?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
kernel - Move mplock to machine-independent C
[dragonfly.git] / sys / kern / vfs_aio.c
blob2e9dcc281d04035b34bbb869559a8b6bddd214a8
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 * $FreeBSD: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $
17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.42 2007/07/20 17:21:52 dillon Exp $
18 */
19
20 /*
21 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
22 */
23
24 #include <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/buf.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/fcntl.h>
31 #include <sys/file.h>
32 #include <sys/lock.h>
33 #include <sys/unistd.h>
34 #include <sys/proc.h>
35 #include <sys/resourcevar.h>
36 #include <sys/signalvar.h>
37 #include <sys/protosw.h>
38 #include <sys/socketvar.h>
39 #include <sys/sysctl.h>
40 #include <sys/vnode.h>
41 #include <sys/conf.h>
42 #include <sys/event.h>
43
44 #include <vm/vm.h>
45 #include <vm/vm_extern.h>
46 #include <vm/pmap.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_zone.h>
49 #include <sys/aio.h>
50
51 #include <sys/file2.h>
52 #include <sys/buf2.h>
53 #include <sys/sysref2.h>
54 #include <sys/thread2.h>
55 #include <sys/mplock2.h>
56
57 #include <machine/limits.h>
58 #include "opt_vfs_aio.h"
59
60 #ifdef VFS_AIO
61
62 /*
63 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
64 * overflow.
65 */
66 static long jobrefid;
67
68 #define JOBST_NULL 0x0
69 #define JOBST_JOBQGLOBAL 0x2
70 #define JOBST_JOBRUNNING 0x3
71 #define JOBST_JOBFINISHED 0x4
72 #define JOBST_JOBQBUF 0x5
73 #define JOBST_JOBBFINISHED 0x6
74
75 #ifndef MAX_AIO_PER_PROC
76 #define MAX_AIO_PER_PROC 32
77 #endif
78
79 #ifndef MAX_AIO_QUEUE_PER_PROC
80 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
81 #endif
82
83 #ifndef MAX_AIO_PROCS
84 #define MAX_AIO_PROCS 32
85 #endif
86
87 #ifndef MAX_AIO_QUEUE
88 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
89 #endif
90
91 #ifndef TARGET_AIO_PROCS
92 #define TARGET_AIO_PROCS 4
93 #endif
94
95 #ifndef MAX_BUF_AIO
96 #define MAX_BUF_AIO 16
97 #endif
98
99 #ifndef AIOD_TIMEOUT_DEFAULT
100 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
101 #endif
102
103 #ifndef AIOD_LIFETIME_DEFAULT
104 #define AIOD_LIFETIME_DEFAULT (30 * hz)
105 #endif
106
107 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
108
109 static int max_aio_procs = MAX_AIO_PROCS;
110 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
111 CTLFLAG_RW, &max_aio_procs, 0,
112 "Maximum number of kernel threads to use for handling async IO");
113
114 static int num_aio_procs = 0;
115 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
116 CTLFLAG_RD, &num_aio_procs, 0,
117 "Number of presently active kernel threads for async IO");
118
119 /*
120 * The code will adjust the actual number of AIO processes towards this
121 * number when it gets a chance.
122 */
123 static int target_aio_procs = TARGET_AIO_PROCS;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
125 0, "Preferred number of ready kernel threads for async IO");
126
127 static int max_queue_count = MAX_AIO_QUEUE;
128 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
129 "Maximum number of aio requests to queue, globally");
130
131 static int num_queue_count = 0;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
133 "Number of queued aio requests");
134
135 static int num_buf_aio = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
137 "Number of aio requests presently handled by the buf subsystem");
138
139 /* Number of async I/O thread in the process of being started */
140 /* XXX This should be local to _aio_aqueue() */
141 static int num_aio_resv_start = 0;
142
143 static int aiod_timeout;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
145 "Timeout value for synchronous aio operations");
146
147 static int aiod_lifetime;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
149 "Maximum lifetime for idle aiod");
150
151 static int max_aio_per_proc = MAX_AIO_PER_PROC;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
153 0, "Maximum active aio requests per process (stored in the process)");
154
155 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
157 &max_aio_queue_per_proc, 0,
158 "Maximum queued aio requests per process (stored in the process)");
159
160 static int max_buf_aio = MAX_BUF_AIO;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
162 "Maximum buf aio requests per process (stored in the process)");
163
164 /*
165 * AIO process info
166 */
167 #define AIOP_FREE 0x1 /* proc on free queue */
168 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
169
170 struct aioproclist {
171 int aioprocflags; /* AIO proc flags */
172 TAILQ_ENTRY(aioproclist) list; /* List of processes */
173 struct proc *aioproc; /* The AIO thread */
174 };
175
176 /*
177 * data-structure for lio signal management
178 */
179 struct aio_liojob {
180 int lioj_flags;
181 int lioj_buffer_count;
182 int lioj_buffer_finished_count;
183 int lioj_queue_count;
184 int lioj_queue_finished_count;
185 struct sigevent lioj_signal; /* signal on all I/O done */
186 TAILQ_ENTRY(aio_liojob) lioj_list;
187 struct kaioinfo *lioj_ki;
188 };
189 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
190 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
191
192 /*
193 * per process aio data structure
194 */
195 struct kaioinfo {
196 int kaio_flags; /* per process kaio flags */
197 int kaio_maxactive_count; /* maximum number of AIOs */
198 int kaio_active_count; /* number of currently used AIOs */
199 int kaio_qallowed_count; /* maxiumu size of AIO queue */
200 int kaio_queue_count; /* size of AIO queue */
201 int kaio_ballowed_count; /* maximum number of buffers */
202 int kaio_queue_finished_count; /* number of daemon jobs finished */
203 int kaio_buffer_count; /* number of physio buffers */
204 int kaio_buffer_finished_count; /* count of I/O done */
205 struct proc *kaio_p; /* process that uses this kaio block */
206 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
207 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
208 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
209 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
210 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
211 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
212 };
213
214 #define KAIO_RUNDOWN 0x1 /* process is being run down */
215 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
216
217 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
218 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
219 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
220 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
221
222 static void aio_init_aioinfo(struct proc *p);
223 static void aio_onceonly(void *);
224 static int aio_free_entry(struct aiocblist *aiocbe);
225 static void aio_process(struct aiocblist *aiocbe);
226 static int aio_newproc(void);
227 static int aio_aqueue(struct aiocb *job, int type);
228 static void aio_physwakeup(struct bio *bio);
229 static int aio_fphysio(struct aiocblist *aiocbe);
230 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
231 static void aio_daemon(void *uproc, struct trapframe *frame);
232 static void process_signal(void *aioj);
233
234 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
235
236 /*
237 * Zones for:
238 * kaio Per process async io info
239 * aiop async io thread data
240 * aiocb async io jobs
241 * aiol list io job pointer - internal to aio_suspend XXX
242 * aiolio list io jobs
243 */
244 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
245
246 /*
247 * Startup initialization
248 */
249 static void
250 aio_onceonly(void *na)
251 {
252 TAILQ_INIT(&aio_freeproc);
253 TAILQ_INIT(&aio_activeproc);
254 TAILQ_INIT(&aio_jobs);
255 TAILQ_INIT(&aio_bufjobs);
256 TAILQ_INIT(&aio_freejobs);
257 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
258 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
259 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
260 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
261 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
262 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
263 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
264 jobrefid = 1;
265 }
266
267 /*
268 * Init the per-process aioinfo structure. The aioinfo limits are set
269 * per-process for user limit (resource) management.
