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 * $FreeBSD: releng/5.1/sys/kern/vfs_aio.c 114216 2003-04-29 13:36:06Z kan $
17 */
18
19 /*
20 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21 */
22
23 #include <sys/param.h>
24 #include <sys/systm.h>
25 #include <sys/malloc.h>
26 #include <sys/bio.h>
27 #include <sys/buf.h>
28 #include <sys/eventhandler.h>
29 #include <sys/sysproto.h>
30 #include <sys/filedesc.h>
31 #include <sys/kernel.h>
32 #include <sys/kthread.h>
33 #include <sys/fcntl.h>
34 #include <sys/file.h>
35 #include <sys/limits.h>
36 #include <sys/lock.h>
37 #include <sys/mutex.h>
38 #include <sys/unistd.h>
39 #include <sys/proc.h>
40 #include <sys/resourcevar.h>
41 #include <sys/signalvar.h>
42 #include <sys/protosw.h>
43 #include <sys/socketvar.h>
44 #include <sys/syscall.h>
45 #include <sys/sysent.h>
46 #include <sys/sysctl.h>
47 #include <sys/sx.h>
48 #include <sys/vnode.h>
49 #include <sys/conf.h>
50 #include <sys/event.h>
51
52 #include <posix4/posix4.h>
53 #include <vm/vm.h>
54 #include <vm/vm_extern.h>
55 #include <vm/pmap.h>
56 #include <vm/vm_map.h>
57 #include <vm/uma.h>
58 #include <sys/aio.h>
59
60 #include "opt_vfs_aio.h"
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 unloadable = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
153 "Allow unload of aio (not recommended)");
154
155
156 static int max_aio_per_proc = MAX_AIO_PER_PROC;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
158 0, "Maximum active aio requests per process (stored in the process)");
159
160 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
162 &max_aio_queue_per_proc, 0,
163 "Maximum queued aio requests per process (stored in the process)");
164
165 static int max_buf_aio = MAX_BUF_AIO;
166 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
167 "Maximum buf aio requests per process (stored in the process)");
168
169 struct aiocblist {
170 TAILQ_ENTRY(aiocblist) list; /* List of jobs */
171 TAILQ_ENTRY(aiocblist) plist; /* List of jobs for proc */
172 int jobflags;
173 int jobstate;
174 int inputcharge;
175 int outputcharge;
176 struct callout_handle timeouthandle;
177 struct buf *bp; /* Buffer pointer */
178 struct proc *userproc; /* User process */ /* Not td! */
179 struct ucred *cred; /* Active credential when created */
180 struct file *fd_file; /* Pointer to file structure */
181 struct aio_liojob *lio; /* Optional lio job */
182 struct aiocb *uuaiocb; /* Pointer in userspace of aiocb */
183 struct klist klist; /* list of knotes */
184 struct aiocb uaiocb; /* Kernel I/O control block */
185 };
186
187 /* jobflags */
188 #define AIOCBLIST_RUNDOWN 0x4
189 #define AIOCBLIST_DONE 0x10
190
191 /*
192 * AIO process info
193 */
194 #define AIOP_FREE 0x1 /* proc on free queue */
195 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
196
197 struct aiothreadlist {
198 int aiothreadflags; /* AIO proc flags */
199 TAILQ_ENTRY(aiothreadlist) list; /* List of processes */
200 struct thread *aiothread; /* The AIO thread */
201 };
202
203 /*
204 * data-structure for lio signal management
205 */
206 struct aio_liojob {
207 int lioj_flags;
208 int lioj_buffer_count;
209 int lioj_buffer_finished_count;
210 int lioj_queue_count;
211 int lioj_queue_finished_count;
212 struct sigevent lioj_signal; /* signal on all I/O done */
213 TAILQ_ENTRY(aio_liojob) lioj_list;
214 struct kaioinfo *lioj_ki;
215 };
216 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
217 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
218
219 /*
220 * per process aio data structure
221 */
222 struct kaioinfo {
223 int kaio_flags; /* per process kaio flags */
224 int kaio_maxactive_count; /* maximum number of AIOs */
225 int kaio_active_count; /* number of currently used AIOs */
226 int kaio_qallowed_count; /* maxiumu size of AIO queue */
227 int kaio_queue_count; /* size of AIO queue */
228 int kaio_ballowed_count; /* maximum number of buffers */
229 int kaio_queue_finished_count; /* number of daemon jobs finished */
230 int kaio_buffer_count; /* number of physio buffers */
231 int kaio_buffer_finished_count; /* count of I/O done */
232 struct proc *kaio_p; /* process that uses this kaio block */
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_activeproc; /* Active daemons */
245 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* Idle daemons */
246 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
247 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O 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 void process_signal(void *aioj);
263 static int filt_aioattach(struct knote *kn);
264 static void filt_aiodetach(struct knote *kn);
265 static int filt_aio(struct knote *kn, long hint);
266
267 /*
268 * Zones for:
269 * kaio Per process async io info
270 * aiop async io thread data
271 * aiocb async io jobs
272 * aiol list io job pointer - internal to aio_suspend XXX
273 * aiolio list io jobs
274 */
275 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
276
277 /* kqueue filters for aio */
278 static struct filterops aio_filtops =
279 { 0, filt_aioattach, filt_aiodetach, filt_aio };
280
281 static eventhandler_tag exit_tag, exec_tag;
282
283 /*
284 * Main operations function for use as a kernel module.
285 */
286 static int
287 aio_modload(struct module *module, int cmd, void *arg)
288 {
289 int error = 0;
290
291 switch (cmd) {
292 case MOD_LOAD:
293 aio_onceonly();
294 break;
295 case MOD_UNLOAD:
296 error = aio_unload();
297 break;
298 case MOD_SHUTDOWN:
299 break;
300 default:
301 error = EINVAL;
302 break;
303 }
304 return (error);
305 }
306
307 static moduledata_t aio_mod = {
308 "aio",
309 &aio_modload,
310 NULL
311 };
312
313 SYSCALL_MODULE_HELPER(aio_return);
314 SYSCALL_MODULE_HELPER(aio_suspend);
315 SYSCALL_MODULE_HELPER(aio_cancel);
316 SYSCALL_MODULE_HELPER(aio_error);
317 SYSCALL_MODULE_HELPER(aio_read);
318 SYSCALL_MODULE_HELPER(aio_write);
319 SYSCALL_MODULE_HELPER(aio_waitcomplete);
320 SYSCALL_MODULE_HELPER(lio_listio);
321
322 DECLARE_MODULE(aio, aio_mod,
323 SI_SUB_VFS, SI_ORDER_ANY);
324 MODULE_VERSION(aio, 1);
325
326 /*
327 * Startup initialization
328 */
329 static void
330 aio_onceonly(void)
331 {
332
333 /* XXX: should probably just use so->callback */
334 aio_swake = &aio_swake_cb;
335 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
336 EVENTHANDLER_PRI_ANY);
337 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown, NULL,
338 EVENTHANDLER_PRI_ANY);
339 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
340 TAILQ_INIT(&aio_freeproc);
341 TAILQ_INIT(&aio_activeproc);
342 TAILQ_INIT(&aio_jobs);
343 TAILQ_INIT(&aio_bufjobs);
344 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
345 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
346 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
347 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
348 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
349 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
350 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
351 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
352 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL,
353 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
354 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
355 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
356 jobrefid = 1;
357 async_io_version = _POSIX_VERSION;
358 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
359 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
360 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
361 }
362
363 /*
364 * Callback for unload of AIO when used as a module.
