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/eventhandler.h>
32 #include <sys/sysproto.h>
33 #include <sys/filedesc.h>
34 #include <sys/kernel.h>
35 #include <sys/module.h>
36 #include <sys/kthread.h>
37 #include <sys/fcntl.h>
39 #include <sys/limits.h>
41 #include <sys/mutex.h>
42 #include <sys/unistd.h>
43 #include <sys/posix4.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/syscall.h>
52 #include <sys/sysent.h>
53 #include <sys/sysctl.h>
55 #include <sys/taskqueue.h>
56 #include <sys/vnode.h>
58 #include <sys/event.h>
59 #include <sys/mount.h>
61 #include <machine/atomic.h>
64 #include <vm/vm_extern.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_object.h>
71 #include "opt_vfs_aio.h"
74 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
75 * overflow. (XXX will be removed soon.)
77 static u_long jobrefid;
80 * Counter for aio_fsync.
82 static uint64_t jobseqno;
85 #define JOBST_JOBQSOCK 1
86 #define JOBST_JOBQGLOBAL 2
87 #define JOBST_JOBRUNNING 3
88 #define JOBST_JOBFINISHED 4
89 #define JOBST_JOBQBUF 5
90 #define JOBST_JOBQSYNC 6
92 #ifndef MAX_AIO_PER_PROC
93 #define MAX_AIO_PER_PROC 32
96 #ifndef MAX_AIO_QUEUE_PER_PROC
97 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
100 #ifndef MAX_AIO_PROCS
101 #define MAX_AIO_PROCS 32
104 #ifndef MAX_AIO_QUEUE
105 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
108 #ifndef TARGET_AIO_PROCS
109 #define TARGET_AIO_PROCS 4
113 #define MAX_BUF_AIO 16
116 #ifndef AIOD_TIMEOUT_DEFAULT
117 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
120 #ifndef AIOD_LIFETIME_DEFAULT
121 #define AIOD_LIFETIME_DEFAULT (30 * hz)
124 FEATURE(aio, "Asynchronous I/O");
126 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
128 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
130 static int max_aio_procs = MAX_AIO_PROCS;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
132 CTLFLAG_RW, &max_aio_procs, 0,
133 "Maximum number of kernel threads to use for handling async IO ");
135 static int num_aio_procs = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
137 CTLFLAG_RD, &num_aio_procs, 0,
138 "Number of presently active kernel threads for async IO");
141 * The code will adjust the actual number of AIO processes towards this
142 * number when it gets a chance.
144 static int target_aio_procs = TARGET_AIO_PROCS;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
146 0, "Preferred number of ready kernel threads for async IO");
148 static int max_queue_count = MAX_AIO_QUEUE;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
150 "Maximum number of aio requests to queue, globally");
152 static int num_queue_count = 0;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
154 "Number of queued aio requests");
156 static int num_buf_aio = 0;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
158 "Number of aio requests presently handled by the buf subsystem");
160 /* Number of async I/O thread in the process of being started */
161 /* XXX This should be local to aio_aqueue() */
162 static int num_aio_resv_start = 0;
164 static int aiod_timeout;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
166 "Timeout value for synchronous aio operations");
168 static int aiod_lifetime;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
170 "Maximum lifetime for idle aiod");
172 static int unloadable = 0;
173 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
174 "Allow unload of aio (not recommended)");
177 static int max_aio_per_proc = MAX_AIO_PER_PROC;
178 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
179 0, "Maximum active aio requests per process (stored in the process)");
181 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
182 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
183 &max_aio_queue_per_proc, 0,
184 "Maximum queued aio requests per process (stored in the process)");
186 static int max_buf_aio = MAX_BUF_AIO;
187 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
188 "Maximum buf aio requests per process (stored in the process)");
190 typedef struct oaiocb {
191 int aio_fildes; /* File descriptor */
192 off_t aio_offset; /* File offset for I/O */
193 volatile void *aio_buf; /* I/O buffer in process space */
194 size_t aio_nbytes; /* Number of bytes for I/O */
195 struct osigevent aio_sigevent; /* Signal to deliver */
196 int aio_lio_opcode; /* LIO opcode */
197 int aio_reqprio; /* Request priority -- ignored */
198 struct __aiocb_private _aiocb_private;
202 * Below is a key of locks used to protect each member of struct aiocblist
203 * aioliojob and kaioinfo and any backends.
205 * * - need not protected
206 * a - locked by kaioinfo lock
207 * b - locked by backend lock, the backend lock can be null in some cases,
208 * for example, BIO belongs to this type, in this case, proc lock is
210 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
214 * Current, there is only two backends: BIO and generic file I/O.
215 * socket I/O is served by generic file I/O, this is not a good idea, since
216 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
217 * threads, if there is no thread to serve socket I/O, the socket I/O will be
218 * delayed too long or starved, we should create some threads dedicated to
219 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
220 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
221 * structure is not safe because there is race between userland and aio
226 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
227 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
228 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
229 int jobflags; /* (a) job flags */
230 int jobstate; /* (b) job state */
231 int inputcharge; /* (*) input blockes */
232 int outputcharge; /* (*) output blockes */
233 struct buf *bp; /* (*) private to BIO backend,
236 struct proc *userproc; /* (*) user process */
237 struct ucred *cred; /* (*) active credential when created */
238 struct file *fd_file; /* (*) pointer to file structure */
239 struct aioliojob *lio; /* (*) optional lio job */
240 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
241 struct knlist klist; /* (a) list of knotes */
242 struct aiocb uaiocb; /* (*) kernel I/O control block */
243 ksiginfo_t ksi; /* (a) realtime signal info */
244 struct task biotask; /* (*) private to BIO backend */
245 uint64_t seqno; /* (*) job number */
246 int pending; /* (a) number of pending I/O, aio_fsync only */
250 #define AIOCBLIST_DONE 0x01
251 #define AIOCBLIST_BUFDONE 0x02
252 #define AIOCBLIST_RUNDOWN 0x04
253 #define AIOCBLIST_CHECKSYNC 0x08
258 #define AIOP_FREE 0x1 /* proc on free queue */
260 struct aiothreadlist {
261 int aiothreadflags; /* (c) AIO proc flags */
262 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
263 struct thread *aiothread; /* (*) the AIO thread */
267 * data-structure for lio signal management
270 int lioj_flags; /* (a) listio flags */
271 int lioj_count; /* (a) listio flags */
272 int lioj_finished_count; /* (a) listio flags */
273 struct sigevent lioj_signal; /* (a) signal on all I/O done */
274 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
275 struct knlist klist; /* (a) list of knotes */
276 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
279 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
280 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
281 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
284 * per process aio data structure
287 struct mtx kaio_mtx; /* the lock to protect this struct */
288 int kaio_flags; /* (a) per process kaio flags */
289 int kaio_maxactive_count; /* (*) maximum number of AIOs */
290 int kaio_active_count; /* (c) number of currently used AIOs */
291 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
292 int kaio_count; /* (a) size of AIO queue */
293 int kaio_ballowed_count; /* (*) maximum number of buffers */
294 int kaio_buffer_count; /* (a) number of physio buffers */
295 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
296 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
297 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
298 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
299 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
300 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
303 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
304 struct task kaio_task; /* (*) task to kick aio threads */
307 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
308 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
309 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
310 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
312 #define KAIO_RUNDOWN 0x1 /* process is being run down */
313 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
316 * Operations used to interact with userland aio control blocks.
317 * Different ABIs provide their own operations.
