdrop(aiocbe->fd_file, curthread);
705 crfree(aiocbe->cred);
706 uma_zfree(aiocb_zone, aiocbe);
713 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
715 aio_proc_rundown(arg, p);
719 * Rundown the jobs for a given process.
722 aio_proc_rundown(void *arg, struct proc *p)
725 struct aioliojob *lj;
726 struct aiocblist *cbe, *cbn;
731 KASSERT(curthread->td_proc == p,
732 ("%s: called on non-curproc", __func__));
738 ki->kaio_flags |= KAIO_RUNDOWN;
743 * Try to cancel all pending requests. This code simulates
744 * aio_cancel on all pending I/O requests.
746 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
748 mtx_lock(&aio_job_mtx);
749 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
750 TAILQ_REMOVE(&aio_jobs, cbe, list);
752 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
754 MPASS(fp->f_type == DTYPE_SOCKET);
756 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
758 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
759 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
762 mtx_unlock(&aio_job_mtx);
765 cbe->jobstate = JOBST_JOBFINISHED;
766 cbe->uaiocb._aiocb_private.status = -1;
767 cbe->uaiocb._aiocb_private.error = ECANCELED;
768 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
769 aio_bio_done_notify(p, cbe, DONE_QUEUE);
773 /* Wait for all running I/O to be finished */
774 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
775 TAILQ_FIRST(&ki->kaio_jobqueue)) {
776 ki->kaio_flags |= KAIO_WAKEUP;
777 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
781 /* Free all completed I/O requests. */
782 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
785 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
786 if (lj->lioj_count == 0) {
787 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
788 knlist_delete(&lj->klist, curthread, 1);
790 sigqueue_take(&lj->lioj_ksi);
792 uma_zfree(aiolio_zone, lj);
794 panic("LIO job not cleaned up: C:%d, FC:%d\n",
795 lj->lioj_count, lj->lioj_finished_count);
799 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
800 mtx_destroy(&ki->kaio_mtx);
801 uma_zfree(kaio_zone, ki);
806 * Select a job to run (called by an AIO daemon).
808 static struct aiocblist *
809 aio_selectjob(struct aiothreadlist *aiop)
811 struct aiocblist *aiocbe;
815 mtx_assert(&aio_job_mtx, MA_OWNED);
816 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
817 userp = aiocbe->userproc;
818 ki = userp->p_aioinfo;
820 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
821 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
822 /* Account for currently active jobs. */
823 ki->kaio_active_count++;
824 aiocbe->jobstate = JOBST_JOBRUNNING;
832 * Move all data to a permanent storage device, this code
833 * simulates fsync syscall.
836 aio_fsync_vnode(struct thread *td, struct vnode *vp)
841 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
843 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
844 if (vp->v_object != NULL) {
845 VM_OBJECT_WLOCK(vp->v_object);
846 vm_object_page_clean(vp->v_object, 0, 0, 0);
847 VM_OBJECT_WUNLOCK(vp->v_object);
849 error = VOP_FSYNC(vp, MNT_WAIT, td);
852 vn_finished_write(mp);
858 * The AIO processing activity. This is the code that does the I/O request for
859 * the non-physio version of the operations. The normal vn operations are used,
860 * and this code should work in all instances for every type of file, including
861 * pipes, sockets, fifos, and regular files.
863 * XXX I don't think it works well for socket, pipe, and fifo.
866 aio_process(struct aiocblist *aiocbe)
868 struct ucred *td_savedcred;
877 int oublock_st, oublock_end;
878 int inblock_st, inblock_end;
881 td_savedcred = td->td_ucred;
882 td->td_ucred = aiocbe->cred;
883 cb = &aiocbe->uaiocb;
884 fp = aiocbe->fd_file;
886 if (cb->aio_lio_opcode == LIO_SYNC) {
889 if (fp->f_vnode != NULL)
890 error = aio_fsync_vnode(td, fp->f_vnode);
891 cb->_aiocb_private.error = error;
892 cb->_aiocb_private.status = 0;
893 td->td_ucred = td_savedcred;
897 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
898 aiov.iov_len = cb->aio_nbytes;
900 auio.uio_iov = &aiov;
902 auio.uio_offset = cb->aio_offset;
903 auio.uio_resid = cb->aio_nbytes;
904 cnt = cb->aio_nbytes;
905 auio.uio_segflg = UIO_USERSPACE;
908 inblock_st = td->td_ru.ru_inblock;
909 oublock_st = td->td_ru.ru_oublock;
911 * aio_aqueue() acquires a reference to the file that is
912 * released in aio_free_entry().
