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: releng/6.0/sys/kern/vfs_aio.c 149312 2005-08-20 06:07:55Z alc $");
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 NET_NEEDS_GIANT("aio");
65
66 /*
67 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
68 * overflow.
69 */
70 static long jobrefid;
71
72 #define JOBST_NULL 0x0
73 #define JOBST_JOBQGLOBAL 0x2
74 #define JOBST_JOBRUNNING 0x3
75 #define JOBST_JOBFINISHED 0x4
76 #define JOBST_JOBQBUF 0x5
77 #define JOBST_JOBBFINISHED 0x6
78
79 #ifndef MAX_AIO_PER_PROC
80 #define MAX_AIO_PER_PROC 32
81 #endif
82
83 #ifndef MAX_AIO_QUEUE_PER_PROC
84 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
85 #endif
86
87 #ifndef MAX_AIO_PROCS
88 #define MAX_AIO_PROCS 32
89 #endif
90
91 #ifndef MAX_AIO_QUEUE
92 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
93 #endif
94
95 #ifndef TARGET_AIO_PROCS
96 #define TARGET_AIO_PROCS 4
97 #endif
98
99 #ifndef MAX_BUF_AIO
100 #define MAX_BUF_AIO 16
101 #endif
102
103 #ifndef AIOD_TIMEOUT_DEFAULT
104 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
105 #endif
106
107 #ifndef AIOD_LIFETIME_DEFAULT
108 #define AIOD_LIFETIME_DEFAULT (30 * hz)
109 #endif
110
111 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
112
113 static int max_aio_procs = MAX_AIO_PROCS;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
115 CTLFLAG_RW, &max_aio_procs, 0,
116 "Maximum number of kernel threads to use for handling async IO ");
117
118 static int num_aio_procs = 0;
119 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
120 CTLFLAG_RD, &num_aio_procs, 0,
121 "Number of presently active kernel threads for async IO");
122
123 /*
124 * The code will adjust the actual number of AIO processes towards this
125 * number when it gets a chance.
126 */
127 static int target_aio_procs = TARGET_AIO_PROCS;
128 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
129 0, "Preferred number of ready kernel threads for async IO");
130
131 static int max_queue_count = MAX_AIO_QUEUE;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
133 "Maximum number of aio requests to queue, globally");
134
135 static int num_queue_count = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
137 "Number of queued aio requests");
138
139 static int num_buf_aio = 0;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
141 "Number of aio requests presently handled by the buf subsystem");
142
143 /* Number of async I/O thread in the process of being started */
144 /* XXX This should be local to _aio_aqueue() */
145 static int num_aio_resv_start = 0;
146
147 static int aiod_timeout;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
149 "Timeout value for synchronous aio operations");
150
151 static int aiod_lifetime;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
153 "Maximum lifetime for idle aiod");
154
155 static int unloadable = 0;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
157 "Allow unload of aio (not recommended)");
158
159
160 static int max_aio_per_proc = MAX_AIO_PER_PROC;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
162 0, "Maximum active aio requests per process (stored in the process)");
163
164 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
166 &max_aio_queue_per_proc, 0,
167 "Maximum queued aio requests per process (stored in the process)");
168
169 static int max_buf_aio = MAX_BUF_AIO;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
171 "Maximum buf aio requests per process (stored in the process)");
172
173 struct aiocblist {
174 TAILQ_ENTRY(aiocblist) list; /* List of jobs */
175 TAILQ_ENTRY(aiocblist) plist; /* List of jobs for proc */
176 int jobflags;
177 int jobstate;
178 int inputcharge;
179 int outputcharge;
180 struct buf *bp; /* Buffer pointer */
181 struct proc *userproc; /* User process */ /* Not td! */
182 struct ucred *cred; /* Active credential when created */
183 struct file *fd_file; /* Pointer to file structure */
184 struct aio_liojob *lio; /* Optional lio job */
185 struct aiocb *uuaiocb; /* Pointer in userspace of aiocb */
186 struct knlist klist; /* list of knotes */
187 struct aiocb uaiocb; /* Kernel I/O control block */
188 };
189
190 /* jobflags */
191 #define AIOCBLIST_RUNDOWN 0x4
192 #define AIOCBLIST_DONE 0x10
193
194 /*
195 * AIO process info
196 */
197 #define AIOP_FREE 0x1 /* proc on free queue */
198
199 struct aiothreadlist {
200 int aiothreadflags; /* AIO proc flags */
201 TAILQ_ENTRY(aiothreadlist) list; /* List of processes */
202 struct thread *aiothread; /* The AIO thread */
203 };
204
205 /*
206 * data-structure for lio signal management
207 */
208 struct aio_liojob {
209 int lioj_flags;
210 int lioj_buffer_count;
211 int lioj_buffer_finished_count;
212 int lioj_queue_count;
213 int lioj_queue_finished_count;
214 struct sigevent lioj_signal; /* signal on all I/O done */
215 TAILQ_ENTRY(aio_liojob) lioj_list;
216 };
217 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
218 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
219
220 /*
221 * per process aio data structure
222 */
223 struct kaioinfo {
224 int kaio_flags; /* per process kaio flags */
225 int kaio_maxactive_count; /* maximum number of AIOs */
226 int kaio_active_count; /* number of currently used AIOs */
227 int kaio_qallowed_count; /* maxiumu size of AIO queue */
228 int kaio_queue_count; /* size of AIO queue */
229 int kaio_ballowed_count; /* maximum number of buffers */
230 int kaio_queue_finished_count; /* number of daemon jobs finished */
231 int kaio_buffer_count; /* number of physio buffers */
232 int kaio_buffer_finished_count; /* count of I/O done */
233 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
234 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
235 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
236 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
237 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
238 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
239 };
240
241 #define KAIO_RUNDOWN 0x1 /* process is being run down */
242 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
243
244 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* Idle daemons */
245 static struct mtx aio_freeproc_mtx;
246
247 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
248
249 static void aio_init_aioinfo(struct proc *p);
250 static void aio_onceonly(void);
251 static int aio_free_entry(struct aiocblist *aiocbe);
252 static void aio_process(struct aiocblist *aiocbe);
253 static int aio_newproc(void);
254 static int aio_aqueue(struct thread *td, struct aiocb *job, int type);
255 static void aio_physwakeup(struct buf *bp);
256 static void aio_proc_rundown(void *arg, struct proc *p);
257 static int aio_fphysio(struct aiocblist *aiocbe);
258 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
259 static void aio_daemon(void *uproc);
260 static void aio_swake_cb(struct socket *, struct sockbuf *);
261 static int aio_unload(void);
262 static int filt_aioattach(struct knote *kn);
263 static void filt_aiodetach(struct knote *kn);
264 static int filt_aio(struct knote *kn, long hint);
265
266 /*
267 * Zones for:
268 * kaio Per process async io info
269 * aiop async io thread data
270 * aiocb async io jobs
271 * aiol list io job pointer - internal to aio_suspend XXX
272 * aiolio list io jobs
273 */
274 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
275
276 /* kqueue filters for aio */
277 static struct filterops aio_filtops =
278 { 0, filt_aioattach, filt_aiodetach, filt_aio };
279
280 static eventhandler_tag exit_tag, exec_tag;
281
282 /*
283 * Main operations function for use as a kernel module.
284 */
285 static int
286 aio_modload(struct module *module, int cmd, void *arg)
287 {
288 int error = 0;
289
290 switch (cmd) {
291 case MOD_LOAD:
292 aio_onceonly();
293 break;
294 case MOD_UNLOAD:
295 error = aio_unload();
296 break;
297 case MOD_SHUTDOWN:
298 break;
299 default:
300 error = EINVAL;
301 break;
302 }
303 return (error);
304 }
305
306 static moduledata_t aio_mod = {
307 "aio",
308 &aio_modload,
309 NULL
310 };
311
312 SYSCALL_MODULE_HELPER(aio_return);
313 SYSCALL_MODULE_HELPER(aio_suspend);
314 SYSCALL_MODULE_HELPER(aio_cancel);
315 SYSCALL_MODULE_HELPER(aio_error);
316 SYSCALL_MODULE_HELPER(aio_read);
317 SYSCALL_MODULE_HELPER(aio_write);
318 SYSCALL_MODULE_HELPER(aio_waitcomplete);
319 SYSCALL_MODULE_HELPER(lio_listio);
320
321 DECLARE_MODULE(aio, aio_mod,
322 SI_SUB_VFS, SI_ORDER_ANY);
323 MODULE_VERSION(aio, 1);
324
325 /*
326 * Startup initialization
327 */
328 static void
329 aio_onceonly(void)
330 {
331
332 /* XXX: should probably just use so->callback */
333 aio_swake = &aio_swake_cb;
334 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
335 EVENTHANDLER_PRI_ANY);
336 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown, NULL,
337 EVENTHANDLER_PRI_ANY);
338 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
339 TAILQ_INIT(&aio_freeproc);
340 mtx_init(&aio_freeproc_mtx, "aio_freeproc", NULL, MTX_DEF);
341 TAILQ_INIT(&aio_jobs);
342 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
343 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
344 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
345 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
346 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
347 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
348 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
349 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
350 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL,
351 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
352 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
353 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
354 jobrefid = 1;
355 async_io_version = _POSIX_VERSION;
356 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
357 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
358 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
359 }
360
361 /*
362 * Callback for unload of AIO when used as a module.