270 */
271 static void
272 aio_init_aioinfo(struct proc *p)
273 {
274 struct kaioinfo *ki;
275 if (p->p_aioinfo == NULL) {
276 ki = zalloc(kaio_zone);
277 p->p_aioinfo = ki;
278 ki->kaio_flags = 0;
279 ki->kaio_maxactive_count = max_aio_per_proc;
280 ki->kaio_active_count = 0;
281 ki->kaio_qallowed_count = max_aio_queue_per_proc;
282 ki->kaio_queue_count = 0;
283 ki->kaio_ballowed_count = max_buf_aio;
284 ki->kaio_buffer_count = 0;
285 ki->kaio_buffer_finished_count = 0;
286 ki->kaio_p = p;
287 TAILQ_INIT(&ki->kaio_jobdone);
288 TAILQ_INIT(&ki->kaio_jobqueue);
289 TAILQ_INIT(&ki->kaio_bufdone);
290 TAILQ_INIT(&ki->kaio_bufqueue);
291 TAILQ_INIT(&ki->kaio_liojoblist);
292 TAILQ_INIT(&ki->kaio_sockqueue);
293 }
294
295 while (num_aio_procs < target_aio_procs)
296 aio_newproc();
297 }
298
299 /*
300 * Free a job entry. Wait for completion if it is currently active, but don't
301 * delay forever. If we delay, we return a flag that says that we have to
302 * restart the queue scan.
303 */
304 static int
305 aio_free_entry(struct aiocblist *aiocbe)
306 {
307 struct kaioinfo *ki;
308 struct aio_liojob *lj;
309 struct proc *p;
310 int error;
311
312 if (aiocbe->jobstate == JOBST_NULL)
313 panic("aio_free_entry: freeing already free job");
314
315 p = aiocbe->userproc;
316 ki = p->p_aioinfo;
317 lj = aiocbe->lio;
318 if (ki == NULL)
319 panic("aio_free_entry: missing p->p_aioinfo");
320
321 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
322 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
323 tsleep(aiocbe, 0, "jobwai", 0);
324 }
325 if (aiocbe->bp == NULL) {
326 if (ki->kaio_queue_count <= 0)
327 panic("aio_free_entry: process queue size <= 0");
328 if (num_queue_count <= 0)
329 panic("aio_free_entry: system wide queue size <= 0");
330
331 if (lj) {
332 lj->lioj_queue_count--;
333 if (aiocbe->jobflags & AIOCBLIST_DONE)
334 lj->lioj_queue_finished_count--;
335 }
336 ki->kaio_queue_count--;
337 if (aiocbe->jobflags & AIOCBLIST_DONE)
338 ki->kaio_queue_finished_count--;
339 num_queue_count--;
340 } else {
341 if (lj) {
342 lj->lioj_buffer_count--;
343 if (aiocbe->jobflags & AIOCBLIST_DONE)
344 lj->lioj_buffer_finished_count--;
345 }
346 if (aiocbe->jobflags & AIOCBLIST_DONE)
347 ki->kaio_buffer_finished_count--;
348 ki->kaio_buffer_count--;
349 num_buf_aio--;
350 }
351
352 /* aiocbe is going away, we need to destroy any knotes */
353 /* XXX lwp knote wants a thread, but only cares about the process */
354 knote_remove(FIRST_LWP_IN_PROC(p)->lwp_thread, &aiocbe->klist);
355
356 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
357 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
358 ki->kaio_flags &= ~KAIO_WAKEUP;
359 wakeup(p);
360 }
361
362 if (aiocbe->jobstate == JOBST_JOBQBUF) {
363 if ((error = aio_fphysio(aiocbe)) != 0)
364 return error;
365 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
366 panic("aio_free_entry: invalid physio finish-up state");
367 crit_enter();
368 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
369 crit_exit();
370 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
371 crit_enter();
372 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
373 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
374 crit_exit();
375 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
376 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
377 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
378 crit_enter();
379 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
380 crit_exit();
381 if (aiocbe->bp) {
382 vunmapbuf(aiocbe->bp);
383 relpbuf(aiocbe->bp, NULL);
384 aiocbe->bp = NULL;
385 }
386 }
387 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
388 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
389 zfree(aiolio_zone, lj);
390 }
391 aiocbe->jobstate = JOBST_NULL;
392 callout_stop(&aiocbe->timeout);
393 fdrop(aiocbe->fd_file);
394 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
395 return 0;
396 }
397 #endif /* VFS_AIO */
398
399 /*
400 * Rundown the jobs for a given process.
401 */
402 void
403 aio_proc_rundown(struct proc *p)
404 {
405 #ifndef VFS_AIO
406 return;
407 #else
408 struct kaioinfo *ki;
409 struct aio_liojob *lj, *ljn;
410 struct aiocblist *aiocbe, *aiocbn;
411 struct file *fp;
412 struct socket *so;
413
414 ki = p->p_aioinfo;
415 if (ki == NULL)
416 return;
417
418 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
419 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
420 ki->kaio_buffer_finished_count)) {
421 ki->kaio_flags |= KAIO_RUNDOWN;
422 if (tsleep(p, 0, "kaiowt", aiod_timeout))
423 break;
424 }
425
426 /*
427 * Move any aio ops that are waiting on socket I/O to the normal job
428 * queues so they are cleaned up with any others.
429 */
430 crit_enter();
431 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
432 aiocbn) {
433 aiocbn = TAILQ_NEXT(aiocbe, plist);
434 fp = aiocbe->fd_file;
435 if (fp != NULL) {
436 so = (struct socket *)fp->f_data;
437 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
438 if (TAILQ_EMPTY(&so->so_aiojobq)) {
439 so->so_snd.ssb_flags &= ~SSB_AIO;
440 so->so_rcv.ssb_flags &= ~SSB_AIO;
441 }
442 }
443 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
444 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
445 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
446 }
447 crit_exit();
448
449 restart1:
450 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
451 aiocbn = TAILQ_NEXT(aiocbe, plist);
452 if (aio_free_entry(aiocbe))
453 goto restart1;
454 }
455
456 restart2:
457 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
458 aiocbn) {
459 aiocbn = TAILQ_NEXT(aiocbe, plist);
460 if (aio_free_entry(aiocbe))
461 goto restart2;
462 }
463
464 restart3:
465 crit_enter();
466 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
467 ki->kaio_flags |= KAIO_WAKEUP;
468 tsleep(p, 0, "aioprn", 0);
469 crit_exit();
470 goto restart3;
471 }
472 crit_exit();
473
474 restart4:
475 crit_enter();
476 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
477 aiocbn = TAILQ_NEXT(aiocbe, plist);
478 if (aio_free_entry(aiocbe)) {
479 crit_exit();
480 goto restart4;
481 }
482 }
483 crit_exit();
484
485 /*
486 * If we've slept, jobs might have moved from one queue to another.
487 * Retry rundown if we didn't manage to empty the queues.
488 */
489 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
490 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
491 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
492 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
493 goto restart1;
494
495 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
496 ljn = TAILQ_NEXT(lj, lioj_list);
497 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
498 0)) {
499 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
500 zfree(aiolio_zone, lj);
501 } else {
502 #ifdef DIAGNOSTIC
503 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
504 "QF:%d\n", lj->lioj_buffer_count,
505 lj->lioj_buffer_finished_count,
506 lj->lioj_queue_count,
507 lj->lioj_queue_finished_count);
508 #endif
509 }
510 }
511
512 zfree(kaio_zone, ki);
513 p->p_aioinfo = NULL;
514 #endif /* VFS_AIO */
515 }
516
517 #ifdef VFS_AIO
518 /*
519 * Select a job to run (called by an AIO daemon).
520 */
521 static struct aiocblist *
522 aio_selectjob(struct aioproclist *aiop)
523 {
524 struct aiocblist *aiocbe;
525 struct kaioinfo *ki;
526 struct proc *userp;
527
528 crit_enter();
529 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
530 TAILQ_NEXT(aiocbe, list)) {
531 userp = aiocbe->userproc;
532 ki = userp->p_aioinfo;
533
534 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
535 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
536 crit_exit();
537 return aiocbe;
538 }
539 }
540 crit_exit();
541
542 return NULL;
543 }
544
545 /*
546 * The AIO processing activity. This is the code that does the I/O request for
547 * the non-physio version of the operations. The normal vn operations are used,
548 * and this code should work in all instances for every type of file, including
549 * pipes, sockets, fifos, and regular files.