365 */
366 static int
367 aio_unload(void)
368 {
369
370 /*
371 * XXX: no unloads by default, it's too dangerous.
372 * perhaps we could do it if locked out callers and then
373 * did an aio_proc_rundown() on each process.
374 */
375 if (!unloadable)
376 return (EOPNOTSUPP);
377
378 async_io_version = 0;
379 aio_swake = NULL;
380 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
381 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
382 kqueue_del_filteropts(EVFILT_AIO);
383 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
384 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
385 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
386 return (0);
387 }
388
389 /*
390 * Init the per-process aioinfo structure. The aioinfo limits are set
391 * per-process for user limit (resource) management.
392 */
393 static void
394 aio_init_aioinfo(struct proc *p)
395 {
396 struct kaioinfo *ki;
397
398 if (p->p_aioinfo == NULL) {
399 ki = uma_zalloc(kaio_zone, M_WAITOK);
400 p->p_aioinfo = ki;
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 ki->kaio_p = p;
410 TAILQ_INIT(&ki->kaio_jobdone);
411 TAILQ_INIT(&ki->kaio_jobqueue);
412 TAILQ_INIT(&ki->kaio_bufdone);
413 TAILQ_INIT(&ki->kaio_bufqueue);
414 TAILQ_INIT(&ki->kaio_liojoblist);
415 TAILQ_INIT(&ki->kaio_sockqueue);
416 }
417
418 while (num_aio_procs < target_aio_procs)
419 aio_newproc();
420 }
421
422 /*
423 * Free a job entry. Wait for completion if it is currently active, but don't
424 * delay forever. If we delay, we return a flag that says that we have to
425 * restart the queue scan.
426 */
427 static int
428 aio_free_entry(struct aiocblist *aiocbe)
429 {
430 struct kaioinfo *ki;
431 struct aio_liojob *lj;
432 struct proc *p;
433 int error;
434 int s;
435
436 if (aiocbe->jobstate == JOBST_NULL)
437 panic("aio_free_entry: freeing already free job");
438
439 p = aiocbe->userproc;
440 ki = p->p_aioinfo;
441 lj = aiocbe->lio;
442 if (ki == NULL)
443 panic("aio_free_entry: missing p->p_aioinfo");
444
445 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
446 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
447 tsleep(aiocbe, PRIBIO, "jobwai", 0);
448 }
449 if (aiocbe->bp == NULL) {
450 if (ki->kaio_queue_count <= 0)
451 panic("aio_free_entry: process queue size <= 0");
452 if (num_queue_count <= 0)
453 panic("aio_free_entry: system wide queue size <= 0");
454
455 if (lj) {
456 lj->lioj_queue_count--;
457 if (aiocbe->jobflags & AIOCBLIST_DONE)
458 lj->lioj_queue_finished_count--;
459 }
460 ki->kaio_queue_count--;
461 if (aiocbe->jobflags & AIOCBLIST_DONE)
462 ki->kaio_queue_finished_count--;
463 num_queue_count--;
464 } else {
465 if (lj) {
466 lj->lioj_buffer_count--;
467 if (aiocbe->jobflags & AIOCBLIST_DONE)
468 lj->lioj_buffer_finished_count--;
469 }
470 if (aiocbe->jobflags & AIOCBLIST_DONE)
471 ki->kaio_buffer_finished_count--;
472 ki->kaio_buffer_count--;
473 num_buf_aio--;
474 }
475
476 /* aiocbe is going away, we need to destroy any knotes */
477 /* XXXKSE Note the thread here is used to eventually find the
478 * owning process again, but it is also used to do a fo_close
479 * and that requires the thread. (but does it require the
480 * OWNING thread? (or maybe the running thread?)
481 * There is a semantic problem here...
482 */
483 knote_remove(FIRST_THREAD_IN_PROC(p), &aiocbe->klist); /* XXXKSE */
484
485 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
486 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
487 ki->kaio_flags &= ~KAIO_WAKEUP;
488 wakeup(p);
489 }
490
491 if (aiocbe->jobstate == JOBST_JOBQBUF) {
492 if ((error = aio_fphysio(aiocbe)) != 0)
493 return (error);
494 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
495 panic("aio_free_entry: invalid physio finish-up state");
496 s = splbio();
497 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
498 splx(s);
499 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
500 s = splnet();
501 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
502 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
503 splx(s);
504 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
505 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
506 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
507 s = splbio();
508 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
509 splx(s);
510 if (aiocbe->bp) {
511 vunmapbuf(aiocbe->bp);
512 relpbuf(aiocbe->bp, NULL);
513 aiocbe->bp = NULL;
514 }
515 }
516 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
517 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
518 uma_zfree(aiolio_zone, lj);
519 }
520 aiocbe->jobstate = JOBST_NULL;
521 untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
522 fdrop(aiocbe->fd_file, curthread);
523 crfree(aiocbe->cred);
524 uma_zfree(aiocb_zone, aiocbe);
525 return (0);
526 }
527
528 /*
529 * Rundown the jobs for a given process.
530 */
531 static void
532 aio_proc_rundown(void *arg, struct proc *p)
533 {
534 int s;
535 struct kaioinfo *ki;
536 struct aio_liojob *lj, *ljn;
537 struct aiocblist *aiocbe, *aiocbn;
538 struct file *fp;
539 struct socket *so;
540
541 ki = p->p_aioinfo;
542 if (ki == NULL)
543 return;
544
545 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
546 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
547 ki->kaio_buffer_finished_count)) {
548 ki->kaio_flags |= KAIO_RUNDOWN;
549 if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
550 break;
551 }
552
553 /*
554 * Move any aio ops that are waiting on socket I/O to the normal job
555 * queues so they are cleaned up with any others.
556 */
557 s = splnet();
558 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
559 aiocbn) {
560 aiocbn = TAILQ_NEXT(aiocbe, plist);
561 fp = aiocbe->fd_file;
562 if (fp != NULL) {
563 so = fp->f_data;
564 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
565 if (TAILQ_EMPTY(&so->so_aiojobq)) {
566 so->so_snd.sb_flags &= ~SB_AIO;
567 so->so_rcv.sb_flags &= ~SB_AIO;
568 }
569 }
570 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
571 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
572 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
573 }
574 splx(s);
575
576 restart1:
577 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
578 aiocbn = TAILQ_NEXT(aiocbe, plist);
579 if (aio_free_entry(aiocbe))
580 goto restart1;
581 }
582
583 restart2:
584 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
585 aiocbn) {
586 aiocbn = TAILQ_NEXT(aiocbe, plist);
587 if (aio_free_entry(aiocbe))
588 goto restart2;
589 }
590
591 /*
592 * Note the use of lots of splbio here, trying to avoid splbio for long chains
593 * of I/O. Probably unnecessary.