363 */
364 static int
365 aio_unload(void)
366 {
367 int error;
368
369 /*
370 * XXX: no unloads by default, it's too dangerous.
371 * perhaps we could do it if locked out callers and then
372 * did an aio_proc_rundown() on each process.
373 */
374 if (!unloadable)
375 return (EOPNOTSUPP);
376
377 error = kqueue_del_filteropts(EVFILT_AIO);
378 if (error)
379 return error;
380
381 async_io_version = 0;
382 aio_swake = NULL;
383 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
384 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
385 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
386 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
387 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
388 return (0);
389 }
390
391 /*
392 * Init the per-process aioinfo structure. The aioinfo limits are set
393 * per-process for user limit (resource) management.
394 */
395 static void
396 aio_init_aioinfo(struct proc *p)
397 {
398 struct kaioinfo *ki;
399
400 ki = uma_zalloc(kaio_zone, M_WAITOK);
401 ki->kaio_flags = 0;
402 ki->kaio_maxactive_count = max_aio_per_proc;
403 ki->kaio_active_count = 0;
404 ki->kaio_qallowed_count = max_aio_queue_per_proc;
405 ki->kaio_queue_count = 0;
406 ki->kaio_ballowed_count = max_buf_aio;
407 ki->kaio_buffer_count = 0;
408 ki->kaio_buffer_finished_count = 0;
409 TAILQ_INIT(&ki->kaio_jobdone);
410 TAILQ_INIT(&ki->kaio_jobqueue);
411 TAILQ_INIT(&ki->kaio_bufdone);
412 TAILQ_INIT(&ki->kaio_bufqueue);
413 TAILQ_INIT(&ki->kaio_liojoblist);
414 TAILQ_INIT(&ki->kaio_sockqueue);
415 PROC_LOCK(p);
416 if (p->p_aioinfo == NULL) {
417 p->p_aioinfo = ki;
418 PROC_UNLOCK(p);
419 } else {
420 PROC_UNLOCK(p);
421 uma_zfree(kaio_zone, ki);
422 }
423
424 while (num_aio_procs < target_aio_procs)
425 aio_newproc();
426 }
427
428 /*
429 * Free a job entry. Wait for completion if it is currently active, but don't
430 * delay forever. If we delay, we return a flag that says that we have to
431 * restart the queue scan.
432 */
433 static int
434 aio_free_entry(struct aiocblist *aiocbe)
435 {
436 struct kaioinfo *ki;
437 struct aio_liojob *lj;
438 struct proc *p;
439 int error;
440 int s;
441
442 if (aiocbe->jobstate == JOBST_NULL)
443 panic("aio_free_entry: freeing already free job");
444
445 p = aiocbe->userproc;
446 ki = p->p_aioinfo;
447 lj = aiocbe->lio;
448 if (ki == NULL)
449 panic("aio_free_entry: missing p->p_aioinfo");
450
451 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
452 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
453 tsleep(aiocbe, PRIBIO, "jobwai", 0);
454 }
455 if (aiocbe->bp == NULL) {
456 if (ki->kaio_queue_count <= 0)
457 panic("aio_free_entry: process queue size <= 0");
458 if (num_queue_count <= 0)
459 panic("aio_free_entry: system wide queue size <= 0");
460
461 if (lj) {
462 lj->lioj_queue_count--;
463 if (aiocbe->jobflags & AIOCBLIST_DONE)
464 lj->lioj_queue_finished_count--;
465 }
466 ki->kaio_queue_count--;
467 if (aiocbe->jobflags & AIOCBLIST_DONE)
468 ki->kaio_queue_finished_count--;
469 num_queue_count--;
470 } else {
471 if (lj) {
472 lj->lioj_buffer_count--;
473 if (aiocbe->jobflags & AIOCBLIST_DONE)
474 lj->lioj_buffer_finished_count--;
475 }
476 if (aiocbe->jobflags & AIOCBLIST_DONE)
477 ki->kaio_buffer_finished_count--;
478 ki->kaio_buffer_count--;
479 num_buf_aio--;
480 }
481
482 /* aiocbe is going away, we need to destroy any knotes */
483 /* XXXKSE Note the thread here is used to eventually find the
484 * owning process again, but it is also used to do a fo_close
485 * and that requires the thread. (but does it require the
486 * OWNING thread? (or maybe the running thread?)
487 * There is a semantic problem here...
488 */
489 knlist_delete(&aiocbe->klist, FIRST_THREAD_IN_PROC(p), 0); /* XXXKSE */
490
491 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
492 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
493 ki->kaio_flags &= ~KAIO_WAKEUP;
494 wakeup(p);
495 }
496
497 if (aiocbe->jobstate == JOBST_JOBQBUF) {
498 if ((error = aio_fphysio(aiocbe)) != 0)
499 return (error);
500 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
501 panic("aio_free_entry: invalid physio finish-up state");
502 s = splbio();
503 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
504 splx(s);
505 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
506 s = splnet();
507 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
508 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
509 splx(s);
510 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
511 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
512 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
513 s = splbio();
514 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
515 splx(s);
516 if (aiocbe->bp) {
517 vunmapbuf(aiocbe->bp);
518 relpbuf(aiocbe->bp, NULL);
519 aiocbe->bp = NULL;
520 }
521 }
522 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
523 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
524 uma_zfree(aiolio_zone, lj);
525 }
526 aiocbe->jobstate = JOBST_NULL;
527 fdrop(aiocbe->fd_file, curthread);
528 crfree(aiocbe->cred);
529 uma_zfree(aiocb_zone, aiocbe);
530 return (0);
531 }
532
533 /*
534 * Rundown the jobs for a given process.
535 */
536 static void
537 aio_proc_rundown(void *arg, struct proc *p)
538 {
539 int s;
540 struct kaioinfo *ki;
541 struct aio_liojob *lj, *ljn;
542 struct aiocblist *aiocbe, *aiocbn;
543 struct file *fp;
544 struct socket *so;
545
546 ki = p->p_aioinfo;
547 if (ki == NULL)
548 return;
549
550 mtx_lock(&Giant);
551 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
552 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
553 ki->kaio_buffer_finished_count)) {
554 ki->kaio_flags |= KAIO_RUNDOWN;
555 if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
556 break;
557 }
558
559 /*
560 * Move any aio ops that are waiting on socket I/O to the normal job
561 * queues so they are cleaned up with any others.
562 */
563 s = splnet();
564 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
565 aiocbn) {
566 aiocbn = TAILQ_NEXT(aiocbe, plist);
567 fp = aiocbe->fd_file;
568 if (fp != NULL) {
569 so = fp->f_data;
570 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
571 if (TAILQ_EMPTY(&so->so_aiojobq)) {
572 SOCKBUF_LOCK(&so->so_snd);
573 so->so_snd.sb_flags &= ~SB_AIO;
574 SOCKBUF_UNLOCK(&so->so_snd);
575 SOCKBUF_LOCK(&so->so_rcv);
576 so->so_rcv.sb_flags &= ~SB_AIO;
577 SOCKBUF_UNLOCK(&so->so_rcv);
578 }
579 }
580 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
581 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
582 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
583 }
584 splx(s);
585
586 restart1:
587 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
588 aiocbn = TAILQ_NEXT(aiocbe, plist);
589 if (aio_free_entry(aiocbe))
590 goto restart1;
591 }
592
593 restart2:
594 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
595 aiocbn) {
596 aiocbn = TAILQ_NEXT(aiocbe, plist);
597 if (aio_free_entry(aiocbe))
598 goto restart2;
599 }
600
601 /*
602 * Note the use of lots of splbio here, trying to avoid splbio for long chains
603 * of I/O. Probably unnecessary.