914 if (cb->aio_lio_opcode == LIO_READ) {
915 auio.uio_rw = UIO_READ;
916 if (auio.uio_resid == 0)
919 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
921 if (fp->f_type == DTYPE_VNODE)
923 auio.uio_rw = UIO_WRITE;
924 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
926 inblock_end = td->td_ru.ru_inblock;
927 oublock_end = td->td_ru.ru_oublock;
929 aiocbe->inputcharge = inblock_end - inblock_st;
930 aiocbe->outputcharge = oublock_end - oublock_st;
932 if ((error) && (auio.uio_resid != cnt)) {
933 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
935 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
937 if (fp->f_type == DTYPE_SOCKET) {
939 if (so->so_options & SO_NOSIGPIPE)
943 PROC_LOCK(aiocbe->userproc);
944 kern_psignal(aiocbe->userproc, SIGPIPE);
945 PROC_UNLOCK(aiocbe->userproc);
950 cnt -= auio.uio_resid;
951 cb->_aiocb_private.error = error;
952 cb->_aiocb_private.status = cnt;
953 td->td_ucred = td_savedcred;
957 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
959 struct aioliojob *lj;
961 struct aiocblist *scb, *scbn;
964 ki = userp->p_aioinfo;
965 AIO_LOCK_ASSERT(ki, MA_OWNED);
969 lj->lioj_finished_count++;
970 if (lj->lioj_count == lj->lioj_finished_count)
973 if (type == DONE_QUEUE) {
974 aiocbe->jobflags |= AIOCBLIST_DONE;
976 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
978 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
979 aiocbe->jobstate = JOBST_JOBFINISHED;
981 if (ki->kaio_flags & KAIO_RUNDOWN)
982 goto notification_done;
984 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
985 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
986 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
988 KNOTE_LOCKED(&aiocbe->klist, 1);
991 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
992 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
993 KNOTE_LOCKED(&lj->klist, 1);
995 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
997 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
998 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
999 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
1000 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1005 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1006 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1007 if (aiocbe->fd_file == scb->fd_file &&
1008 aiocbe->seqno < scb->seqno) {
1009 if (--scb->pending == 0) {
1010 mtx_lock(&aio_job_mtx);
1011 scb->jobstate = JOBST_JOBQGLOBAL;
1012 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1013 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1014 aio_kick_nowait(userp);
1015 mtx_unlock(&aio_job_mtx);
1020 if (ki->kaio_flags & KAIO_WAKEUP) {
1021 ki->kaio_flags &= ~KAIO_WAKEUP;
1022 wakeup(&userp->p_aioinfo);
1027 * The AIO daemon, most of the actual work is done in aio_process,
1028 * but the setup (and address space mgmt) is done in this routine.
1031 aio_daemon(void *_id)
1033 struct aiocblist *aiocbe;
1034 struct aiothreadlist *aiop;
1035 struct kaioinfo *ki;
1036 struct proc *curcp, *mycp, *userp;
1037 struct vmspace *myvm, *tmpvm;
1038 struct thread *td = curthread;
1039 int id = (intptr_t)_id;
1042 * Local copies of curproc (cp) and vmspace (myvm)
1045 myvm = mycp->p_vmspace;
1047 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1050 * Allocate and ready the aio control info. There is one aiop structure
1053 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1054 aiop->aiothread = td;
1055 aiop->aiothreadflags = 0;
1057 /* The daemon resides in its own pgrp. */
1058 sys_setsid(td, NULL);
1061 * Wakeup parent process. (Parent sleeps to keep from blasting away
1062 * and creating too many daemons.)
1064 sema_post(&aio_newproc_sem);
1066 mtx_lock(&aio_job_mtx);
1069 * curcp is the current daemon process context.
1070 * userp is the current user process context.
1075 * Take daemon off of free queue
1077 if (aiop->aiothreadflags & AIOP_FREE) {
1078 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1079 aiop->aiothreadflags &= ~AIOP_FREE;
1085 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1086 mtx_unlock(&aio_job_mtx);
1087 userp = aiocbe->userproc;
1090 * Connect to process address space for user program.
1092 if (userp != curcp) {
1094 * Save the current address space that we are
1097 tmpvm = mycp->p_vmspace;
1100 * Point to the new user address space, and
1103 mycp->p_vmspace = userp->p_vmspace;
1104 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1106 /* Activate the new mapping. */
1107 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1110 * If the old address space wasn't the daemons
1111 * own address space, then we need to remove the
1112 * daemon's reference from the other process
1113 * that it was acting on behalf of.
1115 if (tmpvm != myvm) {
1116 vmspace_free(tmpvm);
1121 ki = userp->p_aioinfo;
1123 /* Do the I/O function. */
1124 aio_process(aiocbe);
1126 mtx_lock(&aio_job_mtx);
1127 /* Decrement the active job count. */
1128 ki->kaio_active_count--;
1129 mtx_unlock(&aio_job_mtx);
1132 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1133 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1136 mtx_lock(&aio_job_mtx);
1140 * Disconnect from user address space.
1142 if (curcp != mycp) {
1144 mtx_unlock(&aio_job_mtx);
1146 /* Get the user address space to disconnect from. */
1147 tmpvm = mycp->p_vmspace;
1149 /* Get original address space for daemon. */
1150 mycp->p_vmspace = myvm;
1152 /* Activate the daemon's address space. */
1153 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1155 if (tmpvm == myvm) {
1156 printf("AIOD: vmspace problem -- %d\n",
1160 /* Remove our vmspace reference. */
1161 vmspace_free(tmpvm);
1165 mtx_lock(&aio_job_mtx);
1167 * We have to restart to avoid race, we only sleep if
1168 * no job can be selected, that should be
1174 mtx_assert(&aio_job_mtx, MA_OWNED);
1176 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1177 aiop->aiothreadflags |= AIOP_FREE;
1180 * If daemon is inactive for a long time, allow it to exit,
1181 * thereby freeing resources.
1183 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1185 if (TAILQ_EMPTY(&aio_jobs)) {
1186 if ((aiop->aiothreadflags & AIOP_FREE) &&
1187 (num_aio_procs > target_aio_procs)) {
1188 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1190 mtx_unlock(&aio_job_mtx);
1191 uma_zfree(aiop_zone, aiop);
1192 free_unr(aiod_unr, id);
1194 if (mycp->p_vmspace->vm_refcnt <= 1) {
1195 printf("AIOD: bad vm refcnt for"
1196 " exiting daemon: %d\n",
1197 mycp->p_vmspace->vm_refcnt);
1205 mtx_unlock(&aio_job_mtx);
1206 panic("shouldn't be here\n");
1210 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1211 * AIO daemon modifies its environment itself.