320 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
321 long (*fetch_status)(struct aiocb *ujob);
322 long (*fetch_error)(struct aiocb *ujob);
323 int (*store_status)(struct aiocb *ujob, long status);
324 int (*store_error)(struct aiocb *ujob, long error);
325 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
326 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
329 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
330 static struct sema aio_newproc_sem;
331 static struct mtx aio_job_mtx;
332 static struct mtx aio_sock_mtx;
333 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
334 static struct unrhdr *aiod_unr;
336 void aio_init_aioinfo(struct proc *p);
337 static void aio_onceonly(void);
338 static int aio_free_entry(struct aiocblist *aiocbe);
339 static void aio_process(struct aiocblist *aiocbe);
340 static int aio_newproc(int *);
341 int aio_aqueue(struct thread *td, struct aiocb *job,
342 struct aioliojob *lio, int type, struct aiocb_ops *ops);
343 static void aio_physwakeup(struct buf *bp);
344 static void aio_proc_rundown(void *arg, struct proc *p);
345 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
346 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
347 static void biohelper(void *, int);
348 static void aio_daemon(void *param);
349 static void aio_swake_cb(struct socket *, struct sockbuf *);
350 static int aio_unload(void);
351 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
354 static int aio_kick(struct proc *userp);
355 static void aio_kick_nowait(struct proc *userp);
356 static void aio_kick_helper(void *context, int pending);
357 static int filt_aioattach(struct knote *kn);
358 static void filt_aiodetach(struct knote *kn);
359 static int filt_aio(struct knote *kn, long hint);
360 static int filt_lioattach(struct knote *kn);
361 static void filt_liodetach(struct knote *kn);
362 static int filt_lio(struct knote *kn, long hint);
366 * kaio Per process async io info
367 * aiop async io thread data
368 * aiocb async io jobs
369 * aiol list io job pointer - internal to aio_suspend XXX
370 * aiolio list io jobs
372 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
374 /* kqueue filters for aio */
375 static struct filterops aio_filtops = {
377 .f_attach = filt_aioattach,
378 .f_detach = filt_aiodetach,
381 static struct filterops lio_filtops = {
383 .f_attach = filt_lioattach,
384 .f_detach = filt_liodetach,
388 static eventhandler_tag exit_tag, exec_tag;
390 TASKQUEUE_DEFINE_THREAD(aiod_bio);
393 * Main operations function for use as a kernel module.
396 aio_modload(struct module *module, int cmd, void *arg)
405 error = aio_unload();
416 static moduledata_t aio_mod = {
422 SYSCALL_MODULE_HELPER(aio_cancel);
423 SYSCALL_MODULE_HELPER(aio_error);
424 SYSCALL_MODULE_HELPER(aio_fsync);
425 SYSCALL_MODULE_HELPER(aio_read);
426 SYSCALL_MODULE_HELPER(aio_return);
427 SYSCALL_MODULE_HELPER(aio_suspend);
428 SYSCALL_MODULE_HELPER(aio_waitcomplete);
429 SYSCALL_MODULE_HELPER(aio_write);
430 SYSCALL_MODULE_HELPER(lio_listio);
431 SYSCALL_MODULE_HELPER(oaio_read);
432 SYSCALL_MODULE_HELPER(oaio_write);
433 SYSCALL_MODULE_HELPER(olio_listio);
435 DECLARE_MODULE(aio, aio_mod,
436 SI_SUB_VFS, SI_ORDER_ANY);
437 MODULE_VERSION(aio, 1);
440 * Startup initialization
446 /* XXX: should probably just use so->callback */
447 aio_swake = &aio_swake_cb;
448 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
449 EVENTHANDLER_PRI_ANY);
450 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
451 EVENTHANDLER_PRI_ANY);
452 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
453 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
454 TAILQ_INIT(&aio_freeproc);
455 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
456 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
457 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
458 TAILQ_INIT(&aio_jobs);
459 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
460 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
461 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
462 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
463 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
464 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
465 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
466 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
467 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
468 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
469 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
470 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
471 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
473 async_io_version = _POSIX_VERSION;
474 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
475 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
476 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
480 * Callback for unload of AIO when used as a module.
488 * XXX: no unloads by default, it's too dangerous.
489 * perhaps we could do it if locked out callers and then
490 * did an aio_proc_rundown() on each process.
492 * jhb: aio_proc_rundown() needs to run on curproc though,
493 * so I don't think that would fly.
498 error = kqueue_del_filteropts(EVFILT_AIO);
501 error = kqueue_del_filteropts(EVFILT_LIO);
504 async_io_version = 0;
506 taskqueue_free(taskqueue_aiod_bio);
507 delete_unrhdr(aiod_unr);
508 uma_zdestroy(kaio_zone);
509 uma_zdestroy(aiop_zone);
510 uma_zdestroy(aiocb_zone);
511 uma_zdestroy(aiol_zone);
512 uma_zdestroy(aiolio_zone);
513 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
514 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
515 mtx_destroy(&aio_job_mtx);
516 mtx_destroy(&aio_sock_mtx);
517 sema_destroy(&aio_newproc_sem);
518 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
519 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
520 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
525 * Init the per-process aioinfo structure. The aioinfo limits are set
526 * per-process for user limit (resource) management.
529 aio_init_aioinfo(struct proc *p)
533 ki = uma_zalloc(kaio_zone, M_WAITOK);
534 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
536 ki->kaio_maxactive_count = max_aio_per_proc;
537 ki->kaio_active_count = 0;
538 ki->kaio_qallowed_count = max_aio_queue_per_proc;
540 ki->kaio_ballowed_count = max_buf_aio;
541 ki->kaio_buffer_count = 0;
542 TAILQ_INIT(&ki->kaio_all);
543 TAILQ_INIT(&ki->kaio_done);
544 TAILQ_INIT(&ki->kaio_jobqueue);
545 TAILQ_INIT(&ki->kaio_bufqueue);
546 TAILQ_INIT(&ki->kaio_liojoblist);
547 TAILQ_INIT(&ki->kaio_sockqueue);
548 TAILQ_INIT(&ki->kaio_syncqueue);
549 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
551 if (p->p_aioinfo == NULL) {
556 mtx_destroy(&ki->kaio_mtx);
557 uma_zfree(kaio_zone, ki);
560 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
565 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
571 ksi->ksi_code = SI_ASYNCIO;
572 ksi->ksi_flags |= KSI_EXT | KSI_INS;
573 ret = psignal_event(p, sigev, ksi);
580 * Free a job entry. Wait for completion if it is currently active, but don't
581 * delay forever. If we delay, we return a flag that says that we have to
582 * restart the queue scan.
585 aio_free_entry(struct aiocblist *aiocbe)
588 struct aioliojob *lj;
591 p = aiocbe->userproc;
596 AIO_LOCK_ASSERT(ki, MA_OWNED);
597 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
599 atomic_subtract_int(&num_queue_count, 1);
602 MPASS(ki->kaio_count >= 0);
604 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
605 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
610 lj->lioj_finished_count--;
612 if (lj->lioj_count == 0) {
613 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
614 /* lio is going away, we need to destroy any knotes */
615 knlist_delete(&lj->klist, curthread, 1);
617 sigqueue_take(&lj->lioj_ksi);
619 uma_zfree(aiolio_zone, lj);
623 /* aiocbe is going away, we need to destroy any knotes */
624 knlist_delete(&aiocbe->klist, curthread, 1);
626 sigqueue_take(&aiocbe->ksi);
629 MPASS(aiocbe->bp == NULL);
630 aiocbe->jobstate = JOBST_NULL;
634 * The thread argument here is used to find the owning process
635 * and is also passed to fo_close() which may pass it to various
636 * places such as devsw close() routines. Because of that, we
637 * need a thread pointer from the process owning the job that is
638 * persistent and won't disappear out from under us or move to
641 * Currently, all the callers of this function call it to remove
642 * an aiocblist from the current process' job list either via a
643 * syscall or due to the current process calling exit() or
644 * execve(). Thus, we know that p == curproc. We also know that
645 * curthread can't exit since we are curthread.