594 */
595 restart3:
596 s = splbio();
597 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
598 ki->kaio_flags |= KAIO_WAKEUP;
599 tsleep(p, PRIBIO, "aioprn", 0);
600 splx(s);
601 goto restart3;
602 }
603 splx(s);
604
605 restart4:
606 s = splbio();
607 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
608 aiocbn = TAILQ_NEXT(aiocbe, plist);
609 if (aio_free_entry(aiocbe)) {
610 splx(s);
611 goto restart4;
612 }
613 }
614 splx(s);
615
616 /*
617 * If we've slept, jobs might have moved from one queue to another.
618 * Retry rundown if we didn't manage to empty the queues.
619 */
620 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
621 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
622 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
623 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
624 goto restart1;
625
626 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
627 ljn = TAILQ_NEXT(lj, lioj_list);
628 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
629 0)) {
630 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
631 uma_zfree(aiolio_zone, lj);
632 } else {
633 #ifdef DIAGNOSTIC
634 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
635 "QF:%d\n", lj->lioj_buffer_count,
636 lj->lioj_buffer_finished_count,
637 lj->lioj_queue_count,
638 lj->lioj_queue_finished_count);
639 #endif
640 }
641 }
642
643 uma_zfree(kaio_zone, ki);
644 p->p_aioinfo = NULL;
645 }
646
647 /*
648 * Select a job to run (called by an AIO daemon).
649 */
650 static struct aiocblist *
651 aio_selectjob(struct aiothreadlist *aiop)
652 {
653 int s;
654 struct aiocblist *aiocbe;
655 struct kaioinfo *ki;
656 struct proc *userp;
657
658 s = splnet();
659 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
660 TAILQ_NEXT(aiocbe, list)) {
661 userp = aiocbe->userproc;
662 ki = userp->p_aioinfo;
663
664 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
665 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
666 splx(s);
667 return (aiocbe);
668 }
669 }
670 splx(s);
671
672 return (NULL);
673 }
674
675 /*
676 * The AIO processing activity. This is the code that does the I/O request for
677 * the non-physio version of the operations. The normal vn operations are used,
678 * and this code should work in all instances for every type of file, including
679 * pipes, sockets, fifos, and regular files.
680 */
681 static void
682 aio_process(struct aiocblist *aiocbe)
683 {
684 struct ucred *td_savedcred;
685 struct thread *td;
686 struct proc *mycp;
687 struct aiocb *cb;
688 struct file *fp;
689 struct uio auio;
690 struct iovec aiov;
691 int cnt;
692 int error;
693 int oublock_st, oublock_end;
694 int inblock_st, inblock_end;
695
696 td = curthread;
697 td_savedcred = td->td_ucred;
698 td->td_ucred = aiocbe->cred;
699 mycp = td->td_proc;
700 cb = &aiocbe->uaiocb;
701 fp = aiocbe->fd_file;
702
703 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
704 aiov.iov_len = cb->aio_nbytes;
705
706 auio.uio_iov = &aiov;
707 auio.uio_iovcnt = 1;
708 auio.uio_offset = cb->aio_offset;
709 auio.uio_resid = cb->aio_nbytes;
710 cnt = cb->aio_nbytes;
711 auio.uio_segflg = UIO_USERSPACE;
712 auio.uio_td = td;
713
714 inblock_st = mycp->p_stats->p_ru.ru_inblock;
715 oublock_st = mycp->p_stats->p_ru.ru_oublock;
716 /*
717 * _aio_aqueue() acquires a reference to the file that is
718 * released in aio_free_entry().
719 */
720 if (cb->aio_lio_opcode == LIO_READ) {
721 auio.uio_rw = UIO_READ;
722 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
723 } else {
724 auio.uio_rw = UIO_WRITE;
725 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
726 }
727 inblock_end = mycp->p_stats->p_ru.ru_inblock;
728 oublock_end = mycp->p_stats->p_ru.ru_oublock;
729
730 aiocbe->inputcharge = inblock_end - inblock_st;
731 aiocbe->outputcharge = oublock_end - oublock_st;
732
733 if ((error) && (auio.uio_resid != cnt)) {
734 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
735 error = 0;
736 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
737 PROC_LOCK(aiocbe->userproc);
738 psignal(aiocbe->userproc, SIGPIPE);
739 PROC_UNLOCK(aiocbe->userproc);
740 }
741 }
742
743 cnt -= auio.uio_resid;
744 cb->_aiocb_private.error = error;
745 cb->_aiocb_private.status = cnt;
746 td->td_ucred = td_savedcred;
747 }
748
749 /*
750 * The AIO daemon, most of the actual work is done in aio_process,
751 * but the setup (and address space mgmt) is done in this routine.
752 */
753 static void
754 aio_daemon(void *uproc)
755 {
756 int s;
757 struct aio_liojob *lj;
758 struct aiocb *cb;
759 struct aiocblist *aiocbe;
760 struct aiothreadlist *aiop;
761 struct kaioinfo *ki;
762 struct proc *curcp, *mycp, *userp;
763 struct vmspace *myvm, *tmpvm;
764 struct thread *td = curthread;
765 struct pgrp *newpgrp;
766 struct session *newsess;
767
768 mtx_lock(&Giant);
769 /*
770 * Local copies of curproc (cp) and vmspace (myvm)
771 */
772 mycp = td->td_proc;
773 myvm = mycp->p_vmspace;
774
775 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
776
777 /*
778 * Allocate and ready the aio control info. There is one aiop structure
779 * per daemon.
780 */
781 aiop = uma_zalloc(aiop_zone, M_WAITOK);
782 aiop->aiothread = td;
783 aiop->aiothreadflags |= AIOP_FREE;
784
785 s = splnet();
786
787 /*
788 * Place thread (lightweight process) onto the AIO free thread list.
789 */
790 if (TAILQ_EMPTY(&aio_freeproc))
791 wakeup(&aio_freeproc);
792 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
793
794 splx(s);
795
796 /*
797 * Get rid of our current filedescriptors. AIOD's don't need any
798 * filedescriptors, except as temporarily inherited from the client.
799 */
800 fdfree(td);
801
802 mtx_unlock(&Giant);
803 /* The daemon resides in its own pgrp. */
804 MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
805 M_WAITOK | M_ZERO);
806 MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
807 M_WAITOK | M_ZERO);
808
809 sx_xlock(&proctree_lock);
810 enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
811 sx_xunlock(&proctree_lock);
812 mtx_lock(&Giant);
813
814 /*
815 * Wakeup parent process. (Parent sleeps to keep from blasting away
816 * and creating too many daemons.)