1214 aio_newproc(int *start)
1220 id = alloc_unr(aiod_unr);
1221 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1222 RFNOWAIT, 0, "aiod%d", id);
1225 * Wait until daemon is started.
1227 sema_wait(&aio_newproc_sem);
1228 mtx_lock(&aio_job_mtx);
1232 mtx_unlock(&aio_job_mtx);
1234 free_unr(aiod_unr, id);
1240 * Try the high-performance, low-overhead physio method for eligible
1241 * VCHR devices. This method doesn't use an aio helper thread, and
1242 * thus has very low overhead.
1244 * Assumes that the caller, aio_aqueue(), has incremented the file
1245 * structure's reference count, preventing its deallocation for the
1246 * duration of this call.
1249 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1257 struct kaioinfo *ki;
1258 struct aioliojob *lj;
1261 cb = &aiocbe->uaiocb;
1262 fp = aiocbe->fd_file;
1264 if (fp->f_type != DTYPE_VNODE)
1270 * If its not a disk, we don't want to return a positive error.
1271 * It causes the aio code to not fall through to try the thread
1272 * way when you're talking to a regular file.
1274 if (!vn_isdisk(vp, &error)) {
1275 if (error == ENOTBLK)
1281 if (vp->v_bufobj.bo_bsize == 0)
1284 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1287 if (cb->aio_nbytes >
1288 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1292 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1296 csw = devvn_refthread(vp, &dev, &ref);
1299 if (cb->aio_nbytes > dev->si_iosize_max) {
1304 /* Create and build a buffer header for a transfer. */
1305 bp = (struct buf *)getpbuf(NULL);
1310 ki->kaio_buffer_count++;
1317 * Get a copy of the kva from the physical buffer.
1321 bp->b_bcount = cb->aio_nbytes;
1322 bp->b_bufsize = cb->aio_nbytes;
1323 bp->b_iodone = aio_physwakeup;
1324 bp->b_saveaddr = bp->b_data;
1325 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1326 bp->b_offset = cb->aio_offset;
1327 bp->b_iooffset = cb->aio_offset;
1328 bp->b_blkno = btodb(cb->aio_offset);
1329 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1332 * Bring buffer into kernel space.
1334 if (vmapbuf(bp, (csw->d_flags & D_UNMAPPED_IO) == 0) < 0) {
1341 bp->b_caller1 = (void *)aiocbe;
1342 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1343 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1344 aiocbe->jobstate = JOBST_JOBQBUF;
1345 cb->_aiocb_private.status = cb->aio_nbytes;
1348 atomic_add_int(&num_queue_count, 1);
1349 atomic_add_int(&num_buf_aio, 1);
1353 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1355 /* Perform transfer. */
1356 dev_strategy_csw(dev, csw, bp);
1357 dev_relthread(dev, ref);
1363 ki->kaio_buffer_count--;
1370 dev_relthread(dev, ref);
1375 * Wake up aio requests that may be serviceable now.
1378 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1380 struct aiocblist *cb, *cbn;
1383 SOCKBUF_LOCK_ASSERT(sb);
1384 if (sb == &so->so_snd)
1389 sb->sb_flags &= ~SB_AIO;
1390 mtx_lock(&aio_job_mtx);
1391 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1392 if (opcode == cb->uaiocb.aio_lio_opcode) {
1393 if (cb->jobstate != JOBST_JOBQSOCK)
1394 panic("invalid queue value");
1396 * We don't have actual sockets backend yet,
1397 * so we simply move the requests to the generic
1400 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1401 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1402 aio_kick_nowait(cb->userproc);
1405 mtx_unlock(&aio_job_mtx);
1409 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1413 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1414 * supported by AIO with the old sigevent structure.
1416 nsig->sigev_notify = osig->sigev_notify;
1417 switch (nsig->sigev_notify) {
1421 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1424 nsig->sigev_notify_kqueue =
1425 osig->__sigev_u.__sigev_notify_kqueue;
1426 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1435 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1437 struct oaiocb *ojob;
1440 bzero(kjob, sizeof(struct aiocb));
1441 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1444 ojob = (struct oaiocb *)kjob;
1445 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1449 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1452 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1456 aiocb_fetch_status(struct aiocb *ujob)
1459 return (fuword(&ujob->_aiocb_private.status));
1463 aiocb_fetch_error(struct aiocb *ujob)
1466 return (fuword(&ujob->_aiocb_private.error));
1470 aiocb_store_status(struct aiocb *ujob, long status)
1473 return (suword(&ujob->_aiocb_private.status, status));
1477 aiocb_store_error(struct aiocb *ujob, long error)
1480 return (suword(&ujob->_aiocb_private.error, error));
1484 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1487 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1491 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1494 return (suword(ujobp, (long)ujob));
1497 static struct aiocb_ops aiocb_ops = {
1498 .copyin = aiocb_copyin,
1499 .fetch_status = aiocb_fetch_status,
1500 .fetch_error = aiocb_fetch_error,
1501 .store_status = aiocb_store_status,
1502 .store_error = aiocb_store_error,
1503 .store_kernelinfo = aiocb_store_kernelinfo,
1504 .store_aiocb = aiocb_store_aiocb,
1507 static struct aiocb_ops aiocb_ops_osigevent = {
1508 .copyin = aiocb_copyin_old_sigevent,
1509 .fetch_status = aiocb_fetch_status,
1510 .fetch_error = aiocb_fetch_error,
1511 .store_status = aiocb_store_status,
1512 .store_error = aiocb_store_error,
1513 .store_kernelinfo = aiocb_store_kernelinfo,
1514 .store_aiocb = aiocb_store_aiocb,
1518 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1519 * technique is done in this code.