647 * Therefore, we use curthread as the thread to pass to
648 * knlist_delete(). This does mean that it is possible for the
649 * thread pointer at close time to differ from the thread pointer
650 * at open time, but this is already true of file descriptors in
651 * a multithreaded process.
653 fdrop(aiocbe->fd_file, curthread);
654 crfree(aiocbe->cred);
655 uma_zfree(aiocb_zone, aiocbe);
662 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
664 aio_proc_rundown(arg, p);
668 * Rundown the jobs for a given process.
671 aio_proc_rundown(void *arg, struct proc *p)
674 struct aioliojob *lj;
675 struct aiocblist *cbe, *cbn;
680 KASSERT(curthread->td_proc == p,
681 ("%s: called on non-curproc", __func__));
687 ki->kaio_flags |= KAIO_RUNDOWN;
692 * Try to cancel all pending requests. This code simulates
693 * aio_cancel on all pending I/O requests.
695 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
697 mtx_lock(&aio_job_mtx);
698 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
699 TAILQ_REMOVE(&aio_jobs, cbe, list);
701 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
703 MPASS(fp->f_type == DTYPE_SOCKET);
705 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
707 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
708 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
711 mtx_unlock(&aio_job_mtx);
714 cbe->jobstate = JOBST_JOBFINISHED;
715 cbe->uaiocb._aiocb_private.status = -1;
716 cbe->uaiocb._aiocb_private.error = ECANCELED;
717 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
718 aio_bio_done_notify(p, cbe, DONE_QUEUE);
722 /* Wait for all running I/O to be finished */
723 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
724 TAILQ_FIRST(&ki->kaio_jobqueue)) {
725 ki->kaio_flags |= KAIO_WAKEUP;
726 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
730 /* Free all completed I/O requests. */
731 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
734 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
735 if (lj->lioj_count == 0) {
736 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
737 knlist_delete(&lj->klist, curthread, 1);
739 sigqueue_take(&lj->lioj_ksi);
741 uma_zfree(aiolio_zone, lj);
743 panic("LIO job not cleaned up: C:%d, FC:%d\n",
744 lj->lioj_count, lj->lioj_finished_count);
748 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
749 mtx_destroy(&ki->kaio_mtx);
750 uma_zfree(kaio_zone, ki);
755 * Select a job to run (called by an AIO daemon).
757 static struct aiocblist *
758 aio_selectjob(struct aiothreadlist *aiop)
760 struct aiocblist *aiocbe;
764 mtx_assert(&aio_job_mtx, MA_OWNED);
765 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
766 userp = aiocbe->userproc;
767 ki = userp->p_aioinfo;
769 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
770 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
771 /* Account for currently active jobs. */
772 ki->kaio_active_count++;
773 aiocbe->jobstate = JOBST_JOBRUNNING;
781 * Move all data to a permanent storage device, this code
782 * simulates fsync syscall.
785 aio_fsync_vnode(struct thread *td, struct vnode *vp)
791 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
792 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
794 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
795 if (vp->v_object != NULL) {
796 VM_OBJECT_LOCK(vp->v_object);
797 vm_object_page_clean(vp->v_object, 0, 0, 0);
798 VM_OBJECT_UNLOCK(vp->v_object);
800 error = VOP_FSYNC(vp, MNT_WAIT, td);
803 vn_finished_write(mp);
805 VFS_UNLOCK_GIANT(vfslocked);
810 * The AIO processing activity. This is the code that does the I/O request for
811 * the non-physio version of the operations. The normal vn operations are used,
812 * and this code should work in all instances for every type of file, including
813 * pipes, sockets, fifos, and regular files.
815 * XXX I don't think it works well for socket, pipe, and fifo.
818 aio_process(struct aiocblist *aiocbe)
820 struct ucred *td_savedcred;
829 int oublock_st, oublock_end;
830 int inblock_st, inblock_end;
833 td_savedcred = td->td_ucred;
834 td->td_ucred = aiocbe->cred;
835 cb = &aiocbe->uaiocb;
836 fp = aiocbe->fd_file;
838 if (cb->aio_lio_opcode == LIO_SYNC) {
841 if (fp->f_vnode != NULL)
842 error = aio_fsync_vnode(td, fp->f_vnode);
843 cb->_aiocb_private.error = error;
844 cb->_aiocb_private.status = 0;
845 td->td_ucred = td_savedcred;
849 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
850 aiov.iov_len = cb->aio_nbytes;
852 auio.uio_iov = &aiov;
854 auio.uio_offset = cb->aio_offset;
855 auio.uio_resid = cb->aio_nbytes;
856 cnt = cb->aio_nbytes;
857 auio.uio_segflg = UIO_USERSPACE;
860 inblock_st = td->td_ru.ru_inblock;
861 oublock_st = td->td_ru.ru_oublock;
863 * aio_aqueue() acquires a reference to the file that is
864 * released in aio_free_entry().
866 if (cb->aio_lio_opcode == LIO_READ) {
867 auio.uio_rw = UIO_READ;
868 if (auio.uio_resid == 0)
871 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
873 if (fp->f_type == DTYPE_VNODE)
875 auio.uio_rw = UIO_WRITE;
876 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
878 inblock_end = td->td_ru.ru_inblock;
879 oublock_end = td->td_ru.ru_oublock;
881 aiocbe->inputcharge = inblock_end - inblock_st;
882 aiocbe->outputcharge = oublock_end - oublock_st;
884 if ((error) && (auio.uio_resid != cnt)) {
885 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
887 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
889 if (fp->f_type == DTYPE_SOCKET) {
891 if (so->so_options & SO_NOSIGPIPE)
895 PROC_LOCK(aiocbe->userproc);
896 psignal(aiocbe->userproc, SIGPIPE);
897 PROC_UNLOCK(aiocbe->userproc);
902 cnt -= auio.uio_resid;
903 cb->_aiocb_private.error = error;
904 cb->_aiocb_private.status = cnt;
905 td->td_ucred = td_savedcred;
909 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
911 struct aioliojob *lj;
913 struct aiocblist *scb, *scbn;
916 ki = userp->p_aioinfo;
917 AIO_LOCK_ASSERT(ki, MA_OWNED);
921 lj->lioj_finished_count++;
922 if (lj->lioj_count == lj->lioj_finished_count)
925 if (type == DONE_QUEUE) {
926 aiocbe->jobflags |= AIOCBLIST_DONE;
928 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
930 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
931 aiocbe->jobstate = JOBST_JOBFINISHED;
933 if (ki->kaio_flags & KAIO_RUNDOWN)
934 goto notification_done;
936 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
937 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
938 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
940 KNOTE_LOCKED(&aiocbe->klist, 1);
943 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
944 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
945 KNOTE_LOCKED(&lj->klist, 1);
947 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
949 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
950 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
951 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
952 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
957 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
958 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
959 if (aiocbe->fd_file == scb->fd_file &&
960 aiocbe->seqno < scb->seqno) {
961 if (--scb->pending == 0) {
962 mtx_lock(&aio_job_mtx);
963 scb->jobstate = JOBST_JOBQGLOBAL;
964 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
965 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
966 aio_kick_nowait(userp);
967 mtx_unlock(&aio_job_mtx);
972 if (ki->kaio_flags & KAIO_WAKEUP) {
973 ki->kaio_flags &= ~KAIO_WAKEUP;
974 wakeup(&userp->p_aioinfo);
979 * The AIO daemon, most of the actual work is done in aio_process,
980 * but the setup (and address space mgmt) is done in this routine.