550 */
551 static void
552 aio_process(struct aiocblist *aiocbe)
553 {
554 struct thread *mytd;
555 struct aiocb *cb;
556 struct file *fp;
557 struct uio auio;
558 struct iovec aiov;
559 int cnt;
560 int error;
561 int oublock_st, oublock_end;
562 int inblock_st, inblock_end;
563
564 mytd = curthread;
565 cb = &aiocbe->uaiocb;
566 fp = aiocbe->fd_file;
567
568 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
569 aiov.iov_len = cb->aio_nbytes;
570
571 auio.uio_iov = &aiov;
572 auio.uio_iovcnt = 1;
573 auio.uio_offset = cb->aio_offset;
574 auio.uio_resid = cb->aio_nbytes;
575 cnt = cb->aio_nbytes;
576 auio.uio_segflg = UIO_USERSPACE;
577 auio.uio_td = mytd;
578
579 inblock_st = mytd->td_lwp->lwp_ru.ru_inblock;
580 oublock_st = mytd->td_lwp->lwp_ru.ru_oublock;
581 /*
582 * _aio_aqueue() acquires a reference to the file that is
583 * released in aio_free_entry().
584 */
585 if (cb->aio_lio_opcode == LIO_READ) {
586 auio.uio_rw = UIO_READ;
587 error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET);
588 } else {
589 auio.uio_rw = UIO_WRITE;
590 error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET);
591 }
592 inblock_end = mytd->td_lwp->lwp_ru.ru_inblock;
593 oublock_end = mytd->td_lwp->lwp_ru.ru_oublock;
594
595 aiocbe->inputcharge = inblock_end - inblock_st;
596 aiocbe->outputcharge = oublock_end - oublock_st;
597
598 if ((error) && (auio.uio_resid != cnt)) {
599 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
600 error = 0;
601 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
602 ksignal(aiocbe->userproc, SIGPIPE);
603 }
604
605 cnt -= auio.uio_resid;
606 cb->_aiocb_private.error = error;
607 cb->_aiocb_private.status = cnt;
608 }
609
610 /*
611 * The AIO daemon, most of the actual work is done in aio_process,
612 * but the setup (and address space mgmt) is done in this routine.
613 *
614 * The MP lock is held on entry.
615 */
616 static void
617 aio_daemon(void *uproc, struct trapframe *frame)
618 {
619 struct aio_liojob *lj;
620 struct aiocb *cb;
621 struct aiocblist *aiocbe;
622 struct aioproclist *aiop;
623 struct kaioinfo *ki;
624 struct proc *mycp, *userp;
625 struct vmspace *curvm;
626 struct lwp *mylwp;
627 struct ucred *cr;
628
629 mylwp = curthread->td_lwp;
630 mycp = mylwp->lwp_proc;
631
632 if (mycp->p_textvp) {
633 vrele(mycp->p_textvp);
634 mycp->p_textvp = NULL;
635 }
636
637 /*
638 * Allocate and ready the aio control info. There is one aiop structure
639 * per daemon.
640 */
641 aiop = zalloc(aiop_zone);
642 aiop->aioproc = mycp;
643 aiop->aioprocflags |= AIOP_FREE;
644
645 crit_enter();
646
647 /*
648 * Place thread (lightweight process) onto the AIO free thread list.
649 */
650 if (TAILQ_EMPTY(&aio_freeproc))
651 wakeup(&aio_freeproc);
652 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
653
654 crit_exit();
655
656 /* Make up a name for the daemon. */
657 strcpy(mycp->p_comm, "aiod");
658
659 /*
660 * Get rid of our current filedescriptors. AIOD's don't need any
661 * filedescriptors, except as temporarily inherited from the client.
662 * Credentials are also cloned, and made equivalent to "root".
663 */
664 fdfree(mycp, NULL);
665 cr = cratom(&mycp->p_ucred);
666 cr->cr_uid = 0;
667 uireplace(&cr->cr_uidinfo, uifind(0));
668 cr->cr_ngroups = 1;
669 cr->cr_groups[0] = 1;
670
671 /* The daemon resides in its own pgrp. */
672 enterpgrp(mycp, mycp->p_pid, 1);
673
674 /* Mark special process type. */
675 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
676
677 /*
678 * Wakeup parent process. (Parent sleeps to keep from blasting away
679 * and creating too many daemons.)
680 */
681 wakeup(mycp);
682 curvm = NULL;
683
684 for (;;) {
685 /*
686 * Take daemon off of free queue
687 */
688 if (aiop->aioprocflags & AIOP_FREE) {
689 crit_enter();
690 TAILQ_REMOVE(&aio_freeproc, aiop, list);
691 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
692 aiop->aioprocflags &= ~AIOP_FREE;
693 crit_exit();
694 }
695 aiop->aioprocflags &= ~AIOP_SCHED;
696
697 /*
698 * Check for jobs.
699 */
700 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
701 cb = &aiocbe->uaiocb;
702 userp = aiocbe->userproc;
703
704 aiocbe->jobstate = JOBST_JOBRUNNING;
705
706 /*
707 * Connect to process address space for user program.
708 */
709 if (curvm != userp->p_vmspace) {
710 pmap_setlwpvm(mylwp, userp->p_vmspace);
711 if (curvm)
712 sysref_put(&curvm->vm_sysref);
713 curvm = userp->p_vmspace;
714 sysref_get(&curvm->vm_sysref);
715 }
716
717 ki = userp->p_aioinfo;
718 lj = aiocbe->lio;
719
720 /* Account for currently active jobs. */
721 ki->kaio_active_count++;
722
723 /* Do the I/O function. */
724 aio_process(aiocbe);
725
726 /* Decrement the active job count. */
727 ki->kaio_active_count--;
728
729 /*
730 * Increment the completion count for wakeup/signal
731 * comparisons.
732 */
733 aiocbe->jobflags |= AIOCBLIST_DONE;
734 ki->kaio_queue_finished_count++;
735 if (lj)
736 lj->lioj_queue_finished_count++;
737 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
738 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
739 ki->kaio_flags &= ~KAIO_WAKEUP;
740 wakeup(userp);
741 }
742
743 crit_enter();
744 if (lj && (lj->lioj_flags &
745 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
746 if ((lj->lioj_queue_finished_count ==
747 lj->lioj_queue_count) &&
748 (lj->lioj_buffer_finished_count ==
749 lj->lioj_buffer_count)) {
750 ksignal(userp,
751 lj->lioj_signal.sigev_signo);
752 lj->lioj_flags |=
753 LIOJ_SIGNAL_POSTED;
754 }
755 }
756 crit_exit();
757
758 aiocbe->jobstate = JOBST_JOBFINISHED;
759
760 crit_enter();
761 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
762 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
763 crit_exit();
764 KNOTE(&aiocbe->klist, 0);
765
766 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
767 wakeup(aiocbe);
768 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
769 }
770
771 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
772 ksignal(userp, cb->aio_sigevent.sigev_signo);
773 }
774 }
775
776 /*
777 * Disconnect from user address space.
778 */
779 if (curvm) {
780 /* swap our original address space back in */
781 pmap_setlwpvm(mylwp, mycp->p_vmspace);
782 sysref_put(&curvm->vm_sysref);
783 curvm = NULL;
784 }
785
786 /*
787 * If we are the first to be put onto the free queue, wakeup
788 * anyone waiting for a daemon.
789 */
790 crit_enter();
791 TAILQ_REMOVE(&aio_activeproc, aiop, list);
792 if (TAILQ_EMPTY(&aio_freeproc))
793 wakeup(&aio_freeproc);
794 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
795 aiop->aioprocflags |= AIOP_FREE;
796 crit_exit();
797
798 /*
799 * If daemon is inactive for a long time, allow it to exit,
800 * thereby freeing resources.
