FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_aio.c
1 /*-
2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
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.
11 *
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.
15 */
16
17 /*
18 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
19 */
20
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
23
24 #include <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/malloc.h>
27 #include <sys/bio.h>
28 #include <sys/buf.h>
29 #include <sys/eventhandler.h>
30 #include <sys/sysproto.h>
31 #include <sys/filedesc.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/kthread.h>
35 #include <sys/fcntl.h>
36 #include <sys/file.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/unistd.h>
41 #include <sys/proc.h>
42 #include <sys/resourcevar.h>
43 #include <sys/signalvar.h>
44 #include <sys/protosw.h>
45 #include <sys/socketvar.h>
46 #include <sys/syscall.h>
47 #include <sys/sysent.h>
48 #include <sys/sysctl.h>
49 #include <sys/sx.h>
50 #include <sys/vnode.h>
51 #include <sys/conf.h>
52 #include <sys/event.h>
53
54 #include <posix4/posix4.h>
55 #include <vm/vm.h>
56 #include <vm/vm_extern.h>
57 #include <vm/pmap.h>
58 #include <vm/vm_map.h>
59 #include <vm/uma.h>
60 #include <sys/aio.h>
61
62 #include "opt_vfs_aio.h"
63
64
65 /*
66 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
67 * overflow.
68 */
69 static long jobrefid;
70
71 #define JOBST_NULL 0x0
72 #define JOBST_JOBQGLOBAL 0x2
73 #define JOBST_JOBRUNNING 0x3
74 #define JOBST_JOBFINISHED 0x4
75 #define JOBST_JOBQBUF 0x5
76 #define JOBST_JOBBFINISHED 0x6
77
78 #ifndef MAX_AIO_PER_PROC
79 #define MAX_AIO_PER_PROC 32
80 #endif
81
82 #ifndef MAX_AIO_QUEUE_PER_PROC
83 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
84 #endif
85
86 #ifndef MAX_AIO_PROCS
87 #define MAX_AIO_PROCS 32
88 #endif
89
90 #ifndef MAX_AIO_QUEUE
91 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
92 #endif
93
94 #ifndef TARGET_AIO_PROCS
95 #define TARGET_AIO_PROCS 4
96 #endif
97
98 #ifndef MAX_BUF_AIO
99 #define MAX_BUF_AIO 16
100 #endif
101
102 #ifndef AIOD_TIMEOUT_DEFAULT
103 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
104 #endif
105
106 #ifndef AIOD_LIFETIME_DEFAULT
107 #define AIOD_LIFETIME_DEFAULT (30 * hz)
108 #endif
109
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
111
112 static int max_aio_procs = MAX_AIO_PROCS;
113 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
114 CTLFLAG_RW, &max_aio_procs, 0,
115 "Maximum number of kernel threads to use for handling async IO ");
116
117 static int num_aio_procs = 0;
118 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
119 CTLFLAG_RD, &num_aio_procs, 0,
120 "Number of presently active kernel threads for async IO");
121
122 /*
123 * The code will adjust the actual number of AIO processes towards this
124 * number when it gets a chance.
125 */
126 static int target_aio_procs = TARGET_AIO_PROCS;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
128 0, "Preferred number of ready kernel threads for async IO");
129
130 static int max_queue_count = MAX_AIO_QUEUE;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
132 "Maximum number of aio requests to queue, globally");
133
134 static int num_queue_count = 0;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
136 "Number of queued aio requests");
137
138 static int num_buf_aio = 0;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
140 "Number of aio requests presently handled by the buf subsystem");
141
142 /* Number of async I/O thread in the process of being started */
143 /* XXX This should be local to _aio_aqueue() */
144 static int num_aio_resv_start = 0;
145
146 static int aiod_timeout;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
148 "Timeout value for synchronous aio operations");
149
150 static int aiod_lifetime;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
152 "Maximum lifetime for idle aiod");
153
154 static int unloadable = 0;
155 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
156 "Allow unload of aio (not recommended)");
157
158
159 static int max_aio_per_proc = MAX_AIO_PER_PROC;
160 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
161 0, "Maximum active aio requests per process (stored in the process)");
162
163 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
164 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
165 &max_aio_queue_per_proc, 0,
166 "Maximum queued aio requests per process (stored in the process)");
167
168 static int max_buf_aio = MAX_BUF_AIO;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
170 "Maximum buf aio requests per process (stored in the process)");
171
172 struct aiocblist {
173 TAILQ_ENTRY(aiocblist) list; /* List of jobs */
174 TAILQ_ENTRY(aiocblist) plist; /* List of jobs for proc */
175 int jobflags;
176 int jobstate;
177 int inputcharge;
178 int outputcharge;
179 struct buf *bp; /* Buffer pointer */
180 struct proc *userproc; /* User process */ /* Not td! */
181 struct ucred *cred; /* Active credential when created */
182 struct file *fd_file; /* Pointer to file structure */
183 struct aio_liojob *lio; /* Optional lio job */
184 struct aiocb *uuaiocb; /* Pointer in userspace of aiocb */
185 struct knlist klist; /* list of knotes */
186 struct aiocb uaiocb; /* Kernel I/O control block */
187 };
188
189 /* jobflags */
190 #define AIOCBLIST_RUNDOWN 0x4
191 #define AIOCBLIST_DONE 0x10
192
193 /*
194 * AIO process info
195 */
196 #define AIOP_FREE 0x1 /* proc on free queue */
197
198 struct aiothreadlist {
199 int aiothreadflags; /* AIO proc flags */
200 TAILQ_ENTRY(aiothreadlist) list; /* List of processes */
201 struct thread *aiothread; /* The AIO thread */
202 };
203
204 /*
205 * data-structure for lio signal management
206 */
207 struct aio_liojob {
208 int lioj_flags;
209 int lioj_buffer_count;
210 int lioj_buffer_finished_count;
211 int lioj_queue_count;
212 int lioj_queue_finished_count;
213 struct sigevent lioj_signal; /* signal on all I/O done */
214 TAILQ_ENTRY(aio_liojob) lioj_list;
215 };
216 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
217 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
218
219 /*
220 * per process aio data structure
221 */
222 struct kaioinfo {
223 int kaio_flags; /* per process kaio flags */
224 int kaio_maxactive_count; /* maximum number of AIOs */
225 int kaio_active_count; /* number of currently used AIOs */
226 int kaio_qallowed_count; /* maxiumu size of AIO queue */
227 int kaio_queue_count; /* size of AIO queue */
228 int kaio_ballowed_count; /* maximum number of buffers */
229 int kaio_queue_finished_count; /* number of daemon jobs finished */
230 int kaio_buffer_count; /* number of physio buffers */
231 int kaio_buffer_finished_count; /* count of I/O done */
232 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
233 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
234 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
235 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
236 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
237 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
238 };
239
240 #define KAIO_RUNDOWN 0x1 /* process is being run down */
241 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
242
243 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* Idle daemons */
244 static struct mtx aio_freeproc_mtx;
245
246 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
247
248 static void aio_init_aioinfo(struct proc *p);
249 static void aio_onceonly(void);
250 static int aio_free_entry(struct aiocblist *aiocbe);
251 static void aio_process(struct aiocblist *aiocbe);
252 static int aio_newproc(void);
253 static int aio_aqueue(struct thread *td, struct aiocb *job, int type);
254 static void aio_physwakeup(struct buf *bp);
255 static void aio_proc_rundown(void *arg, struct proc *p);
256 static int aio_fphysio(struct aiocblist *aiocbe);
257 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
258 static void aio_daemon(void *uproc);
259 static void aio_swake_cb(struct socket *, struct sockbuf *);
260 static int aio_unload(void);
261 static int filt_aioattach(struct knote *kn);
262 static void filt_aiodetach(struct knote *kn);
263 static int filt_aio(struct knote *kn, long hint);
264
265 /*
266 * Zones for:
267 * kaio Per process async io info
268 * aiop async io thread data
269 * aiocb async io jobs
270 * aiol list io job pointer - internal to aio_suspend XXX
271 * aiolio list io jobs
272 */
273 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
274
275 /* kqueue filters for aio */
276 static struct filterops aio_filtops =
277 { 0, filt_aioattach, filt_aiodetach, filt_aio };
278
279 static eventhandler_tag exit_tag, exec_tag;
280
281 /*
282 * Main operations function for use as a kernel module.
