Replace the linear search in file descriptor allocation with an O(log N)
[dragonfly.git] / sys / kern / vfs_aio.c
CommitLineData
984263bc
MD
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 $
831f78e5 17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.16 2005/06/09 20:47:37 swildner Exp $
984263bc
MD
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>
7b95be2a
MD
50#include <sys/file2.h>
51#include <sys/buf2.h>
831f78e5 52#include <sys/thread2.h>
984263bc
MD
53
54#include <machine/limits.h>
55#include "opt_vfs_aio.h"
56
57#ifdef VFS_AIO
58
59/*
60 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
61 * overflow.
62 */
63static long jobrefid;
64
65#define JOBST_NULL 0x0
66#define JOBST_JOBQGLOBAL 0x2
67#define JOBST_JOBRUNNING 0x3
68#define JOBST_JOBFINISHED 0x4
69#define JOBST_JOBQBUF 0x5
70#define JOBST_JOBBFINISHED 0x6
71
72#ifndef MAX_AIO_PER_PROC
73#define MAX_AIO_PER_PROC 32
74#endif
75
76#ifndef MAX_AIO_QUEUE_PER_PROC
77#define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
78#endif
79
80#ifndef MAX_AIO_PROCS
81#define MAX_AIO_PROCS 32
82#endif
83
84#ifndef MAX_AIO_QUEUE
85#define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
86#endif
87
88#ifndef TARGET_AIO_PROCS
89#define TARGET_AIO_PROCS 4
90#endif
91
92#ifndef MAX_BUF_AIO
93#define MAX_BUF_AIO 16
94#endif
95
96#ifndef AIOD_TIMEOUT_DEFAULT
97#define AIOD_TIMEOUT_DEFAULT (10 * hz)
98#endif
99
100#ifndef AIOD_LIFETIME_DEFAULT
101#define AIOD_LIFETIME_DEFAULT (30 * hz)
102#endif
103
104SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
105
106static int max_aio_procs = MAX_AIO_PROCS;
107SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
108 CTLFLAG_RW, &max_aio_procs, 0,
109 "Maximum number of kernel threads to use for handling async IO");
110
111static int num_aio_procs = 0;
112SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
113 CTLFLAG_RD, &num_aio_procs, 0,
114 "Number of presently active kernel threads for async IO");
115
116/*
117 * The code will adjust the actual number of AIO processes towards this
118 * number when it gets a chance.
119 */
120static int target_aio_procs = TARGET_AIO_PROCS;
121SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
122 0, "Preferred number of ready kernel threads for async IO");
123
124static int max_queue_count = MAX_AIO_QUEUE;
125SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
126 "Maximum number of aio requests to queue, globally");
127
128static int num_queue_count = 0;
129SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
130 "Number of queued aio requests");
131
132static int num_buf_aio = 0;
133SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
134 "Number of aio requests presently handled by the buf subsystem");
135
136/* Number of async I/O thread in the process of being started */
137/* XXX This should be local to _aio_aqueue() */
138static int num_aio_resv_start = 0;
139
140static int aiod_timeout;
141SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
142 "Timeout value for synchronous aio operations");
143
144static int aiod_lifetime;
145SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
146 "Maximum lifetime for idle aiod");
147
148static int max_aio_per_proc = MAX_AIO_PER_PROC;
149SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
150 0, "Maximum active aio requests per process (stored in the process)");
151
152static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
153SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
154 &max_aio_queue_per_proc, 0,
155 "Maximum queued aio requests per process (stored in the process)");
156
157static int max_buf_aio = MAX_BUF_AIO;
158SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
159 "Maximum buf aio requests per process (stored in the process)");
160
161/*
162 * AIO process info
163 */
164#define AIOP_FREE 0x1 /* proc on free queue */
165#define AIOP_SCHED 0x2 /* proc explicitly scheduled */
166
167struct aioproclist {
168 int aioprocflags; /* AIO proc flags */
169 TAILQ_ENTRY(aioproclist) list; /* List of processes */
170 struct proc *aioproc; /* The AIO thread */
171};
172
173/*
174 * data-structure for lio signal management
175 */
176struct aio_liojob {
177 int lioj_flags;
178 int lioj_buffer_count;
179 int lioj_buffer_finished_count;
180 int lioj_queue_count;
181 int lioj_queue_finished_count;
182 struct sigevent lioj_signal; /* signal on all I/O done */