801 */
802 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
803 0, "aiordy", aiod_lifetime)) {
804 crit_enter();
805 if (TAILQ_EMPTY(&aio_jobs)) {
806 if ((aiop->aioprocflags & AIOP_FREE) &&
807 (num_aio_procs > target_aio_procs)) {
808 TAILQ_REMOVE(&aio_freeproc, aiop, list);
809 crit_exit();
810 zfree(aiop_zone, aiop);
811 num_aio_procs--;
812 #ifdef DIAGNOSTIC
813 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) {
814 kprintf("AIOD: bad vm refcnt for"
815 " exiting daemon: %d\n",
816 mycp->p_vmspace->vm_sysref.refcnt);
817 }
818 #endif
819 exit1(0);
820 }
821 }
822 crit_exit();
823 }
824 }
825 }
826
827 /*
828 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
829 * AIO daemon modifies its environment itself.
830 */
831 static int
832 aio_newproc(void)
833 {
834 int error;
835 struct lwp *lp, *nlp;
836 struct proc *np;
837
838 lp = &lwp0;
839 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
840 if (error)
841 return error;
842 nlp = ONLY_LWP_IN_PROC(np);
843 cpu_set_fork_handler(nlp, aio_daemon, curproc);
844 start_forked_proc(lp, np);
845
846 /*
847 * Wait until daemon is started, but continue on just in case to
848 * handle error conditions.
849 */
850 error = tsleep(np, 0, "aiosta", aiod_timeout);
851 num_aio_procs++;
852
853 return error;
854 }
855
856 /*
857 * Try the high-performance, low-overhead physio method for eligible
858 * VCHR devices. This method doesn't use an aio helper thread, and
859 * thus has very low overhead.
860 *
861 * Assumes that the caller, _aio_aqueue(), has incremented the file
862 * structure's reference count, preventing its deallocation for the
863 * duration of this call.
864 */
865 static int
866 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
867 {
868 int error;
869 struct aiocb *cb;
870 struct file *fp;
871 struct buf *bp;
872 struct vnode *vp;
873 struct kaioinfo *ki;
874 struct aio_liojob *lj;
875 int notify;
876
877 cb = &aiocbe->uaiocb;
878 fp = aiocbe->fd_file;
879
880 if (fp->f_type != DTYPE_VNODE)
881 return (-1);
882
883 vp = (struct vnode *)fp->f_data;
884
885 /*
886 * If its not a disk, we don't want to return a positive error.
887 * It causes the aio code to not fall through to try the thread
888 * way when you're talking to a regular file.
889 */
890 if (!vn_isdisk(vp, &error)) {
891 if (error == ENOTBLK)
892 return (-1);
893 else
894 return (error);
895 }
896
897 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
898 return (-1);
899
900 if (cb->aio_nbytes >
901 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
902 return (-1);
903
904 ki = p->p_aioinfo;
905 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
906 return (-1);
907
908 ki->kaio_buffer_count++;
909
910 lj = aiocbe->lio;
911 if (lj)
912 lj->lioj_buffer_count++;
913
914 /* Create and build a buffer header for a transfer. */
915 bp = getpbuf(NULL);
916 BUF_KERNPROC(bp);
917
918 /*
919 * Get a copy of the kva from the physical buffer.
920 */
921 bp->b_bio1.bio_caller_info1.ptr = p;
922 error = 0;
923
924 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
925 BUF_CMD_WRITE : BUF_CMD_READ;
926 bp->b_bio1.bio_done = aio_physwakeup;
927 bp->b_bio1.bio_flags |= BIO_SYNC;
928 bp->b_bio1.bio_offset = cb->aio_offset;
929
930 /* Bring buffer into kernel space. */
931 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
932 error = EFAULT;
933 goto doerror;
934 }
935
936 crit_enter();
937
938 aiocbe->bp = bp;
939 bp->b_bio1.bio_caller_info2.ptr = aiocbe;
940 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
941 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
942 aiocbe->jobstate = JOBST_JOBQBUF;
943 cb->_aiocb_private.status = cb->aio_nbytes;
944 num_buf_aio++;
945 bp->b_error = 0;
946
947 crit_exit();
948
949 /*
950 * Perform the transfer. vn_strategy must be used even though we
951 * know we have a device in order to deal with requests which exceed
952 * device DMA limitations.
953 */
954 vn_strategy(vp, &bp->b_bio1);
955
956 notify = 0;
957 crit_enter();
958
959 #if 0
960 /*
961 * If we had an error invoking the request, or an error in processing
962 * the request before we have returned, we process it as an error in
963 * transfer. Note that such an I/O error is not indicated immediately,
964 * but is returned using the aio_error mechanism. In this case,
965 * aio_suspend will return immediately.
966 */
967 if (bp->b_error || (bp->b_flags & B_ERROR)) {
968 struct aiocb *job = aiocbe->uuaiocb;
969
970 aiocbe->uaiocb._aiocb_private.status = 0;
971 suword(&job->_aiocb_private.status, 0);
972 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
973 suword(&job->_aiocb_private.error, bp->b_error);
974
975 ki->kaio_buffer_finished_count++;
976
977 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
978 aiocbe->jobstate = JOBST_JOBBFINISHED;
979 aiocbe->jobflags |= AIOCBLIST_DONE;
980 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
981 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
982 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
983 notify = 1;
984 }
985 }
986 #endif
987 crit_exit();
988 if (notify)
989 KNOTE(&aiocbe->klist, 0);
990 return 0;
991
992 doerror:
993 ki->kaio_buffer_count--;
994 if (lj)
995 lj->lioj_buffer_count--;
996 aiocbe->bp = NULL;
997 relpbuf(bp, NULL);
998 return error;
999 }
1000
1001 /*
1002 * This waits/tests physio completion.
1003 */
1004 static int
1005 aio_fphysio(struct aiocblist *iocb)
1006 {
1007 struct buf *bp;
1008 int error;
1009
1010 bp = iocb->bp;
1011
1012 error = biowait_timeout(&bp->b_bio1, "physstr", aiod_timeout);
1013 if (error == EWOULDBLOCK)
1014 return EINPROGRESS;
1015
1016 /* Release mapping into kernel space. */
1017 vunmapbuf(bp);
1018 iocb->bp = 0;
1019
1020 error = 0;
1021
1022 /* Check for an error. */
1023 if (bp->b_flags & B_ERROR)
1024 error = bp->b_error;
1025
1026 relpbuf(bp, NULL);
1027 return (error);
1028 }
1029 #endif /* VFS_AIO */
1030
1031 /*
1032 * Wake up aio requests that may be serviceable now.
1033 */
1034 void
1035 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1036 {
1037 #ifndef VFS_AIO
1038 return;
1039 #else
1040 struct aiocblist *cb,*cbn;
1041 struct proc *p;
1042 struct kaioinfo *ki = NULL;
1043 int opcode, wakecount = 0;
1044 struct aioproclist *aiop;
1045
1046 if (ssb == &so->so_snd) {
1047 opcode = LIO_WRITE;
1048 so->so_snd.ssb_flags &= ~SSB_AIO;
1049 } else {
1050 opcode = LIO_READ;
1051 so->so_rcv.ssb_flags &= ~SSB_AIO;
1052 }
1053
1054 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1055 cbn = TAILQ_NEXT(cb, list);
1056 if (opcode == cb->uaiocb.aio_lio_opcode) {
1057 p = cb->userproc;
1058 ki = p->p_aioinfo;
1059 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1060 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1061 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1062 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1063 wakecount++;
1064 if (cb->jobstate != JOBST_JOBQGLOBAL)
1065 panic("invalid queue value");
1066 }
1067 }
1068
1069 while (wakecount--) {
1070 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1071 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1072 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1073 aiop->aioprocflags &= ~AIOP_FREE;
1074 wakeup(aiop->aioproc);
1075 }
1076 }
1077 #endif /* VFS_AIO */
1078 }
1079
1080 #ifdef VFS_AIO
1081 /*
1082 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1083 * technique is done in this code.