283 */
284 static int
285 aio_modload(struct module *module, int cmd, void *arg)
286 {
287 int error = 0;
288
289 switch (cmd) {
290 case MOD_LOAD:
291 aio_onceonly();
292 break;
293 case MOD_UNLOAD:
294 error = aio_unload();
295 break;
296 case MOD_SHUTDOWN:
297 break;
298 default:
299 error = EINVAL;
300 break;
301 }
302 return (error);
303 }
304
305 static moduledata_t aio_mod = {
306 "aio",
307 &aio_modload,
308 NULL
309 };
310
311 SYSCALL_MODULE_HELPER(aio_return);
312 SYSCALL_MODULE_HELPER(aio_suspend);
313 SYSCALL_MODULE_HELPER(aio_cancel);
314 SYSCALL_MODULE_HELPER(aio_error);
315 SYSCALL_MODULE_HELPER(aio_read);
316 SYSCALL_MODULE_HELPER(aio_write);
317 SYSCALL_MODULE_HELPER(aio_waitcomplete);
318 SYSCALL_MODULE_HELPER(lio_listio);
319
320 DECLARE_MODULE(aio, aio_mod,
321 SI_SUB_VFS, SI_ORDER_ANY);
322 MODULE_VERSION(aio, 1);
323
324 /*
325 * Startup initialization
326 */
327 static void
328 aio_onceonly(void)
329 {
330
331 /* XXX: should probably just use so->callback */
332 aio_swake = &aio_swake_cb;
333 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
334 EVENTHANDLER_PRI_ANY);
335 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown, NULL,
336 EVENTHANDLER_PRI_ANY);
337 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
338 TAILQ_INIT(&aio_freeproc);
339 mtx_init(&aio_freeproc_mtx, "aio_freeproc", NULL, MTX_DEF);
340 TAILQ_INIT(&aio_jobs);
341 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
342 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
343 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
344 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
345 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
346 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
347 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
348 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
349 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL,
350 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
351 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
352 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
353 jobrefid = 1;
354 async_io_version = _POSIX_VERSION;
355 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
356 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
357 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
358 }
359
360 /*
361 * Callback for unload of AIO when used as a module.
362 */
363 static int
364 aio_unload(void)
365 {
366 int error;
367
368 /*
369 * XXX: no unloads by default, it's too dangerous.
370 * perhaps we could do it if locked out callers and then
371 * did an aio_proc_rundown() on each process.
372 */
373 if (!unloadable)
374 return (EOPNOTSUPP);
375
376 error = kqueue_del_filteropts(EVFILT_AIO);
377 if (error)
378 return error;
379
380 async_io_version = 0;
381 aio_swake = NULL;
382 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
383 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
384 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
385 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
386 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
387 return (0);
388 }
389
390 /*
391 * Init the per-process aioinfo structure. The aioinfo limits are set
392 * per-process for user limit (resource) management.
393 */
394 static void
395 aio_init_aioinfo(struct proc *p)
396 {
397 struct kaioinfo *ki;
398
399 ki = uma_zalloc(kaio_zone, M_WAITOK);
400 ki->kaio_flags = 0;
401 ki->kaio_maxactive_count = max_aio_per_proc;
402 ki->kaio_active_count = 0;
403 ki->kaio_qallowed_count = max_aio_queue_per_proc;
404 ki->kaio_queue_count = 0;
405 ki->kaio_ballowed_count = max_buf_aio;
406 ki->kaio_buffer_count = 0;
407 ki->kaio_buffer_finished_count = 0;
408 TAILQ_INIT(&ki->kaio_jobdone);
409 TAILQ_INIT(&ki->kaio_jobqueue);
410 TAILQ_INIT(&ki->kaio_bufdone);
411 TAILQ_INIT(&ki->kaio_bufqueue);
412 TAILQ_INIT(&ki->kaio_liojoblist);
413 TAILQ_INIT(&ki->kaio_sockqueue);
414 PROC_LOCK(p);
415 if (p->p_aioinfo == NULL) {
416 p->p_aioinfo = ki;
417 PROC_UNLOCK(p);
418 } else {
419 PROC_UNLOCK(p);
420 uma_zfree(kaio_zone, ki);
421 }
422
423 while (num_aio_procs < target_aio_procs)
424 aio_newproc();
425 }
426
427 /*
428 * Free a job entry. Wait for completion if it is currently active, but don't
429 * delay forever. If we delay, we return a flag that says that we have to
430 * restart the queue scan.
431 */
432 static int
433 aio_free_entry(struct aiocblist *aiocbe)
434 {
435 struct kaioinfo *ki;
436 struct aio_liojob *lj;
437 struct proc *p;
438 int error;
439 int s;
440
441 if (aiocbe->jobstate == JOBST_NULL)
442 panic("aio_free_entry: freeing already free job");
443
444 p = aiocbe->userproc;
445 ki = p->p_aioinfo;
446 lj = aiocbe->lio;
447 if (ki == NULL)
448 panic("aio_free_entry: missing p->p_aioinfo");
449
450 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
451 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
452 tsleep(aiocbe, PRIBIO, "jobwai", 0);
453 }
454 if (aiocbe->bp == NULL) {
455 if (ki->kaio_queue_count <= 0)
456 panic("aio_free_entry: process queue size <= 0");
457 if (num_queue_count <= 0)
458 panic("aio_free_entry: system wide queue size <= 0");
459
460 if (lj) {
461 lj->lioj_queue_count--;
462 if (aiocbe->jobflags & AIOCBLIST_DONE)
463 lj->lioj_queue_finished_count--;
464 }
465 ki->kaio_queue_count--;
466 if (aiocbe->jobflags & AIOCBLIST_DONE)
467 ki->kaio_queue_finished_count--;
468 num_queue_count--;
469 } else {
470 if (lj) {
471 lj->lioj_buffer_count--;
472 if (aiocbe->jobflags & AIOCBLIST_DONE)
473 lj->lioj_buffer_finished_count--;
474 }
475 if (aiocbe->jobflags & AIOCBLIST_DONE)
476 ki->kaio_buffer_finished_count--;
477 ki->kaio_buffer_count--;
478 num_buf_aio--;
479 }
480
481 /* aiocbe is going away, we need to destroy any knotes */
482 /* XXXKSE Note the thread here is used to eventually find the
483 * owning process again, but it is also used to do a fo_close
484 * and that requires the thread. (but does it require the
485 * OWNING thread? (or maybe the running thread?)
486 * There is a semantic problem here...
487 */
488 knlist_delete(&aiocbe->klist, FIRST_THREAD_IN_PROC(p), 0); /* XXXKSE */
489
490 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
491 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
492 ki->kaio_flags &= ~KAIO_WAKEUP;
493 wakeup(p);
494 }
495
496 if (aiocbe->jobstate == JOBST_JOBQBUF) {
497 if ((error = aio_fphysio(aiocbe)) != 0)
498 return (error);
499 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
500 panic("aio_free_entry: invalid physio finish-up state");
501 s = splbio();
502 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
503 splx(s);
504 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
505 s = splnet();
506 mtx_lock(&aio_freeproc_mtx);
507 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
508 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
509 mtx_unlock(&aio_freeproc_mtx);
510 splx(s);
511 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
512 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
513 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
514 s = splbio();
515 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
516 splx(s);
517 if (aiocbe->bp) {
518 vunmapbuf(aiocbe->bp);
519 relpbuf(aiocbe->bp, NULL);
520 aiocbe->bp = NULL;
521 }
522 }
523 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
524 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
525 uma_zfree(aiolio_zone, lj);
526 }
527 aiocbe->jobstate = JOBST_NULL;
528 fdrop(aiocbe->fd_file, curthread);
529 crfree(aiocbe->cred);
530 uma_zfree(aiocb_zone, aiocbe);
531 return (0);
532 }
533
534 /*
535 * Rundown the jobs for a given process.
536 */
537 static void
538 aio_proc_rundown(void *arg, struct proc *p)
539 {
540 int s;
541 struct kaioinfo *ki;
542 struct aio_liojob *lj, *ljn;
543 struct aiocblist *aiocbe, *aiocbn;
544 struct file *fp;
545 struct socket *so;
546
547 ki = p->p_aioinfo;
548 if (ki == NULL)
549 return;
550
551 mtx_lock(&Giant);
552 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
553 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
554 ki->kaio_buffer_finished_count)) {
555 ki->kaio_flags |= KAIO_RUNDOWN;
556 if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
557 break;
558 }
559
560 /*
561 * Move any aio ops that are waiting on socket I/O to the normal job
562 * queues so they are cleaned up with any others.