183 TAILQ_ENTRY(aio_liojob) lioj_list;
184 struct kaioinfo *lioj_ki;
185};
186#define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
187#define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
188
189/*
190 * per process aio data structure
191 */
192struct kaioinfo {
193 int kaio_flags; /* per process kaio flags */
194 int kaio_maxactive_count; /* maximum number of AIOs */
195 int kaio_active_count; /* number of currently used AIOs */
196 int kaio_qallowed_count; /* maxiumu size of AIO queue */
197 int kaio_queue_count; /* size of AIO queue */
198 int kaio_ballowed_count; /* maximum number of buffers */
199 int kaio_queue_finished_count; /* number of daemon jobs finished */
200 int kaio_buffer_count; /* number of physio buffers */
201 int kaio_buffer_finished_count; /* count of I/O done */
202 struct proc *kaio_p; /* process that uses this kaio block */
203 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
204 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
205 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
206 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
207 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
208 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
209};
210
211#define KAIO_RUNDOWN 0x1 /* process is being run down */
212#define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
213
214static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
215static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
216static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
217static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
218
219static void aio_init_aioinfo(struct proc *p);
220static void aio_onceonly(void *);
221static int aio_free_entry(struct aiocblist *aiocbe);
222static void aio_process(struct aiocblist *aiocbe);
223static int aio_newproc(void);
41c20dac 224static int aio_aqueue(struct aiocb *job, int type);
984263bc
MD
225static void aio_physwakeup(struct buf *bp);
226static int aio_fphysio(struct aiocblist *aiocbe);
227static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
228static void aio_daemon(void *uproc);
229static void process_signal(void *aioj);
230
231SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
232
233/*
234 * Zones for:
235 * kaio Per process async io info
236 * aiop async io thread data
237 * aiocb async io jobs
238 * aiol list io job pointer - internal to aio_suspend XXX
239 * aiolio list io jobs
240 */
241static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
242
243/*
244 * Startup initialization
245 */
246static void
247aio_onceonly(void *na)
248{
249 TAILQ_INIT(&aio_freeproc);
250 TAILQ_INIT(&aio_activeproc);
251 TAILQ_INIT(&aio_jobs);
252 TAILQ_INIT(&aio_bufjobs);
253 TAILQ_INIT(&aio_freejobs);
254 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
255 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
256 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
257 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
258 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
259 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
260 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
261 jobrefid = 1;
262}
263
264/*
265 * Init the per-process aioinfo structure. The aioinfo limits are set
266 * per-process for user limit (resource) management.
267 */
268static void
269aio_init_aioinfo(struct proc *p)
270{
271 struct kaioinfo *ki;
272 if (p->p_aioinfo == NULL) {
273 ki = zalloc(kaio_zone);
274 p->p_aioinfo = ki;
275 ki->kaio_flags = 0;
276 ki->kaio_maxactive_count = max_aio_per_proc;
277 ki->kaio_active_count = 0;
278 ki->kaio_qallowed_count = max_aio_queue_per_proc;
279 ki->kaio_queue_count = 0;
280 ki->kaio_ballowed_count = max_buf_aio;
281 ki->kaio_buffer_count = 0;
282 ki->kaio_buffer_finished_count = 0;
283 ki->kaio_p = p;
284 TAILQ_INIT(&ki->kaio_jobdone);
285 TAILQ_INIT(&ki->kaio_jobqueue);
286 TAILQ_INIT(&ki->kaio_bufdone);
287 TAILQ_INIT(&ki->kaio_bufqueue);
288 TAILQ_INIT(&ki->kaio_liojoblist);
289 TAILQ_INIT(&ki->kaio_sockqueue);
290 }
291
292 while (num_aio_procs < target_aio_procs)
293 aio_newproc();
294}
295
296/*
297 * Free a job entry. Wait for completion if it is currently active, but don't
298 * delay forever. If we delay, we return a flag that says that we have to
299 * restart the queue scan.