983 aio_daemon(void *_id)
985 struct aiocblist *aiocbe;
986 struct aiothreadlist *aiop;
988 struct proc *curcp, *mycp, *userp;
989 struct vmspace *myvm, *tmpvm;
990 struct thread *td = curthread;
991 int id = (intptr_t)_id;
994 * Local copies of curproc (cp) and vmspace (myvm)
997 myvm = mycp->p_vmspace;
999 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1002 * Allocate and ready the aio control info. There is one aiop structure
1005 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1006 aiop->aiothread = td;
1007 aiop->aiothreadflags = 0;
1009 /* The daemon resides in its own pgrp. */
1013 * Wakeup parent process. (Parent sleeps to keep from blasting away
1014 * and creating too many daemons.)
1016 sema_post(&aio_newproc_sem);
1018 mtx_lock(&aio_job_mtx);
1021 * curcp is the current daemon process context.
1022 * userp is the current user process context.
1027 * Take daemon off of free queue
1029 if (aiop->aiothreadflags & AIOP_FREE) {
1030 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1031 aiop->aiothreadflags &= ~AIOP_FREE;
1037 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1038 mtx_unlock(&aio_job_mtx);
1039 userp = aiocbe->userproc;
1042 * Connect to process address space for user program.
1044 if (userp != curcp) {
1046 * Save the current address space that we are
1049 tmpvm = mycp->p_vmspace;
1052 * Point to the new user address space, and
1055 mycp->p_vmspace = userp->p_vmspace;
1056 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1058 /* Activate the new mapping. */
1059 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1062 * If the old address space wasn't the daemons
1063 * own address space, then we need to remove the
1064 * daemon's reference from the other process
1065 * that it was acting on behalf of.
1067 if (tmpvm != myvm) {
1068 vmspace_free(tmpvm);
1073 ki = userp->p_aioinfo;
1075 /* Do the I/O function. */
1076 aio_process(aiocbe);
1078 mtx_lock(&aio_job_mtx);
1079 /* Decrement the active job count. */
1080 ki->kaio_active_count--;
1081 mtx_unlock(&aio_job_mtx);
1084 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1085 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1088 mtx_lock(&aio_job_mtx);
1092 * Disconnect from user address space.
1094 if (curcp != mycp) {
1096 mtx_unlock(&aio_job_mtx);
1098 /* Get the user address space to disconnect from. */
1099 tmpvm = mycp->p_vmspace;
1101 /* Get original address space for daemon. */
1102 mycp->p_vmspace = myvm;
1104 /* Activate the daemon's address space. */
1105 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1107 if (tmpvm == myvm) {
1108 printf("AIOD: vmspace problem -- %d\n",
1112 /* Remove our vmspace reference. */
1113 vmspace_free(tmpvm);
1117 mtx_lock(&aio_job_mtx);
1119 * We have to restart to avoid race, we only sleep if
1120 * no job can be selected, that should be
1126 mtx_assert(&aio_job_mtx, MA_OWNED);
1128 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1129 aiop->aiothreadflags |= AIOP_FREE;
1132 * If daemon is inactive for a long time, allow it to exit,
1133 * thereby freeing resources.
1135 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1137 if (TAILQ_EMPTY(&aio_jobs)) {
1138 if ((aiop->aiothreadflags & AIOP_FREE) &&
1139 (num_aio_procs > target_aio_procs)) {
1140 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1142 mtx_unlock(&aio_job_mtx);
1143 uma_zfree(aiop_zone, aiop);
1144 free_unr(aiod_unr, id);
1146 if (mycp->p_vmspace->vm_refcnt <= 1) {
1147 printf("AIOD: bad vm refcnt for"
1148 " exiting daemon: %d\n",
1149 mycp->p_vmspace->vm_refcnt);
1157 mtx_unlock(&aio_job_mtx);
1158 panic("shouldn't be here\n");
1162 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1163 * AIO daemon modifies its environment itself.
1166 aio_newproc(int *start)
1172 id = alloc_unr(aiod_unr);
1173 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1174 RFNOWAIT, 0, "aiod%d", id);
1177 * Wait until daemon is started.
1179 sema_wait(&aio_newproc_sem);
1180 mtx_lock(&aio_job_mtx);
1184 mtx_unlock(&aio_job_mtx);
1186 free_unr(aiod_unr, id);
1192 * Try the high-performance, low-overhead physio method for eligible
1193 * VCHR devices. This method doesn't use an aio helper thread, and
1194 * thus has very low overhead.
1196 * Assumes that the caller, aio_aqueue(), has incremented the file
1197 * structure's reference count, preventing its deallocation for the
1198 * duration of this call.
1201 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1207 struct kaioinfo *ki;
1208 struct aioliojob *lj;
1211 cb = &aiocbe->uaiocb;
1212 fp = aiocbe->fd_file;
1214 if (fp->f_type != DTYPE_VNODE)
1220 * If its not a disk, we don't want to return a positive error.
1221 * It causes the aio code to not fall through to try the thread
1222 * way when you're talking to a regular file.
1224 if (!vn_isdisk(vp, &error)) {
1225 if (error == ENOTBLK)
1231 if (vp->v_bufobj.bo_bsize == 0)
1234 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1237 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1240 if (cb->aio_nbytes >
1241 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1245 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1248 /* Create and build a buffer header for a transfer. */
1249 bp = (struct buf *)getpbuf(NULL);
1254 ki->kaio_buffer_count++;
1261 * Get a copy of the kva from the physical buffer.
1265 bp->b_bcount = cb->aio_nbytes;
1266 bp->b_bufsize = cb->aio_nbytes;
1267 bp->b_iodone = aio_physwakeup;
1268 bp->b_saveaddr = bp->b_data;
1269 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1270 bp->b_offset = cb->aio_offset;
1271 bp->b_iooffset = cb->aio_offset;
1272 bp->b_blkno = btodb(cb->aio_offset);
1273 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1276 * Bring buffer into kernel space.
1278 if (vmapbuf(bp) < 0) {
1285 bp->b_caller1 = (void *)aiocbe;
1286 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1287 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1288 aiocbe->jobstate = JOBST_JOBQBUF;
1289 cb->_aiocb_private.status = cb->aio_nbytes;
1292 atomic_add_int(&num_queue_count, 1);
1293 atomic_add_int(&num_buf_aio, 1);
1297 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1299 /* Perform transfer. */
1300 dev_strategy(vp->v_rdev, bp);
1306 ki->kaio_buffer_count--;
1316 * Wake up aio requests that may be serviceable now.
1319 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1321 struct aiocblist *cb, *cbn;
1324 SOCKBUF_LOCK_ASSERT(sb);
1325 if (sb == &so->so_snd)
1330 sb->sb_flags &= ~SB_AIO;
1331 mtx_lock(&aio_job_mtx);
1332 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1333 if (opcode == cb->uaiocb.aio_lio_opcode) {
1334 if (cb->jobstate != JOBST_JOBQSOCK)
1335 panic("invalid queue value");
1337 * We don't have actual sockets backend yet,
1338 * so we simply move the requests to the generic
1341 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1342 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1343 aio_kick_nowait(cb->userproc);
1346 mtx_unlock(&aio_job_mtx);
1350 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1354 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1355 * supported by AIO with the old sigevent structure.