563 */
564 s = splnet();
565 mtx_lock(&aio_freeproc_mtx);
566 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
567 aiocbn) {
568 aiocbn = TAILQ_NEXT(aiocbe, plist);
569 fp = aiocbe->fd_file;
570 if (fp != NULL) {
571 so = fp->f_data;
572 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
573 if (TAILQ_EMPTY(&so->so_aiojobq)) {
574 SOCKBUF_LOCK(&so->so_snd);
575 so->so_snd.sb_flags &= ~SB_AIO;
576 SOCKBUF_UNLOCK(&so->so_snd);
577 SOCKBUF_LOCK(&so->so_rcv);
578 so->so_rcv.sb_flags &= ~SB_AIO;
579 SOCKBUF_UNLOCK(&so->so_rcv);
580 }
581 }
582 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
583 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
584 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
585 }
586 mtx_unlock(&aio_freeproc_mtx);
587 splx(s);
588
589 restart1:
590 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
591 aiocbn = TAILQ_NEXT(aiocbe, plist);
592 if (aio_free_entry(aiocbe))
593 goto restart1;
594 }
595
596 restart2:
597 mtx_lock(&aio_freeproc_mtx);
598 while ((aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue)) != NULL) {
599 mtx_unlock(&aio_freeproc_mtx);
600 if (aio_free_entry(aiocbe))
601 goto restart2;
602 mtx_lock(&aio_freeproc_mtx);
603 }
604 mtx_unlock(&aio_freeproc_mtx);
605
606 /*
607 * Note the use of lots of splbio here, trying to avoid splbio for long chains
608 * of I/O. Probably unnecessary.
609 */
610 restart3:
611 s = splbio();
612 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
613 ki->kaio_flags |= KAIO_WAKEUP;
614 tsleep(p, PRIBIO, "aioprn", 0);
615 splx(s);
616 goto restart3;
617 }
618 splx(s);
619
620 restart4:
621 s = splbio();
622 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
623 aiocbn = TAILQ_NEXT(aiocbe, plist);
624 if (aio_free_entry(aiocbe)) {
625 splx(s);
626 goto restart4;
627 }
628 }
629 splx(s);
630
631 /*
632 * If we've slept, jobs might have moved from one queue to another.
633 * Retry rundown if we didn't manage to empty the queues.
634 */
635 mtx_lock(&aio_freeproc_mtx);
636 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
637 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
638 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
639 TAILQ_FIRST(&ki->kaio_bufdone) != NULL) {
640 mtx_unlock(&aio_freeproc_mtx);
641 goto restart1;
642 }
643 mtx_unlock(&aio_freeproc_mtx);
644
645 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
646 ljn = TAILQ_NEXT(lj, lioj_list);
647 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
648 0)) {
649 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
650 uma_zfree(aiolio_zone, lj);
651 } else {
652 #ifdef DIAGNOSTIC
653 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
654 "QF:%d\n", lj->lioj_buffer_count,
655 lj->lioj_buffer_finished_count,
656 lj->lioj_queue_count,
657 lj->lioj_queue_finished_count);
658 #endif
659 }
660 }
661
662 uma_zfree(kaio_zone, ki);
663 p->p_aioinfo = NULL;
664 mtx_unlock(&Giant);
665 }
666
667 /*
668 * Select a job to run (called by an AIO daemon).
669 */
670 static struct aiocblist *
671 aio_selectjob(struct aiothreadlist *aiop)
672 {
673 int s;
674 struct aiocblist *aiocbe;
675 struct kaioinfo *ki;
676 struct proc *userp;
677
678 s = splnet();
679 mtx_lock(&aio_freeproc_mtx);
680 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
681 TAILQ_NEXT(aiocbe, list)) {
682 userp = aiocbe->userproc;
683 ki = userp->p_aioinfo;
684
685 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
686 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
687 mtx_unlock(&aio_freeproc_mtx);
688 splx(s);
689 return (aiocbe);
690 }
691 }
692 mtx_unlock(&aio_freeproc_mtx);
693 splx(s);
694
695 return (NULL);
696 }
697
698 /*
699 * The AIO processing activity. This is the code that does the I/O request for
700 * the non-physio version of the operations. The normal vn operations are used,
701 * and this code should work in all instances for every type of file, including
702 * pipes, sockets, fifos, and regular files.
703 */
704 static void
705 aio_process(struct aiocblist *aiocbe)
706 {
707 struct ucred *td_savedcred;
708 struct thread *td;
709 struct proc *mycp;
710 struct aiocb *cb;
711 struct file *fp;
712 struct uio auio;
713 struct iovec aiov;
714 int cnt;
715 int error;
716 int oublock_st, oublock_end;
717 int inblock_st, inblock_end;
718
719 td = curthread;
720 td_savedcred = td->td_ucred;
721 td->td_ucred = aiocbe->cred;
722 mycp = td->td_proc;
723 cb = &aiocbe->uaiocb;
724 fp = aiocbe->fd_file;
725
726 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
727 aiov.iov_len = cb->aio_nbytes;
728
729 auio.uio_iov = &aiov;
730 auio.uio_iovcnt = 1;
731 auio.uio_offset = cb->aio_offset;
732 auio.uio_resid = cb->aio_nbytes;
733 cnt = cb->aio_nbytes;
734 auio.uio_segflg = UIO_USERSPACE;
735 auio.uio_td = td;
736
737 inblock_st = mycp->p_stats->p_ru.ru_inblock;
738 oublock_st = mycp->p_stats->p_ru.ru_oublock;
739 /*
740 * _aio_aqueue() acquires a reference to the file that is
741 * released in aio_free_entry().
742 */
743 if (cb->aio_lio_opcode == LIO_READ) {
744 auio.uio_rw = UIO_READ;
745 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
746 } else {
747 auio.uio_rw = UIO_WRITE;
748 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
749 }
750 inblock_end = mycp->p_stats->p_ru.ru_inblock;
751 oublock_end = mycp->p_stats->p_ru.ru_oublock;
752
753 aiocbe->inputcharge = inblock_end - inblock_st;
754 aiocbe->outputcharge = oublock_end - oublock_st;
755
756 if ((error) && (auio.uio_resid != cnt)) {
757 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
758 error = 0;
759 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
760 PROC_LOCK(aiocbe->userproc);
761 psignal(aiocbe->userproc, SIGPIPE);
762 PROC_UNLOCK(aiocbe->userproc);
763 }
764 }
765
766 cnt -= auio.uio_resid;
767 cb->_aiocb_private.error = error;
768 cb->_aiocb_private.status = cnt;
769 td->td_ucred = td_savedcred;
770 }
771
772 /*
773 * The AIO daemon, most of the actual work is done in aio_process,
774 * but the setup (and address space mgmt) is done in this routine.
775 */
776 static void
777 aio_daemon(void *uproc)
778 {
779 int s;
780 struct aio_liojob *lj;
781 struct aiocb *cb;
782 struct aiocblist *aiocbe;
783 struct aiothreadlist *aiop;
784 struct kaioinfo *ki;
785 struct proc *curcp, *mycp, *userp;
786 struct vmspace *myvm, *tmpvm;
787 struct thread *td = curthread;
788 struct pgrp *newpgrp;
789 struct session *newsess;
790
791 /*
792 * Local copies of curproc (cp) and vmspace (myvm)
793 */
794 mycp = td->td_proc;
795 myvm = mycp->p_vmspace;
796
797 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
798
799 /*
800 * Allocate and ready the aio control info. There is one aiop structure
801 * per daemon.
802 */
803 aiop = uma_zalloc(aiop_zone, M_WAITOK);
804 aiop->aiothread = td;
805 aiop->aiothreadflags |= AIOP_FREE;
806
807 /*
808 * Place thread (lightweight process) onto the AIO free thread list.
809 */
810 mtx_lock(&aio_freeproc_mtx);
811 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
812 mtx_unlock(&aio_freeproc_mtx);
813
814 /* The daemon resides in its own pgrp. */
815 MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
816 M_WAITOK | M_ZERO);
817 MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
818 M_WAITOK | M_ZERO);
819
820 sx_xlock(&proctree_lock);
821 enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
822 sx_xunlock(&proctree_lock);
823 mtx_lock(&Giant);
824
825 /*
826 * Wakeup parent process. (Parent sleeps to keep from blasting away
827 * and creating too many daemons.)
828 */
829 wakeup(mycp);
830
831 mtx_lock(&aio_freeproc_mtx);
832 for (;;) {
833 /*
834 * curcp is the current daemon process context.
835 * userp is the current user process context.
836 */
837 curcp = mycp;
838
839 /*
840 * Take daemon off of free queue
841 */
842 if (aiop->aiothreadflags & AIOP_FREE) {
843 TAILQ_REMOVE(&aio_freeproc, aiop, list);
844 aiop->aiothreadflags &= ~AIOP_FREE;
845 }
846 mtx_unlock(&aio_freeproc_mtx);
847
848 /*
849 * Check for jobs.