300 */
301static int
302aio_free_entry(struct aiocblist *aiocbe)
303{
304 struct kaioinfo *ki;
305 struct aio_liojob *lj;
306 struct proc *p;
307 int error;
984263bc
MD
308
309 if (aiocbe->jobstate == JOBST_NULL)
310 panic("aio_free_entry: freeing already free job");
311
312 p = aiocbe->userproc;
313 ki = p->p_aioinfo;
314 lj = aiocbe->lio;
315 if (ki == NULL)
316 panic("aio_free_entry: missing p->p_aioinfo");
317
318 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
319 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
377d4740 320 tsleep(aiocbe, 0, "jobwai", 0);
984263bc
MD
321 }
322 if (aiocbe->bp == NULL) {
323 if (ki->kaio_queue_count <= 0)
324 panic("aio_free_entry: process queue size <= 0");
325 if (num_queue_count <= 0)
326 panic("aio_free_entry: system wide queue size <= 0");
327
328 if (lj) {
329 lj->lioj_queue_count--;
330 if (aiocbe->jobflags & AIOCBLIST_DONE)
331 lj->lioj_queue_finished_count--;
332 }
333 ki->kaio_queue_count--;
334 if (aiocbe->jobflags & AIOCBLIST_DONE)
335 ki->kaio_queue_finished_count--;
336 num_queue_count--;
337 } else {
338 if (lj) {
339 lj->lioj_buffer_count--;
340 if (aiocbe->jobflags & AIOCBLIST_DONE)
341 lj->lioj_buffer_finished_count--;
342 }
343 if (aiocbe->jobflags & AIOCBLIST_DONE)
344 ki->kaio_buffer_finished_count--;
345 ki->kaio_buffer_count--;
346 num_buf_aio--;
347 }
348
349 /* aiocbe is going away, we need to destroy any knotes */
7b95be2a 350 knote_remove(p->p_thread, &aiocbe->klist);
984263bc
MD
351
352 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
353 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
354 ki->kaio_flags &= ~KAIO_WAKEUP;
355 wakeup(p);
356 }
357
358 if (aiocbe->jobstate == JOBST_JOBQBUF) {
359 if ((error = aio_fphysio(aiocbe)) != 0)
360 return error;
361 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
362 panic("aio_free_entry: invalid physio finish-up state");
e43a034f 363 crit_enter();
984263bc 364 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
e43a034f 365 crit_exit();
984263bc 366 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
e43a034f 367 crit_enter();
984263bc
MD
368 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
369 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
e43a034f 370 crit_exit();
984263bc
MD
371 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
372 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
373 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
e43a034f 374 crit_enter();
984263bc 375 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
e43a034f 376 crit_exit();
984263bc
MD
377 if (aiocbe->bp) {
378 vunmapbuf(aiocbe->bp);
379 relpbuf(aiocbe->bp, NULL);
380 aiocbe->bp = NULL;
381 }
382 }
383 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
384 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
385 zfree(aiolio_zone, lj);
386 }
387 aiocbe->jobstate = JOBST_NULL;
a5eb27b6 388 callout_stop(&aiocbe->timeout);
7b95be2a 389 fdrop(aiocbe->fd_file, curthread);
984263bc
MD
390 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
391 return 0;
392}
393#endif /* VFS_AIO */
394
395/*
396 * Rundown the jobs for a given process.
397 */
398void
399aio_proc_rundown(struct proc *p)
400{
401#ifndef VFS_AIO
402 return;
403#else
984263bc
MD
404 struct kaioinfo *ki;
405 struct aio_liojob *lj, *ljn;
406 struct aiocblist *aiocbe, *aiocbn;
407 struct file *fp;
408 struct socket *so;
409
410 ki = p->p_aioinfo;
411 if (ki == NULL)
412 return;
413
414 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
415 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
416 ki->kaio_buffer_finished_count)) {
417 ki->kaio_flags |= KAIO_RUNDOWN;
377d4740 418 if (tsleep(p, 0, "kaiowt", aiod_timeout))
984263bc
MD
419 break;
420 }
421
422 /*
423 * Move any aio ops that are waiting on socket I/O to the normal job
424 * queues so they are cleaned up with any others.