1357 nsig->sigev_notify = osig->sigev_notify;
1358 switch (nsig->sigev_notify) {
1362 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1365 nsig->sigev_notify_kqueue =
1366 osig->__sigev_u.__sigev_notify_kqueue;
1367 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1376 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1378 struct oaiocb *ojob;
1381 bzero(kjob, sizeof(struct aiocb));
1382 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1385 ojob = (struct oaiocb *)kjob;
1386 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1390 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1393 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1397 aiocb_fetch_status(struct aiocb *ujob)
1400 return (fuword(&ujob->_aiocb_private.status));
1404 aiocb_fetch_error(struct aiocb *ujob)
1407 return (fuword(&ujob->_aiocb_private.error));
1411 aiocb_store_status(struct aiocb *ujob, long status)
1414 return (suword(&ujob->_aiocb_private.status, status));
1418 aiocb_store_error(struct aiocb *ujob, long error)
1421 return (suword(&ujob->_aiocb_private.error, error));
1425 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1428 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1432 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1435 return (suword(ujobp, (long)ujob));
1438 static struct aiocb_ops aiocb_ops = {
1439 .copyin = aiocb_copyin,
1440 .fetch_status = aiocb_fetch_status,
1441 .fetch_error = aiocb_fetch_error,
1442 .store_status = aiocb_store_status,
1443 .store_error = aiocb_store_error,
1444 .store_kernelinfo = aiocb_store_kernelinfo,
1445 .store_aiocb = aiocb_store_aiocb,
1448 static struct aiocb_ops aiocb_ops_osigevent = {
1449 .copyin = aiocb_copyin_old_sigevent,
1450 .fetch_status = aiocb_fetch_status,
1451 .fetch_error = aiocb_fetch_error,
1452 .store_status = aiocb_store_status,
1453 .store_error = aiocb_store_error,
1454 .store_kernelinfo = aiocb_store_kernelinfo,
1455 .store_aiocb = aiocb_store_aiocb,
1459 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1460 * technique is done in this code.
1463 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1464 int type, struct aiocb_ops *ops)
1466 struct proc *p = td->td_proc;
1469 struct aiocblist *aiocbe, *cb;
1470 struct kaioinfo *ki;
1478 if (p->p_aioinfo == NULL)
1479 aio_init_aioinfo(p);
1483 ops->store_status(job, -1);
1484 ops->store_error(job, 0);
1485 ops->store_kernelinfo(job, -1);
1487 if (num_queue_count >= max_queue_count ||
1488 ki->kaio_count >= ki->kaio_qallowed_count) {
1489 ops->store_error(job, EAGAIN);
1493 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1494 aiocbe->inputcharge = 0;
1495 aiocbe->outputcharge = 0;
1496 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1498 error = ops->copyin(job, &aiocbe->uaiocb);
1500 ops->store_error(job, error);
1501 uma_zfree(aiocb_zone, aiocbe);
1505 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1506 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1507 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1508 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1509 ops->store_error(job, EINVAL);
1510 uma_zfree(aiocb_zone, aiocbe);
1514 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1515 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1516 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1517 uma_zfree(aiocb_zone, aiocbe);
1521 ksiginfo_init(&aiocbe->ksi);
1523 /* Save userspace address of the job info. */
1524 aiocbe->uuaiocb = job;
1526 /* Get the opcode. */
1527 if (type != LIO_NOP)
1528 aiocbe->uaiocb.aio_lio_opcode = type;
1529 opcode = aiocbe->uaiocb.aio_lio_opcode;
1531 /* Fetch the file object for the specified file descriptor. */
1532 fd = aiocbe->uaiocb.aio_fildes;
1535 error = fget_write(td, fd, &fp);
1538 error = fget_read(td, fd, &fp);
1541 error = fget(td, fd, &fp);
1544 uma_zfree(aiocb_zone, aiocbe);
1545 ops->store_error(job, error);
1549 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1554 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1559 aiocbe->fd_file = fp;
1561 mtx_lock(&aio_job_mtx);
1563 aiocbe->seqno = jobseqno++;
1564 mtx_unlock(&aio_job_mtx);
1565 error = ops->store_kernelinfo(job, jid);
1570 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1572 if (opcode == LIO_NOP) {
1574 uma_zfree(aiocb_zone, aiocbe);
1577 if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1578 (opcode != LIO_SYNC)) {
1583 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1585 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1586 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1587 kev.filter = EVFILT_AIO;
1588 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1589 kev.data = (intptr_t)aiocbe;
1590 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1591 error = kqfd_register(kqfd, &kev, td, 1);
1595 uma_zfree(aiocb_zone, aiocbe);
1596 ops->store_error(job, error);
1601 ops->store_error(job, EINPROGRESS);
1602 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1603 aiocbe->userproc = p;
1604 aiocbe->cred = crhold(td->td_ucred);
1605 aiocbe->jobflags = 0;
1608 if (opcode == LIO_SYNC)
1611 if (fp->f_type == DTYPE_SOCKET) {
1613 * Alternate queueing for socket ops: Reach down into the
1614 * descriptor to get the socket data. Then check to see if the
1615 * socket is ready to be read or written (based on the requested
1618 * If it is not ready for io, then queue the aiocbe on the
1619 * socket, and set the flags so we get a call when sbnotify()
1622 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1623 * and unlock the snd sockbuf for no reason.
1626 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1628 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1629 LIO_WRITE) && (!sowriteable(so)))) {
1630 sb->sb_flags |= SB_AIO;
1632 mtx_lock(&aio_job_mtx);
1633 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1634 mtx_unlock(&aio_job_mtx);
1637 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1638 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1639 aiocbe->jobstate = JOBST_JOBQSOCK;
1645 atomic_add_int(&num_queue_count, 1);
1652 if ((error = aio_qphysio(p, aiocbe)) == 0)
1656 aiocbe->uaiocb._aiocb_private.error = error;
1657 ops->store_error(job, error);
1662 /* No buffer for daemon I/O. */
1664 atomic_add_int(&num_queue_count, 1);
1670 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1671 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1672 if (opcode == LIO_SYNC) {
1673 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1674 if (cb->fd_file == aiocbe->fd_file &&
1675 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1676 cb->seqno < aiocbe->seqno) {
1677 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1681 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1682 if (cb->fd_file == aiocbe->fd_file &&
1683 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1684 cb->seqno < aiocbe->seqno) {
1685 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1689 if (aiocbe->pending != 0) {
1690 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1691 aiocbe->jobstate = JOBST_JOBQSYNC;
1696 mtx_lock(&aio_job_mtx);
1697 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1698 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1700 mtx_unlock(&aio_job_mtx);
1708 aio_kick_nowait(struct proc *userp)
1710 struct kaioinfo *ki = userp->p_aioinfo;
1711 struct aiothreadlist *aiop;
1713 mtx_assert(&aio_job_mtx, MA_OWNED);
1714 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1715 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1716 aiop->aiothreadflags &= ~AIOP_FREE;
1717 wakeup(aiop->aiothread);
1718 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1719 ((ki->kaio_active_count + num_aio_resv_start) <
1720 ki->kaio_maxactive_count)) {
1721 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1726 aio_kick(struct proc *userp)
1728 struct kaioinfo *ki = userp->p_aioinfo;
1729 struct aiothreadlist *aiop;
1732 mtx_assert(&aio_job_mtx, MA_OWNED);
1734 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1735 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1736 aiop->aiothreadflags &= ~AIOP_FREE;
1737 wakeup(aiop->aiothread);
1738 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1739 ((ki->kaio_active_count + num_aio_resv_start) <
1740 ki->kaio_maxactive_count)) {
1741 num_aio_resv_start++;
1742 mtx_unlock(&aio_job_mtx);
1743 error = aio_newproc(&num_aio_resv_start);
1744 mtx_lock(&aio_job_mtx);
1746 num_aio_resv_start--;
1756 aio_kick_helper(void *context, int pending)
1758 struct proc *userp = context;
1760 mtx_lock(&aio_job_mtx);
1761 while (--pending >= 0) {
1762 if (aio_kick(userp))
1765 mtx_unlock(&aio_job_mtx);
1769 * Support the aio_return system call, as a side-effect, kernel resources are
1773 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1775 struct proc *p = td->td_proc;
1776 struct aiocblist *cb;
1777 struct kaioinfo *ki;
1784 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1785 if (cb->uuaiocb == uaiocb)
1789 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1790 status = cb->uaiocb._aiocb_private.status;
1791 error = cb->uaiocb._aiocb_private.error;
1792 td->td_retval[0] = status;
1793 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1794 td->td_ru.ru_oublock += cb->outputcharge;
1795 cb->outputcharge = 0;
1796 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1797 td->td_ru.ru_inblock += cb->inputcharge;
1798 cb->inputcharge = 0;
1802 ops->store_error(uaiocb, error);
1803 ops->store_status(uaiocb, status);
1812 aio_return(struct thread *td, struct aio_return_args *uap)
1815 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1819 * Allow a process to wakeup when any of the I/O requests are completed.