850 */
851 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
852 cb = &aiocbe->uaiocb;
853 userp = aiocbe->userproc;
854
855 aiocbe->jobstate = JOBST_JOBRUNNING;
856
857 /*
858 * Connect to process address space for user program.
859 */
860 if (userp != curcp) {
861 /*
862 * Save the current address space that we are
863 * connected to.
864 */
865 tmpvm = mycp->p_vmspace;
866
867 /*
868 * Point to the new user address space, and
869 * refer to it.
870 */
871 mycp->p_vmspace = userp->p_vmspace;
872 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
873
874 /* Activate the new mapping. */
875 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
876
877 /*
878 * If the old address space wasn't the daemons
879 * own address space, then we need to remove the
880 * daemon's reference from the other process
881 * that it was acting on behalf of.
882 */
883 if (tmpvm != myvm) {
884 vmspace_free(tmpvm);
885 }
886 curcp = userp;
887 }
888
889 ki = userp->p_aioinfo;
890 lj = aiocbe->lio;
891
892 /* Account for currently active jobs. */
893 ki->kaio_active_count++;
894
895 /* Do the I/O function. */
896 aio_process(aiocbe);
897
898 /* Decrement the active job count. */
899 ki->kaio_active_count--;
900
901 /*
902 * Increment the completion count for wakeup/signal
903 * comparisons.
904 */
905 aiocbe->jobflags |= AIOCBLIST_DONE;
906 ki->kaio_queue_finished_count++;
907 if (lj)
908 lj->lioj_queue_finished_count++;
909 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
910 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
911 ki->kaio_flags &= ~KAIO_WAKEUP;
912 wakeup(userp);
913 }
914
915 s = splbio();
916 if (lj && (lj->lioj_flags &
917 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
918 if ((lj->lioj_queue_finished_count ==
919 lj->lioj_queue_count) &&
920 (lj->lioj_buffer_finished_count ==
921 lj->lioj_buffer_count)) {
922 PROC_LOCK(userp);
923 psignal(userp,
924 lj->lioj_signal.sigev_signo);
925 PROC_UNLOCK(userp);
926 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
927 }
928 }
929 splx(s);
930
931 aiocbe->jobstate = JOBST_JOBFINISHED;
932
933 s = splnet();
934 mtx_lock(&aio_freeproc_mtx);
935 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
936 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
937 mtx_unlock(&aio_freeproc_mtx);
938 splx(s);
939 KNOTE_UNLOCKED(&aiocbe->klist, 0);
940
941 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
942 wakeup(aiocbe);
943 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
944 }
945
946 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
947 PROC_LOCK(userp);
948 psignal(userp, cb->aio_sigevent.sigev_signo);
949 PROC_UNLOCK(userp);
950 }
951 }
952
953 /*
954 * Disconnect from user address space.
955 */
956 if (curcp != mycp) {
957 /* Get the user address space to disconnect from. */
958 tmpvm = mycp->p_vmspace;
959
960 /* Get original address space for daemon. */
961 mycp->p_vmspace = myvm;
962
963 /* Activate the daemon's address space. */
964 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
965 #ifdef DIAGNOSTIC
966 if (tmpvm == myvm) {
967 printf("AIOD: vmspace problem -- %d\n",
968 mycp->p_pid);
969 }
970 #endif
971 /* Remove our vmspace reference. */
972 vmspace_free(tmpvm);
973
974 curcp = mycp;
975 }
976
977 mtx_lock(&aio_freeproc_mtx);
978 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
979 aiop->aiothreadflags |= AIOP_FREE;
980
981 /*
982 * If daemon is inactive for a long time, allow it to exit,
983 * thereby freeing resources.
984 */
985 if (msleep(aiop->aiothread, &aio_freeproc_mtx, PRIBIO,
986 "aiordy", aiod_lifetime)) {
987 s = splnet();
988 if (TAILQ_EMPTY(&aio_jobs)) {
989 if ((aiop->aiothreadflags & AIOP_FREE) &&
990 (num_aio_procs > target_aio_procs)) {
991 TAILQ_REMOVE(&aio_freeproc, aiop, list);
992 mtx_unlock(&aio_freeproc_mtx);
993 splx(s);
994 uma_zfree(aiop_zone, aiop);
995 num_aio_procs--;
996 #ifdef DIAGNOSTIC
997 if (mycp->p_vmspace->vm_refcnt <= 1) {
998 printf("AIOD: bad vm refcnt for"
999 " exiting daemon: %d\n",
1000 mycp->p_vmspace->vm_refcnt);
1001 }
1002 #endif
1003 kthread_exit(0);
1004 }
1005 }
1006 splx(s);
1007 }
1008 }
1009 }
1010
1011 /*
1012 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1013 * AIO daemon modifies its environment itself.
1014 */
1015 static int
1016 aio_newproc(void)
1017 {
1018 int error;
1019 struct proc *p;
1020
1021 error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, 0, "aiod%d",
1022 num_aio_procs);
1023 if (error)
1024 return (error);
1025
1026 /*
1027 * Wait until daemon is started, but continue on just in case to
1028 * handle error conditions.
1029 */
1030 error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1031
1032 num_aio_procs++;
1033
1034 return (error);
1035 }
1036
1037 /*
1038 * Try the high-performance, low-overhead physio method for eligible
1039 * VCHR devices. This method doesn't use an aio helper thread, and
1040 * thus has very low overhead.
1041 *
1042 * Assumes that the caller, _aio_aqueue(), has incremented the file
1043 * structure's reference count, preventing its deallocation for the
1044 * duration of this call.
1045 */
1046 static int
1047 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1048 {
1049 int error;
1050 struct aiocb *cb;
1051 struct file *fp;
1052 struct buf *bp;
1053 struct vnode *vp;
1054 struct kaioinfo *ki;
1055 struct aio_liojob *lj;
1056 int s;
1057 int notify;
1058
1059 cb = &aiocbe->uaiocb;
1060 fp = aiocbe->fd_file;
1061
1062 if (fp->f_type != DTYPE_VNODE)
1063 return (-1);
1064
1065 vp = fp->f_vnode;
1066
1067 /*
1068 * If its not a disk, we don't want to return a positive error.
1069 * It causes the aio code to not fall through to try the thread
1070 * way when you're talking to a regular file.
1071 */
1072 if (!vn_isdisk(vp, &error)) {
1073 if (error == ENOTBLK)
1074 return (-1);
1075 else
1076 return (error);
1077 }
1078
1079 if (vp->v_bufobj.bo_bsize == 0)
1080 return (-1);
1081
1082 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1083 return (-1);
1084
1085 if (cb->aio_nbytes >
1086 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1087 return (-1);
1088
1089 ki = p->p_aioinfo;
1090 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1091 return (-1);
1092
1093 ki->kaio_buffer_count++;
1094
1095 lj = aiocbe->lio;
1096 if (lj)
1097 lj->lioj_buffer_count++;
1098
1099 /* Create and build a buffer header for a transfer. */
1100 bp = (struct buf *)getpbuf(NULL);
1101 BUF_KERNPROC(bp);
1102
1103 /*
1104 * Get a copy of the kva from the physical buffer.
1105 */
1106 error = 0;
1107
1108 bp->b_bcount = cb->aio_nbytes;
1109 bp->b_bufsize = cb->aio_nbytes;
1110 bp->b_iodone = aio_physwakeup;
1111 bp->b_saveaddr = bp->b_data;
1112 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1113 bp->b_offset = cb->aio_offset;
1114 bp->b_iooffset = cb->aio_offset;
1115 bp->b_blkno = btodb(cb->aio_offset);
1116 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1117
1118 /*
1119 * Bring buffer into kernel space.
1120 */
1121 if (vmapbuf(bp) < 0) {
1122 error = EFAULT;
1123 goto doerror;
1124 }
1125
1126 s = splbio();
1127 aiocbe->bp = bp;
1128 bp->b_caller1 = (void *)aiocbe;
1129 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1130 aiocbe->jobstate = JOBST_JOBQBUF;
1131 cb->_aiocb_private.status = cb->aio_nbytes;
1132 num_buf_aio++;
1133 bp->b_error = 0;
1134
1135 splx(s);
1136
1137 /* Perform transfer. */
1138 dev_strategy(vp->v_rdev, bp);
1139
1140 notify = 0;
1141 s = splbio();
1142
1143 /*
1144 * If we had an error invoking the request, or an error in processing
1145 * the request before we have returned, we process it as an error in
1146 * transfer. Note that such an I/O error is not indicated immediately,
1147 * but is returned using the aio_error mechanism. In this case,
1148 * aio_suspend will return immediately.