1522 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1523 int type, struct aiocb_ops *ops)
1525 struct proc *p = td->td_proc;
1528 struct aiocblist *aiocbe, *cb;
1529 struct kaioinfo *ki;
1538 if (p->p_aioinfo == NULL)
1539 aio_init_aioinfo(p);
1543 ops->store_status(job, -1);
1544 ops->store_error(job, 0);
1545 ops->store_kernelinfo(job, -1);
1547 if (num_queue_count >= max_queue_count ||
1548 ki->kaio_count >= ki->kaio_qallowed_count) {
1549 ops->store_error(job, EAGAIN);
1553 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1554 aiocbe->inputcharge = 0;
1555 aiocbe->outputcharge = 0;
1556 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1558 error = ops->copyin(job, &aiocbe->uaiocb);
1560 ops->store_error(job, error);
1561 uma_zfree(aiocb_zone, aiocbe);
1565 /* XXX: aio_nbytes is later casted to signed types. */
1566 if (aiocbe->uaiocb.aio_nbytes > INT_MAX) {
1567 uma_zfree(aiocb_zone, aiocbe);
1571 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1572 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1573 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1574 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1575 ops->store_error(job, EINVAL);
1576 uma_zfree(aiocb_zone, aiocbe);
1580 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1581 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1582 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1583 uma_zfree(aiocb_zone, aiocbe);
1587 ksiginfo_init(&aiocbe->ksi);
1589 /* Save userspace address of the job info. */
1590 aiocbe->uuaiocb = job;
1592 /* Get the opcode. */
1593 if (type != LIO_NOP)
1594 aiocbe->uaiocb.aio_lio_opcode = type;
1595 opcode = aiocbe->uaiocb.aio_lio_opcode;
1598 * Validate the opcode and fetch the file object for the specified
1601 * XXXRW: Moved the opcode validation up here so that we don't
1602 * retrieve a file descriptor without knowing what the capabiltity
1605 fd = aiocbe->uaiocb.aio_fildes;
1608 error = fget_write(td, fd, CAP_PWRITE, &fp);
1611 error = fget_read(td, fd, CAP_PREAD, &fp);
1614 error = fget(td, fd, CAP_FSYNC, &fp);
1617 error = fget(td, fd, CAP_NONE, &fp);
1623 uma_zfree(aiocb_zone, aiocbe);
1624 ops->store_error(job, error);
1628 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1633 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1638 aiocbe->fd_file = fp;
1640 mtx_lock(&aio_job_mtx);
1642 aiocbe->seqno = jobseqno++;
1643 mtx_unlock(&aio_job_mtx);
1644 error = ops->store_kernelinfo(job, jid);
1649 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1651 if (opcode == LIO_NOP) {
1653 uma_zfree(aiocb_zone, aiocbe);
1657 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1659 evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1660 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1664 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1665 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1666 kev.filter = EVFILT_AIO;
1667 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1668 kev.data = (intptr_t)aiocbe;
1669 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1670 error = kqfd_register(kqfd, &kev, td, 1);
1674 uma_zfree(aiocb_zone, aiocbe);
1675 ops->store_error(job, error);
1680 ops->store_error(job, EINPROGRESS);
1681 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1682 aiocbe->userproc = p;
1683 aiocbe->cred = crhold(td->td_ucred);
1684 aiocbe->jobflags = 0;
1687 if (opcode == LIO_SYNC)
1690 if (fp->f_type == DTYPE_SOCKET) {
1692 * Alternate queueing for socket ops: Reach down into the
1693 * descriptor to get the socket data. Then check to see if the
1694 * socket is ready to be read or written (based on the requested
1697 * If it is not ready for io, then queue the aiocbe on the
1698 * socket, and set the flags so we get a call when sbnotify()
1701 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1702 * and unlock the snd sockbuf for no reason.
1705 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1707 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1708 LIO_WRITE) && (!sowriteable(so)))) {
1709 sb->sb_flags |= SB_AIO;
1711 mtx_lock(&aio_job_mtx);
1712 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1713 mtx_unlock(&aio_job_mtx);
1716 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1717 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1718 aiocbe->jobstate = JOBST_JOBQSOCK;
1724 atomic_add_int(&num_queue_count, 1);
1731 if ((error = aio_qphysio(p, aiocbe)) == 0)
1735 aiocbe->uaiocb._aiocb_private.error = error;
1736 ops->store_error(job, error);
1741 /* No buffer for daemon I/O. */
1743 atomic_add_int(&num_queue_count, 1);
1749 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1750 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1751 if (opcode == LIO_SYNC) {
1752 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1753 if (cb->fd_file == aiocbe->fd_file &&
1754 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1755 cb->seqno < aiocbe->seqno) {
1756 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1760 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1761 if (cb->fd_file == aiocbe->fd_file &&
1762 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1763 cb->seqno < aiocbe->seqno) {
1764 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1768 if (aiocbe->pending != 0) {
1769 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1770 aiocbe->jobstate = JOBST_JOBQSYNC;
1775 mtx_lock(&aio_job_mtx);
1776 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1777 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1779 mtx_unlock(&aio_job_mtx);
1787 aio_kick_nowait(struct proc *userp)
1789 struct kaioinfo *ki = userp->p_aioinfo;