425 */
e43a034f 426 crit_enter();
984263bc
MD
427 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
428 aiocbn) {
429 aiocbn = TAILQ_NEXT(aiocbe, plist);
430 fp = aiocbe->fd_file;
431 if (fp != NULL) {
432 so = (struct socket *)fp->f_data;
433 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
434 if (TAILQ_EMPTY(&so->so_aiojobq)) {
435 so->so_snd.sb_flags &= ~SB_AIO;
436 so->so_rcv.sb_flags &= ~SB_AIO;
437 }
438 }
439 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
440 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
441 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
442 }
e43a034f 443 crit_exit();
984263bc
MD
444
445restart1:
446 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
447 aiocbn = TAILQ_NEXT(aiocbe, plist);
448 if (aio_free_entry(aiocbe))
449 goto restart1;
450 }
451
452restart2:
453 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
454 aiocbn) {
455 aiocbn = TAILQ_NEXT(aiocbe, plist);
456 if (aio_free_entry(aiocbe))
457 goto restart2;
458 }
459
984263bc 460restart3:
e43a034f 461 crit_enter();
984263bc
MD
462 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
463 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 464 tsleep(p, 0, "aioprn", 0);
e43a034f 465 crit_exit();
984263bc
MD
466 goto restart3;
467 }
e43a034f 468 crit_exit();
984263bc
MD
469
470restart4:
e43a034f 471 crit_enter();
984263bc
MD
472 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
473 aiocbn = TAILQ_NEXT(aiocbe, plist);
474 if (aio_free_entry(aiocbe)) {
e43a034f 475 crit_exit();
984263bc
MD
476 goto restart4;
477 }
478 }
e43a034f 479 crit_exit();
984263bc
MD
480
481 /*
482 * If we've slept, jobs might have moved from one queue to another.
483 * Retry rundown if we didn't manage to empty the queues.
484 */
485 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
486 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
487 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
488 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
489 goto restart1;
490
491 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
492 ljn = TAILQ_NEXT(lj, lioj_list);
493 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
494 0)) {
495 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
496 zfree(aiolio_zone, lj);
497 } else {
498#ifdef DIAGNOSTIC
499 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
500 "QF:%d\n", lj->lioj_buffer_count,
501 lj->lioj_buffer_finished_count,
502 lj->lioj_queue_count,
503 lj->lioj_queue_finished_count);
504#endif
505 }
506 }
507
508 zfree(kaio_zone, ki);
509 p->p_aioinfo = NULL;
510#endif /* VFS_AIO */
511}
512
513#ifdef VFS_AIO
514/*
515 * Select a job to run (called by an AIO daemon).
516 */
517static struct aiocblist *
518aio_selectjob(struct aioproclist *aiop)
519{
984263bc
MD
520 struct aiocblist *aiocbe;
521 struct kaioinfo *ki;
522 struct proc *userp;
523
e43a034f 524 crit_enter();
984263bc
MD
525 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
526 TAILQ_NEXT(aiocbe, list)) {
527 userp = aiocbe->userproc;
528 ki = userp->p_aioinfo;
529
530 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
531 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
e43a034f 532 crit_exit();
984263bc
MD
533 return aiocbe;
534 }
535 }
e43a034f 536 crit_exit();
984263bc
MD
537
538 return NULL;
539}
540
541/*
542 * The AIO processing activity. This is the code that does the I/O request for
543 * the non-physio version of the operations. The normal vn operations are used,
544 * and this code should work in all instances for every type of file, including
545 * pipes, sockets, fifos, and regular files.
546 */
547static void
548aio_process(struct aiocblist *aiocbe)
549{
7b95be2a 550 struct thread *mytd;
984263bc
MD
551 struct aiocb *cb;
552 struct file *fp;
553 struct uio auio;
554 struct iovec aiov;
555 int cnt;
556 int error;
557 int oublock_st, oublock_end;
558 int inblock_st, inblock_end;
559
7b95be2a 560 mytd = curthread;
984263bc
MD
561 cb = &aiocbe->uaiocb;
562 fp = aiocbe->fd_file;
563
564 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
565 aiov.iov_len = cb->aio_nbytes;
566
567 auio.uio_iov = &aiov;
568 auio.uio_iovcnt = 1;
569 auio.uio_offset = cb->aio_offset;
570 auio.uio_resid = cb->aio_nbytes;
571 cnt = cb->aio_nbytes;
572 auio.uio_segflg = UIO_USERSPACE;
7b95be2a 573 auio.uio_td = mytd;
984263bc 574
7b95be2a
MD
575 inblock_st = mytd->td_proc->p_stats->p_ru.ru_inblock;
576 oublock_st = mytd->td_proc->p_stats->p_ru.ru_oublock;
984263bc
MD
577 /*
578 * _aio_aqueue() acquires a reference to the file that is
579 * released in aio_free_entry().