817 */
818 wakeup(mycp);
819
820 for (;;) {
821 /*
822 * curcp is the current daemon process context.
823 * userp is the current user process context.
824 */
825 curcp = mycp;
826
827 /*
828 * Take daemon off of free queue
829 */
830 if (aiop->aiothreadflags & AIOP_FREE) {
831 s = splnet();
832 TAILQ_REMOVE(&aio_freeproc, aiop, list);
833 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
834 aiop->aiothreadflags &= ~AIOP_FREE;
835 splx(s);
836 }
837 aiop->aiothreadflags &= ~AIOP_SCHED;
838
839 /*
840 * Check for jobs.
841 */
842 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
843 cb = &aiocbe->uaiocb;
844 userp = aiocbe->userproc;
845
846 aiocbe->jobstate = JOBST_JOBRUNNING;
847
848 /*
849 * Connect to process address space for user program.
850 */
851 if (userp != curcp) {
852 /*
853 * Save the current address space that we are
854 * connected to.
855 */
856 tmpvm = mycp->p_vmspace;
857
858 /*
859 * Point to the new user address space, and
860 * refer to it.
861 */
862 mycp->p_vmspace = userp->p_vmspace;
863 mycp->p_vmspace->vm_refcnt++;
864
865 /* Activate the new mapping. */
866 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
867
868 /*
869 * If the old address space wasn't the daemons
870 * own address space, then we need to remove the
871 * daemon's reference from the other process
872 * that it was acting on behalf of.
873 */
874 if (tmpvm != myvm) {
875 vmspace_free(tmpvm);
876 }
877 curcp = userp;
878 }
879
880 ki = userp->p_aioinfo;
881 lj = aiocbe->lio;
882
883 /* Account for currently active jobs. */
884 ki->kaio_active_count++;
885
886 /* Do the I/O function. */
887 aio_process(aiocbe);
888
889 /* Decrement the active job count. */
890 ki->kaio_active_count--;
891
892 /*
893 * Increment the completion count for wakeup/signal
894 * comparisons.
895 */
896 aiocbe->jobflags |= AIOCBLIST_DONE;
897 ki->kaio_queue_finished_count++;
898 if (lj)
899 lj->lioj_queue_finished_count++;
900 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
901 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
902 ki->kaio_flags &= ~KAIO_WAKEUP;
903 wakeup(userp);
904 }
905
906 s = splbio();
907 if (lj && (lj->lioj_flags &
908 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
909 if ((lj->lioj_queue_finished_count ==
910 lj->lioj_queue_count) &&
911 (lj->lioj_buffer_finished_count ==
912 lj->lioj_buffer_count)) {
913 PROC_LOCK(userp);
914 psignal(userp,
915 lj->lioj_signal.sigev_signo);
916 PROC_UNLOCK(userp);
917 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
918 }
919 }
920 splx(s);
921
922 aiocbe->jobstate = JOBST_JOBFINISHED;
923
924 s = splnet();
925 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
926 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
927 splx(s);
928 KNOTE(&aiocbe->klist, 0);
929
930 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
931 wakeup(aiocbe);
932 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
933 }
934
935 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
936 PROC_LOCK(userp);
937 psignal(userp, cb->aio_sigevent.sigev_signo);
938 PROC_UNLOCK(userp);
939 }
940 }
941
942 /*
943 * Disconnect from user address space.
944 */
945 if (curcp != mycp) {
946 /* Get the user address space to disconnect from. */
947 tmpvm = mycp->p_vmspace;
948
949 /* Get original address space for daemon. */
950 mycp->p_vmspace = myvm;
951
952 /* Activate the daemon's address space. */
953 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
954 #ifdef DIAGNOSTIC
955 if (tmpvm == myvm) {
956 printf("AIOD: vmspace problem -- %d\n",
957 mycp->p_pid);
958 }
959 #endif
960 /* Remove our vmspace reference. */
961 vmspace_free(tmpvm);
962
963 curcp = mycp;
964 }
965
966 /*
967 * If we are the first to be put onto the free queue, wakeup
968 * anyone waiting for a daemon.
969 */
970 s = splnet();
971 TAILQ_REMOVE(&aio_activeproc, aiop, list);
972 if (TAILQ_EMPTY(&aio_freeproc))
973 wakeup(&aio_freeproc);
974 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
975 aiop->aiothreadflags |= AIOP_FREE;
976 splx(s);
977
978 /*
979 * If daemon is inactive for a long time, allow it to exit,
980 * thereby freeing resources.
981 */
982 if ((aiop->aiothreadflags & AIOP_SCHED) == 0 &&
983 tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) {
984 s = splnet();
985 if (TAILQ_EMPTY(&aio_jobs)) {
986 if ((aiop->aiothreadflags & AIOP_FREE) &&
987 (num_aio_procs > target_aio_procs)) {
988 TAILQ_REMOVE(&aio_freeproc, aiop, list);
989 splx(s);
990 uma_zfree(aiop_zone, aiop);
991 num_aio_procs--;
992 #ifdef DIAGNOSTIC
993 if (mycp->p_vmspace->vm_refcnt <= 1) {
994 printf("AIOD: bad vm refcnt for"
995 " exiting daemon: %d\n",
996 mycp->p_vmspace->vm_refcnt);
997 }
998 #endif
999 kthread_exit(0);
1000 }
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_data;
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_rdev->si_bsize_phys)
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 bp->b_caller1 = p;
1100 bp->b_dev = vp->v_rdev;
1101 error = 0;
1102
1103 bp->b_bcount = cb->aio_nbytes;
1104 bp->b_bufsize = cb->aio_nbytes;
1105 bp->b_flags = B_PHYS;
1106 bp->b_iodone = aio_physwakeup;
1107 bp->b_saveaddr = bp->b_data;
1108 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1109 bp->b_blkno = btodb(cb->aio_offset);
1110 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1111
1112 /*
1113 * Bring buffer into kernel space.
1114 */
1115 if (vmapbuf(bp) < 0) {
1116 error = EFAULT;
1117 goto doerror;
1118 }
1119
1120 s = splbio();
1121 aiocbe->bp = bp;
1122 bp->b_spc = (void *)aiocbe;
1123 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1124 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1125 aiocbe->jobstate = JOBST_JOBQBUF;
1126 cb->_aiocb_private.status = cb->aio_nbytes;
1127 num_buf_aio++;
1128 bp->b_error = 0;
1129
1130 splx(s);
1131
1132 /* Perform transfer. */
1133 DEV_STRATEGY(bp);
1134
1135 notify = 0;
1136 s = splbio();
1137
1138 /*
1139 * If we had an error invoking the request, or an error in processing
1140 * the request before we have returned, we process it as an error in
1141 * transfer. Note that such an I/O error is not indicated immediately,
1142 * but is returned using the aio_error mechanism. In this case,
1143 * aio_suspend will return immediately.