604 */
605 restart3:
606 s = splbio();
607 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
608 ki->kaio_flags |= KAIO_WAKEUP;
609 tsleep(p, PRIBIO, "aioprn", 0);
610 splx(s);
611 goto restart3;
612 }
613 splx(s);
614
615 restart4:
616 s = splbio();
617 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
618 aiocbn = TAILQ_NEXT(aiocbe, plist);
619 if (aio_free_entry(aiocbe)) {
620 splx(s);
621 goto restart4;
622 }
623 }
624 splx(s);
625
626 /*
627 * If we've slept, jobs might have moved from one queue to another.
628 * Retry rundown if we didn't manage to empty the queues.
629 */
630 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
631 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
632 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
633 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
634 goto restart1;
635
636 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
637 ljn = TAILQ_NEXT(lj, lioj_list);
638 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
639 0)) {
640 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
641 uma_zfree(aiolio_zone, lj);
642 } else {
643 #ifdef DIAGNOSTIC
644 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
645 "QF:%d\n", lj->lioj_buffer_count,
646 lj->lioj_buffer_finished_count,
647 lj->lioj_queue_count,
648 lj->lioj_queue_finished_count);
649 #endif
650 }
651 }
652
653 uma_zfree(kaio_zone, ki);
654 p->p_aioinfo = NULL;
655 mtx_unlock(&Giant);
656 }
657
658 /*
659 * Select a job to run (called by an AIO daemon).
660 */
661 static struct aiocblist *
662 aio_selectjob(struct aiothreadlist *aiop)
663 {
664 int s;
665 struct aiocblist *aiocbe;
666 struct kaioinfo *ki;
667 struct proc *userp;
668
669 s = splnet();
670 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
671 TAILQ_NEXT(aiocbe, list)) {
672 userp = aiocbe->userproc;
673 ki = userp->p_aioinfo;
674
675 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
676 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
677 splx(s);
678 return (aiocbe);
679 }
680 }
681 splx(s);
682
683 return (NULL);
684 }
685
686 /*
687 * The AIO processing activity. This is the code that does the I/O request for
688 * the non-physio version of the operations. The normal vn operations are used,
689 * and this code should work in all instances for every type of file, including
690 * pipes, sockets, fifos, and regular files.
691 */
692 static void
693 aio_process(struct aiocblist *aiocbe)
694 {
695 struct ucred *td_savedcred;
696 struct thread *td;
697 struct proc *mycp;
698 struct aiocb *cb;
699 struct file *fp;
700 struct uio auio;
701 struct iovec aiov;
702 int cnt;
703 int error;
704 int oublock_st, oublock_end;
705 int inblock_st, inblock_end;
706
707 td = curthread;
708 td_savedcred = td->td_ucred;
709 td->td_ucred = aiocbe->cred;
710 mycp = td->td_proc;
711 cb = &aiocbe->uaiocb;
712 fp = aiocbe->fd_file;
713
714 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
715 aiov.iov_len = cb->aio_nbytes;
716
717 auio.uio_iov = &aiov;
718 auio.uio_iovcnt = 1;
719 auio.uio_offset = cb->aio_offset;
720 auio.uio_resid = cb->aio_nbytes;
721 cnt = cb->aio_nbytes;
722 auio.uio_segflg = UIO_USERSPACE;
723 auio.uio_td = td;
724
725 inblock_st = mycp->p_stats->p_ru.ru_inblock;
726 oublock_st = mycp->p_stats->p_ru.ru_oublock;
727 /*
728 * _aio_aqueue() acquires a reference to the file that is
729 * released in aio_free_entry().
730 */
731 if (cb->aio_lio_opcode == LIO_READ) {
732 auio.uio_rw = UIO_READ;
733 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
734 } else {
735 auio.uio_rw = UIO_WRITE;
736 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
737 }
738 inblock_end = mycp->p_stats->p_ru.ru_inblock;
739 oublock_end = mycp->p_stats->p_ru.ru_oublock;
740
741 aiocbe->inputcharge = inblock_end - inblock_st;
742 aiocbe->outputcharge = oublock_end - oublock_st;
743
744 if ((error) && (auio.uio_resid != cnt)) {
745 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
746 error = 0;
747 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
748 PROC_LOCK(aiocbe->userproc);
749 psignal(aiocbe->userproc, SIGPIPE);
750 PROC_UNLOCK(aiocbe->userproc);
751 }
752 }
753
754 cnt -= auio.uio_resid;
755 cb->_aiocb_private.error = error;
756 cb->_aiocb_private.status = cnt;
757 td->td_ucred = td_savedcred;
758 }
759
760 /*
761 * The AIO daemon, most of the actual work is done in aio_process,
762 * but the setup (and address space mgmt) is done in this routine.
763 */
764 static void
765 aio_daemon(void *uproc)
766 {
767 int s;
768 struct aio_liojob *lj;
769 struct aiocb *cb;
770 struct aiocblist *aiocbe;
771 struct aiothreadlist *aiop;
772 struct kaioinfo *ki;
773 struct proc *curcp, *mycp, *userp;
774 struct vmspace *myvm, *tmpvm;
775 struct thread *td = curthread;
776 struct pgrp *newpgrp;
777 struct session *newsess;
778
779 /*
780 * Local copies of curproc (cp) and vmspace (myvm)
781 */
782 mycp = td->td_proc;
783 myvm = mycp->p_vmspace;
784
785 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
786
787 /*
788 * Allocate and ready the aio control info. There is one aiop structure
789 * per daemon.
790 */
791 aiop = uma_zalloc(aiop_zone, M_WAITOK);
792 aiop->aiothread = td;
793 aiop->aiothreadflags |= AIOP_FREE;
794
795 /*
796 * Place thread (lightweight process) onto the AIO free thread list.
797 */
798 mtx_lock(&aio_freeproc_mtx);
799 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
800 mtx_unlock(&aio_freeproc_mtx);
801
802 /*
803 * Get rid of our current filedescriptors. AIOD's don't need any
804 * filedescriptors, except as temporarily inherited from the client.
805 */
806 mtx_lock(&Giant);
807 fdfree(td);
808
809 mtx_unlock(&Giant);
810 /* The daemon resides in its own pgrp. */
811 MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
812 M_WAITOK | M_ZERO);
813 MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
814 M_WAITOK | M_ZERO);
815
816 sx_xlock(&proctree_lock);
817 enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
818 sx_xunlock(&proctree_lock);
819 mtx_lock(&Giant);
820
821 /*
822 * Wakeup parent process. (Parent sleeps to keep from blasting away
823 * and creating too many daemons.)
824 */
825 wakeup(mycp);
826
827 for (;;) {
828 /*
829 * curcp is the current daemon process context.
830 * userp is the current user process context.
831 */
832 curcp = mycp;
833
834 /*
835 * Take daemon off of free queue
836 */
837 mtx_lock(&aio_freeproc_mtx);
838 if (aiop->aiothreadflags & AIOP_FREE) {
839 TAILQ_REMOVE(&aio_freeproc, aiop, list);
840 aiop->aiothreadflags &= ~AIOP_FREE;
841 }
842 mtx_unlock(&aio_freeproc_mtx);
843
844 /*
845 * Check for jobs.
846 */
847 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
848 cb = &aiocbe->uaiocb;
849 userp = aiocbe->userproc;
850
851 aiocbe->jobstate = JOBST_JOBRUNNING;
852
853 /*
854 * Connect to process address space for user program.
855 */
856 if (userp != curcp) {
857 /*
858 * Save the current address space that we are
859 * connected to.
860 */
861 tmpvm = mycp->p_vmspace;
862
863 /*
864 * Point to the new user address space, and
865 * refer to it.
866 */
867 mycp->p_vmspace = userp->p_vmspace;
868 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
869
870 /* Activate the new mapping. */
871 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
872
873 /*
874 * If the old address space wasn't the daemons
875 * own address space, then we need to remove the
876 * daemon's reference from the other process
877 * that it was acting on behalf of.