580 */
581 if (cb->aio_lio_opcode == LIO_READ) {
582 auio.uio_rw = UIO_READ;
7b95be2a 583 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, mytd);
984263bc
MD
584 } else {
585 auio.uio_rw = UIO_WRITE;
7b95be2a 586 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, mytd);
984263bc 587 }
7b95be2a
MD
588 inblock_end = mytd->td_proc->p_stats->p_ru.ru_inblock;
589 oublock_end = mytd->td_proc->p_stats->p_ru.ru_oublock;
984263bc
MD
590
591 aiocbe->inputcharge = inblock_end - inblock_st;
592 aiocbe->outputcharge = oublock_end - oublock_st;
593
594 if ((error) && (auio.uio_resid != cnt)) {
595 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
596 error = 0;
597 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
598 psignal(aiocbe->userproc, SIGPIPE);
599 }
600
601 cnt -= auio.uio_resid;
602 cb->_aiocb_private.error = error;
603 cb->_aiocb_private.status = cnt;
604}
605
606/*
607 * The AIO daemon, most of the actual work is done in aio_process,
608 * but the setup (and address space mgmt) is done in this routine.
8a8d5d85
MD
609 *
610 * The MP lock is held on entry.
984263bc
MD
611 */
612static void
613aio_daemon(void *uproc)
614{
984263bc
MD
615 struct aio_liojob *lj;
616 struct aiocb *cb;
617 struct aiocblist *aiocbe;
618 struct aioproclist *aiop;
619 struct kaioinfo *ki;
620 struct proc *curcp, *mycp, *userp;
621 struct vmspace *myvm, *tmpvm;
e9a372eb 622 struct ucred *cr;
984263bc
MD
623
624 /*
625 * Local copies of curproc (cp) and vmspace (myvm)
626 */
627 mycp = curproc;
628 myvm = mycp->p_vmspace;
629
630 if (mycp->p_textvp) {
631 vrele(mycp->p_textvp);
632 mycp->p_textvp = NULL;
633 }
634
635 /*
636 * Allocate and ready the aio control info. There is one aiop structure
637 * per daemon.
638 */
639 aiop = zalloc(aiop_zone);
640 aiop->aioproc = mycp;
641 aiop->aioprocflags |= AIOP_FREE;
642
e43a034f 643 crit_enter();
984263bc
MD
644
645 /*
646 * Place thread (lightweight process) onto the AIO free thread list.
647 */
648 if (TAILQ_EMPTY(&aio_freeproc))
649 wakeup(&aio_freeproc);
650 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
651
e43a034f 652 crit_exit();
984263bc
MD
653
654 /* Make up a name for the daemon. */
655 strcpy(mycp->p_comm, "aiod");
656
657 /*
658 * Get rid of our current filedescriptors. AIOD's don't need any
659 * filedescriptors, except as temporarily inherited from the client.
660 * Credentials are also cloned, and made equivalent to "root".
661 */
662 fdfree(mycp);
663 mycp->p_fd = NULL;
e9a372eb
MD
664 cr = cratom(&mycp->p_ucred);
665 cr->cr_uid = 0;
6eedb489 666 uireplace(&cr->cr_uidinfo, uifind(0));
e9a372eb
MD
667 cr->cr_ngroups = 1;
668 cr->cr_groups[0] = 1;
984263bc
MD
669
670 /* The daemon resides in its own pgrp. */
671 enterpgrp(mycp, mycp->p_pid, 1);
672
673 /* Mark special process type. */
674 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
675
676 /*
677 * Wakeup parent process. (Parent sleeps to keep from blasting away
678 * and creating too many daemons.)
679 */
680 wakeup(mycp);
681
682 for (;;) {
683 /*
684 * curcp is the current daemon process context.
685 * userp is the current user process context.
686 */
687 curcp = mycp;
688
689 /*
690 * Take daemon off of free queue
691 */
692 if (aiop->aioprocflags & AIOP_FREE) {
e43a034f 693 crit_enter();
984263bc
MD
694 TAILQ_REMOVE(&aio_freeproc, aiop, list);
695 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
696 aiop->aioprocflags &= ~AIOP_FREE;
e43a034f 697 crit_exit();
984263bc
MD
698 }
699 aiop->aioprocflags &= ~AIOP_SCHED;
700
701 /*
702 * Check for jobs.
703 */
704 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
705 cb = &aiocbe->uaiocb;
706 userp = aiocbe->userproc;
707
708 aiocbe->jobstate = JOBST_JOBRUNNING;
709
710 /*
711 * Connect to process address space for user program.
712 */
713 if (userp != curcp) {
714 /*
715 * Save the current address space that we are
716 * connected to.
717 */
718 tmpvm = mycp->p_vmspace;
719
720 /*
721 * Point to the new user address space, and
722 * refer to it.