1144 */
1145 if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1146 struct aiocb *job = aiocbe->uuaiocb;
1147
1148 aiocbe->uaiocb._aiocb_private.status = 0;
1149 suword(&job->_aiocb_private.status, 0);
1150 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1151 suword(&job->_aiocb_private.error, bp->b_error);
1152
1153 ki->kaio_buffer_finished_count++;
1154
1155 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1156 aiocbe->jobstate = JOBST_JOBBFINISHED;
1157 aiocbe->jobflags |= AIOCBLIST_DONE;
1158 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1159 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1160 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1161 notify = 1;
1162 }
1163 }
1164 splx(s);
1165 if (notify)
1166 KNOTE(&aiocbe->klist, 0);
1167 return (0);
1168
1169 doerror:
1170 ki->kaio_buffer_count--;
1171 if (lj)
1172 lj->lioj_buffer_count--;
1173 aiocbe->bp = NULL;
1174 relpbuf(bp, NULL);
1175 return (error);
1176 }
1177
1178 /*
1179 * This waits/tests physio completion.
1180 */
1181 static int
1182 aio_fphysio(struct aiocblist *iocb)
1183 {
1184 int s;
1185 struct buf *bp;
1186 int error;
1187
1188 bp = iocb->bp;
1189
1190 s = splbio();
1191 while ((bp->b_flags & B_DONE) == 0) {
1192 if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1193 if ((bp->b_flags & B_DONE) == 0) {
1194 splx(s);
1195 return (EINPROGRESS);
1196 } else
1197 break;
1198 }
1199 }
1200 splx(s);
1201
1202 /* Release mapping into kernel space. */
1203 vunmapbuf(bp);
1204 iocb->bp = 0;
1205
1206 error = 0;
1207
1208 /* Check for an error. */
1209 if (bp->b_ioflags & BIO_ERROR)
1210 error = bp->b_error;
1211
1212 relpbuf(bp, NULL);
1213 return (error);
1214 }
1215
1216 /*
1217 * Wake up aio requests that may be serviceable now.
1218 */
1219 static void
1220 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1221 {
1222 struct aiocblist *cb,*cbn;
1223 struct proc *p;
1224 struct kaioinfo *ki = NULL;
1225 int opcode, wakecount = 0;
1226 struct aiothreadlist *aiop;
1227
1228 if (sb == &so->so_snd) {
1229 opcode = LIO_WRITE;
1230 so->so_snd.sb_flags &= ~SB_AIO;
1231 } else {
1232 opcode = LIO_READ;
1233 so->so_rcv.sb_flags &= ~SB_AIO;
1234 }
1235
1236 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1237 cbn = TAILQ_NEXT(cb, list);
1238 if (opcode == cb->uaiocb.aio_lio_opcode) {
1239 p = cb->userproc;
1240 ki = p->p_aioinfo;
1241 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1242 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1243 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1244 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1245 wakecount++;
1246 if (cb->jobstate != JOBST_JOBQGLOBAL)
1247 panic("invalid queue value");
1248 }
1249 }
1250
1251 while (wakecount--) {
1252 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1253 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1254 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1255 aiop->aiothreadflags &= ~AIOP_FREE;
1256 wakeup(aiop->aiothread);
1257 }
1258 }
1259 }
1260
1261 /*
1262 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1263 * technique is done in this code.
1264 */
1265 static int
1266 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1267 {
1268 struct proc *p = td->td_proc;
1269 struct filedesc *fdp;
1270 struct file *fp;
1271 unsigned int fd;
1272 struct socket *so;
1273 int s;
1274 int error;
1275 int opcode, user_opcode;
1276 struct aiocblist *aiocbe;
1277 struct aiothreadlist *aiop;
1278 struct kaioinfo *ki;
1279 struct kevent kev;
1280 struct kqueue *kq;
1281 struct file *kq_fp;
1282
1283 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1284 aiocbe->inputcharge = 0;
1285 aiocbe->outputcharge = 0;
1286 callout_handle_init(&aiocbe->timeouthandle);
1287 SLIST_INIT(&aiocbe->klist);
1288
1289 suword(&job->_aiocb_private.status, -1);
1290 suword(&job->_aiocb_private.error, 0);
1291 suword(&job->_aiocb_private.kernelinfo, -1);
1292
1293 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1294 if (error) {
1295 suword(&job->_aiocb_private.error, error);
1296 uma_zfree(aiocb_zone, aiocbe);
1297 return (error);
1298 }
1299 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1300 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1301 uma_zfree(aiocb_zone, aiocbe);
1302 return (EINVAL);
1303 }
1304
1305 /* Save userspace address of the job info. */
1306 aiocbe->uuaiocb = job;
1307
1308 /* Get the opcode. */
1309 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1310 if (type != LIO_NOP)
1311 aiocbe->uaiocb.aio_lio_opcode = type;
1312 opcode = aiocbe->uaiocb.aio_lio_opcode;
1313
1314 /* Get the fd info for process. */
1315 fdp = p->p_fd;
1316
1317 /*
1318 * Range check file descriptor.
1319 */
1320 FILEDESC_LOCK(fdp);
1321 fd = aiocbe->uaiocb.aio_fildes;
1322 if (fd >= fdp->fd_nfiles) {
1323 FILEDESC_UNLOCK(fdp);
1324 uma_zfree(aiocb_zone, aiocbe);
1325 if (type == 0)
1326 suword(&job->_aiocb_private.error, EBADF);
1327 return (EBADF);
1328 }
1329
1330 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1331 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1332 0))) {
1333 FILEDESC_UNLOCK(fdp);
1334 uma_zfree(aiocb_zone, aiocbe);
1335 if (type == 0)
1336 suword(&job->_aiocb_private.error, EBADF);
1337 return (EBADF);
1338 }
1339 fhold(fp);
1340 FILEDESC_UNLOCK(fdp);
1341
1342 if (aiocbe->uaiocb.aio_offset == -1LL) {
1343 error = EINVAL;
1344 goto aqueue_fail;
1345 }
1346 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1347 if (error) {
1348 error = EINVAL;
1349 goto aqueue_fail;
1350 }
1351 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1352 if (jobrefid == LONG_MAX)
1353 jobrefid = 1;
1354 else
1355 jobrefid++;
1356
1357 if (opcode == LIO_NOP) {
1358 fdrop(fp, td);
1359 uma_zfree(aiocb_zone, aiocbe);
1360 if (type == 0) {
1361 suword(&job->_aiocb_private.error, 0);
1362 suword(&job->_aiocb_private.status, 0);
1363 suword(&job->_aiocb_private.kernelinfo, 0);
1364 }
1365 return (0);
1366 }
1367 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1368 if (type == 0)
1369 suword(&job->_aiocb_private.status, 0);
1370 error = EINVAL;
1371 goto aqueue_fail;
1372 }
1373
1374 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1375 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1376 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1377 }
1378 else {
1379 /*
1380 * This method for requesting kevent-based notification won't
1381 * work on the alpha, since we're passing in a pointer
1382 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1383 * based method instead.