878 */
879 if (tmpvm != myvm) {
880 vmspace_free(tmpvm);
881 }
882 curcp = userp;
883 }
884
885 ki = userp->p_aioinfo;
886 lj = aiocbe->lio;
887
888 /* Account for currently active jobs. */
889 ki->kaio_active_count++;
890
891 /* Do the I/O function. */
892 aio_process(aiocbe);
893
894 /* Decrement the active job count. */
895 ki->kaio_active_count--;
896
897 /*
898 * Increment the completion count for wakeup/signal
899 * comparisons.
900 */
901 aiocbe->jobflags |= AIOCBLIST_DONE;
902 ki->kaio_queue_finished_count++;
903 if (lj)
904 lj->lioj_queue_finished_count++;
905 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
906 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
907 ki->kaio_flags &= ~KAIO_WAKEUP;
908 wakeup(userp);
909 }
910
911 s = splbio();
912 if (lj && (lj->lioj_flags &
913 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
914 if ((lj->lioj_queue_finished_count ==
915 lj->lioj_queue_count) &&
916 (lj->lioj_buffer_finished_count ==
917 lj->lioj_buffer_count)) {
918 PROC_LOCK(userp);
919 psignal(userp,
920 lj->lioj_signal.sigev_signo);
921 PROC_UNLOCK(userp);
922 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
923 }
924 }
925 splx(s);
926
927 aiocbe->jobstate = JOBST_JOBFINISHED;
928
929 s = splnet();
930 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
931 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
932 splx(s);
933 KNOTE_UNLOCKED(&aiocbe->klist, 0);
934
935 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
936 wakeup(aiocbe);
937 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
938 }
939
940 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
941 PROC_LOCK(userp);
942 psignal(userp, cb->aio_sigevent.sigev_signo);
943 PROC_UNLOCK(userp);
944 }
945 }
946
947 /*
948 * Disconnect from user address space.
949 */
950 if (curcp != mycp) {
951 /* Get the user address space to disconnect from. */
952 tmpvm = mycp->p_vmspace;
953
954 /* Get original address space for daemon. */
955 mycp->p_vmspace = myvm;
956
957 /* Activate the daemon's address space. */
958 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
959 #ifdef DIAGNOSTIC
960 if (tmpvm == myvm) {
961 printf("AIOD: vmspace problem -- %d\n",
962 mycp->p_pid);
963 }
964 #endif
965 /* Remove our vmspace reference. */
966 vmspace_free(tmpvm);
967
968 curcp = mycp;
969 }
970
971 mtx_lock(&aio_freeproc_mtx);
972 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
973 aiop->aiothreadflags |= AIOP_FREE;
974
975 /*
976 * If daemon is inactive for a long time, allow it to exit,
977 * thereby freeing resources.
978 */
979 if (msleep(aiop->aiothread, &aio_freeproc_mtx, PDROP | PRIBIO,
980 "aiordy", aiod_lifetime)) {
981 s = splnet();
982 if (TAILQ_EMPTY(&aio_jobs)) {
983 mtx_lock(&aio_freeproc_mtx);
984 if ((aiop->aiothreadflags & AIOP_FREE) &&
985 (num_aio_procs > target_aio_procs)) {
986 TAILQ_REMOVE(&aio_freeproc, aiop, list);
987 mtx_unlock(&aio_freeproc_mtx);
988 splx(s);
989 uma_zfree(aiop_zone, aiop);
990 num_aio_procs--;
991 #ifdef DIAGNOSTIC
992 if (mycp->p_vmspace->vm_refcnt <= 1) {
993 printf("AIOD: bad vm refcnt for"
994 " exiting daemon: %d\n",
995 mycp->p_vmspace->vm_refcnt);
996 }
997 #endif
998 kthread_exit(0);
999 }
1000 mtx_unlock(&aio_freeproc_mtx);
1001 }
1002 splx(s);
1003 }
1004 }
1005 }
1006
1007 /*
1008 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1009 * AIO daemon modifies its environment itself.
1010 */
1011 static int
1012 aio_newproc(void)
1013 {
1014 int error;
1015 struct proc *p;
1016
1017 error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, 0, "aiod%d",
1018 num_aio_procs);
1019 if (error)
1020 return (error);
1021
1022 /*
1023 * Wait until daemon is started, but continue on just in case to
1024 * handle error conditions.
1025 */
1026 error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1027
1028 num_aio_procs++;
1029
1030 return (error);
1031 }
1032
1033 /*
1034 * Try the high-performance, low-overhead physio method for eligible
1035 * VCHR devices. This method doesn't use an aio helper thread, and
1036 * thus has very low overhead.
1037 *
1038 * Assumes that the caller, _aio_aqueue(), has incremented the file
1039 * structure's reference count, preventing its deallocation for the
1040 * duration of this call.
1041 */
1042 static int
1043 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1044 {
1045 int error;
1046 struct aiocb *cb;
1047 struct file *fp;
1048 struct buf *bp;
1049 struct vnode *vp;
1050 struct kaioinfo *ki;
1051 struct aio_liojob *lj;
1052 int s;
1053 int notify;
1054
1055 cb = &aiocbe->uaiocb;
1056 fp = aiocbe->fd_file;
1057
1058 if (fp->f_type != DTYPE_VNODE)
1059 return (-1);
1060
1061 vp = fp->f_vnode;
1062
1063 /*
1064 * If its not a disk, we don't want to return a positive error.
1065 * It causes the aio code to not fall through to try the thread
1066 * way when you're talking to a regular file.
1067 */
1068 if (!vn_isdisk(vp, &error)) {
1069 if (error == ENOTBLK)
1070 return (-1);
1071 else
1072 return (error);
1073 }
1074
1075 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1076 return (-1);
1077
1078 if (cb->aio_nbytes >
1079 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1080 return (-1);
1081
1082 ki = p->p_aioinfo;
1083 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1084 return (-1);
1085
1086 ki->kaio_buffer_count++;
1087
1088 lj = aiocbe->lio;
1089 if (lj)
1090 lj->lioj_buffer_count++;
1091
1092 /* Create and build a buffer header for a transfer. */
1093 bp = (struct buf *)getpbuf(NULL);
1094 BUF_KERNPROC(bp);
1095
1096 /*
1097 * Get a copy of the kva from the physical buffer.
1098 */
1099 error = 0;
1100
1101 bp->b_bcount = cb->aio_nbytes;
1102 bp->b_bufsize = cb->aio_nbytes;
1103 bp->b_iodone = aio_physwakeup;
1104 bp->b_saveaddr = bp->b_data;
1105 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1106 bp->b_offset = cb->aio_offset;
1107 bp->b_iooffset = cb->aio_offset;
1108 bp->b_blkno = btodb(cb->aio_offset);
1109 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1110
1111 /*
1112 * Bring buffer into kernel space.
1113 */
1114 if (vmapbuf(bp) < 0) {
1115 error = EFAULT;
1116 goto doerror;
1117 }
1118
1119 s = splbio();
1120 aiocbe->bp = bp;
1121 bp->b_caller1 = (void *)aiocbe;
1122 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1123 aiocbe->jobstate = JOBST_JOBQBUF;
1124 cb->_aiocb_private.status = cb->aio_nbytes;
1125 num_buf_aio++;
1126 bp->b_error = 0;
1127
1128 splx(s);
1129
1130 /* Perform transfer. */
1131 dev_strategy(vp->v_rdev, bp);
1132
1133 notify = 0;
1134 s = splbio();
1135
1136 /*
1137 * If we had an error invoking the request, or an error in processing
1138 * the request before we have returned, we process it as an error in
1139 * transfer. Note that such an I/O error is not indicated immediately,
1140 * but is returned using the aio_error mechanism. In this case,
1141 * aio_suspend will return immediately.
1142 */
1143 if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1144 struct aiocb *job = aiocbe->uuaiocb;
1145
1146 aiocbe->uaiocb._aiocb_private.status = 0;
1147 suword(&job->_aiocb_private.status, 0);
1148 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1149 suword(&job->_aiocb_private.error, bp->b_error);
1150
1151 ki->kaio_buffer_finished_count++;
1152
1153 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1154 aiocbe->jobstate = JOBST_JOBBFINISHED;
1155 aiocbe->jobflags |= AIOCBLIST_DONE;
1156 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1157 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1158 notify = 1;
1159 }
1160 }
1161 splx(s);
1162 if (notify)
1163 KNOTE_UNLOCKED(&aiocbe->klist, 0);
1164 return (0);
1165
1166 doerror:
1167 ki->kaio_buffer_count--;
1168 if (lj)
1169 lj->lioj_buffer_count--;
1170 aiocbe->bp = NULL;
1171 relpbuf(bp, NULL);
1172 return (error);
1173 }
1174
1175 /*
1176 * This waits/tests physio completion.