1149 */
1150 if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1151 struct aiocb *job = aiocbe->uuaiocb;
1152
1153 aiocbe->uaiocb._aiocb_private.status = 0;
1154 suword(&job->_aiocb_private.status, 0);
1155 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1156 suword(&job->_aiocb_private.error, bp->b_error);
1157
1158 ki->kaio_buffer_finished_count++;
1159
1160 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1161 aiocbe->jobstate = JOBST_JOBBFINISHED;
1162 aiocbe->jobflags |= AIOCBLIST_DONE;
1163 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1164 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1165 notify = 1;
1166 }
1167 }
1168 splx(s);
1169 if (notify)
1170 KNOTE_UNLOCKED(&aiocbe->klist, 0);
1171 return (0);
1172
1173 doerror:
1174 ki->kaio_buffer_count--;
1175 if (lj)
1176 lj->lioj_buffer_count--;
1177 aiocbe->bp = NULL;
1178 relpbuf(bp, NULL);
1179 return (error);
1180 }
1181
1182 /*
1183 * This waits/tests physio completion.
1184 */
1185 static int
1186 aio_fphysio(struct aiocblist *iocb)
1187 {
1188 int s;
1189 struct buf *bp;
1190 int error;
1191
1192 bp = iocb->bp;
1193
1194 s = splbio();
1195 while ((bp->b_flags & B_DONE) == 0) {
1196 if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1197 if ((bp->b_flags & B_DONE) == 0) {
1198 splx(s);
1199 return (EINPROGRESS);
1200 } else
1201 break;
1202 }
1203 }
1204 splx(s);
1205
1206 /* Release mapping into kernel space. */
1207 vunmapbuf(bp);
1208 iocb->bp = 0;
1209
1210 error = 0;
1211
1212 /* Check for an error. */
1213 if (bp->b_ioflags & BIO_ERROR)
1214 error = bp->b_error;
1215
1216 relpbuf(bp, NULL);
1217 return (error);
1218 }
1219
1220 /*
1221 * Wake up aio requests that may be serviceable now.
1222 */
1223 static void
1224 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1225 {
1226 struct aiocblist *cb,*cbn;
1227 struct proc *p;
1228 struct kaioinfo *ki = NULL;
1229 int opcode, wakecount = 0;
1230 struct aiothreadlist *aiop;
1231
1232 if (sb == &so->so_snd) {
1233 opcode = LIO_WRITE;
1234 SOCKBUF_LOCK(&so->so_snd);
1235 so->so_snd.sb_flags &= ~SB_AIO;
1236 SOCKBUF_UNLOCK(&so->so_snd);
1237 } else {
1238 opcode = LIO_READ;
1239 SOCKBUF_LOCK(&so->so_rcv);
1240 so->so_rcv.sb_flags &= ~SB_AIO;
1241 SOCKBUF_UNLOCK(&so->so_rcv);
1242 }
1243
1244 mtx_lock(&aio_freeproc_mtx);
1245 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1246 cbn = TAILQ_NEXT(cb, list);
1247 if (opcode == cb->uaiocb.aio_lio_opcode) {
1248 p = cb->userproc;
1249 ki = p->p_aioinfo;
1250 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1251 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1252 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1253 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1254 wakecount++;
1255 if (cb->jobstate != JOBST_JOBQGLOBAL)
1256 panic("invalid queue value");
1257 }
1258 }
1259
1260 while (wakecount--) {
1261 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1262 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1263 aiop->aiothreadflags &= ~AIOP_FREE;
1264 wakeup(aiop->aiothread);
1265 }
1266 }
1267 mtx_unlock(&aio_freeproc_mtx);
1268 }
1269
1270 /*
1271 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1272 * technique is done in this code.
1273 */
1274 static int
1275 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1276 {
1277 struct proc *p = td->td_proc;
1278 struct filedesc *fdp;
1279 struct file *fp;
1280 unsigned int fd;
1281 struct socket *so;
1282 int s;
1283 int error;
1284 int opcode;
1285 struct aiocblist *aiocbe;
1286 struct aiothreadlist *aiop;
1287 struct kaioinfo *ki;
1288 struct kevent kev;
1289 struct kqueue *kq;
1290 struct file *kq_fp;
1291 struct sockbuf *sb;
1292
1293 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1294 aiocbe->inputcharge = 0;
1295 aiocbe->outputcharge = 0;
1296 /* XXX - need a lock */
1297 knlist_init(&aiocbe->klist, NULL, NULL, NULL, NULL);
1298
1299 suword(&job->_aiocb_private.status, -1);
1300 suword(&job->_aiocb_private.error, 0);
1301 suword(&job->_aiocb_private.kernelinfo, -1);
1302
1303 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1304 if (error) {
1305 suword(&job->_aiocb_private.error, error);
1306 uma_zfree(aiocb_zone, aiocbe);
1307 return (error);
1308 }
1309 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1310 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1311 uma_zfree(aiocb_zone, aiocbe);
1312 return (EINVAL);
1313 }
1314
1315 /* Save userspace address of the job info. */
1316 aiocbe->uuaiocb = job;
1317
1318 /* Get the opcode. */
1319 if (type != LIO_NOP)
1320 aiocbe->uaiocb.aio_lio_opcode = type;
1321 opcode = aiocbe->uaiocb.aio_lio_opcode;
1322
1323 /* Get the fd info for process. */
1324 fdp = p->p_fd;
1325
1326 /*
1327 * Range check file descriptor.
1328 */
1329 FILEDESC_LOCK(fdp);
1330 fd = aiocbe->uaiocb.aio_fildes;
1331 if (fd >= fdp->fd_nfiles) {
1332 FILEDESC_UNLOCK(fdp);
1333 uma_zfree(aiocb_zone, aiocbe);
1334 if (type == 0)
1335 suword(&job->_aiocb_private.error, EBADF);
1336 return (EBADF);
1337 }
1338
1339 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1340 if ((fp == NULL) ||
1341 ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 0)) ||
1342 ((opcode == LIO_READ) && ((fp->f_flag & FREAD) == 0))) {
1343 FILEDESC_UNLOCK(fdp);
1344 uma_zfree(aiocb_zone, aiocbe);
1345 if (type == 0)
1346 suword(&job->_aiocb_private.error, EBADF);
1347 return (EBADF);
1348 }
1349 fhold(fp);
1350 FILEDESC_UNLOCK(fdp);
1351
1352 if (aiocbe->uaiocb.aio_offset == -1LL) {
1353 error = EINVAL;
1354 goto aqueue_fail;
1355 }
1356 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1357 if (error) {
1358 error = EINVAL;
1359 goto aqueue_fail;
1360 }
1361 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1362 if (jobrefid == LONG_MAX)
1363 jobrefid = 1;
1364 else
1365 jobrefid++;
1366
1367 if (opcode == LIO_NOP) {
1368 fdrop(fp, td);
1369 uma_zfree(aiocb_zone, aiocbe);
1370 if (type == 0) {
1371 suword(&job->_aiocb_private.error, 0);
1372 suword(&job->_aiocb_private.status, 0);
1373 suword(&job->_aiocb_private.kernelinfo, 0);
1374 }
1375 return (0);
1376 }
1377 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1378 if (type == 0)
1379 suword(&job->_aiocb_private.status, 0);
1380 error = EINVAL;
1381 goto aqueue_fail;
1382 }
1383
1384 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1385 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1386 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1387 } else
1388 goto no_kqueue;
1389 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1390 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1391 (kq_fp->f_type != DTYPE_KQUEUE)) {
1392 error = EBADF;
1393 goto aqueue_fail;
1394 }
1395 kq = kq_fp->f_data;
1396 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1397 kev.filter = EVFILT_AIO;
1398 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1399 kev.data = (intptr_t)aiocbe;
1400 error = kqueue_register(kq, &kev, td, 1);
1401 aqueue_fail:
1402 if (error) {
1403 fdrop(fp, td);
1404 uma_zfree(aiocb_zone, aiocbe);
1405 if (type == 0)
1406 suword(&job->_aiocb_private.error, error);
1407 goto done;
1408 }
1409 no_kqueue:
1410
1411 suword(&job->_aiocb_private.error, EINPROGRESS);
1412 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1413 aiocbe->userproc = p;
1414 aiocbe->cred = crhold(td->td_ucred);
1415 aiocbe->jobflags = 0;
1416 aiocbe->lio = lj;
1417 ki = p->p_aioinfo;
1418
1419 if (fp->f_type == DTYPE_SOCKET) {
1420 /*
1421 * Alternate queueing for socket ops: Reach down into the
1422 * descriptor to get the socket data. Then check to see if the
1423 * socket is ready to be read or written (based on the requested
1424 * operation).