1384 */
1385 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1386 user_opcode == LIO_WRITE)
1387 goto no_kqueue;
1388
1389 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1390 &kev, sizeof(kev));
1391 if (error)
1392 goto aqueue_fail;
1393 }
1394 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1395 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1396 (kq_fp->f_type != DTYPE_KQUEUE)) {
1397 error = EBADF;
1398 goto aqueue_fail;
1399 }
1400 kq = kq_fp->f_data;
1401 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1402 kev.filter = EVFILT_AIO;
1403 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1404 kev.data = (intptr_t)aiocbe;
1405 error = kqueue_register(kq, &kev, td);
1406 aqueue_fail:
1407 if (error) {
1408 fdrop(fp, td);
1409 uma_zfree(aiocb_zone, aiocbe);
1410 if (type == 0)
1411 suword(&job->_aiocb_private.error, error);
1412 goto done;
1413 }
1414 no_kqueue:
1415
1416 suword(&job->_aiocb_private.error, EINPROGRESS);
1417 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1418 aiocbe->userproc = p;
1419 aiocbe->cred = crhold(td->td_ucred);
1420 aiocbe->jobflags = 0;
1421 aiocbe->lio = lj;
1422 ki = p->p_aioinfo;
1423
1424 if (fp->f_type == DTYPE_SOCKET) {
1425 /*
1426 * Alternate queueing for socket ops: Reach down into the
1427 * descriptor to get the socket data. Then check to see if the
1428 * socket is ready to be read or written (based on the requested
1429 * operation).
1430 *
1431 * If it is not ready for io, then queue the aiocbe on the
1432 * socket, and set the flags so we get a call when sbnotify()
1433 * happens.
1434 */
1435 so = fp->f_data;
1436 s = splnet();
1437 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1438 LIO_WRITE) && (!sowriteable(so)))) {
1439 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1440 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1441 if (opcode == LIO_READ)
1442 so->so_rcv.sb_flags |= SB_AIO;
1443 else
1444 so->so_snd.sb_flags |= SB_AIO;
1445 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1446 ki->kaio_queue_count++;
1447 num_queue_count++;
1448 splx(s);
1449 error = 0;
1450 goto done;
1451 }
1452 splx(s);
1453 }
1454
1455 if ((error = aio_qphysio(p, aiocbe)) == 0)
1456 goto done;
1457 if (error > 0) {
1458 suword(&job->_aiocb_private.status, 0);
1459 aiocbe->uaiocb._aiocb_private.error = error;
1460 suword(&job->_aiocb_private.error, error);
1461 goto done;
1462 }
1463
1464 /* No buffer for daemon I/O. */
1465 aiocbe->bp = NULL;
1466
1467 ki->kaio_queue_count++;
1468 if (lj)
1469 lj->lioj_queue_count++;
1470 s = splnet();
1471 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1472 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1473 splx(s);
1474 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1475
1476 num_queue_count++;
1477 error = 0;
1478
1479 /*
1480 * If we don't have a free AIO process, and we are below our quota, then
1481 * start one. Otherwise, depend on the subsequent I/O completions to
1482 * pick-up this job. If we don't sucessfully create the new process
1483 * (thread) due to resource issues, we return an error for now (EAGAIN),
1484 * which is likely not the correct thing to do.
1485 */
1486 s = splnet();
1487 retryproc:
1488 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1489 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1490 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1491 aiop->aiothreadflags &= ~AIOP_FREE;
1492 wakeup(aiop->aiothread);
1493 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1494 ((ki->kaio_active_count + num_aio_resv_start) <
1495 ki->kaio_maxactive_count)) {
1496 num_aio_resv_start++;
1497 if ((error = aio_newproc()) == 0) {
1498 num_aio_resv_start--;
1499 goto retryproc;
1500 }
1501 num_aio_resv_start--;
1502 }
1503 splx(s);
1504 done:
1505 return (error);
1506 }
1507
1508 /*
1509 * This routine queues an AIO request, checking for quotas.
1510 */
1511 static int
1512 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1513 {
1514 struct proc *p = td->td_proc;
1515 struct kaioinfo *ki;
1516
1517 if (p->p_aioinfo == NULL)
1518 aio_init_aioinfo(p);
1519
1520 if (num_queue_count >= max_queue_count)
1521 return (EAGAIN);
1522
1523 ki = p->p_aioinfo;
1524 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1525 return (EAGAIN);
1526
1527 return _aio_aqueue(td, job, NULL, type);
1528 }
1529
1530 /*
1531 * Support the aio_return system call, as a side-effect, kernel resources are
1532 * released.
1533 */
1534 int
1535 aio_return(struct thread *td, struct aio_return_args *uap)
1536 {
1537 struct proc *p = td->td_proc;
1538 int s;
1539 long jobref;
1540 struct aiocblist *cb, *ncb;
1541 struct aiocb *ujob;
1542 struct kaioinfo *ki;
1543
1544 ujob = uap->aiocbp;
1545 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1546 if (jobref == -1 || jobref == 0)
1547 return (EINVAL);
1548
1549 ki = p->p_aioinfo;
1550 if (ki == NULL)
1551 return (EINVAL);
1552 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1553 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1554 jobref) {
1555 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1556 p->p_stats->p_ru.ru_oublock +=
1557 cb->outputcharge;
1558 cb->outputcharge = 0;
1559 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1560 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1561 cb->inputcharge = 0;
1562 }
1563 goto done;
1564 }
1565 }
1566 s = splbio();
1567 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1568 ncb = TAILQ_NEXT(cb, plist);
1569 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1570 == jobref) {
1571 break;
1572 }
1573 }
1574 splx(s);
1575 done:
1576 if (cb != NULL) {
1577 if (ujob == cb->uuaiocb) {
1578 td->td_retval[0] =
1579 cb->uaiocb._aiocb_private.status;
1580 } else
1581 td->td_retval[0] = EFAULT;
1582 aio_free_entry(cb);
1583 return (0);
1584 }
1585 return (EINVAL);
1586 }
1587
1588 /*
1589 * Allow a process to wakeup when any of the I/O requests are completed.