1790 struct aiothreadlist *aiop;
1792 mtx_assert(&aio_job_mtx, MA_OWNED);
1793 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1794 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1795 aiop->aiothreadflags &= ~AIOP_FREE;
1796 wakeup(aiop->aiothread);
1797 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1798 ((ki->kaio_active_count + num_aio_resv_start) <
1799 ki->kaio_maxactive_count)) {
1800 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1805 aio_kick(struct proc *userp)
1807 struct kaioinfo *ki = userp->p_aioinfo;
1808 struct aiothreadlist *aiop;
1811 mtx_assert(&aio_job_mtx, MA_OWNED);
1813 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1814 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1815 aiop->aiothreadflags &= ~AIOP_FREE;
1816 wakeup(aiop->aiothread);
1817 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1818 ((ki->kaio_active_count + num_aio_resv_start) <
1819 ki->kaio_maxactive_count)) {
1820 num_aio_resv_start++;
1821 mtx_unlock(&aio_job_mtx);
1822 error = aio_newproc(&num_aio_resv_start);
1823 mtx_lock(&aio_job_mtx);
1825 num_aio_resv_start--;
1835 aio_kick_helper(void *context, int pending)
1837 struct proc *userp = context;
1839 mtx_lock(&aio_job_mtx);
1840 while (--pending >= 0) {
1841 if (aio_kick(userp))
1844 mtx_unlock(&aio_job_mtx);
1848 * Support the aio_return system call, as a side-effect, kernel resources are
1852 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1854 struct proc *p = td->td_proc;
1855 struct aiocblist *cb;
1856 struct kaioinfo *ki;
1863 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1864 if (cb->uuaiocb == uaiocb)
1868 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1869 status = cb->uaiocb._aiocb_private.status;
1870 error = cb->uaiocb._aiocb_private.error;
1871 td->td_retval[0] = status;
1872 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1873 td->td_ru.ru_oublock += cb->outputcharge;
1874 cb->outputcharge = 0;
1875 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1876 td->td_ru.ru_inblock += cb->inputcharge;
1877 cb->inputcharge = 0;
1881 ops->store_error(uaiocb, error);
1882 ops->store_status(uaiocb, status);
1891 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1894 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1898 * Allow a process to wakeup when any of the I/O requests are completed.
1901 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1902 struct timespec *ts)
1904 struct proc *p = td->td_proc;
1906 struct kaioinfo *ki;
1907 struct aiocblist *cb, *cbfirst;
1912 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1915 TIMESPEC_TO_TIMEVAL(&atv, ts);
1916 if (itimerfix(&atv))
1918 timo = tvtohz(&atv);
1932 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1933 for (i = 0; i < njoblist; i++) {
1934 if (cb->uuaiocb == ujoblist[i]) {
1935 if (cbfirst == NULL)
1937 if (cb->jobstate == JOBST_JOBFINISHED)
1942 /* All tasks were finished. */
1943 if (cbfirst == NULL)
1946 ki->kaio_flags |= KAIO_WAKEUP;
1947 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1949 if (error == ERESTART)
1960 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1962 struct timespec ts, *tsp;
1963 struct aiocb **ujoblist;
1966 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1970 /* Get timespec struct. */
1971 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1977 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1978 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1980 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1981 uma_zfree(aiol_zone, ujoblist);
1986 * aio_cancel cancels any non-physio aio operations not currently in
1990 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1992 struct proc *p = td->td_proc;
1993 struct kaioinfo *ki;
1994 struct aiocblist *cbe, *cbn;
2000 int notcancelled = 0;
2003 /* Lookup file object. */
2004 error = fget(td, uap->fd, 0, &fp);
2012 if (fp->f_type == DTYPE_VNODE) {
2014 if (vn_isdisk(vp, &error)) {
2016 td->td_retval[0] = AIO_NOTCANCELED;
2022 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
2023 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
2024 ((uap->aiocbp == NULL) ||
2025 (uap->aiocbp == cbe->uuaiocb))) {
2028 mtx_lock(&aio_job_mtx);
2029 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
2030 TAILQ_REMOVE(&aio_jobs, cbe, list);
2032 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
2033 MPASS(fp->f_type == DTYPE_SOCKET);
2035 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2037 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
2038 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2041 mtx_unlock(&aio_job_mtx);
2044 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2045 cbe->uaiocb._aiocb_private.status = -1;
2046 cbe->uaiocb._aiocb_private.error = ECANCELED;
2047 aio_bio_done_notify(p, cbe, DONE_QUEUE);
2052 if (uap->aiocbp != NULL)
2061 if (uap->aiocbp != NULL) {
2063 td->td_retval[0] = AIO_CANCELED;
2069 td->td_retval[0] = AIO_NOTCANCELED;
2074 td->td_retval[0] = AIO_CANCELED;
2078 td->td_retval[0] = AIO_ALLDONE;
2084 * aio_error is implemented in the kernel level for compatibility purposes
2085 * only. For a user mode async implementation, it would be best to do it in
2086 * a userland subroutine.
2089 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2091 struct proc *p = td->td_proc;
2092 struct aiocblist *cb;
2093 struct kaioinfo *ki;
2098 td->td_retval[0] = EINVAL;
2103 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2104 if (cb->uuaiocb == aiocbp) {
2105 if (cb->jobstate == JOBST_JOBFINISHED)
2107 cb->uaiocb._aiocb_private.error;
2109 td->td_retval[0] = EINPROGRESS;
2117 * Hack for failure of aio_aqueue.