723 */
724 mycp->p_vmspace = userp->p_vmspace;
725 mycp->p_vmspace->vm_refcnt++;
726
727 /* Activate the new mapping. */
728 pmap_activate(mycp);
729
730 /*
731 * If the old address space wasn't the daemons
732 * own address space, then we need to remove the
733 * daemon's reference from the other process
734 * that it was acting on behalf of.
735 */
736 if (tmpvm != myvm) {
737 vmspace_free(tmpvm);
738 }
739 curcp = userp;
740 }
741
742 ki = userp->p_aioinfo;
743 lj = aiocbe->lio;
744
745 /* Account for currently active jobs. */
746 ki->kaio_active_count++;
747
748 /* Do the I/O function. */
749 aio_process(aiocbe);
750
751 /* Decrement the active job count. */
752 ki->kaio_active_count--;
753
754 /*
755 * Increment the completion count for wakeup/signal
756 * comparisons.
757 */
758 aiocbe->jobflags |= AIOCBLIST_DONE;
759 ki->kaio_queue_finished_count++;
760 if (lj)
761 lj->lioj_queue_finished_count++;
762 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
763 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
764 ki->kaio_flags &= ~KAIO_WAKEUP;
765 wakeup(userp);
766 }
767
e43a034f 768 crit_enter();
984263bc
MD
769 if (lj && (lj->lioj_flags &
770 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
771 if ((lj->lioj_queue_finished_count ==
772 lj->lioj_queue_count) &&
773 (lj->lioj_buffer_finished_count ==
774 lj->lioj_buffer_count)) {
775 psignal(userp,
776 lj->lioj_signal.sigev_signo);
777 lj->lioj_flags |=
778 LIOJ_SIGNAL_POSTED;
779 }
780 }
e43a034f 781 crit_exit();
984263bc
MD
782
783 aiocbe->jobstate = JOBST_JOBFINISHED;
784
e43a034f 785 crit_enter();
984263bc
MD
786 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
787 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
e43a034f 788 crit_exit();
984263bc
MD
789 KNOTE(&aiocbe->klist, 0);
790
791 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
792 wakeup(aiocbe);
793 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
794 }
795
796 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
797 psignal(userp, cb->aio_sigevent.sigev_signo);
798 }
799 }
800
801 /*
802 * Disconnect from user address space.
803 */
804 if (curcp != mycp) {
805 /* Get the user address space to disconnect from. */
806 tmpvm = mycp->p_vmspace;
807
808 /* Get original address space for daemon. */
809 mycp->p_vmspace = myvm;
810
811 /* Activate the daemon's address space. */
812 pmap_activate(mycp);
813#ifdef DIAGNOSTIC
814 if (tmpvm == myvm) {
815 printf("AIOD: vmspace problem -- %d\n",
816 mycp->p_pid);
817 }
818#endif
819 /* Remove our vmspace reference. */
820 vmspace_free(tmpvm);
821
822 curcp = mycp;
823 }
824
825 /*
826 * If we are the first to be put onto the free queue, wakeup
827 * anyone waiting for a daemon.
828 */
e43a034f 829 crit_enter();
984263bc
MD
830 TAILQ_REMOVE(&aio_activeproc, aiop, list);
831 if (TAILQ_EMPTY(&aio_freeproc))
832 wakeup(&aio_freeproc);
833 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
834 aiop->aioprocflags |= AIOP_FREE;
e43a034f 835 crit_exit();
984263bc
MD
836
837 /*
838 * If daemon is inactive for a long time, allow it to exit,
839 * thereby freeing resources.
840 */
841 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
377d4740 842 0, "aiordy", aiod_lifetime)) {
e43a034f 843 crit_enter();
984263bc
MD
844 if (TAILQ_EMPTY(&aio_jobs)) {
845 if ((aiop->aioprocflags & AIOP_FREE) &&
846 (num_aio_procs > target_aio_procs)) {
847 TAILQ_REMOVE(&aio_freeproc, aiop, list);
e43a034f 848 crit_exit();
984263bc
MD
849 zfree(aiop_zone, aiop);
850 num_aio_procs--;
851#ifdef DIAGNOSTIC
852 if (mycp->p_vmspace->vm_refcnt <= 1) {
853 printf("AIOD: bad vm refcnt for"
854 " exiting daemon: %d\n",
855 mycp->p_vmspace->vm_refcnt);
856 }
857#endif
7b95be2a 858 exit1(0);
984263bc
MD
859 }
860 }
e43a034f 861 crit_exit();
984263bc
MD
862 }
863 }
864}
865
866/*
867 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
868 * AIO daemon modifies its environment itself.