1425 *
1426 * If it is not ready for io, then queue the aiocbe on the
1427 * socket, and set the flags so we get a call when sbnotify()
1428 * happens.
1429 *
1430 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1431 * and unlock the snd sockbuf for no reason.
1432 */
1433 so = fp->f_data;
1434 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1435 mtx_lock(&aio_freeproc_mtx);
1436 SOCKBUF_LOCK(sb);
1437 s = splnet();
1438 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1439 LIO_WRITE) && (!sowriteable(so)))) {
1440 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1441 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1442 sb->sb_flags |= SB_AIO;
1443 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1444 ki->kaio_queue_count++;
1445 num_queue_count++;
1446 SOCKBUF_UNLOCK(sb);
1447 mtx_unlock(&aio_freeproc_mtx);
1448 splx(s);
1449 error = 0;
1450 goto done;
1451 }
1452 SOCKBUF_UNLOCK(sb);
1453 mtx_unlock(&aio_freeproc_mtx);
1454 splx(s);
1455 }
1456
1457 if ((error = aio_qphysio(p, aiocbe)) == 0)
1458 goto done;
1459 if (error > 0) {
1460 suword(&job->_aiocb_private.status, 0);
1461 aiocbe->uaiocb._aiocb_private.error = error;
1462 suword(&job->_aiocb_private.error, error);
1463 goto done;
1464 }
1465
1466 /* No buffer for daemon I/O. */
1467 aiocbe->bp = NULL;
1468
1469 ki->kaio_queue_count++;
1470 if (lj)
1471 lj->lioj_queue_count++;
1472 s = splnet();
1473 mtx_lock(&aio_freeproc_mtx);
1474 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1475 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1476 splx(s);
1477 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1478
1479 num_queue_count++;
1480 error = 0;
1481
1482 /*
1483 * If we don't have a free AIO process, and we are below our quota, then
1484 * start one. Otherwise, depend on the subsequent I/O completions to
1485 * pick-up this job. If we don't sucessfully create the new process
1486 * (thread) due to resource issues, we return an error for now (EAGAIN),
1487 * which is likely not the correct thing to do.
1488 */
1489 retryproc:
1490 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1491 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1492 aiop->aiothreadflags &= ~AIOP_FREE;
1493 wakeup(aiop->aiothread);
1494 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1495 ((ki->kaio_active_count + num_aio_resv_start) <
1496 ki->kaio_maxactive_count)) {
1497 num_aio_resv_start++;
1498 mtx_unlock(&aio_freeproc_mtx);
1499 if ((error = aio_newproc()) == 0) {
1500 mtx_lock(&aio_freeproc_mtx);
1501 num_aio_resv_start--;
1502 goto retryproc;
1503 }
1504 mtx_lock(&aio_freeproc_mtx);
1505 num_aio_resv_start--;
1506 }
1507 mtx_unlock(&aio_freeproc_mtx);
1508 done:
1509 return (error);
1510 }
1511
1512 /*
1513 * This routine queues an AIO request, checking for quotas.
1514 */
1515 static int
1516 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1517 {
1518 struct proc *p = td->td_proc;
1519 struct kaioinfo *ki;
1520
1521 if (p->p_aioinfo == NULL)
1522 aio_init_aioinfo(p);
1523
1524 if (num_queue_count >= max_queue_count)
1525 return (EAGAIN);
1526
1527 ki = p->p_aioinfo;
1528 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1529 return (EAGAIN);
1530
1531 return _aio_aqueue(td, job, NULL, type);
1532 }
1533
1534 /*
1535 * Support the aio_return system call, as a side-effect, kernel resources are
1536 * released.
1537 */
1538 int
1539 aio_return(struct thread *td, struct aio_return_args *uap)
1540 {
1541 struct proc *p = td->td_proc;
1542 int s;
1543 long jobref;
1544 struct aiocblist *cb, *ncb;
1545 struct aiocb *ujob;
1546 struct kaioinfo *ki;
1547
1548 ujob = uap->aiocbp;
1549 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1550 if (jobref == -1 || jobref == 0)
1551 return (EINVAL);
1552
1553 ki = p->p_aioinfo;
1554 if (ki == NULL)
1555 return (EINVAL);
1556 PROC_LOCK(p);
1557 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1558 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1559 jobref) {
1560 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1561 p->p_stats->p_ru.ru_oublock +=
1562 cb->outputcharge;
1563 cb->outputcharge = 0;
1564 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1565 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1566 cb->inputcharge = 0;
1567 }
1568 goto done;
1569 }
1570 }
1571 s = splbio();
1572 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1573 ncb = TAILQ_NEXT(cb, plist);
1574 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1575 == jobref) {
1576 break;
1577 }
1578 }
1579 splx(s);
1580 done:
1581 PROC_UNLOCK(p);
1582 if (cb != NULL) {
1583 if (ujob == cb->uuaiocb) {
1584 td->td_retval[0] =
1585 cb->uaiocb._aiocb_private.status;
1586 } else
1587 td->td_retval[0] = EFAULT;
1588 aio_free_entry(cb);
1589 return (0);
1590 }
1591 return (EINVAL);
1592 }
1593
1594 /*
1595 * Allow a process to wakeup when any of the I/O requests are completed.
1596 */
1597 int
1598 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1599 {
1600 struct proc *p = td->td_proc;
1601 struct timeval atv;
1602 struct timespec ts;
1603 struct aiocb *const *cbptr, *cbp;
1604 struct kaioinfo *ki;
1605 struct aiocblist *cb;
1606 int i;
1607 int njoblist;
1608 int error, s, timo;
1609 long *ijoblist;
1610 struct aiocb **ujoblist;
1611
1612 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1613 return (EINVAL);
1614
1615 timo = 0;
1616 if (uap->timeout) {
1617 /* Get timespec struct. */
1618 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1619 return (error);
1620
1621 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1622 return (EINVAL);
1623
1624 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1625 if (itimerfix(&atv))
1626 return (EINVAL);
1627 timo = tvtohz(&atv);
1628 }
1629
1630 ki = p->p_aioinfo;
1631 if (ki == NULL)
1632 return (EAGAIN);
1633
1634 njoblist = 0;
1635 ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1636 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1637 cbptr = uap->aiocbp;
1638
1639 for (i = 0; i < uap->nent; i++) {
1640 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1641 if (cbp == 0)
1642 continue;
1643 ujoblist[njoblist] = cbp;
1644 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1645 njoblist++;
1646 }
1647
1648 if (njoblist == 0) {
1649 uma_zfree(aiol_zone, ijoblist);
1650 uma_zfree(aiol_zone, ujoblist);
1651 return (0);
1652 }
1653
1654 error = 0;
1655 for (;;) {
1656 PROC_LOCK(p);
1657 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1658 for (i = 0; i < njoblist; i++) {
1659 if (((intptr_t)
1660 cb->uaiocb._aiocb_private.kernelinfo) ==
1661 ijoblist[i]) {
1662 PROC_UNLOCK(p);
1663 if (ujoblist[i] != cb->uuaiocb)
1664 error = EINVAL;
1665 uma_zfree(aiol_zone, ijoblist);
1666 uma_zfree(aiol_zone, ujoblist);
1667 return (error);
1668 }
1669 }
1670 }
1671
1672 s = splbio();
1673 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1674 TAILQ_NEXT(cb, plist)) {
1675 for (i = 0; i < njoblist; i++) {
1676 if (((intptr_t)
1677 cb->uaiocb._aiocb_private.kernelinfo) ==
1678 ijoblist[i]) {
1679 PROC_UNLOCK(p);
1680 splx(s);
1681 if (ujoblist[i] != cb->uuaiocb)
1682 error = EINVAL;
1683 uma_zfree(aiol_zone, ijoblist);
1684 uma_zfree(aiol_zone, ujoblist);
1685 return (error);
1686 }
1687 }
1688 }
1689
1690 ki->kaio_flags |= KAIO_WAKEUP;
1691 error = msleep(p, &p->p_mtx, PDROP | PRIBIO | PCATCH, "aiospn",
1692 timo);
1693 splx(s);
1694
1695 if (error == ERESTART || error == EINTR) {
1696 uma_zfree(aiol_zone, ijoblist);
1697 uma_zfree(aiol_zone, ujoblist);
1698 return (EINTR);
1699 } else if (error == EWOULDBLOCK) {
1700 uma_zfree(aiol_zone, ijoblist);
1701 uma_zfree(aiol_zone, ujoblist);
1702 return (EAGAIN);
1703 }
1704 }
1705
1706 /* NOTREACHED */
1707 return (EINVAL);
1708 }
1709
1710 /*
1711 * aio_cancel cancels any non-physio aio operations not currently in
1712 * progress.