1590 */
1591 int
1592 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1593 {
1594 struct proc *p = td->td_proc;
1595 struct timeval atv;
1596 struct timespec ts;
1597 struct aiocb *const *cbptr, *cbp;
1598 struct kaioinfo *ki;
1599 struct aiocblist *cb;
1600 int i;
1601 int njoblist;
1602 int error, s, timo;
1603 long *ijoblist;
1604 struct aiocb **ujoblist;
1605
1606 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1607 return (EINVAL);
1608
1609 timo = 0;
1610 if (uap->timeout) {
1611 /* Get timespec struct. */
1612 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1613 return (error);
1614
1615 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1616 return (EINVAL);
1617
1618 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1619 if (itimerfix(&atv))
1620 return (EINVAL);
1621 timo = tvtohz(&atv);
1622 }
1623
1624 ki = p->p_aioinfo;
1625 if (ki == NULL)
1626 return (EAGAIN);
1627
1628 njoblist = 0;
1629 ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1630 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1631 cbptr = uap->aiocbp;
1632
1633 for (i = 0; i < uap->nent; i++) {
1634 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1635 if (cbp == 0)
1636 continue;
1637 ujoblist[njoblist] = cbp;
1638 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1639 njoblist++;
1640 }
1641
1642 if (njoblist == 0) {
1643 uma_zfree(aiol_zone, ijoblist);
1644 uma_zfree(aiol_zone, ujoblist);
1645 return (0);
1646 }
1647
1648 error = 0;
1649 for (;;) {
1650 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1651 for (i = 0; i < njoblist; i++) {
1652 if (((intptr_t)
1653 cb->uaiocb._aiocb_private.kernelinfo) ==
1654 ijoblist[i]) {
1655 if (ujoblist[i] != cb->uuaiocb)
1656 error = EINVAL;
1657 uma_zfree(aiol_zone, ijoblist);
1658 uma_zfree(aiol_zone, ujoblist);
1659 return (error);
1660 }
1661 }
1662 }
1663
1664 s = splbio();
1665 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1666 TAILQ_NEXT(cb, plist)) {
1667 for (i = 0; i < njoblist; i++) {
1668 if (((intptr_t)
1669 cb->uaiocb._aiocb_private.kernelinfo) ==
1670 ijoblist[i]) {
1671 splx(s);
1672 if (ujoblist[i] != cb->uuaiocb)
1673 error = EINVAL;
1674 uma_zfree(aiol_zone, ijoblist);
1675 uma_zfree(aiol_zone, ujoblist);
1676 return (error);
1677 }
1678 }
1679 }
1680
1681 ki->kaio_flags |= KAIO_WAKEUP;
1682 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 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_data;
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 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1842 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1843 jobref) {
1844 td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1845 return (0);
1846 }
1847 }
1848
1849 s = splnet();
1850
1851 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1852 plist)) {
1853 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1854 jobref) {
1855 td->td_retval[0] = EINPROGRESS;
1856 splx(s);
1857 return (0);
1858 }
1859 }
1860
1861 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1862 plist)) {
1863 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1864 jobref) {
1865 td->td_retval[0] = EINPROGRESS;
1866 splx(s);
1867 return (0);
1868 }
1869 }
1870 splx(s);
1871
1872 s = splbio();
1873 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1874 plist)) {
1875 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1876 jobref) {
1877 td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1878 splx(s);
1879 return (0);
1880 }
1881 }
1882
1883 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1884 plist)) {
1885 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1886 jobref) {
1887 td->td_retval[0] = EINPROGRESS;
1888 splx(s);
1889 return (0);
1890 }
1891 }
1892 splx(s);
1893
1894 #if (0)
1895 /*
1896 * Hack for lio.
1897 */
1898 status = fuword(&uap->aiocbp->_aiocb_private.status);
1899 if (status == -1)
1900 return fuword(&uap->aiocbp->_aiocb_private.error);
1901 #endif
1902 return (EINVAL);
1903 }
1904
1905 /* syscall - asynchronous read from a file (REALTIME) */
1906 int
1907 aio_read(struct thread *td, struct aio_read_args *uap)
1908 {
1909
1910 return aio_aqueue(td, uap->aiocbp, LIO_READ);
1911 }
1912
1913 /* syscall - asynchronous write to a file (REALTIME) */
1914 int
1915 aio_write(struct thread *td, struct aio_write_args *uap)
1916 {
1917
1918 return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1919 }
1920
1921 /* syscall - list directed I/O (REALTIME) */
1922 int
1923 lio_listio(struct thread *td, struct lio_listio_args *uap)
1924 {
1925 struct proc *p = td->td_proc;
1926 int nent, nentqueued;
1927 struct aiocb *iocb, * const *cbptr;
1928 struct aiocblist *cb;
1929 struct kaioinfo *ki;
1930 struct aio_liojob *lj;
1931 int error, runningcode;
1932 int nerror;
1933 int i;
1934 int s;
1935
1936 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1937 return (EINVAL);
1938
1939 nent = uap->nent;
1940 if (nent < 0 || nent > AIO_LISTIO_MAX)
1941 return (EINVAL);
1942
1943 if (p->p_aioinfo == NULL)
1944 aio_init_aioinfo(p);
1945
1946 if ((nent + num_queue_count) > max_queue_count)
1947 return (EAGAIN);
1948
1949 ki = p->p_aioinfo;
1950 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1951 return (EAGAIN);
1952
1953 lj = uma_zalloc(aiolio_zone, M_WAITOK);
1954 if (!lj)
1955 return (EAGAIN);
1956
1957 lj->lioj_flags = 0;
1958 lj->lioj_buffer_count = 0;
1959 lj->lioj_buffer_finished_count = 0;
1960 lj->lioj_queue_count = 0;
1961 lj->lioj_queue_finished_count = 0;
1962 lj->lioj_ki = ki;
1963
1964 /*
1965 * Setup signal.
1966 */
1967 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1968 error = copyin(uap->sig, &lj->lioj_signal,
1969 sizeof(lj->lioj_signal));
1970 if (error) {
1971 uma_zfree(aiolio_zone, lj);
1972 return (error);
1973 }
1974 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1975 uma_zfree(aiolio_zone, lj);
1976 return (EINVAL);
1977 }
1978 lj->lioj_flags |= LIOJ_SIGNAL;
1979 }
1980 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1981 /*
1982 * Get pointers to the list of I/O requests.
1983 */
1984 nerror = 0;
1985 nentqueued = 0;
1986 cbptr = uap->acb_list;
1987 for (i = 0; i < uap->nent; i++) {
1988 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1989 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1990 error = _aio_aqueue(td, iocb, lj, 0);
1991 if (error == 0)
1992 nentqueued++;
1993 else
1994 nerror++;
1995 }
1996 }
1997
1998 /*
1999 * If we haven't queued any, then just return error.
2000 */
2001 if (nentqueued == 0)
2002 return (0);
2003
2004 /*
2005 * Calculate the appropriate error return.
2006 */
2007 runningcode = 0;
2008 if (nerror)
2009 runningcode = EIO;
2010
2011 if (uap->mode == LIO_WAIT) {
2012 int command, found, jobref;
2013
2014 for (;;) {
2015 found = 0;
2016 for (i = 0; i < uap->nent; i++) {
2017 /*
2018 * Fetch address of the control buf pointer in
2019 * user space.
2020 */
2021 iocb = (struct aiocb *)
2022 (intptr_t)fuword(&cbptr[i]);
2023 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2024 == 0))
2025 continue;
2026
2027 /*
2028 * Fetch the associated command from user space.