2119 status = ops->fetch_status(aiocbp);
2121 td->td_retval[0] = ops->fetch_error(aiocbp);
2125 td->td_retval[0] = EINVAL;
2130 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2133 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2136 /* syscall - asynchronous read from a file (REALTIME) */
2138 sys_oaio_read(struct thread *td, struct oaio_read_args *uap)
2141 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2142 &aiocb_ops_osigevent));
2146 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2149 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2152 /* syscall - asynchronous write to a file (REALTIME) */
2154 sys_oaio_write(struct thread *td, struct oaio_write_args *uap)
2157 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2158 &aiocb_ops_osigevent));
2162 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2165 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2169 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2170 struct aiocb **acb_list, int nent, struct sigevent *sig,
2171 struct aiocb_ops *ops)
2173 struct proc *p = td->td_proc;
2175 struct kaioinfo *ki;
2176 struct aioliojob *lj;
2182 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2185 if (nent < 0 || nent > AIO_LISTIO_MAX)
2188 if (p->p_aioinfo == NULL)
2189 aio_init_aioinfo(p);
2193 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2196 lj->lioj_finished_count = 0;
2197 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2198 ksiginfo_init(&lj->lioj_ksi);
2203 if (sig && (mode == LIO_NOWAIT)) {
2204 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2205 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2206 /* Assume only new style KEVENT */
2207 kev.filter = EVFILT_LIO;
2208 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2209 kev.ident = (uintptr_t)uacb_list; /* something unique */
2210 kev.data = (intptr_t)lj;
2211 /* pass user defined sigval data */
2212 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2213 error = kqfd_register(
2214 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2216 uma_zfree(aiolio_zone, lj);
2219 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2221 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2222 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2223 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2224 uma_zfree(aiolio_zone, lj);
2227 lj->lioj_flags |= LIOJ_SIGNAL;
2229 uma_zfree(aiolio_zone, lj);
2235 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2237 * Add extra aiocb count to avoid the lio to be freed
2238 * by other threads doing aio_waitcomplete or aio_return,
2239 * and prevent event from being sent until we have queued
2246 * Get pointers to the list of I/O requests.
2249 for (i = 0; i < nent; i++) {
2252 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2260 if (mode == LIO_WAIT) {
2261 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2262 ki->kaio_flags |= KAIO_WAKEUP;
2263 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2264 PRIBIO | PCATCH, "aiospn", 0);
2265 if (error == ERESTART)
2271 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2272 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2273 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2274 KNOTE_LOCKED(&lj->klist, 1);
2276 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2278 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2279 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2280 aio_sendsig(p, &lj->lioj_signal,
2282 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2287 if (lj->lioj_count == 0) {
2288 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2289 knlist_delete(&lj->klist, curthread, 1);
2291 sigqueue_take(&lj->lioj_ksi);
2294 uma_zfree(aiolio_zone, lj);
2303 /* syscall - list directed I/O (REALTIME) */
2305 sys_olio_listio(struct thread *td, struct olio_listio_args *uap)
2307 struct aiocb **acb_list;
2308 struct sigevent *sigp, sig;
2309 struct osigevent osig;
2312 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2316 if (nent < 0 || nent > AIO_LISTIO_MAX)
2319 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2320 error = copyin(uap->sig, &osig, sizeof(osig));
2323 error = convert_old_sigevent(&osig, &sig);
2330 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2331 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2333 error = kern_lio_listio(td, uap->mode,
2334 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2335 &aiocb_ops_osigevent);
2336 free(acb_list, M_LIO);
2340 /* syscall - list directed I/O (REALTIME) */
2342 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2344 struct aiocb **acb_list;
2345 struct sigevent *sigp, sig;
2348 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2352 if (nent < 0 || nent > AIO_LISTIO_MAX)
2355 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2356 error = copyin(uap->sig, &sig, sizeof(sig));
2363 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2364 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2366 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2367 nent, sigp, &aiocb_ops);
2368 free(acb_list, M_LIO);
2373 * Called from interrupt thread for physio, we should return as fast
2374 * as possible, so we schedule a biohelper task.
2377 aio_physwakeup(struct buf *bp)
2379 struct aiocblist *aiocbe;
2381 aiocbe = (struct aiocblist *)bp->b_caller1;
2382 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2386 * Task routine to perform heavy tasks, process wakeup, and signals.
2389 biohelper(void *context, int pending)
2391 struct aiocblist *aiocbe = context;
2394 struct kaioinfo *ki;
2398 userp = aiocbe->userproc;