1084 */
1085 static int
1086 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1087 {
1088 struct proc *p = curproc;
1089 struct file *fp;
1090 unsigned int fd;
1091 struct socket *so;
1092 int error;
1093 int opcode, user_opcode;
1094 struct aiocblist *aiocbe;
1095 struct aioproclist *aiop;
1096 struct kaioinfo *ki;
1097 struct kevent kev;
1098 struct kqueue *kq;
1099 struct file *kq_fp;
1100 int fflags;
1101
1102 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1103 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1104 else
1105 aiocbe = zalloc (aiocb_zone);
1106
1107 aiocbe->inputcharge = 0;
1108 aiocbe->outputcharge = 0;
1109 callout_init(&aiocbe->timeout);
1110 SLIST_INIT(&aiocbe->klist);
1111
1112 suword(&job->_aiocb_private.status, -1);
1113 suword(&job->_aiocb_private.error, 0);
1114 suword(&job->_aiocb_private.kernelinfo, -1);
1115
1116 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1117 if (error) {
1118 suword(&job->_aiocb_private.error, error);
1119 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1120 return error;
1121 }
1122 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1123 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1124 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1125 return EINVAL;
1126 }
1127
1128 /* Save userspace address of the job info. */
1129 aiocbe->uuaiocb = job;
1130
1131 /* Get the opcode. */
1132 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1133 if (type != LIO_NOP)
1134 aiocbe->uaiocb.aio_lio_opcode = type;
1135 opcode = aiocbe->uaiocb.aio_lio_opcode;
1136
1137 /*
1138 * Range check file descriptor.
1139 */
1140 fflags = (opcode == LIO_WRITE) ? FWRITE : FREAD;
1141 fd = aiocbe->uaiocb.aio_fildes;
1142 fp = holdfp(p->p_fd, fd, fflags);
1143 if (fp == NULL) {
1144 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1145 if (type == 0)
1146 suword(&job->_aiocb_private.error, EBADF);
1147 return EBADF;
1148 }
1149
1150 aiocbe->fd_file = fp;
1151
1152 if (aiocbe->uaiocb.aio_offset == -1LL) {
1153 error = EINVAL;
1154 goto aqueue_fail;
1155 }
1156 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1157 if (error) {
1158 error = EINVAL;
1159 goto aqueue_fail;
1160 }
1161 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1162 if (jobrefid == LONG_MAX)
1163 jobrefid = 1;
1164 else
1165 jobrefid++;
1166
1167 if (opcode == LIO_NOP) {
1168 fdrop(fp);
1169 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1170 if (type == 0) {
1171 suword(&job->_aiocb_private.error, 0);
1172 suword(&job->_aiocb_private.status, 0);
1173 suword(&job->_aiocb_private.kernelinfo, 0);
1174 }
1175 return 0;
1176 }
1177 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1178 if (type == 0)
1179 suword(&job->_aiocb_private.status, 0);
1180 error = EINVAL;
1181 goto aqueue_fail;
1182 }
1183
1184 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1185 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1186 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1187 }
1188 else {
1189 /*
1190 * This method for requesting kevent-based notification won't
1191 * work on the alpha, since we're passing in a pointer
1192 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1193 * based method instead.
1194 */
1195 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1196 user_opcode == LIO_WRITE)
1197 goto no_kqueue;
1198
1199 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1200 &kev, sizeof(kev));
1201 if (error)
1202 goto aqueue_fail;
1203 }
1204 kq_fp = holdfp(p->p_fd, (int)kev.ident, -1);
1205 if (kq_fp == NULL || kq_fp->f_type != DTYPE_KQUEUE) {
1206 if (kq_fp) {
1207 fdrop(kq_fp);
1208 kq_fp = NULL;
1209 }
1210 error = EBADF;
1211 goto aqueue_fail;
1212 }
1213 kq = (struct kqueue *)kq_fp->f_data;
1214 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1215 kev.filter = EVFILT_AIO;
1216 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1217 kev.data = (intptr_t)aiocbe;
1218 /* XXX lwp kqueue_register takes a thread, but only uses its proc */
1219 error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread);
1220 fdrop(kq_fp);
1221 aqueue_fail:
1222 if (error) {
1223 fdrop(fp);
1224 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1225 if (type == 0)
1226 suword(&job->_aiocb_private.error, error);
1227 goto done;
1228 }
1229 no_kqueue:
1230
1231 suword(&job->_aiocb_private.error, EINPROGRESS);
1232 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1233 aiocbe->userproc = p;
1234 aiocbe->jobflags = 0;
1235 aiocbe->lio = lj;
1236 ki = p->p_aioinfo;
1237
1238 if (fp->f_type == DTYPE_SOCKET) {
1239 /*
1240 * Alternate queueing for socket ops: Reach down into the
1241 * descriptor to get the socket data. Then check to see if the
1242 * socket is ready to be read or written (based on the requested
1243 * operation).
1244 *
1245 * If it is not ready for io, then queue the aiocbe on the
1246 * socket, and set the flags so we get a call when ssb_notify()
1247 * happens.
1248 */
1249 so = (struct socket *)fp->f_data;
1250 crit_enter();
1251 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1252 LIO_WRITE) && (!sowriteable(so)))) {
1253 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1254 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1255 if (opcode == LIO_READ)
1256 so->so_rcv.ssb_flags |= SSB_AIO;
1257 else
1258 so->so_snd.ssb_flags |= SSB_AIO;
1259 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1260 ki->kaio_queue_count++;
1261 num_queue_count++;
1262 crit_exit();
1263 error = 0;
1264 goto done;
1265 }
1266 crit_exit();
1267 }
1268
1269 if ((error = aio_qphysio(p, aiocbe)) == 0)
1270 goto done;
1271 if (error > 0) {
1272 suword(&job->_aiocb_private.status, 0);
1273 aiocbe->uaiocb._aiocb_private.error = error;
1274 suword(&job->_aiocb_private.error, error);
1275 goto done;
1276 }
1277
1278 /* No buffer for daemon I/O. */
1279 aiocbe->bp = NULL;
1280
1281 ki->kaio_queue_count++;
1282 if (lj)
1283 lj->lioj_queue_count++;
1284 crit_enter();
1285 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1286 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1287 crit_exit();
1288 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1289
1290 num_queue_count++;
1291 error = 0;
1292
1293 /*
1294 * If we don't have a free AIO process, and we are below our quota, then
1295 * start one. Otherwise, depend on the subsequent I/O completions to
1296 * pick-up this job. If we don't successfully create the new process
1297 * (thread) due to resource issues, we return an error for now (EAGAIN),
1298 * which is likely not the correct thing to do.
1299 */
1300 crit_enter();
1301 retryproc:
1302 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1303 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1304 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1305 aiop->aioprocflags &= ~AIOP_FREE;
1306 wakeup(aiop->aioproc);
1307 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1308 ((ki->kaio_active_count + num_aio_resv_start) <
1309 ki->kaio_maxactive_count)) {
1310 num_aio_resv_start++;
1311 if ((error = aio_newproc()) == 0) {
1312 num_aio_resv_start--;
1313 goto retryproc;
1314 }
1315 num_aio_resv_start--;
1316 }
1317 crit_exit();
1318 done:
1319 return error;
1320 }
1321
1322 /*
1323 * This routine queues an AIO request, checking for quotas.
1324 */
1325 static int
1326 aio_aqueue(struct aiocb *job, int type)
1327 {
1328 struct proc *p = curproc;
1329 struct kaioinfo *ki;
1330
1331 if (p->p_aioinfo == NULL)
1332 aio_init_aioinfo(p);
1333
1334 if (num_queue_count >= max_queue_count)
1335 return EAGAIN;
1336
1337 ki = p->p_aioinfo;
1338 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1339 return EAGAIN;
1340
1341 return _aio_aqueue(job, NULL, type);
1342 }
1343 #endif /* VFS_AIO */
1344
1345 /*
1346 * Support the aio_return system call, as a side-effect, kernel resources are
1347 * released.