869 */
870static int
871aio_newproc()
872{
873 int error;
874 struct proc *p, *np;
875
876 p = &proc0;
877 error = fork1(p, RFPROC|RFMEM|RFNOWAIT, &np);
878 if (error)
879 return error;
880 cpu_set_fork_handler(np, aio_daemon, curproc);
7d0bac62 881 start_forked_proc(p, np);
984263bc
MD
882
883 /*
884 * Wait until daemon is started, but continue on just in case to
885 * handle error conditions.
886 */
377d4740 887 error = tsleep(np, 0, "aiosta", aiod_timeout);
984263bc
MD
888 num_aio_procs++;
889
890 return error;
891}
892
893/*
894 * Try the high-performance, low-overhead physio method for eligible
895 * VCHR devices. This method doesn't use an aio helper thread, and
896 * thus has very low overhead.
897 *
898 * Assumes that the caller, _aio_aqueue(), has incremented the file
899 * structure's reference count, preventing its deallocation for the
900 * duration of this call.
901 */
902static int
903aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
904{
905 int error;
906 struct aiocb *cb;
907 struct file *fp;
908 struct buf *bp;
909 struct vnode *vp;
910 struct kaioinfo *ki;
911 struct aio_liojob *lj;
984263bc
MD
912 int notify;
913
914 cb = &aiocbe->uaiocb;
915 fp = aiocbe->fd_file;
916
917 if (fp->f_type != DTYPE_VNODE)
918 return (-1);
919
920 vp = (struct vnode *)fp->f_data;
921
922 /*
923 * If its not a disk, we don't want to return a positive error.
924 * It causes the aio code to not fall through to try the thread
925 * way when you're talking to a regular file.
926 */
927 if (!vn_isdisk(vp, &error)) {
928 if (error == ENOTBLK)
929 return (-1);
930 else
931 return (error);
932 }
933
934 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
935 return (-1);
936
937 if (cb->aio_nbytes >
938 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
939 return (-1);
940
941 ki = p->p_aioinfo;
942 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
943 return (-1);
944
945 ki->kaio_buffer_count++;
946
947 lj = aiocbe->lio;
948 if (lj)
949 lj->lioj_buffer_count++;
950
951 /* Create and build a buffer header for a transfer. */
952 bp = (struct buf *)getpbuf(NULL);
953 BUF_KERNPROC(bp);
954
955 /*
956 * Get a copy of the kva from the physical buffer.
957 */
958 bp->b_caller1 = p;
959 bp->b_dev = vp->v_rdev;
960 error = 0;
961
962 bp->b_bcount = cb->aio_nbytes;
963 bp->b_bufsize = cb->aio_nbytes;
964 bp->b_flags = B_PHYS | B_CALL | (cb->aio_lio_opcode == LIO_WRITE ?
965 B_WRITE : B_READ);
966 bp->b_iodone = aio_physwakeup;
967 bp->b_saveaddr = bp->b_data;
968 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
969 bp->b_blkno = btodb(cb->aio_offset);
970
971 /* Bring buffer into kernel space. */
972 if (vmapbuf(bp) < 0) {
973 error = EFAULT;
974 goto doerror;
975 }
976
e43a034f
MD
977 crit_enter();
978
984263bc
MD
979 aiocbe->bp = bp;
980 bp->b_spc = (void *)aiocbe;
981 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
982 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
983 aiocbe->jobstate = JOBST_JOBQBUF;
984 cb->_aiocb_private.status = cb->aio_nbytes;
985 num_buf_aio++;
986 bp->b_error = 0;
987
e43a034f 988 crit_exit();
984263bc
MD
989
990 /* Perform transfer. */
991 BUF_STRATEGY(bp, 0);
992
993 notify = 0;
e43a034f 994 crit_enter();
984263bc
MD
995
996 /*
997 * If we had an error invoking the request, or an error in processing
998 * the request before we have returned, we process it as an error in
999 * transfer. Note that such an I/O error is not indicated immediately,
1000 * but is returned using the aio_error mechanism. In this case,
1001 * aio_suspend will return immediately.