1713 */
1714 int
1715 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1716 {
1717 struct proc *p = td->td_proc;
1718 struct kaioinfo *ki;
1719 struct aiocblist *cbe, *cbn;
1720 struct file *fp;
1721 struct filedesc *fdp;
1722 struct socket *so;
1723 struct proc *po;
1724 int s,error;
1725 int cancelled=0;
1726 int notcancelled=0;
1727 struct vnode *vp;
1728
1729 fdp = p->p_fd;
1730 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1731 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1732 return (EBADF);
1733
1734 if (fp->f_type == DTYPE_VNODE) {
1735 vp = fp->f_vnode;
1736
1737 if (vn_isdisk(vp,&error)) {
1738 td->td_retval[0] = AIO_NOTCANCELED;
1739 return (0);
1740 }
1741 } else if (fp->f_type == DTYPE_SOCKET) {
1742 so = fp->f_data;
1743
1744 s = splnet();
1745
1746 mtx_lock(&aio_freeproc_mtx);
1747 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1748 cbn = TAILQ_NEXT(cbe, list);
1749 if ((uap->aiocbp == NULL) ||
1750 (uap->aiocbp == cbe->uuaiocb) ) {
1751 po = cbe->userproc;
1752 ki = po->p_aioinfo;
1753 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1754 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1755 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1756 if (ki->kaio_flags & KAIO_WAKEUP) {
1757 wakeup(po);
1758 }
1759 cbe->jobstate = JOBST_JOBFINISHED;
1760 cbe->uaiocb._aiocb_private.status=-1;
1761 cbe->uaiocb._aiocb_private.error=ECANCELED;
1762 cancelled++;
1763 /* XXX cancelled, knote? */
1764 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1765 SIGEV_SIGNAL) {
1766 PROC_LOCK(cbe->userproc);
1767 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1768 PROC_UNLOCK(cbe->userproc);
1769 }
1770 if (uap->aiocbp)
1771 break;
1772 }
1773 }
1774 mtx_unlock(&aio_freeproc_mtx);
1775 splx(s);
1776
1777 if ((cancelled) && (uap->aiocbp)) {
1778 td->td_retval[0] = AIO_CANCELED;
1779 return (0);
1780 }
1781 }
1782 ki=p->p_aioinfo;
1783 if (ki == NULL)
1784 goto done;
1785 s = splnet();
1786
1787 mtx_lock(&aio_freeproc_mtx);
1788 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1789 cbn = TAILQ_NEXT(cbe, plist);
1790
1791 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1792 ((uap->aiocbp == NULL ) ||
1793 (uap->aiocbp == cbe->uuaiocb))) {
1794
1795 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1796 TAILQ_REMOVE(&aio_jobs, cbe, list);
1797 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1798 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1799 plist);
1800 cancelled++;
1801 ki->kaio_queue_finished_count++;
1802 cbe->jobstate = JOBST_JOBFINISHED;
1803 cbe->uaiocb._aiocb_private.status = -1;
1804 cbe->uaiocb._aiocb_private.error = ECANCELED;
1805 /* XXX cancelled, knote? */
1806 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1807 SIGEV_SIGNAL) {
1808 PROC_LOCK(cbe->userproc);
1809 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1810 PROC_UNLOCK(cbe->userproc);
1811 }
1812 } else {
1813 notcancelled++;
1814 }
1815 }
1816 }
1817 mtx_unlock(&aio_freeproc_mtx);
1818 splx(s);
1819 done:
1820 if (notcancelled) {
1821 td->td_retval[0] = AIO_NOTCANCELED;
1822 return (0);
1823 }
1824 if (cancelled) {
1825 td->td_retval[0] = AIO_CANCELED;
1826 return (0);
1827 }
1828 td->td_retval[0] = AIO_ALLDONE;
1829
1830 return (0);
1831 }
1832
1833 /*
1834 * aio_error is implemented in the kernel level for compatibility purposes only.
1835 * For a user mode async implementation, it would be best to do it in a userland
1836 * subroutine.
1837 */
1838 int
1839 aio_error(struct thread *td, struct aio_error_args *uap)
1840 {
1841 struct proc *p = td->td_proc;
1842 int s;
1843 struct aiocblist *cb;
1844 struct kaioinfo *ki;
1845 long jobref;
1846
1847 ki = p->p_aioinfo;
1848 if (ki == NULL)
1849 return (EINVAL);
1850
1851 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1852 if ((jobref == -1) || (jobref == 0))
1853 return (EINVAL);
1854
1855 mtx_lock(&aio_freeproc_mtx);
1856 PROC_LOCK(p);
1857 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1858 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1859 jobref) {
1860 PROC_UNLOCK(p);
1861 mtx_unlock(&aio_freeproc_mtx);
1862 td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1863 return (0);
1864 }
1865 }
1866
1867 s = splnet();
1868
1869 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1870 plist)) {
1871 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1872 jobref) {
1873 PROC_UNLOCK(p);
1874 mtx_unlock(&aio_freeproc_mtx);
1875 td->td_retval[0] = EINPROGRESS;
1876 splx(s);
1877 return (0);
1878 }
1879 }
1880
1881 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1882 plist)) {
1883 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1884 jobref) {
1885 PROC_UNLOCK(p);
1886 mtx_unlock(&aio_freeproc_mtx);
1887 td->td_retval[0] = EINPROGRESS;
1888 splx(s);
1889 return (0);
1890 }
1891 }
1892 splx(s);
1893
1894 s = splbio();
1895 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1896 plist)) {
1897 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1898 jobref) {
1899 PROC_UNLOCK(p);
1900 mtx_unlock(&aio_freeproc_mtx);
1901 td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1902 splx(s);
1903 return (0);
1904 }
1905 }
1906
1907 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1908 plist)) {
1909 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1910 jobref) {
1911 PROC_UNLOCK(p);
1912 mtx_unlock(&aio_freeproc_mtx);
1913 td->td_retval[0] = EINPROGRESS;
1914 splx(s);
1915 return (0);
1916 }
1917 }
1918 splx(s);
1919 PROC_UNLOCK(p);
1920 mtx_unlock(&aio_freeproc_mtx);
1921
1922 #if (0)
1923 /*
1924 * Hack for lio.
1925 */
1926 status = fuword(&uap->aiocbp->_aiocb_private.status);
1927 if (status == -1)
1928 return fuword(&uap->aiocbp->_aiocb_private.error);
1929 #endif
1930 return (EINVAL);
1931 }
1932
1933 /* syscall - asynchronous read from a file (REALTIME) */
1934 int
1935 aio_read(struct thread *td, struct aio_read_args *uap)
1936 {
1937
1938 return aio_aqueue(td, uap->aiocbp, LIO_READ);
1939 }
1940
1941 /* syscall - asynchronous write to a file (REALTIME) */
1942 int
1943 aio_write(struct thread *td, struct aio_write_args *uap)
1944 {
1945
1946 return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1947 }
1948
1949 /* syscall - list directed I/O (REALTIME) */
1950 int
1951 lio_listio(struct thread *td, struct lio_listio_args *uap)
1952 {
1953 struct proc *p = td->td_proc;
1954 int nent, nentqueued;
1955 struct aiocb *iocb, * const *cbptr;
1956 struct aiocblist *cb;
1957 struct kaioinfo *ki;
1958 struct aio_liojob *lj;
1959 int error, runningcode;
1960 int nerror;
1961 int i;
1962 int s;
1963
1964 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1965 return (EINVAL);
1966
1967 nent = uap->nent;
1968 if (nent < 0 || nent > AIO_LISTIO_MAX)
1969 return (EINVAL);
1970
1971 if (p->p_aioinfo == NULL)
1972 aio_init_aioinfo(p);
1973
1974 if ((nent + num_queue_count) > max_queue_count)
1975 return (EAGAIN);
1976
1977 ki = p->p_aioinfo;
1978 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1979 return (EAGAIN);
1980
1981 lj = uma_zalloc(aiolio_zone, M_WAITOK);
1982 if (!lj)
1983 return (EAGAIN);
1984
1985 lj->lioj_flags = 0;
1986 lj->lioj_buffer_count = 0;
1987 lj->lioj_buffer_finished_count = 0;
1988 lj->lioj_queue_count = 0;
1989 lj->lioj_queue_finished_count = 0;
1990
1991 /*
1992 * Setup signal.
1993 */
1994 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1995 error = copyin(uap->sig, &lj->lioj_signal,
1996 sizeof(lj->lioj_signal));
1997 if (error) {
1998 uma_zfree(aiolio_zone, lj);
1999 return (error);
2000 }
2001 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2002 uma_zfree(aiolio_zone, lj);
2003 return (EINVAL);
2004 }
2005 lj->lioj_flags |= LIOJ_SIGNAL;
2006 }
2007 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2008 /*
2009 * Get pointers to the list of I/O requests.