2029 */
2030 command = fuword(&iocb->aio_lio_opcode);
2031 if (command == LIO_NOP) {
2032 found++;
2033 continue;
2034 }
2035
2036 jobref =
2037 fuword(&iocb->_aiocb_private.kernelinfo);
2038
2039 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2040 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2041 == jobref) {
2042 if (cb->uaiocb.aio_lio_opcode
2043 == LIO_WRITE) {
2044 p->p_stats->p_ru.ru_oublock
2045 +=
2046 cb->outputcharge;
2047 cb->outputcharge = 0;
2048 } else if (cb->uaiocb.aio_lio_opcode
2049 == LIO_READ) {
2050 p->p_stats->p_ru.ru_inblock
2051 += cb->inputcharge;
2052 cb->inputcharge = 0;
2053 }
2054 found++;
2055 break;
2056 }
2057 }
2058
2059 s = splbio();
2060 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2061 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2062 == jobref) {
2063 found++;
2064 break;
2065 }
2066 }
2067 splx(s);
2068 }
2069
2070 /*
2071 * If all I/Os have been disposed of, then we can
2072 * return.
2073 */
2074 if (found == nentqueued)
2075 return (runningcode);
2076
2077 ki->kaio_flags |= KAIO_WAKEUP;
2078 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2079
2080 if (error == EINTR)
2081 return (EINTR);
2082 else if (error == EWOULDBLOCK)
2083 return (EAGAIN);
2084 }
2085 }
2086
2087 return (runningcode);
2088 }
2089
2090 /*
2091 * This is a weird hack so that we can post a signal. It is safe to do so from
2092 * a timeout routine, but *not* from an interrupt routine.
2093 */
2094 static void
2095 process_signal(void *aioj)
2096 {
2097 struct aiocblist *aiocbe = aioj;
2098 struct aio_liojob *lj = aiocbe->lio;
2099 struct aiocb *cb = &aiocbe->uaiocb;
2100
2101 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2102 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2103 PROC_LOCK(lj->lioj_ki->kaio_p);
2104 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2105 PROC_UNLOCK(lj->lioj_ki->kaio_p);
2106 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2107 }
2108
2109 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2110 PROC_LOCK(aiocbe->userproc);
2111 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2112 PROC_UNLOCK(aiocbe->userproc);
2113 }
2114 }
2115
2116 /*
2117 * Interrupt handler for physio, performs the necessary process wakeups, and
2118 * signals.
2119 */
2120 static void
2121 aio_physwakeup(struct buf *bp)
2122 {
2123 struct aiocblist *aiocbe;
2124 struct proc *p;
2125 struct kaioinfo *ki;
2126 struct aio_liojob *lj;
2127
2128 wakeup(bp);
2129
2130 aiocbe = (struct aiocblist *)bp->b_spc;
2131 if (aiocbe) {
2132 p = bp->b_caller1;
2133
2134 aiocbe->jobstate = JOBST_JOBBFINISHED;
2135 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2136 aiocbe->uaiocb._aiocb_private.error = 0;
2137 aiocbe->jobflags |= AIOCBLIST_DONE;
2138
2139 if (bp->b_ioflags & BIO_ERROR)
2140 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2141
2142 lj = aiocbe->lio;
2143 if (lj) {
2144 lj->lioj_buffer_finished_count++;
2145
2146 /*
2147 * wakeup/signal if all of the interrupt jobs are done.
2148 */
2149 if (lj->lioj_buffer_finished_count ==
2150 lj->lioj_buffer_count) {
2151 /*
2152 * Post a signal if it is called for.
2153 */
2154 if ((lj->lioj_flags &
2155 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2156 LIOJ_SIGNAL) {
2157 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2158 aiocbe->timeouthandle =
2159 timeout(process_signal,
2160 aiocbe, 0);
2161 }
2162 }
2163 }
2164
2165 ki = p->p_aioinfo;
2166 if (ki) {
2167 ki->kaio_buffer_finished_count++;
2168 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2169 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2170 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2171
2172 KNOTE(&aiocbe->klist, 0);
2173 /* Do the wakeup. */
2174 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2175 ki->kaio_flags &= ~KAIO_WAKEUP;
2176 wakeup(p);
2177 }
2178 }
2179
2180 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2181 aiocbe->timeouthandle =
2182 timeout(process_signal, aiocbe, 0);
2183 }
2184 }
2185
2186 /* syscall - wait for the next completion of an aio request */
2187 int
2188 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2189 {
2190 struct proc *p = td->td_proc;
2191 struct timeval atv;
2192 struct timespec ts;
2193 struct kaioinfo *ki;
2194 struct aiocblist *cb = NULL;
2195 int error, s, timo;
2196
2197 suword(uap->aiocbp, (int)NULL);
2198
2199 timo = 0;
2200 if (uap->timeout) {
2201 /* Get timespec struct. */
2202 error = copyin(uap->timeout, &ts, sizeof(ts));
2203 if (error)
2204 return (error);
2205
2206 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2207 return (EINVAL);
2208
2209 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2210 if (itimerfix(&atv))
2211 return (EINVAL);
2212 timo = tvtohz(&atv);
2213 }
2214
2215 ki = p->p_aioinfo;
2216 if (ki == NULL)
2217 return (EAGAIN);
2218
2219 for (;;) {
2220 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2221 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2222 td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2223 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2224 p->p_stats->p_ru.ru_oublock +=
2225 cb->outputcharge;
2226 cb->outputcharge = 0;
2227 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2228 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2229 cb->inputcharge = 0;
2230 }
2231 aio_free_entry(cb);
2232 return (cb->uaiocb._aiocb_private.error);
2233 }
2234
2235 s = splbio();
2236 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2237 splx(s);
2238 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2239 td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2240 aio_free_entry(cb);
2241 return (cb->uaiocb._aiocb_private.error);
2242 }
2243
2244 ki->kaio_flags |= KAIO_WAKEUP;
2245 error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2246 splx(s);
2247
2248 if (error == ERESTART)
2249 return (EINTR);
2250 else if (error < 0)
2251 return (error);
2252 else if (error == EINTR)
2253 return (EINTR);
2254 else if (error == EWOULDBLOCK)
2255 return (EAGAIN);
2256 }
2257 }
2258
2259 /* kqueue attach function */
2260 static int
2261 filt_aioattach(struct knote *kn)
2262 {
2263 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2264
2265 /*
2266 * The aiocbe pointer must be validated before using it, so
2267 * registration is restricted to the kernel; the user cannot
2268 * set EV_FLAG1.
2269 */
2270 if ((kn->kn_flags & EV_FLAG1) == 0)
2271 return (EPERM);
2272 kn->kn_flags &= ~EV_FLAG1;
2273
2274 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2275
2276 return (0);
2277 }
2278
2279 /* kqueue detach function */
2280 static void
2281 filt_aiodetach(struct knote *kn)
2282 {
2283 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2284
2285 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2286 }
2287
2288 /* kqueue filter function */
2289 /*ARGSUSED*/
2290 static int
2291 filt_aio(struct knote *kn, long hint)
2292 {
2293 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2294
2295 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2296 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2297 aiocbe->jobstate != JOBST_JOBBFINISHED)
2298 return (0);
2299 kn->kn_flags |= EV_EOF;
2300 return (1);
2301 }
Cache object: f0ddcd2531667957bcebeaa6ea5a77c8
|