1177 */
1178 static int
1179 aio_fphysio(struct aiocblist *iocb)
1180 {
1181 int s;
1182 struct buf *bp;
1183 int error;
1184
1185 bp = iocb->bp;
1186
1187 s = splbio();
1188 while ((bp->b_flags & B_DONE) == 0) {
1189 if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1190 if ((bp->b_flags & B_DONE) == 0) {
1191 splx(s);
1192 return (EINPROGRESS);
1193 } else
1194 break;
1195 }
1196 }
1197 splx(s);
1198
1199 /* Release mapping into kernel space. */
1200 vunmapbuf(bp);
1201 iocb->bp = 0;
1202
1203 error = 0;
1204
1205 /* Check for an error. */
1206 if (bp->b_ioflags & BIO_ERROR)
1207 error = bp->b_error;
1208
1209 relpbuf(bp, NULL);
1210 return (error);
1211 }
1212
1213 /*
1214 * Wake up aio requests that may be serviceable now.
1215 */
1216 static void
1217 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1218 {
1219 struct aiocblist *cb,*cbn;
1220 struct proc *p;
1221 struct kaioinfo *ki = NULL;
1222 int opcode, wakecount = 0;
1223 struct aiothreadlist *aiop;
1224
1225 if (sb == &so->so_snd) {
1226 opcode = LIO_WRITE;
1227 SOCKBUF_LOCK(&so->so_snd);
1228 so->so_snd.sb_flags &= ~SB_AIO;
1229 SOCKBUF_UNLOCK(&so->so_snd);
1230 } else {
1231 opcode = LIO_READ;
1232 SOCKBUF_LOCK(&so->so_rcv);
1233 so->so_rcv.sb_flags &= ~SB_AIO;
1234 SOCKBUF_UNLOCK(&so->so_rcv);
1235 }
1236
1237 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1238 cbn = TAILQ_NEXT(cb, list);
1239 if (opcode == cb->uaiocb.aio_lio_opcode) {
1240 p = cb->userproc;
1241 ki = p->p_aioinfo;
1242 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1243 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1244 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1245 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1246 wakecount++;
1247 if (cb->jobstate != JOBST_JOBQGLOBAL)
1248 panic("invalid queue value");
1249 }
1250 }
1251
1252 while (wakecount--) {
1253 mtx_lock(&aio_freeproc_mtx);
1254 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1255 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1256 aiop->aiothreadflags &= ~AIOP_FREE;
1257 wakeup(aiop->aiothread);
1258 }
1259 mtx_unlock(&aio_freeproc_mtx);
1260 }
1261 }
1262
1263 /*
1264 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1265 * technique is done in this code.
1266 */
1267 static int
1268 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1269 {
1270 struct proc *p = td->td_proc;
1271 struct filedesc *fdp;
1272 struct file *fp;
1273 unsigned int fd;
1274 struct socket *so;
1275 int s;
1276 int error;
1277 int opcode;
1278 struct aiocblist *aiocbe;
1279 struct aiothreadlist *aiop;
1280 struct kaioinfo *ki;
1281 struct kevent kev;
1282 struct kqueue *kq;
1283 struct file *kq_fp;
1284 struct sockbuf *sb;
1285
1286 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1287 aiocbe->inputcharge = 0;
1288 aiocbe->outputcharge = 0;
1289 /* XXX - need a lock */
1290 knlist_init(&aiocbe->klist, NULL, NULL, NULL, NULL);
1291
1292 suword(&job->_aiocb_private.status, -1);
1293 suword(&job->_aiocb_private.error, 0);
1294 suword(&job->_aiocb_private.kernelinfo, -1);
1295
1296 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1297 if (error) {
1298 suword(&job->_aiocb_private.error, error);
1299 uma_zfree(aiocb_zone, aiocbe);
1300 return (error);
1301 }
1302 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1303 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1304 uma_zfree(aiocb_zone, aiocbe);
1305 return (EINVAL);
1306 }
1307
1308 /* Save userspace address of the job info. */
1309 aiocbe->uuaiocb = job;
1310
1311 /* Get the opcode. */
1312 if (type != LIO_NOP)
1313 aiocbe->uaiocb.aio_lio_opcode = type;
1314 opcode = aiocbe->uaiocb.aio_lio_opcode;
1315
1316 /* Get the fd info for process. */
1317 fdp = p->p_fd;
1318
1319 /*
1320 * Range check file descriptor.
1321 */
1322 FILEDESC_LOCK(fdp);
1323 fd = aiocbe->uaiocb.aio_fildes;
1324 if (fd >= fdp->fd_nfiles) {
1325 FILEDESC_UNLOCK(fdp);
1326 uma_zfree(aiocb_zone, aiocbe);
1327 if (type == 0)
1328 suword(&job->_aiocb_private.error, EBADF);
1329 return (EBADF);
1330 }
1331
1332 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1333 if ((fp == NULL) ||
1334 ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 0)) ||
1335 ((opcode == LIO_READ) && ((fp->f_flag & FREAD) == 0))) {
1336 FILEDESC_UNLOCK(fdp);
1337 uma_zfree(aiocb_zone, aiocbe);
1338 if (type == 0)
1339 suword(&job->_aiocb_private.error, EBADF);
1340 return (EBADF);
1341 }
1342 fhold(fp);
1343 FILEDESC_UNLOCK(fdp);
1344
1345 if (aiocbe->uaiocb.aio_offset == -1LL) {
1346 error = EINVAL;
1347 goto aqueue_fail;
1348 }
1349 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1350 if (error) {
1351 error = EINVAL;
1352 goto aqueue_fail;
1353 }
1354 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1355 if (jobrefid == LONG_MAX)
1356 jobrefid = 1;
1357 else
1358 jobrefid++;
1359
1360 if (opcode == LIO_NOP) {
1361 fdrop(fp, td);
1362 uma_zfree(aiocb_zone, aiocbe);
1363 if (type == 0) {
1364 suword(&job->_aiocb_private.error, 0);
1365 suword(&job->_aiocb_private.status, 0);
1366 suword(&job->_aiocb_private.kernelinfo, 0);
1367 }
1368 return (0);
1369 }
1370 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1371 if (type == 0)
1372 suword(&job->_aiocb_private.status, 0);
1373 error = EINVAL;
1374 goto aqueue_fail;
1375 }
1376
1377 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1378 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1379 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1380 } else
1381 goto no_kqueue;
1382 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1383 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1384 (kq_fp->f_type != DTYPE_KQUEUE)) {
1385 error = EBADF;
1386 goto aqueue_fail;
1387 }
1388 kq = kq_fp->f_data;
1389 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1390 kev.filter = EVFILT_AIO;
1391 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1392 kev.data = (intptr_t)aiocbe;
1393 error = kqueue_register(kq, &kev, td, 1);
1394 aqueue_fail:
1395 if (error) {
1396 fdrop(fp, td);
1397 uma_zfree(aiocb_zone, aiocbe);
1398 if (type == 0)
1399 suword(&job->_aiocb_private.error, error);
1400 goto done;
1401 }
1402 no_kqueue:
1403
1404 suword(&job->_aiocb_private.error, EINPROGRESS);
1405 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1406 aiocbe->userproc = p;
1407 aiocbe->cred = crhold(td->td_ucred);
1408 aiocbe->jobflags = 0;
1409 aiocbe->lio = lj;
1410 ki = p->p_aioinfo;
1411
1412 if (fp->f_type == DTYPE_SOCKET) {
1413 /*
1414 * Alternate queueing for socket ops: Reach down into the
1415 * descriptor to get the socket data. Then check to see if the
1416 * socket is ready to be read or written (based on the requested
1417 * operation).
1418 *
1419 * If it is not ready for io, then queue the aiocbe on the
1420 * socket, and set the flags so we get a call when sbnotify()
1421 * happens.