2010 */
2011 nerror = 0;
2012 nentqueued = 0;
2013 cbptr = uap->acb_list;
2014 for (i = 0; i < uap->nent; i++) {
2015 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
2016 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
2017 error = _aio_aqueue(td, iocb, lj, 0);
2018 if (error == 0)
2019 nentqueued++;
2020 else
2021 nerror++;
2022 }
2023 }
2024
2025 /*
2026 * If we haven't queued any, then just return error.
2027 */
2028 if (nentqueued == 0)
2029 return (0);
2030
2031 /*
2032 * Calculate the appropriate error return.
2033 */
2034 runningcode = 0;
2035 if (nerror)
2036 runningcode = EIO;
2037
2038 if (uap->mode == LIO_WAIT) {
2039 int command, found;
2040 long jobref;
2041
2042 for (;;) {
2043 found = 0;
2044 for (i = 0; i < uap->nent; i++) {
2045 /*
2046 * Fetch address of the control buf pointer in
2047 * user space.
2048 */
2049 iocb = (struct aiocb *)
2050 (intptr_t)fuword(&cbptr[i]);
2051 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2052 == 0))
2053 continue;
2054
2055 /*
2056 * Fetch the associated command from user space.
2057 */
2058 command = fuword(&iocb->aio_lio_opcode);
2059 if (command == LIO_NOP) {
2060 found++;
2061 continue;
2062 }
2063
2064 jobref =
2065 fuword(&iocb->_aiocb_private.kernelinfo);
2066
2067 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2068 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2069 == jobref) {
2070 if (cb->uaiocb.aio_lio_opcode
2071 == LIO_WRITE) {
2072 p->p_stats->p_ru.ru_oublock
2073 +=
2074 cb->outputcharge;
2075 cb->outputcharge = 0;
2076 } else if (cb->uaiocb.aio_lio_opcode
2077 == LIO_READ) {
2078 p->p_stats->p_ru.ru_inblock
2079 += cb->inputcharge;
2080 cb->inputcharge = 0;
2081 }
2082 found++;
2083 break;
2084 }
2085 }
2086
2087 s = splbio();
2088 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2089 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2090 == jobref) {
2091 found++;
2092 break;
2093 }
2094 }
2095 splx(s);
2096 }
2097
2098 /*
2099 * If all I/Os have been disposed of, then we can
2100 * return.
2101 */
2102 if (found == nentqueued)
2103 return (runningcode);
2104
2105 ki->kaio_flags |= KAIO_WAKEUP;
2106 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2107
2108 if (error == EINTR)
2109 return (EINTR);
2110 else if (error == EWOULDBLOCK)
2111 return (EAGAIN);
2112 }
2113 }
2114
2115 return (runningcode);
2116 }
2117
2118 /*
2119 * Interrupt handler for physio, performs the necessary process wakeups, and
2120 * signals.
2121 */
2122 static void
2123 aio_physwakeup(struct buf *bp)
2124 {
2125 struct aiocblist *aiocbe;
2126 struct proc *p;
2127 struct kaioinfo *ki;
2128 struct aio_liojob *lj;
2129
2130 mtx_lock(&Giant);
2131 bp->b_flags |= B_DONE;
2132 wakeup(bp);
2133
2134 aiocbe = (struct aiocblist *)bp->b_caller1;
2135 if (aiocbe) {
2136 p = aiocbe->userproc;
2137
2138 aiocbe->jobstate = JOBST_JOBBFINISHED;
2139 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2140 aiocbe->uaiocb._aiocb_private.error = 0;
2141 aiocbe->jobflags |= AIOCBLIST_DONE;
2142
2143 if (bp->b_ioflags & BIO_ERROR)
2144 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2145
2146 lj = aiocbe->lio;
2147 if (lj) {
2148 lj->lioj_buffer_finished_count++;
2149
2150 /*
2151 * wakeup/signal if all of the interrupt jobs are done.
2152 */
2153 if (lj->lioj_buffer_finished_count ==
2154 lj->lioj_buffer_count &&
2155 lj->lioj_queue_finished_count ==
2156 lj->lioj_queue_count) {
2157 /*
2158 * Post a signal if it is called for.
2159 */
2160 if ((lj->lioj_flags &
2161 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2162 LIOJ_SIGNAL) {
2163 PROC_LOCK(p);
2164 psignal(p, lj->lioj_signal.sigev_signo);
2165 PROC_UNLOCK(p);
2166 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2167 }
2168 }
2169 }
2170
2171 ki = p->p_aioinfo;
2172 if (ki) {
2173 ki->kaio_buffer_finished_count++;
2174 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2175 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2176
2177 KNOTE_UNLOCKED(&aiocbe->klist, 0);
2178 /* Do the wakeup. */
2179 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2180 ki->kaio_flags &= ~KAIO_WAKEUP;
2181 wakeup(p);
2182 }
2183 }
2184
2185 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2186 PROC_LOCK(p);
2187 psignal(p, aiocbe->uaiocb.aio_sigevent.sigev_signo);
2188 PROC_UNLOCK(p);
2189 }
2190 }
2191 mtx_unlock(&Giant);
2192 }
2193
2194 /* syscall - wait for the next completion of an aio request */
2195 int
2196 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2197 {
2198 struct proc *p = td->td_proc;
2199 struct timeval atv;
2200 struct timespec ts;
2201 struct kaioinfo *ki;
2202 struct aiocblist *cb = NULL;
2203 int error, s, timo;
2204
2205 suword(uap->aiocbp, (int)NULL);
2206
2207 timo = 0;
2208 if (uap->timeout) {
2209 /* Get timespec struct. */
2210 error = copyin(uap->timeout, &ts, sizeof(ts));
2211 if (error)
2212 return (error);
2213
2214 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2215 return (EINVAL);
2216
2217 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2218 if (itimerfix(&atv))
2219 return (EINVAL);
2220 timo = tvtohz(&atv);
2221 }
2222
2223 ki = p->p_aioinfo;
2224 if (ki == NULL)
2225 return (EAGAIN);
2226
2227 for (;;) {
2228 PROC_LOCK(p);
2229 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2230 PROC_UNLOCK(p);
2231 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2232 td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2233 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2234 p->p_stats->p_ru.ru_oublock +=
2235 cb->outputcharge;
2236 cb->outputcharge = 0;
2237 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2238 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2239 cb->inputcharge = 0;
2240 }
2241 error = cb->uaiocb._aiocb_private.error;
2242 aio_free_entry(cb);
2243 return (error);
2244 }
2245
2246 s = splbio();
2247 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2248 PROC_UNLOCK(p);
2249 splx(s);
2250 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2251 error = cb->uaiocb._aiocb_private.error;
2252 td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2253 aio_free_entry(cb);
2254 return (error);
2255 }
2256
2257 ki->kaio_flags |= KAIO_WAKEUP;
2258 error = msleep(p, &p->p_mtx, PDROP | PRIBIO | PCATCH, "aiowc",
2259 timo);
2260 splx(s);
2261
2262 if (error == ERESTART)
2263 return (EINTR);
2264 else if (error < 0)
2265 return (error);
2266 else if (error == EINTR)
2267 return (EINTR);
2268 else if (error == EWOULDBLOCK)
2269 return (EAGAIN);
2270 }
2271 }
2272
2273 /* kqueue attach function */
2274 static int
2275 filt_aioattach(struct knote *kn)
2276 {
2277 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2278
2279 /*
2280 * The aiocbe pointer must be validated before using it, so
2281 * registration is restricted to the kernel; the user cannot
2282 * set EV_FLAG1.
2283 */
2284 if ((kn->kn_flags & EV_FLAG1) == 0)
2285 return (EPERM);
2286 kn->kn_flags &= ~EV_FLAG1;
2287
2288 knlist_add(&aiocbe->klist, kn, 0);
2289
2290 return (0);
2291 }
2292
2293 /* kqueue detach function */
2294 static void
2295 filt_aiodetach(struct knote *kn)
2296 {
2297 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2298
2299 knlist_remove(&aiocbe->klist, kn, 0);
2300 }
2301
2302 /* kqueue filter function */
2303 /*ARGSUSED*/
2304 static int
2305 filt_aio(struct knote *kn, long hint)
2306 {
2307 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2308
2309 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2310 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2311 aiocbe->jobstate != JOBST_JOBBFINISHED)
2312 return (0);
2313 kn->kn_flags |= EV_EOF;
2314 return (1);
2315 }
Cache object: 32651ba7a1e336579d353c63b2582706
|