1348 *
1349 * MPALMOSTSAFE
1350 */
1351 int
1352 sys_aio_return(struct aio_return_args *uap)
1353 {
1354 #ifndef VFS_AIO
1355 return (ENOSYS);
1356 #else
1357 struct proc *p = curproc;
1358 struct lwp *lp = curthread->td_lwp;
1359 long jobref;
1360 struct aiocblist *cb, *ncb;
1361 struct aiocb *ujob;
1362 struct kaioinfo *ki;
1363 int error;
1364
1365 ki = p->p_aioinfo;
1366 if (ki == NULL)
1367 return EINVAL;
1368
1369 ujob = uap->aiocbp;
1370
1371 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1372 if (jobref == -1 || jobref == 0)
1373 return EINVAL;
1374
1375 get_mplock();
1376 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1377 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1378 jobref) {
1379 if (ujob == cb->uuaiocb) {
1380 uap->sysmsg_result =
1381 cb->uaiocb._aiocb_private.status;
1382 } else {
1383 uap->sysmsg_result = EFAULT;
1384 }
1385 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1386 lp->lwp_ru.ru_oublock += cb->outputcharge;
1387 cb->outputcharge = 0;
1388 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1389 lp->lwp_ru.ru_inblock += cb->inputcharge;
1390 cb->inputcharge = 0;
1391 }
1392 aio_free_entry(cb);
1393 error = 0;
1394 goto done;
1395 }
1396 }
1397 crit_enter();
1398 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1399 ncb = TAILQ_NEXT(cb, plist);
1400 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1401 == jobref) {
1402 crit_exit();
1403 if (ujob == cb->uuaiocb) {
1404 uap->sysmsg_result =
1405 cb->uaiocb._aiocb_private.status;
1406 } else {
1407 uap->sysmsg_result = EFAULT;
1408 }
1409 aio_free_entry(cb);
1410 error = 0;
1411 goto done;
1412 }
1413 }
1414 crit_exit();
1415 error = EINVAL;
1416 done:
1417 rel_mplock();
1418 return (error);
1419 #endif /* VFS_AIO */
1420 }
1421
1422 /*
1423 * Allow a process to wakeup when any of the I/O requests are completed.
1424 *
1425 * MPALMOSTSAFE
1426 */
1427 int
1428 sys_aio_suspend(struct aio_suspend_args *uap)
1429 {
1430 #ifndef VFS_AIO
1431 return ENOSYS;
1432 #else
1433 struct proc *p = curproc;
1434 struct timeval atv;
1435 struct timespec ts;
1436 struct aiocb *const *cbptr, *cbp;
1437 struct kaioinfo *ki;
1438 struct aiocblist *cb;
1439 int i;
1440 int njoblist;
1441 int error, timo;
1442 long *ijoblist;
1443 struct aiocb **ujoblist;
1444
1445 if ((u_int)uap->nent > AIO_LISTIO_MAX)
1446 return EINVAL;
1447
1448 timo = 0;
1449 if (uap->timeout) {
1450 /* Get timespec struct. */
1451 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1452 return error;
1453
1454 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1455 return (EINVAL);
1456
1457 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1458 if (itimerfix(&atv))
1459 return (EINVAL);
1460 timo = tvtohz_high(&atv);
1461 }
1462
1463 ki = p->p_aioinfo;
1464 if (ki == NULL)
1465 return EAGAIN;
1466
1467 get_mplock();
1468
1469 njoblist = 0;
1470 ijoblist = zalloc(aiol_zone);
1471 ujoblist = zalloc(aiol_zone);
1472 cbptr = uap->aiocbp;
1473
1474 for (i = 0; i < uap->nent; i++) {
1475 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1476 if (cbp == 0)
1477 continue;
1478 ujoblist[njoblist] = cbp;
1479 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1480 njoblist++;
1481 }
1482
1483 if (njoblist == 0) {
1484 zfree(aiol_zone, ijoblist);
1485 zfree(aiol_zone, ujoblist);
1486 error = 0;
1487 goto done;
1488 }
1489
1490 error = 0;
1491 for (;;) {
1492 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1493 for (i = 0; i < njoblist; i++) {
1494 if (((intptr_t)
1495 cb->uaiocb._aiocb_private.kernelinfo) ==
1496 ijoblist[i]) {
1497 if (ujoblist[i] != cb->uuaiocb)
1498 error = EINVAL;
1499 zfree(aiol_zone, ijoblist);
1500 zfree(aiol_zone, ujoblist);
1501 goto done;
1502 }
1503 }
1504 }
1505
1506 crit_enter();
1507 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1508 TAILQ_NEXT(cb, plist)) {
1509 for (i = 0; i < njoblist; i++) {
1510 if (((intptr_t)
1511 cb->uaiocb._aiocb_private.kernelinfo) ==
1512 ijoblist[i]) {
1513 crit_exit();
1514 if (ujoblist[i] != cb->uuaiocb)
1515 error = EINVAL;
1516 zfree(aiol_zone, ijoblist);
1517 zfree(aiol_zone, ujoblist);
1518 goto done;
1519 }
1520 }
1521 }
1522
1523 ki->kaio_flags |= KAIO_WAKEUP;
1524 error = tsleep(p, PCATCH, "aiospn", timo);
1525 crit_exit();
1526
1527 if (error == ERESTART || error == EINTR) {
1528 zfree(aiol_zone, ijoblist);
1529 zfree(aiol_zone, ujoblist);
1530 error = EINTR;
1531 goto done;
1532 } else if (error == EWOULDBLOCK) {
1533 zfree(aiol_zone, ijoblist);
1534 zfree(aiol_zone, ujoblist);
1535 error = EAGAIN;
1536 goto done;
1537 }
1538 }
1539
1540 /* NOTREACHED */
1541 error = EINVAL;
1542 done:
1543 rel_mplock();
1544 return (error);
1545 #endif /* VFS_AIO */
1546 }
1547
1548 /*
1549 * aio_cancel cancels any non-physio aio operations not currently in
1550 * progress.
1551 *
1552 * MPALMOSTSAFE
1553 */
1554 int
1555 sys_aio_cancel(struct aio_cancel_args *uap)
1556 {
1557 #ifndef VFS_AIO
1558 return ENOSYS;
1559 #else
1560 struct proc *p = curproc;
1561 struct kaioinfo *ki;
1562 struct aiocblist *cbe, *cbn;
1563 struct file *fp;
1564 struct socket *so;
1565 struct proc *po;
1566 int error;
1567 int cancelled=0;
1568 int notcancelled=0;
1569 struct vnode *vp;
1570
1571 fp = holdfp(p->p_fd, uap->fd, -1);
1572 if (fp == NULL)
1573 return (EBADF);
1574
1575 get_mplock();
1576
1577 if (fp->f_type == DTYPE_VNODE) {
1578 vp = (struct vnode *)fp->f_data;
1579
1580 if (vn_isdisk(vp,&error)) {
1581 uap->sysmsg_result = AIO_NOTCANCELED;
1582 error = 0;
1583 goto done2;
1584 }
1585 } else if (fp->f_type == DTYPE_SOCKET) {
1586 so = (struct socket *)fp->f_data;
1587
1588 crit_enter();
1589
1590 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1591 cbn = TAILQ_NEXT(cbe, list);
1592 if ((uap->aiocbp == NULL) ||
1593 (uap->aiocbp == cbe->uuaiocb) ) {
1594 po = cbe->userproc;
1595 ki = po->p_aioinfo;
1596 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1597 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1598 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1599 if (ki->kaio_flags & KAIO_WAKEUP) {
1600 wakeup(po);
1601 }
1602 cbe->jobstate = JOBST_JOBFINISHED;
1603 cbe->uaiocb._aiocb_private.status=-1;
1604 cbe->uaiocb._aiocb_private.error=ECANCELED;
1605 cancelled++;
1606 /* XXX cancelled, knote? */
1607 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1608 SIGEV_SIGNAL)
1609 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1610 if (uap->aiocbp)
1611 break;
1612 }
1613 }
1614 crit_exit();
1615
1616 if ((cancelled) && (uap->aiocbp)) {
1617 uap->sysmsg_result = AIO_CANCELED;
1618 error = 0;
1619 goto done2;
1620 }
1621 }
1622 ki=p->p_aioinfo;
1623 if (ki == NULL)
1624 goto done;
1625 crit_enter();
1626
1627 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1628 cbn = TAILQ_NEXT(cbe, plist);
1629
1630 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1631 ((uap->aiocbp == NULL ) ||
1632 (uap->aiocbp == cbe->uuaiocb))) {
1633
1634 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1635 TAILQ_REMOVE(&aio_jobs, cbe, list);
1636 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1637 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1638 plist);
1639 cancelled++;
1640 ki->kaio_queue_finished_count++;
1641 cbe->jobstate = JOBST_JOBFINISHED;
1642 cbe->uaiocb._aiocb_private.status = -1;
1643 cbe->uaiocb._aiocb_private.error = ECANCELED;
1644 /* XXX cancelled, knote? */
1645 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1646 SIGEV_SIGNAL)
1647 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1648 } else {
1649 notcancelled++;
1650 }
1651 }
1652 }
1653 crit_exit();
1654 done:
1655 if (notcancelled)
1656 uap->sysmsg_result = AIO_NOTCANCELED;
1657 else if (cancelled)
1658 uap->sysmsg_result = AIO_CANCELED;
1659 else
1660 uap->sysmsg_result = AIO_ALLDONE;
1661 error = 0;
1662 done2:
1663 rel_mplock();
1664 fdrop(fp);
1665 return error;
1666 #endif /* VFS_AIO */
1667 }
1668
1669 /*
1670 * aio_error is implemented in the kernel level for compatibility purposes only.