1422 *
1423 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1424 * and unlock the snd sockbuf for no reason.
1425 */
1426 so = fp->f_data;
1427 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1428 SOCKBUF_LOCK(sb);
1429 s = splnet();
1430 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1431 LIO_WRITE) && (!sowriteable(so)))) {
1432 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1433 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1434 sb->sb_flags |= SB_AIO;
1435 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1436 ki->kaio_queue_count++;
1437 num_queue_count++;
1438 SOCKBUF_UNLOCK(sb);
1439 splx(s);
1440 error = 0;
1441 goto done;
1442 }
1443 SOCKBUF_UNLOCK(sb);
1444 splx(s);
1445 }
1446
1447 if ((error = aio_qphysio(p, aiocbe)) == 0)
1448 goto done;
1449 if (error > 0) {
1450 suword(&job->_aiocb_private.status, 0);
1451 aiocbe->uaiocb._aiocb_private.error = error;
1452 suword(&job->_aiocb_private.error, error);
1453 goto done;
1454 }
1455
1456 /* No buffer for daemon I/O. */
1457 aiocbe->bp = NULL;
1458
1459 ki->kaio_queue_count++;
1460 if (lj)
1461 lj->lioj_queue_count++;
1462 s = splnet();
1463 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1464 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1465 splx(s);
1466 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1467
1468 num_queue_count++;
1469 error = 0;
1470
1471 /*
1472 * If we don't have a free AIO process, and we are below our quota, then
1473 * start one. Otherwise, depend on the subsequent I/O completions to
1474 * pick-up this job. If we don't sucessfully create the new process
1475 * (thread) due to resource issues, we return an error for now (EAGAIN),
1476 * which is likely not the correct thing to do.
1477 */
1478 mtx_lock(&aio_freeproc_mtx);
1479 retryproc:
1480 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1481 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1482 aiop->aiothreadflags &= ~AIOP_FREE;
1483 wakeup(aiop->aiothread);
1484 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1485 ((ki->kaio_active_count + num_aio_resv_start) <
1486 ki->kaio_maxactive_count)) {
1487 num_aio_resv_start++;
1488 mtx_unlock(&aio_freeproc_mtx);
1489 if ((error = aio_newproc()) == 0) {
1490 mtx_lock(&aio_freeproc_mtx);
1491 num_aio_resv_start--;
1492 goto retryproc;
1493 }
1494 mtx_lock(&aio_freeproc_mtx);
1495 num_aio_resv_start--;
1496 }
1497 mtx_unlock(&aio_freeproc_mtx);
1498 done:
1499 return (error);
1500 }
1501
1502 /*
1503 * This routine queues an AIO request, checking for quotas.
1504 */
1505 static int
1506 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1507 {
1508 struct proc *p = td->td_proc;
1509 struct kaioinfo *ki;
1510
1511 if (p->p_aioinfo == NULL)
1512 aio_init_aioinfo(p);
1513
1514 if (num_queue_count >= max_queue_count)
1515 return (EAGAIN);
1516
1517 ki = p->p_aioinfo;
1518 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1519 return (EAGAIN);
1520
1521 return _aio_aqueue(td, job, NULL, type);
1522 }
1523
1524 /*
1525 * Support the aio_return system call, as a side-effect, kernel resources are
1526 * released.
1527 */
1528 int
1529 aio_return(struct thread *td, struct aio_return_args *uap)
1530 {
1531 struct proc *p = td->td_proc;
1532 int s;
1533 long jobref;
1534 struct aiocblist *cb, *ncb;
1535 struct aiocb *ujob;
1536 struct kaioinfo *ki;
1537
1538 ujob = uap->aiocbp;
1539 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1540 if (jobref == -1 || jobref == 0)
1541 return (EINVAL);
1542
1543 ki = p->p_aioinfo;
1544 if (ki == NULL)
1545 return (EINVAL);
1546 PROC_LOCK(p);
1547 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1548 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1549 jobref) {
1550 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1551 p->p_stats->p_ru.ru_oublock +=
1552 cb->outputcharge;
1553 cb->outputcharge = 0;
1554 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1555 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1556 cb->inputcharge = 0;
1557 }
1558 goto done;
1559 }
1560 }
1561 s = splbio();
1562 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1563 ncb = TAILQ_NEXT(cb, plist);
1564 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1565 == jobref) {
1566 break;
1567 }
1568 }
1569 splx(s);
1570 done:
1571 PROC_UNLOCK(p);
1572 if (cb != NULL) {
1573 if (ujob == cb->uuaiocb) {
1574 td->td_retval[0] =
1575 cb->uaiocb._aiocb_private.status;
1576 } else
1577 td->td_retval[0] = EFAULT;
1578 aio_free_entry(cb);
1579 return (0);
1580 }
1581 return (EINVAL);
1582 }
1583
1584 /*
1585 * Allow a process to wakeup when any of the I/O requests are completed.
1586 */
1587 int
1588 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1589 {
1590 struct proc *p = td->td_proc;
1591 struct timeval atv;
1592 struct timespec ts;
1593 struct aiocb *const *cbptr, *cbp;
1594 struct kaioinfo *ki;
1595 struct aiocblist *cb;
1596 int i;
1597 int njoblist;
1598 int error, s, timo;
1599 long *ijoblist;
1600 struct aiocb **ujoblist;
1601
1602 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1603 return (EINVAL);
1604
1605 timo = 0;
1606 if (uap->timeout) {
1607 /* Get timespec struct. */
1608 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1609 return (error);
1610
1611 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1612 return (EINVAL);
1613
1614 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1615 if (itimerfix(&atv))
1616 return (EINVAL);
1617 timo = tvtohz(&atv);
1618 }
1619
1620 ki = p->p_aioinfo;
1621 if (ki == NULL)
1622 return (EAGAIN);
1623
1624 njoblist = 0;
1625 ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1626 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1627 cbptr = uap->aiocbp;
1628
1629 for (i = 0; i < uap->nent; i++) {
1630 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1631 if (cbp == 0)
1632 continue;
1633 ujoblist[njoblist] = cbp;
1634 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1635 njoblist++;
1636 }
1637
1638 if (njoblist == 0) {
1639 uma_zfree(aiol_zone, ijoblist);
1640 uma_zfree(aiol_zone, ujoblist);
1641 return (0);
1642 }
1643
1644 error = 0;
1645 for (;;) {
1646 PROC_LOCK(p);
1647 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1648 for (i = 0; i < njoblist; i++) {
1649 if (((intptr_t)
1650 cb->uaiocb._aiocb_private.kernelinfo) ==
1651 ijoblist[i]) {
1652 PROC_UNLOCK(p);
1653 if (ujoblist[i] != cb->uuaiocb)
1654 error = EINVAL;
1655 uma_zfree(aiol_zone, ijoblist);
1656 uma_zfree(aiol_zone, ujoblist);
1657 return (error);
1658 }
1659 }
1660 }
1661
1662 s = splbio();
1663 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1664 TAILQ_NEXT(cb, plist)) {
1665 for (i = 0; i < njoblist; i++) {
1666 if (((intptr_t)
1667 cb->uaiocb._aiocb_private.kernelinfo) ==
1668 ijoblist[i]) {
1669 PROC_UNLOCK(p);
1670 splx(s);
1671 if (ujoblist[i] != cb->uuaiocb)
1672 error = EINVAL;
1673 uma_zfree(aiol_zone, ijoblist);
1674 uma_zfree(aiol_zone, ujoblist);
1675 return (error);
1676 }
1677 }
1678 }
1679
1680 ki->kaio_flags |= KAIO_WAKEUP;
1681 error = msleep(p, &p->p_mtx, PDROP | PRIBIO | PCATCH, "aiospn",
1682 timo);
1683 splx(s);
1684
1685 if (error == ERESTART || error == EINTR) {
1686 uma_zfree(aiol_zone, ijoblist);
1687 uma_zfree(aiol_zone, ujoblist);
1688 return (EINTR);
1689 } else if (error == EWOULDBLOCK) {
1690 uma_zfree(aiol_zone, ijoblist);
1691 uma_zfree(aiol_zone, ujoblist);
1692 return (EAGAIN);
1693 }
1694 }
1695
1696 /* NOTREACHED */
1697 return (EINVAL);
1698 }
1699
1700 /*
1701 * aio_cancel cancels any non-physio aio operations not currently in
1702 * progress.