2399 ki = userp->p_aioinfo;
2401 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2402 aiocbe->uaiocb._aiocb_private.error = 0;
2403 if (bp->b_ioflags & BIO_ERROR)
2404 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2405 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2406 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2407 aiocbe->outputcharge += nblks;
2409 aiocbe->inputcharge += nblks;
2411 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2412 ki->kaio_buffer_count--;
2413 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2416 /* Release mapping into kernel space. */
2419 atomic_subtract_int(&num_buf_aio, 1);
2422 /* syscall - wait for the next completion of an aio request */
2424 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2425 struct timespec *ts, struct aiocb_ops *ops)
2427 struct proc *p = td->td_proc;
2429 struct kaioinfo *ki;
2430 struct aiocblist *cb;
2431 struct aiocb *uuaiocb;
2432 int error, status, timo;
2434 ops->store_aiocb(aiocbp, NULL);
2438 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2441 TIMESPEC_TO_TIMEVAL(&atv, ts);
2442 if (itimerfix(&atv))
2444 timo = tvtohz(&atv);
2447 if (p->p_aioinfo == NULL)
2448 aio_init_aioinfo(p);
2454 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2455 ki->kaio_flags |= KAIO_WAKEUP;
2456 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2458 if (timo && error == ERESTART)
2465 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2466 uuaiocb = cb->uuaiocb;
2467 status = cb->uaiocb._aiocb_private.status;
2468 error = cb->uaiocb._aiocb_private.error;
2469 td->td_retval[0] = status;
2470 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2471 td->td_ru.ru_oublock += cb->outputcharge;
2472 cb->outputcharge = 0;
2473 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2474 td->td_ru.ru_inblock += cb->inputcharge;
2475 cb->inputcharge = 0;
2479 ops->store_aiocb(aiocbp, uuaiocb);
2480 ops->store_error(uuaiocb, error);
2481 ops->store_status(uuaiocb, status);
2489 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2491 struct timespec ts, *tsp;
2495 /* Get timespec struct. */
2496 error = copyin(uap->timeout, &ts, sizeof(ts));
2503 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2507 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2508 struct aiocb_ops *ops)
2510 struct proc *p = td->td_proc;
2511 struct kaioinfo *ki;
2513 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2517 aio_init_aioinfo(p);
2518 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2522 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2525 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2528 /* kqueue attach function */
2530 filt_aioattach(struct knote *kn)
2532 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2535 * The aiocbe pointer must be validated before using it, so
2536 * registration is restricted to the kernel; the user cannot
2539 if ((kn->kn_flags & EV_FLAG1) == 0)
2541 kn->kn_ptr.p_aio = aiocbe;
2542 kn->kn_flags &= ~EV_FLAG1;
2544 knlist_add(&aiocbe->klist, kn, 0);
2549 /* kqueue detach function */
2551 filt_aiodetach(struct knote *kn)
2555 knl = &kn->kn_ptr.p_aio->klist;
2556 knl->kl_lock(knl->kl_lockarg);
2557 if (!knlist_empty(knl))
2558 knlist_remove(knl, kn, 1);
2559 knl->kl_unlock(knl->kl_lockarg);
2562 /* kqueue filter function */
2565 filt_aio(struct knote *kn, long hint)
2567 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2569 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2570 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2572 kn->kn_flags |= EV_EOF;
2576 /* kqueue attach function */
2578 filt_lioattach(struct knote *kn)
2580 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2583 * The aioliojob pointer must be validated before using it, so
2584 * registration is restricted to the kernel; the user cannot
2587 if ((kn->kn_flags & EV_FLAG1) == 0)
2589 kn->kn_ptr.p_lio = lj;
2590 kn->kn_flags &= ~EV_FLAG1;
2592 knlist_add(&lj->klist, kn, 0);
2597 /* kqueue detach function */
2599 filt_liodetach(struct knote *kn)
2603 knl = &kn->kn_ptr.p_lio->klist;
2604 knl->kl_lock(knl->kl_lockarg);
2605 if (!knlist_empty(knl))
2606 knlist_remove(knl, kn, 1);
2607 knl->kl_unlock(knl->kl_lockarg);
2610 /* kqueue filter function */
2613 filt_lio(struct knote *kn, long hint)
2615 struct aioliojob * lj = kn->kn_ptr.p_lio;
2617 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2620 #ifdef COMPAT_FREEBSD32
2622 struct __aiocb_private32 {
2625 uint32_t kernelinfo;
2628 typedef struct oaiocb32 {
2629 int aio_fildes; /* File descriptor */
2630 uint64_t aio_offset __packed; /* File offset for I/O */
2631 uint32_t aio_buf; /* I/O buffer in process space */
2632 uint32_t aio_nbytes; /* Number of bytes for I/O */
2633 struct osigevent32 aio_sigevent; /* Signal to deliver */
2634 int aio_lio_opcode; /* LIO opcode */
2635 int aio_reqprio; /* Request priority -- ignored */
2636 struct __aiocb_private32 _aiocb_private;
2639 typedef struct aiocb32 {
2640 int32_t aio_fildes; /* File descriptor */
2641 uint64_t aio_offset __packed; /* File offset for I/O */
2642 uint32_t aio_buf; /* I/O buffer in process space */
2643 uint32_t aio_nbytes; /* Number of bytes for I/O */
2645 uint32_t __spare2__;
2646 int aio_lio_opcode; /* LIO opcode */
2647 int aio_reqprio; /* Request priority -- ignored */
2648 struct __aiocb_private32 _aiocb_private;
2649 struct sigevent32 aio_sigevent; /* Signal to deliver */
2653 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2657 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2658 * supported by AIO with the old sigevent structure.