1671 * For a user mode async implementation, it would be best to do it in a userland
1672 * subroutine.
1673 *
1674 * MPALMOSTSAFE
1675 */
1676 int
1677 sys_aio_error(struct aio_error_args *uap)
1678 {
1679 #ifndef VFS_AIO
1680 return ENOSYS;
1681 #else
1682 struct proc *p = curproc;
1683 struct aiocblist *cb;
1684 struct kaioinfo *ki;
1685 long jobref;
1686 int error;
1687
1688 ki = p->p_aioinfo;
1689 if (ki == NULL)
1690 return EINVAL;
1691
1692 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1693 if ((jobref == -1) || (jobref == 0))
1694 return EINVAL;
1695
1696 get_mplock();
1697 error = 0;
1698
1699 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1700 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1701 jobref) {
1702 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1703 goto done;
1704 }
1705 }
1706
1707 crit_enter();
1708
1709 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1710 plist)) {
1711 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1712 jobref) {
1713 uap->sysmsg_result = EINPROGRESS;
1714 crit_exit();
1715 goto done;
1716 }
1717 }
1718
1719 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1720 plist)) {
1721 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1722 jobref) {
1723 uap->sysmsg_result = EINPROGRESS;
1724 crit_exit();
1725 goto done;
1726 }
1727 }
1728 crit_exit();
1729
1730 crit_enter();
1731 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1732 plist)) {
1733 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1734 jobref) {
1735 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1736 crit_exit();
1737 goto done;
1738 }
1739 }
1740
1741 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1742 plist)) {
1743 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1744 jobref) {
1745 uap->sysmsg_result = EINPROGRESS;
1746 crit_exit();
1747 goto done;
1748 }
1749 }
1750 crit_exit();
1751 error = EINVAL;
1752 done:
1753 rel_mplock();
1754 return (error);
1755 #endif /* VFS_AIO */
1756 }
1757
1758 /*
1759 * syscall - asynchronous read from a file (REALTIME)
1760 *
1761 * MPALMOSTSAFE
1762 */
1763 int
1764 sys_aio_read(struct aio_read_args *uap)
1765 {
1766 #ifndef VFS_AIO
1767 return ENOSYS;
1768 #else
1769 int error;
1770
1771 get_mplock();
1772 error = aio_aqueue(uap->aiocbp, LIO_READ);
1773 rel_mplock();
1774 return (error);
1775 #endif /* VFS_AIO */
1776 }
1777
1778 /*
1779 * syscall - asynchronous write to a file (REALTIME)
1780 *
1781 * MPALMOSTSAFE
1782 */
1783 int
1784 sys_aio_write(struct aio_write_args *uap)
1785 {
1786 #ifndef VFS_AIO
1787 return ENOSYS;
1788 #else
1789 int error;
1790
1791 get_mplock();
1792 error = aio_aqueue(uap->aiocbp, LIO_WRITE);
1793 rel_mplock();
1794 return (error);
1795 #endif /* VFS_AIO */
1796 }
1797
1798 /*
1799 * syscall - XXX undocumented
1800 *
1801 * MPALMOSTSAFE
1802 */
1803 int
1804 sys_lio_listio(struct lio_listio_args *uap)
1805 {
1806 #ifndef VFS_AIO
1807 return ENOSYS;
1808 #else
1809 struct proc *p = curproc;
1810 struct lwp *lp = curthread->td_lwp;
1811 int nent, nentqueued;
1812 struct aiocb *iocb, * const *cbptr;
1813 struct aiocblist *cb;
1814 struct kaioinfo *ki;
1815 struct aio_liojob *lj;
1816 int error, runningcode;
1817 int nerror;
1818 int i;
1819
1820 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1821 return EINVAL;
1822
1823 nent = uap->nent;
1824 if (nent > AIO_LISTIO_MAX)
1825 return EINVAL;
1826
1827 get_mplock();
1828
1829 if (p->p_aioinfo == NULL)
1830 aio_init_aioinfo(p);
1831
1832 if ((nent + num_queue_count) > max_queue_count) {
1833 error = EAGAIN;
1834 goto done;
1835 }
1836
1837 ki = p->p_aioinfo;
1838 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) {
1839 error = EAGAIN;
1840 goto done;
1841 }
1842
1843 lj = zalloc(aiolio_zone);
1844 if (lj == NULL) {
1845 error = EAGAIN;
1846 goto done;
1847 }
1848
1849 lj->lioj_flags = 0;
1850 lj->lioj_buffer_count = 0;
1851 lj->lioj_buffer_finished_count = 0;
1852 lj->lioj_queue_count = 0;
1853 lj->lioj_queue_finished_count = 0;
1854 lj->lioj_ki = ki;
1855
1856 /*
1857 * Setup signal.
1858 */
1859 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1860 error = copyin(uap->sig, &lj->lioj_signal,
1861 sizeof(lj->lioj_signal));
1862 if (error) {
1863 zfree(aiolio_zone, lj);
1864 goto done;
1865 }
1866 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1867 zfree(aiolio_zone, lj);
1868 error = EINVAL;
1869 goto done;
1870 }
1871 lj->lioj_flags |= LIOJ_SIGNAL;
1872 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1873 } else
1874 lj->lioj_flags &= ~LIOJ_SIGNAL;
1875
1876 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1877 /*
1878 * Get pointers to the list of I/O requests.
1879 */
1880 nerror = 0;
1881 nentqueued = 0;
1882 cbptr = uap->acb_list;
1883 for (i = 0; i < uap->nent; i++) {
1884 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1885 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1886 error = _aio_aqueue(iocb, lj, 0);
1887 if (error == 0)
1888 nentqueued++;
1889 else
1890 nerror++;
1891 }
1892 }
1893
1894 /*
1895 * If we haven't queued any, then just return error.
1896 */
1897 if (nentqueued == 0) {
1898 error = 0;
1899 goto done;
1900 }
1901
1902 /*
1903 * Calculate the appropriate error return.
1904 */
1905 runningcode = 0;
1906 if (nerror)
1907 runningcode = EIO;
1908
1909 if (uap->mode == LIO_WAIT) {
1910 int command, found, jobref;
1911
1912 for (;;) {
1913 found = 0;
1914 for (i = 0; i < uap->nent; i++) {
1915 /*
1916 * Fetch address of the control buf pointer in
1917 * user space.
1918 */
1919 iocb = (struct aiocb *)
1920 (intptr_t)fuword(&cbptr[i]);
1921 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1922 == 0))
1923 continue;
1924
1925 /*
1926 * Fetch the associated command from user space.