1703 */
1704 int
1705 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1706 {
1707 struct proc *p = td->td_proc;
1708 struct kaioinfo *ki;
1709 struct aiocblist *cbe, *cbn;
1710 struct file *fp;
1711 struct filedesc *fdp;
1712 struct socket *so;
1713 struct proc *po;
1714 int s,error;
1715 int cancelled=0;
1716 int notcancelled=0;
1717 struct vnode *vp;
1718
1719 fdp = p->p_fd;
1720 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1721 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1722 return (EBADF);
1723
1724 if (fp->f_type == DTYPE_VNODE) {
1725 vp = fp->f_vnode;
1726
1727 if (vn_isdisk(vp,&error)) {
1728 td->td_retval[0] = AIO_NOTCANCELED;
1729 return (0);
1730 }
1731 } else if (fp->f_type == DTYPE_SOCKET) {
1732 so = fp->f_data;
1733
1734 s = splnet();
1735
1736 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1737 cbn = TAILQ_NEXT(cbe, list);
1738 if ((uap->aiocbp == NULL) ||
1739 (uap->aiocbp == cbe->uuaiocb) ) {
1740 po = cbe->userproc;
1741 ki = po->p_aioinfo;
1742 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1743 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1744 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1745 if (ki->kaio_flags & KAIO_WAKEUP) {
1746 wakeup(po);
1747 }
1748 cbe->jobstate = JOBST_JOBFINISHED;
1749 cbe->uaiocb._aiocb_private.status=-1;
1750 cbe->uaiocb._aiocb_private.error=ECANCELED;
1751 cancelled++;
1752 /* XXX cancelled, knote? */
1753 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1754 SIGEV_SIGNAL) {
1755 PROC_LOCK(cbe->userproc);
1756 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1757 PROC_UNLOCK(cbe->userproc);
1758 }
1759 if (uap->aiocbp)
1760 break;
1761 }
1762 }
1763 splx(s);
1764
1765 if ((cancelled) && (uap->aiocbp)) {
1766 td->td_retval[0] = AIO_CANCELED;
1767 return (0);
1768 }
1769 }
1770 ki=p->p_aioinfo;
1771 if (ki == NULL)
1772 goto done;
1773 s = splnet();
1774
1775 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1776 cbn = TAILQ_NEXT(cbe, plist);
1777
1778 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1779 ((uap->aiocbp == NULL ) ||
1780 (uap->aiocbp == cbe->uuaiocb))) {
1781
1782 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1783 TAILQ_REMOVE(&aio_jobs, cbe, list);
1784 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1785 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1786 plist);
1787 cancelled++;
1788 ki->kaio_queue_finished_count++;
1789 cbe->jobstate = JOBST_JOBFINISHED;
1790 cbe->uaiocb._aiocb_private.status = -1;
1791 cbe->uaiocb._aiocb_private.error = ECANCELED;
1792 /* XXX cancelled, knote? */
1793 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1794 SIGEV_SIGNAL) {
1795 PROC_LOCK(cbe->userproc);
1796 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1797 PROC_UNLOCK(cbe->userproc);
1798 }
1799 } else {
1800 notcancelled++;
1801 }
1802 }
1803 }
1804 splx(s);
1805 done:
1806 if (notcancelled) {
1807 td->td_retval[0] = AIO_NOTCANCELED;
1808 return (0);
1809 }
1810 if (cancelled) {
1811 td->td_retval[0] = AIO_CANCELED;
1812 return (0);
1813 }
1814 td->td_retval[0] = AIO_ALLDONE;
1815
1816 return (0);
1817 }
1818
1819 /*
1820 * aio_error is implemented in the kernel level for compatibility purposes only.
1821 * For a user mode async implementation, it would be best to do it in a userland
1822 * subroutine.
1823 */
1824 int
1825 aio_error(struct thread *td, struct aio_error_args *uap)
1826 {
1827 struct proc *p = td->td_proc;
1828 int s;
1829 struct aiocblist *cb;
1830 struct kaioinfo *ki;
1831 long jobref;
1832
1833 ki = p->p_aioinfo;
1834 if (ki == NULL)
1835 return (EINVAL);
1836
1837 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1838 if ((jobref == -1) || (jobref == 0))
1839 return (EINVAL);
1840
1841 PROC_LOCK(p);
1842 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1843 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1844 jobref) {
1845 PROC_UNLOCK(p);
1846 td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1847 return (0);
1848 }
1849 }
1850
1851 s = splnet();
1852
1853 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1854 plist)) {
1855 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1856 jobref) {
1857 PROC_UNLOCK(p);
1858 td->td_retval[0] = EINPROGRESS;
1859 splx(s);
1860 return (0);
1861 }
1862 }
1863
1864 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1865 plist)) {
1866 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1867 jobref) {
1868 PROC_UNLOCK(p);
1869 td->td_retval[0] = EINPROGRESS;
1870 splx(s);
1871 return (0);
1872 }
1873 }
1874 splx(s);
1875
1876 s = splbio();
1877 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1878 plist)) {
1879 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1880 jobref) {
1881 PROC_UNLOCK(p);
1882 td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1883 splx(s);
1884 return (0);
1885 }
1886 }
1887
1888 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1889 plist)) {
1890 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1891 jobref) {
1892 PROC_UNLOCK(p);
1893 td->td_retval[0] = EINPROGRESS;
1894 splx(s);
1895 return (0);
1896 }
1897 }
1898 splx(s);
1899 PROC_UNLOCK(p);
1900
1901 #if (0)
1902 /*
1903 * Hack for lio.
1904 */
1905 status = fuword(&uap->aiocbp->_aiocb_private.status);
1906 if (status == -1)
1907 return fuword(&uap->aiocbp->_aiocb_private.error);
1908 #endif
1909 return (EINVAL);
1910 }
1911
1912 /* syscall - asynchronous read from a file (REALTIME) */
1913 int
1914 aio_read(struct thread *td, struct aio_read_args *uap)
1915 {
1916
1917 return aio_aqueue(td, uap->aiocbp, LIO_READ);
1918 }
1919
1920 /* syscall - asynchronous write to a file (REALTIME) */
1921 int
1922 aio_write(struct thread *td, struct aio_write_args *uap)
1923 {
1924
1925 return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1926 }
1927
1928 /* syscall - list directed I/O (REALTIME) */
1929 int
1930 lio_listio(struct thread *td, struct lio_listio_args *uap)
1931 {
1932 struct proc *p = td->td_proc;
1933 int nent, nentqueued;
1934 struct aiocb *iocb, * const *cbptr;
1935 struct aiocblist *cb;
1936 struct kaioinfo *ki;
1937 struct aio_liojob *lj;
1938 int error, runningcode;
1939 int nerror;
1940 int i;
1941 int s;
1942
1943 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1944 return (EINVAL);
1945
1946 nent = uap->nent;
1947 if (nent < 0 || nent > AIO_LISTIO_MAX)
1948 return (EINVAL);
1949
1950 if (p->p_aioinfo == NULL)
1951 aio_init_aioinfo(p);
1952
1953 if ((nent + num_queue_count) > max_queue_count)
1954 return (EAGAIN);
1955
1956 ki = p->p_aioinfo;
1957 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1958 return (EAGAIN);
1959
1960 lj = uma_zalloc(aiolio_zone, M_WAITOK);
1961 if (!lj)
1962 return (EAGAIN);
1963
1964 lj->lioj_flags = 0;
1965 lj->lioj_buffer_count = 0;
1966 lj->lioj_buffer_finished_count = 0;
1967 lj->lioj_queue_count = 0;
1968 lj->lioj_queue_finished_count = 0;
1969
1970 /*
1971 * Setup signal.
1972 */
1973 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1974 error = copyin(uap->sig, &lj->lioj_signal,
1975 sizeof(lj->lioj_signal));
1976 if (error) {
1977 uma_zfree(aiolio_zone, lj);
1978 return (error);
1979 }
1980 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1981 uma_zfree(aiolio_zone, lj);
1982 return (EINVAL);
1983 }
1984 lj->lioj_flags |= LIOJ_SIGNAL;
1985 }
1986 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1987 /*
1988 * Get pointers to the list of I/O requests.