2660 CP(*osig, *nsig, sigev_notify);
2661 switch (nsig->sigev_notify) {
2665 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2668 nsig->sigev_notify_kqueue =
2669 osig->__sigev_u.__sigev_notify_kqueue;
2670 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2679 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2681 struct oaiocb32 job32;
2684 bzero(kjob, sizeof(struct aiocb));
2685 error = copyin(ujob, &job32, sizeof(job32));
2689 CP(job32, *kjob, aio_fildes);
2690 CP(job32, *kjob, aio_offset);
2691 PTRIN_CP(job32, *kjob, aio_buf);
2692 CP(job32, *kjob, aio_nbytes);
2693 CP(job32, *kjob, aio_lio_opcode);
2694 CP(job32, *kjob, aio_reqprio);
2695 CP(job32, *kjob, _aiocb_private.status);
2696 CP(job32, *kjob, _aiocb_private.error);
2697 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2698 return (convert_old_sigevent32(&job32.aio_sigevent,
2699 &kjob->aio_sigevent));
2703 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig)
2706 CP(*sig32, *sig, sigev_notify);
2707 switch (sig->sigev_notify) {
2710 case SIGEV_THREAD_ID:
2711 CP(*sig32, *sig, sigev_notify_thread_id);
2714 CP(*sig32, *sig, sigev_signo);
2717 CP(*sig32, *sig, sigev_notify_kqueue);
2718 CP(*sig32, *sig, sigev_notify_kevent_flags);
2719 PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr);
2728 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2730 struct aiocb32 job32;
2733 error = copyin(ujob, &job32, sizeof(job32));
2736 CP(job32, *kjob, aio_fildes);
2737 CP(job32, *kjob, aio_offset);
2738 PTRIN_CP(job32, *kjob, aio_buf);
2739 CP(job32, *kjob, aio_nbytes);
2740 CP(job32, *kjob, aio_lio_opcode);
2741 CP(job32, *kjob, aio_reqprio);
2742 CP(job32, *kjob, _aiocb_private.status);
2743 CP(job32, *kjob, _aiocb_private.error);
2744 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2745 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2749 aiocb32_fetch_status(struct aiocb *ujob)
2751 struct aiocb32 *ujob32;
2753 ujob32 = (struct aiocb32 *)ujob;
2754 return (fuword32(&ujob32->_aiocb_private.status));
2758 aiocb32_fetch_error(struct aiocb *ujob)
2760 struct aiocb32 *ujob32;
2762 ujob32 = (struct aiocb32 *)ujob;
2763 return (fuword32(&ujob32->_aiocb_private.error));
2767 aiocb32_store_status(struct aiocb *ujob, long status)
2769 struct aiocb32 *ujob32;
2771 ujob32 = (struct aiocb32 *)ujob;
2772 return (suword32(&ujob32->_aiocb_private.status, status));
2776 aiocb32_store_error(struct aiocb *ujob, long error)
2778 struct aiocb32 *ujob32;
2780 ujob32 = (struct aiocb32 *)ujob;
2781 return (suword32(&ujob32->_aiocb_private.error, error));
2785 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2787 struct aiocb32 *ujob32;
2789 ujob32 = (struct aiocb32 *)ujob;
2790 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2794 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2797 return (suword32(ujobp, (long)ujob));
2800 static struct aiocb_ops aiocb32_ops = {
2801 .copyin = aiocb32_copyin,
2802 .fetch_status = aiocb32_fetch_status,
2803 .fetch_error = aiocb32_fetch_error,
2804 .store_status = aiocb32_store_status,
2805 .store_error = aiocb32_store_error,
2806 .store_kernelinfo = aiocb32_store_kernelinfo,
2807 .store_aiocb = aiocb32_store_aiocb,
2810 static struct aiocb_ops aiocb32_ops_osigevent = {
2811 .copyin = aiocb32_copyin_old_sigevent,
2812 .fetch_status = aiocb32_fetch_status,
2813 .fetch_error = aiocb32_fetch_error,
2814 .store_status = aiocb32_store_status,
2815 .store_error = aiocb32_store_error,
2816 .store_kernelinfo = aiocb32_store_kernelinfo,
2817 .store_aiocb = aiocb32_store_aiocb,
2821 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2824 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2828 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2830 struct timespec32 ts32;
2831 struct timespec ts, *tsp;
2832 struct aiocb **ujoblist;
2833 uint32_t *ujoblist32;
2836 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2840 /* Get timespec struct. */
2841 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2843 CP(ts32, ts, tv_sec);
2844 CP(ts32, ts, tv_nsec);
2849 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2850 ujoblist32 = (uint32_t *)ujoblist;
2851 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2852 sizeof(ujoblist32[0]));
2854 for (i = uap->nent; i > 0; i--)
2855 ujoblist[i] = PTRIN(ujoblist32[i]);
2857 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2859 uma_zfree(aiol_zone, ujoblist);
2864 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2867 return (sys_aio_cancel(td, (struct aio_cancel_args *)uap));
2871 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2874 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2878 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2881 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2882 &aiocb32_ops_osigevent));
2886 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2889 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2894 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2897 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2898 &aiocb32_ops_osigevent));
2902 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2905 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2910 freebsd32_aio_waitcomplete(struct thread *td,
2911 struct freebsd32_aio_waitcomplete_args *uap)
2913 struct timespec32 ts32;
2914 struct timespec ts, *tsp;
2918 /* Get timespec struct. */
2919 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2922 CP(ts32, ts, tv_sec);
2923 CP(ts32, ts, tv_nsec);
2928 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2933 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2936 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2941 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2943 struct aiocb **acb_list;
2944 struct sigevent *sigp, sig;
2945 struct osigevent32 osig;
2946 uint32_t *acb_list32;
2949 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2953 if (nent < 0 || nent > AIO_LISTIO_MAX)
2956 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2957 error = copyin(uap->sig, &osig, sizeof(osig));
2960 error = convert_old_sigevent32(&osig, &sig);
2967 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2968 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2970 free(acb_list32, M_LIO);
2973 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2974 for (i = 0; i < nent; i++)
2975 acb_list[i] = PTRIN(acb_list32[i]);
2976 free(acb_list32, M_LIO);
2978 error = kern_lio_listio(td, uap->mode,
2979 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2980 &aiocb32_ops_osigevent);
2981 free(acb_list, M_LIO);
2986 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2988 struct aiocb **acb_list;
2989 struct sigevent *sigp, sig;
2990 struct sigevent32 sig32;
2991 uint32_t *acb_list32;
2994 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2998 if (nent < 0 || nent > AIO_LISTIO_MAX)
3001 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3002 error = copyin(uap->sig, &sig32, sizeof(sig32));
3005 error = convert_sigevent32(&sig32, &sig);
3012 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3013 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3015 free(acb_list32, M_LIO);
3018 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3019 for (i = 0; i < nent; i++)
3020 acb_list[i] = PTRIN(acb_list32[i]);
3021 free(acb_list32, M_LIO);
3023 error = kern_lio_listio(td, uap->mode,
3024 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3026 free(acb_list, M_LIO);