1822 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1823 struct timespec *ts)
1825 struct proc *p = td->td_proc;
1827 struct kaioinfo *ki;
1828 struct aiocblist *cb, *cbfirst;
1833 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1836 TIMESPEC_TO_TIMEVAL(&atv, ts);
1837 if (itimerfix(&atv))
1839 timo = tvtohz(&atv);
1853 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1854 for (i = 0; i < njoblist; i++) {
1855 if (cb->uuaiocb == ujoblist[i]) {
1856 if (cbfirst == NULL)
1858 if (cb->jobstate == JOBST_JOBFINISHED)
1863 /* All tasks were finished. */
1864 if (cbfirst == NULL)
1867 ki->kaio_flags |= KAIO_WAKEUP;
1868 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1870 if (error == ERESTART)
1881 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1883 struct timespec ts, *tsp;
1884 struct aiocb **ujoblist;
1887 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1891 /* Get timespec struct. */
1892 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1898 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1899 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1901 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1902 uma_zfree(aiol_zone, ujoblist);
1907 * aio_cancel cancels any non-physio aio operations not currently in
1911 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1913 struct proc *p = td->td_proc;
1914 struct kaioinfo *ki;
1915 struct aiocblist *cbe, *cbn;
1921 int notcancelled = 0;
1924 /* Lookup file object. */
1925 error = fget(td, uap->fd, &fp);
1933 if (fp->f_type == DTYPE_VNODE) {
1935 if (vn_isdisk(vp, &error)) {
1937 td->td_retval[0] = AIO_NOTCANCELED;
1943 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1944 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1945 ((uap->aiocbp == NULL) ||
1946 (uap->aiocbp == cbe->uuaiocb))) {
1949 mtx_lock(&aio_job_mtx);
1950 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1951 TAILQ_REMOVE(&aio_jobs, cbe, list);
1953 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
1954 MPASS(fp->f_type == DTYPE_SOCKET);
1956 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1958 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
1959 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
1962 mtx_unlock(&aio_job_mtx);
1965 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1966 cbe->uaiocb._aiocb_private.status = -1;
1967 cbe->uaiocb._aiocb_private.error = ECANCELED;
1968 aio_bio_done_notify(p, cbe, DONE_QUEUE);
1973 if (uap->aiocbp != NULL)
1982 if (uap->aiocbp != NULL) {
1984 td->td_retval[0] = AIO_CANCELED;
1990 td->td_retval[0] = AIO_NOTCANCELED;
1995 td->td_retval[0] = AIO_CANCELED;
1999 td->td_retval[0] = AIO_ALLDONE;
2005 * aio_error is implemented in the kernel level for compatibility purposes
2006 * only. For a user mode async implementation, it would be best to do it in
2007 * a userland subroutine.
2010 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2012 struct proc *p = td->td_proc;
2013 struct aiocblist *cb;
2014 struct kaioinfo *ki;
2019 td->td_retval[0] = EINVAL;
2024 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2025 if (cb->uuaiocb == aiocbp) {
2026 if (cb->jobstate == JOBST_JOBFINISHED)
2028 cb->uaiocb._aiocb_private.error;
2030 td->td_retval[0] = EINPROGRESS;
2038 * Hack for failure of aio_aqueue.
2040 status = ops->fetch_status(aiocbp);
2042 td->td_retval[0] = ops->fetch_error(aiocbp);
2046 td->td_retval[0] = EINVAL;
2051 aio_error(struct thread *td, struct aio_error_args *uap)
2054 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2057 /* syscall - asynchronous read from a file (REALTIME) */
2059 oaio_read(struct thread *td, struct oaio_read_args *uap)
2062 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2063 &aiocb_ops_osigevent));
2067 aio_read(struct thread *td, struct aio_read_args *uap)
2070 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2073 /* syscall - asynchronous write to a file (REALTIME) */
2075 oaio_write(struct thread *td, struct oaio_write_args *uap)
2078 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2079 &aiocb_ops_osigevent));
2083 aio_write(struct thread *td, struct aio_write_args *uap)
2086 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2090 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2091 struct aiocb **acb_list, int nent, struct sigevent *sig,
2092 struct aiocb_ops *ops)
2094 struct proc *p = td->td_proc;
2096 struct kaioinfo *ki;
2097 struct aioliojob *lj;
2103 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2106 if (nent < 0 || nent > AIO_LISTIO_MAX)
2109 if (p->p_aioinfo == NULL)
2110 aio_init_aioinfo(p);
2114 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2117 lj->lioj_finished_count = 0;
2118 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2119 ksiginfo_init(&lj->lioj_ksi);
2124 if (sig && (mode == LIO_NOWAIT)) {
2125 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2126 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2127 /* Assume only new style KEVENT */
2128 kev.filter = EVFILT_LIO;
2129 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2130 kev.ident = (uintptr_t)uacb_list; /* something unique */
2131 kev.data = (intptr_t)lj;
2132 /* pass user defined sigval data */
2133 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2134 error = kqfd_register(
2135 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2137 uma_zfree(aiolio_zone, lj);
2140 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2142 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2143 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2144 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2145 uma_zfree(aiolio_zone, lj);
2148 lj->lioj_flags |= LIOJ_SIGNAL;
2150 uma_zfree(aiolio_zone, lj);
2156 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2158 * Add extra aiocb count to avoid the lio to be freed
2159 * by other threads doing aio_waitcomplete or aio_return,
2160 * and prevent event from being sent until we have queued
2167 * Get pointers to the list of I/O requests.