1989 */
1990 nerror = 0;
1991 nentqueued = 0;
1992 cbptr = uap->acb_list;
1993 for (i = 0; i < uap->nent; i++) {
1994 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1995 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1996 error = _aio_aqueue(td, iocb, lj, 0);
1997 if (error == 0)
1998 nentqueued++;
1999 else
2000 nerror++;
2001 }
2002 }
2003
2004 /*
2005 * If we haven't queued any, then just return error.
2006 */
2007 if (nentqueued == 0)
2008 return (0);
2009
2010 /*
2011 * Calculate the appropriate error return.
2012 */
2013 runningcode = 0;
2014 if (nerror)
2015 runningcode = EIO;
2016
2017 if (uap->mode == LIO_WAIT) {
2018 int command, found;
2019 long jobref;
2020
2021 for (;;) {
2022 found = 0;
2023 for (i = 0; i < uap->nent; i++) {
2024 /*
2025 * Fetch address of the control buf pointer in
2026 * user space.
2027 */
2028 iocb = (struct aiocb *)
2029 (intptr_t)fuword(&cbptr[i]);
2030 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2031 == 0))
2032 continue;
2033
2034 /*
2035 * Fetch the associated command from user space.
2036 */
2037 command = fuword(&iocb->aio_lio_opcode);
2038 if (command == LIO_NOP) {
2039 found++;
2040 continue;
2041 }
2042
2043 jobref =
2044 fuword(&iocb->_aiocb_private.kernelinfo);
2045
2046 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2047 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2048 == jobref) {
2049 if (cb->uaiocb.aio_lio_opcode
2050 == LIO_WRITE) {
2051 p->p_stats->p_ru.ru_oublock
2052 +=
2053 cb->outputcharge;
2054 cb->outputcharge = 0;
2055 } else if (cb->uaiocb.aio_lio_opcode
2056 == LIO_READ) {
2057 p->p_stats->p_ru.ru_inblock
2058 += cb->inputcharge;
2059 cb->inputcharge = 0;
2060 }
2061 found++;
2062 break;
2063 }
2064 }
2065
2066 s = splbio();
2067 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2068 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2069 == jobref) {
2070 found++;
2071 break;
2072 }
2073 }
2074 splx(s);
2075 }
2076
2077 /*
2078 * If all I/Os have been disposed of, then we can
2079 * return.
2080 */
2081 if (found == nentqueued)
2082 return (runningcode);
2083
2084 ki->kaio_flags |= KAIO_WAKEUP;
2085 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2086
2087 if (error == EINTR)
2088 return (EINTR);
2089 else if (error == EWOULDBLOCK)
2090 return (EAGAIN);
2091 }
2092 }
2093
2094 return (runningcode);
2095 }
2096
2097 /*
2098 * Interrupt handler for physio, performs the necessary process wakeups, and
2099 * signals.
2100 */
2101 static void
2102 aio_physwakeup(struct buf *bp)
2103 {
2104 struct aiocblist *aiocbe;
2105 struct proc *p;
2106 struct kaioinfo *ki;
2107 struct aio_liojob *lj;
2108
2109 mtx_lock(&Giant);
2110 bp->b_flags |= B_DONE;
2111 wakeup(bp);
2112
2113 aiocbe = (struct aiocblist *)bp->b_caller1;
2114 if (aiocbe) {
2115 p = aiocbe->userproc;
2116
2117 aiocbe->jobstate = JOBST_JOBBFINISHED;
2118 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2119 aiocbe->uaiocb._aiocb_private.error = 0;
2120 aiocbe->jobflags |= AIOCBLIST_DONE;
2121
2122 if (bp->b_ioflags & BIO_ERROR)
2123 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2124
2125 lj = aiocbe->lio;
2126 if (lj) {
2127 lj->lioj_buffer_finished_count++;
2128
2129 /*
2130 * wakeup/signal if all of the interrupt jobs are done.
2131 */
2132 if (lj->lioj_buffer_finished_count ==
2133 lj->lioj_buffer_count &&
2134 lj->lioj_queue_finished_count ==
2135 lj->lioj_queue_count) {
2136 /*
2137 * Post a signal if it is called for.
2138 */
2139 if ((lj->lioj_flags &
2140 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2141 LIOJ_SIGNAL) {
2142 PROC_LOCK(p);
2143 psignal(p, lj->lioj_signal.sigev_signo);
2144 PROC_UNLOCK(p);
2145 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2146 }
2147 }
2148 }
2149
2150 ki = p->p_aioinfo;
2151 if (ki) {
2152 ki->kaio_buffer_finished_count++;
2153 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2154 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2155
2156 KNOTE_UNLOCKED(&aiocbe->klist, 0);
2157 /* Do the wakeup. */
2158 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2159 ki->kaio_flags &= ~KAIO_WAKEUP;
2160 wakeup(p);
2161 }
2162 }
2163
2164 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2165 PROC_LOCK(p);
2166 psignal(p, aiocbe->uaiocb.aio_sigevent.sigev_signo);
2167 PROC_UNLOCK(p);
2168 }
2169 }
2170 mtx_unlock(&Giant);
2171 }
2172
2173 /* syscall - wait for the next completion of an aio request */
2174 int
2175 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2176 {
2177 struct proc *p = td->td_proc;
2178 struct timeval atv;
2179 struct timespec ts;
2180 struct kaioinfo *ki;
2181 struct aiocblist *cb = NULL;
2182 int error, s, timo;
2183
2184 suword(uap->aiocbp, (int)NULL);
2185
2186 timo = 0;
2187 if (uap->timeout) {
2188 /* Get timespec struct. */
2189 error = copyin(uap->timeout, &ts, sizeof(ts));
2190 if (error)
2191 return (error);
2192
2193 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2194 return (EINVAL);
2195
2196 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2197 if (itimerfix(&atv))
2198 return (EINVAL);
2199 timo = tvtohz(&atv);
2200 }
2201
2202 ki = p->p_aioinfo;
2203 if (ki == NULL)
2204 return (EAGAIN);
2205
2206 for (;;) {
2207 PROC_LOCK(p);
2208 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2209 PROC_UNLOCK(p);
2210 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2211 td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2212 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2213 p->p_stats->p_ru.ru_oublock +=
2214 cb->outputcharge;
2215 cb->outputcharge = 0;
2216 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2217 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2218 cb->inputcharge = 0;
2219 }
2220 error = cb->uaiocb._aiocb_private.error;
2221 aio_free_entry(cb);
2222 return (error);
2223 }
2224
2225 s = splbio();
2226 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2227 PROC_UNLOCK(p);
2228 splx(s);
2229 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2230 error = cb->uaiocb._aiocb_private.error;
2231 td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2232 aio_free_entry(cb);
2233 return (error);
2234 }
2235
2236 ki->kaio_flags |= KAIO_WAKEUP;
2237 error = msleep(p, &p->p_mtx, PDROP | PRIBIO | PCATCH, "aiowc",
2238 timo);
2239 splx(s);
2240
2241 if (error == ERESTART)
2242 return (EINTR);
2243 else if (error < 0)
2244 return (error);
2245 else if (error == EINTR)
2246 return (EINTR);
2247 else if (error == EWOULDBLOCK)
2248 return (EAGAIN);
2249 }
2250 }
2251
2252 /* kqueue attach function */
2253 static int
2254 filt_aioattach(struct knote *kn)
2255 {
2256 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2257
2258 /*
2259 * The aiocbe pointer must be validated before using it, so
2260 * registration is restricted to the kernel; the user cannot
2261 * set EV_FLAG1.
2262 */
2263 if ((kn->kn_flags & EV_FLAG1) == 0)
2264 return (EPERM);
2265 kn->kn_flags &= ~EV_FLAG1;
2266
2267 knlist_add(&aiocbe->klist, kn, 0);
2268
2269 return (0);
2270 }
2271
2272 /* kqueue detach function */
2273 static void
2274 filt_aiodetach(struct knote *kn)
2275 {
2276 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2277
2278 knlist_remove(&aiocbe->klist, kn, 0);
2279 }
2280
2281 /* kqueue filter function */
2282 /*ARGSUSED*/
2283 static int
2284 filt_aio(struct knote *kn, long hint)
2285 {
2286 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2287
2288 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2289 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2290 aiocbe->jobstate != JOBST_JOBBFINISHED)
2291 return (0);
2292 kn->kn_flags |= EV_EOF;
2293 return (1);
2294 }
Cache object: aeba4b6d726ec930ed2e3b77c39988e9
|