1927 */
1928 command = fuword(&iocb->aio_lio_opcode);
1929 if (command == LIO_NOP) {
1930 found++;
1931 continue;
1932 }
1933
1934 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1935
1936 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1937 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1938 == jobref) {
1939 if (cb->uaiocb.aio_lio_opcode
1940 == LIO_WRITE) {
1941 lp->lwp_ru.ru_oublock +=
1942 cb->outputcharge;
1943 cb->outputcharge = 0;
1944 } else if (cb->uaiocb.aio_lio_opcode
1945 == LIO_READ) {
1946 lp->lwp_ru.ru_inblock +=
1947 cb->inputcharge;
1948 cb->inputcharge = 0;
1949 }
1950 found++;
1951 break;
1952 }
1953 }
1954
1955 crit_enter();
1956 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1957 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1958 == jobref) {
1959 found++;
1960 break;
1961 }
1962 }
1963 crit_exit();
1964 }
1965
1966 /*
1967 * If all I/Os have been disposed of, then we can
1968 * return.
1969 */
1970 if (found == nentqueued) {
1971 error = runningcode;
1972 goto done;
1973 }
1974
1975 ki->kaio_flags |= KAIO_WAKEUP;
1976 error = tsleep(p, PCATCH, "aiospn", 0);
1977
1978 if (error == EINTR) {
1979 goto done;
1980 } else if (error == EWOULDBLOCK) {
1981 error = EAGAIN;
1982 goto done;
1983 }
1984 }
1985 }
1986
1987 error = runningcode;
1988 done:
1989 rel_mplock();
1990 return (error);
1991 #endif /* VFS_AIO */
1992 }
1993
1994 #ifdef VFS_AIO
1995 /*
1996 * This is a weird hack so that we can post a signal. It is safe to do so from
1997 * a timeout routine, but *not* from an interrupt routine.
1998 */
1999 static void
2000 process_signal(void *aioj)
2001 {
2002 struct aiocblist *aiocbe = aioj;
2003 struct aio_liojob *lj = aiocbe->lio;
2004 struct aiocb *cb = &aiocbe->uaiocb;
2005
2006 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2007 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2008 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2009 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2010 }
2011
2012 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2013 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2014 }
2015
2016 /*
2017 * Interrupt handler for physio, performs the necessary process wakeups, and
2018 * signals.
2019 */
2020 static void
2021 aio_physwakeup(struct bio *bio)
2022 {
2023 struct buf *bp = bio->bio_buf;
2024 struct aiocblist *aiocbe;
2025 struct proc *p;
2026 struct kaioinfo *ki;
2027 struct aio_liojob *lj;
2028
2029 aiocbe = bio->bio_caller_info2.ptr;
2030
2031 if (aiocbe) {
2032 p = bio->bio_caller_info1.ptr;
2033
2034 aiocbe->jobstate = JOBST_JOBBFINISHED;
2035 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2036 aiocbe->uaiocb._aiocb_private.error = 0;
2037 aiocbe->jobflags |= AIOCBLIST_DONE;
2038
2039 if (bp->b_flags & B_ERROR)
2040 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2041
2042 lj = aiocbe->lio;
2043 if (lj) {
2044 lj->lioj_buffer_finished_count++;
2045
2046 /*
2047 * wakeup/signal if all of the interrupt jobs are done.
2048 */
2049 if (lj->lioj_buffer_finished_count ==
2050 lj->lioj_buffer_count) {
2051 /*
2052 * Post a signal if it is called for.
2053 */
2054 if ((lj->lioj_flags &
2055 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2056 LIOJ_SIGNAL) {
2057 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2058 callout_reset(&aiocbe->timeout, 0,
2059 process_signal, aiocbe);
2060 }
2061 }
2062 }
2063
2064 ki = p->p_aioinfo;
2065 if (ki) {
2066 ki->kaio_buffer_finished_count++;
2067 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2068 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2069 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2070
2071 KNOTE(&aiocbe->klist, 0);
2072 /* Do the wakeup. */
2073 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2074 ki->kaio_flags &= ~KAIO_WAKEUP;
2075 wakeup(p);
2076 }
2077 }
2078
2079 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2080 callout_reset(&aiocbe->timeout, 0,
2081 process_signal, aiocbe);
2082 }
2083 }
2084 biodone_sync(bio);
2085 }
2086 #endif /* VFS_AIO */
2087
2088 /*
2089 * syscall - wait for the next completion of an aio request
2090 *
2091 * MPALMOSTSAFE
2092 */
2093 int
2094 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2095 {
2096 #ifndef VFS_AIO
2097 return ENOSYS;
2098 #else
2099 struct proc *p = curproc;
2100 struct lwp *lp = curthread->td_lwp;
2101 struct timeval atv;
2102 struct timespec ts;
2103 struct kaioinfo *ki;
2104 struct aiocblist *cb = NULL;
2105 int error, timo;
2106
2107 suword(uap->aiocbp, (int)NULL);
2108
2109 timo = 0;
2110 if (uap->timeout) {
2111 /* Get timespec struct. */
2112 error = copyin(uap->timeout, &ts, sizeof(ts));
2113 if (error)
2114 return error;
2115
2116 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2117 return (EINVAL);
2118
2119 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2120 if (itimerfix(&atv))
2121 return (EINVAL);
2122 timo = tvtohz_high(&atv);
2123 }
2124
2125 ki = p->p_aioinfo;
2126 if (ki == NULL)
2127 return EAGAIN;
2128
2129 get_mplock();
2130
2131 for (;;) {
2132 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2133 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2134 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2135 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2136 lp->lwp_ru.ru_oublock +=
2137 cb->outputcharge;
2138 cb->outputcharge = 0;
2139 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2140 lp->lwp_ru.ru_inblock += cb->inputcharge;
2141 cb->inputcharge = 0;
2142 }
2143 aio_free_entry(cb);
2144 error = cb->uaiocb._aiocb_private.error;
2145 break;
2146 }
2147
2148 crit_enter();
2149 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2150 crit_exit();
2151 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2152 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2153 aio_free_entry(cb);
2154 error = cb->uaiocb._aiocb_private.error;
2155 break;
2156 }
2157
2158 ki->kaio_flags |= KAIO_WAKEUP;
2159 error = tsleep(p, PCATCH, "aiowc", timo);
2160 crit_exit();
2161
2162 if (error == ERESTART) {
2163 error = EINTR;
2164 break;
2165 }
2166 if (error < 0)
2167 break;
2168 if (error == EINTR)
2169 break;
2170 if (error == EWOULDBLOCK) {
2171 error = EAGAIN;
2172 break;
2173 }
2174 }
2175 rel_mplock();
2176 return (error);
2177 #endif /* VFS_AIO */
2178 }
2179
2180 #ifndef VFS_AIO
2181 static int
2182 filt_aioattach(struct knote *kn)
2183 {
2184
2185 return (ENXIO);
2186 }
2187
2188 struct filterops aio_filtops =
2189 { 0, filt_aioattach, NULL, NULL };
2190
2191 #else
2192 /* kqueue attach function */
2193 static int
2194 filt_aioattach(struct knote *kn)
2195 {
2196 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2197
2198 /*
2199 * The aiocbe pointer must be validated before using it, so
2200 * registration is restricted to the kernel; the user cannot
2201 * set EV_FLAG1.
2202 */
2203 if ((kn->kn_flags & EV_FLAG1) == 0)
2204 return (EPERM);
2205 kn->kn_flags &= ~EV_FLAG1;
2206
2207 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2208
2209 return (0);
2210 }
2211
2212 /* kqueue detach function */
2213 static void
2214 filt_aiodetach(struct knote *kn)
2215 {
2216 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2217
2218 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2219 }
2220
2221 /* kqueue filter function */
2222 /*ARGSUSED*/
2223 static int
2224 filt_aio(struct knote *kn, long hint)
2225 {
2226 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2227
2228 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2229 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2230 aiocbe->jobstate != JOBST_JOBBFINISHED)
2231 return (0);
2232 kn->kn_flags |= EV_EOF;
2233 return (1);
2234 }
2235
2236 struct filterops aio_filtops =
2237 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2238 #endif /* VFS_AIO */
DragonFly BSD