2170 for (i = 0; i < nent; i++) {
2173 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2181 if (mode == LIO_WAIT) {
2182 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2183 ki->kaio_flags |= KAIO_WAKEUP;
2184 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2185 PRIBIO | PCATCH, "aiospn", 0);
2186 if (error == ERESTART)
2192 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2193 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2194 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2195 KNOTE_LOCKED(&lj->klist, 1);
2197 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2199 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2200 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2201 aio_sendsig(p, &lj->lioj_signal,
2203 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2208 if (lj->lioj_count == 0) {
2209 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2210 knlist_delete(&lj->klist, curthread, 1);
2212 sigqueue_take(&lj->lioj_ksi);
2215 uma_zfree(aiolio_zone, lj);
2224 /* syscall - list directed I/O (REALTIME) */
2226 olio_listio(struct thread *td, struct olio_listio_args *uap)
2228 struct aiocb **acb_list;
2229 struct sigevent *sigp, sig;
2230 struct osigevent osig;
2233 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2237 if (nent < 0 || nent > AIO_LISTIO_MAX)
2240 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2241 error = copyin(uap->sig, &osig, sizeof(osig));
2244 error = convert_old_sigevent(&osig, &sig);
2251 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2252 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2254 error = kern_lio_listio(td, uap->mode,
2255 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2256 &aiocb_ops_osigevent);
2257 free(acb_list, M_LIO);
2261 /* syscall - list directed I/O (REALTIME) */
2263 lio_listio(struct thread *td, struct lio_listio_args *uap)
2265 struct aiocb **acb_list;
2266 struct sigevent *sigp, sig;
2269 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2273 if (nent < 0 || nent > AIO_LISTIO_MAX)
2276 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2277 error = copyin(uap->sig, &sig, sizeof(sig));
2284 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2285 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2287 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2288 nent, sigp, &aiocb_ops);
2289 free(acb_list, M_LIO);
2294 * Called from interrupt thread for physio, we should return as fast
2295 * as possible, so we schedule a biohelper task.
2298 aio_physwakeup(struct buf *bp)
2300 struct aiocblist *aiocbe;
2302 aiocbe = (struct aiocblist *)bp->b_caller1;
2303 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2307 * Task routine to perform heavy tasks, process wakeup, and signals.
2310 biohelper(void *context, int pending)
2312 struct aiocblist *aiocbe = context;
2315 struct kaioinfo *ki;
2319 userp = aiocbe->userproc;
2320 ki = userp->p_aioinfo;
2322 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2323 aiocbe->uaiocb._aiocb_private.error = 0;
2324 if (bp->b_ioflags & BIO_ERROR)
2325 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2326 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2327 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2328 aiocbe->outputcharge += nblks;
2330 aiocbe->inputcharge += nblks;
2332 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2333 ki->kaio_buffer_count--;
2334 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2337 /* Release mapping into kernel space. */
2340 atomic_subtract_int(&num_buf_aio, 1);
2343 /* syscall - wait for the next completion of an aio request */
2345 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2346 struct timespec *ts, struct aiocb_ops *ops)
2348 struct proc *p = td->td_proc;
2350 struct kaioinfo *ki;
2351 struct aiocblist *cb;
2352 struct aiocb *uuaiocb;
2353 int error, status, timo;
2355 ops->store_aiocb(aiocbp, NULL);
2359 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2362 TIMESPEC_TO_TIMEVAL(&atv, ts);
2363 if (itimerfix(&atv))
2365 timo = tvtohz(&atv);
2368 if (p->p_aioinfo == NULL)
2369 aio_init_aioinfo(p);
2375 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2376 ki->kaio_flags |= KAIO_WAKEUP;
2377 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2379 if (timo && error == ERESTART)
2386 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2387 uuaiocb = cb->uuaiocb;
2388 status = cb->uaiocb._aiocb_private.status;
2389 error = cb->uaiocb._aiocb_private.error;
2390 td->td_retval[0] = status;
2391 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2392 td->td_ru.ru_oublock += cb->outputcharge;
2393 cb->outputcharge = 0;
2394 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2395 td->td_ru.ru_inblock += cb->inputcharge;
2396 cb->inputcharge = 0;
2400 ops->store_aiocb(aiocbp, uuaiocb);
2401 ops->store_error(uuaiocb, error);
2402 ops->store_status(uuaiocb, status);
2410 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2412 struct timespec ts, *tsp;
2416 /* Get timespec struct. */
2417 error = copyin(uap->timeout, &ts, sizeof(ts));
2424 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2428 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2429 struct aiocb_ops *ops)
2431 struct proc *p = td->td_proc;
2432 struct kaioinfo *ki;
2434 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2438 aio_init_aioinfo(p);
2439 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2443 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2446 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2449 /* kqueue attach function */
2451 filt_aioattach(struct knote *kn)
2453 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2456 * The aiocbe pointer must be validated before using it, so
2457 * registration is restricted to the kernel; the user cannot
2460 if ((kn->kn_flags & EV_FLAG1) == 0)
2462 kn->kn_ptr.p_aio = aiocbe;
2463 kn->kn_flags &= ~EV_FLAG1;
2465 knlist_add(&aiocbe->klist, kn, 0);
2470 /* kqueue detach function */
2472 filt_aiodetach(struct knote *kn)
2474 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2476 if (!knlist_empty(&aiocbe->klist))
2477 knlist_remove(&aiocbe->klist, kn, 0);
2480 /* kqueue filter function */
2483 filt_aio(struct knote *kn, long hint)
2485 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2487 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2488 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2490 kn->kn_flags |= EV_EOF;
2494 /* kqueue attach function */
2496 filt_lioattach(struct knote *kn)
2498 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2501 * The aioliojob pointer must be validated before using it, so
2502 * registration is restricted to the kernel; the user cannot
2505 if ((kn->kn_flags & EV_FLAG1) == 0)
2507 kn->kn_ptr.p_lio = lj;
2508 kn->kn_flags &= ~EV_FLAG1;
2510 knlist_add(&lj->klist, kn, 0);
2515 /* kqueue detach function */
2517 filt_liodetach(struct knote *kn)
2519 struct aioliojob * lj = kn->kn_ptr.p_lio;
2521 if (!knlist_empty(&lj->klist))
2522 knlist_remove(&lj->klist, kn, 0);
2525 /* kqueue filter function */
2528 filt_lio(struct knote *kn, long hint)
2530 struct aioliojob * lj = kn->kn_ptr.p_lio;
2532 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2536 #include <sys/mount.h>
2537 #include <sys/socket.h>
2538 #include <compat/freebsd32/freebsd32.h>
2539 #include <compat/freebsd32/freebsd32_proto.h>
2540 #include <compat/freebsd32/freebsd32_signal.h>
2541 #include <compat/freebsd32/freebsd32_syscall.h>
2542 #include <compat/freebsd32/freebsd32_util.h>
2544 struct __aiocb_private32 {
2547 uint32_t kernelinfo;
2550 typedef struct oaiocb32 {
2551 int aio_fildes; /* File descriptor */
2552 uint64_t aio_offset __packed; /* File offset for I/O */
2553 uint32_t aio_buf; /* I/O buffer in process space */
2554 uint32_t aio_nbytes; /* Number of bytes for I/O */
2555 struct osigevent32 aio_sigevent; /* Signal to deliver */
2556 int aio_lio_opcode; /* LIO opcode */
2557 int aio_reqprio; /* Request priority -- ignored */
2558 struct __aiocb_private32 _aiocb_private;
2561 typedef struct aiocb32 {
2562 int32_t aio_fildes; /* File descriptor */
2563 uint64_t aio_offset __packed; /* File offset for I/O */
2564 uint32_t aio_buf; /* I/O buffer in process space */
2565 uint32_t aio_nbytes; /* Number of bytes for I/O */
2567 uint32_t __spare2__;
2568 int aio_lio_opcode; /* LIO opcode */
2569 int aio_reqprio; /* Request priority -- ignored */
2570 struct __aiocb_private32 _aiocb_private;
2571 struct sigevent32 aio_sigevent; /* Signal to deliver */
2575 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2579 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2580 * supported by AIO with the old sigevent structure.
2582 CP(*osig, *nsig, sigev_notify);
2583 switch (nsig->sigev_notify) {
2587 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2590 nsig->sigev_notify_kqueue =
2591 osig->__sigev_u.__sigev_notify_kqueue;
2592 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2601 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2603 struct oaiocb32 job32;
2606 bzero(kjob, sizeof(struct aiocb));
2607 error = copyin(ujob, &job32, sizeof(job32));
2611 CP(job32, *kjob, aio_fildes);
2612 CP(job32, *kjob, aio_offset);
2613 PTRIN_CP(job32, *kjob, aio_buf);
2614 CP(job32, *kjob, aio_nbytes);
2615 CP(job32, *kjob, aio_lio_opcode);
2616 CP(job32, *kjob, aio_reqprio);
2617 CP(job32, *kjob, _aiocb_private.status);
2618 CP(job32, *kjob, _aiocb_private.error);
2619 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2620 return (convert_old_sigevent32(&job32.aio_sigevent,
2621 &kjob->aio_sigevent));
2625 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig)
2628 CP(*sig32, *sig, sigev_notify);
2629 switch (sig->sigev_notify) {
2632 case SIGEV_THREAD_ID:
2633 CP(*sig32, *sig, sigev_notify_thread_id);
2636 CP(*sig32, *sig, sigev_signo);
2639 CP(*sig32, *sig, sigev_notify_kqueue);
2640 PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr);
2649 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2651 struct aiocb32 job32;
2654 error = copyin(ujob, &job32, sizeof(job32));
2657 CP(job32, *kjob, aio_fildes);
2658 CP(job32, *kjob, aio_offset);
2659 PTRIN_CP(job32, *kjob, aio_buf);
2660 CP(job32, *kjob, aio_nbytes);
2661 CP(job32, *kjob, aio_lio_opcode);
2662 CP(job32, *kjob, aio_reqprio);
2663 CP(job32, *kjob, _aiocb_private.status);
2664 CP(job32, *kjob, _aiocb_private.error);
2665 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2666 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2670 aiocb32_fetch_status(struct aiocb *ujob)
2672 struct aiocb32 *ujob32;
2674 ujob32 = (struct aiocb32 *)ujob;
2675 return (fuword32(&ujob32->_aiocb_private.status));
2679 aiocb32_fetch_error(struct aiocb *ujob)
2681 struct aiocb32 *ujob32;
2683 ujob32 = (struct aiocb32 *)ujob;
2684 return (fuword32(&ujob32->_aiocb_private.error));
2688 aiocb32_store_status(struct aiocb *ujob, long status)
2690 struct aiocb32 *ujob32;
2692 ujob32 = (struct aiocb32 *)ujob;
2693 return (suword32(&ujob32->_aiocb_private.status, status));
2697 aiocb32_store_error(struct aiocb *ujob, long error)
2699 struct aiocb32 *ujob32;
2701 ujob32 = (struct aiocb32 *)ujob;
2702 return (suword32(&ujob32->_aiocb_private.error, error));
2706 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2708 struct aiocb32 *ujob32;
2710 ujob32 = (struct aiocb32 *)ujob;
2711 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2715 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2718 return (suword32(ujobp, (long)ujob));
2721 static struct aiocb_ops aiocb32_ops = {
2722 .copyin = aiocb32_copyin,
2723 .fetch_status = aiocb32_fetch_status,
2724 .fetch_error = aiocb32_fetch_error,
2725 .store_status = aiocb32_store_status,
2726 .store_error = aiocb32_store_error,
2727 .store_kernelinfo = aiocb32_store_kernelinfo,
2728 .store_aiocb = aiocb32_store_aiocb,
2731 static struct aiocb_ops aiocb32_ops_osigevent = {
2732 .copyin = aiocb32_copyin_old_sigevent,
2733 .fetch_status = aiocb32_fetch_status,
2734 .fetch_error = aiocb32_fetch_error,
2735 .store_status = aiocb32_store_status,
2736 .store_error = aiocb32_store_error,
2737 .store_kernelinfo = aiocb32_store_kernelinfo,
2738 .store_aiocb = aiocb32_store_aiocb,
2742 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2745 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2749 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2751 struct timespec32 ts32;
2752 struct timespec ts, *tsp;
2753 struct aiocb **ujoblist;
2754 uint32_t *ujoblist32;
2757 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2761 /* Get timespec struct. */
2762 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2764 CP(ts32, ts, tv_sec);
2765 CP(ts32, ts, tv_nsec);
2770 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2771 ujoblist32 = (uint32_t *)ujoblist;
2772 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2773 sizeof(ujoblist32[0]));
2775 for (i = uap->nent; i > 0; i--)
2776 ujoblist[i] = PTRIN(ujoblist32[i]);
2778 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2780 uma_zfree(aiol_zone, ujoblist);
2785 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2788 return (aio_cancel(td, (struct aio_cancel_args *)uap));
2792 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2795 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2799 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2802 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2803 &aiocb32_ops_osigevent));
2807 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2810 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2815 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2818 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2819 &aiocb32_ops_osigevent));
2823 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2826 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2831 freebsd32_aio_waitcomplete(struct thread *td,
2832 struct freebsd32_aio_waitcomplete_args *uap)
2834 struct timespec32 ts32;
2835 struct timespec ts, *tsp;
2839 /* Get timespec struct. */
2840 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2843 CP(ts32, ts, tv_sec);
2844 CP(ts32, ts, tv_nsec);
2849 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2854 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2857 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2862 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2864 struct aiocb **acb_list;
2865 struct sigevent *sigp, sig;
2866 struct osigevent32 osig;
2867 uint32_t *acb_list32;
2870 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2874 if (nent < 0 || nent > AIO_LISTIO_MAX)
2877 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2878 error = copyin(uap->sig, &osig, sizeof(osig));
2881 error = convert_old_sigevent32(&osig, &sig);
2888 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2889 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2891 free(acb_list32, M_LIO);
2894 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2895 for (i = 0; i < nent; i++)
2896 acb_list[i] = PTRIN(acb_list32[i]);
2897 free(acb_list32, M_LIO);
2899 error = kern_lio_listio(td, uap->mode,
2900 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2901 &aiocb32_ops_osigevent);
2902 free(acb_list, M_LIO);
2907 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2909 struct aiocb **acb_list;
2910 struct sigevent *sigp, sig;
2911 struct sigevent32 sig32;
2912 uint32_t *acb_list32;
2915 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2919 if (nent < 0 || nent > AIO_LISTIO_MAX)
2922 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2923 error = copyin(uap->sig, &sig32, sizeof(sig32));
2926 error = convert_sigevent32(&sig32, &sig);
2933 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2934 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2936 free(acb_list32, M_LIO);
2939 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2940 for (i = 0; i < nent; i++)
2941 acb_list[i] = PTRIN(acb_list32[i]);
2942 free(acb_list32, M_LIO);
2944 error = kern_lio_listio(td, uap->mode,
2945 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2947 free(acb_list, M_LIO);
2951 SYSCALL32_MODULE_HELPER(freebsd32_aio_return);
2952 SYSCALL32_MODULE_HELPER(freebsd32_aio_suspend);
2953 SYSCALL32_MODULE_HELPER(freebsd32_aio_cancel);
2954 SYSCALL32_MODULE_HELPER(freebsd32_aio_error);
2955 SYSCALL32_MODULE_HELPER(freebsd32_aio_fsync);
2956 SYSCALL32_MODULE_HELPER(freebsd32_aio_read);
2957 SYSCALL32_MODULE_HELPER(freebsd32_aio_write);
2958 SYSCALL32_MODULE_HELPER(freebsd32_aio_waitcomplete);
2959 SYSCALL32_MODULE_HELPER(freebsd32_lio_listio);
2960 SYSCALL32_MODULE_HELPER(freebsd32_oaio_read);
2961 SYSCALL32_MODULE_HELPER(freebsd32_oaio_write);
2962 SYSCALL32_MODULE_HELPER(freebsd32_olio_listio);