1002 */
1003 if (bp->b_error || (bp->b_flags & B_ERROR)) {
1004 struct aiocb *job = aiocbe->uuaiocb;
1005
1006 aiocbe->uaiocb._aiocb_private.status = 0;
1007 suword(&job->_aiocb_private.status, 0);
1008 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1009 suword(&job->_aiocb_private.error, bp->b_error);
1010
1011 ki->kaio_buffer_finished_count++;
1012
1013 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1014 aiocbe->jobstate = JOBST_JOBBFINISHED;
1015 aiocbe->jobflags |= AIOCBLIST_DONE;
1016 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1017 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1018 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1019 notify = 1;
1020 }
1021 }
e43a034f 1022 crit_exit();
984263bc
MD
1023 if (notify)
1024 KNOTE(&aiocbe->klist, 0);
1025 return 0;
1026
1027doerror:
1028 ki->kaio_buffer_count--;
1029 if (lj)
1030 lj->lioj_buffer_count--;
1031 aiocbe->bp = NULL;
1032 relpbuf(bp, NULL);
1033 return error;
1034}
1035
1036/*
1037 * This waits/tests physio completion.
1038 */
1039static int
1040aio_fphysio(struct aiocblist *iocb)
1041{
984263bc
MD
1042 struct buf *bp;
1043 int error;
1044
1045 bp = iocb->bp;
1046
e43a034f 1047 crit_enter();
984263bc 1048 while ((bp->b_flags & B_DONE) == 0) {
377d4740 1049 if (tsleep(bp, 0, "physstr", aiod_timeout)) {
984263bc 1050 if ((bp->b_flags & B_DONE) == 0) {
e43a034f 1051 crit_exit();
984263bc
MD
1052 return EINPROGRESS;
1053 } else
1054 break;
1055 }
1056 }
e43a034f 1057 crit_exit();
984263bc
MD
1058
1059 /* Release mapping into kernel space. */
1060 vunmapbuf(bp);
1061 iocb->bp = 0;
1062
1063 error = 0;
1064
1065 /* Check for an error. */
1066 if (bp->b_flags & B_ERROR)
1067 error = bp->b_error;
1068
1069 relpbuf(bp, NULL);
1070 return (error);
1071}
1072#endif /* VFS_AIO */
1073
1074/*
1075 * Wake up aio requests that may be serviceable now.
1076 */
1077void
1078aio_swake(struct socket *so, struct sockbuf *sb)
1079{
1080#ifndef VFS_AIO
1081 return;
1082#else
1083 struct aiocblist *cb,*cbn;
1084 struct proc *p;
1085 struct kaioinfo *ki = NULL;
1086 int opcode, wakecount = 0;
1087 struct aioproclist *aiop;
1088
1089 if (sb == &so->so_snd) {
1090 opcode = LIO_WRITE;
1091 so->so_snd.sb_flags &= ~SB_AIO;
1092 } else {
1093 opcode = LIO_READ;
1094 so->so_rcv.sb_flags &= ~SB_AIO;
1095 }
1096
1097 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1098 cbn = TAILQ_NEXT(cb, list);
1099 if (opcode == cb->uaiocb.aio_lio_opcode) {
1100 p = cb->userproc;
1101 ki = p->p_aioinfo;
1102 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1103 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1104 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1105 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1106 wakecount++;
1107 if (cb->jobstate != JOBST_JOBQGLOBAL)
1108 panic("invalid queue value");
1109 }
1110 }
1111
1112 while (wakecount--) {
1113 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1114 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1115 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1116 aiop->aioprocflags &= ~AIOP_FREE;
1117 wakeup(aiop->aioproc);
1118 }
1119 }
1120#endif /* VFS_AIO */
1121}
1122
1123#ifdef VFS_AIO
1124/*
1125 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1126 * technique is done in this code.
1127 */
1128static int
41c20dac 1129_aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
984263bc 1130{
7b95be2a 1131 struct proc *p = curproc;
984263bc
MD
1132 struct filedesc *fdp;
1133 struct file *fp;
1134 unsigned int fd;
1135 struct socket *so;
984263bc
MD
1136 int error;
1137 int opcode, user_opcode;
1138 struct aiocblist *aiocbe;
1139 struct aioproclist *aiop;
1140 struct kaioinfo *ki;
1141 struct kevent kev;
1142 struct kqueue *kq;
1143 struct file *kq_fp;
1144
1145 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1146 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1147 else
1148 aiocbe = zalloc (aiocb_zone);
1149
1150 aiocbe->inputcharge = 0;
1151 aiocbe->outputcharge = 0;
a5eb27b6 1152 callout_init(&aiocbe->timeout);
984263bc
MD
1153 SLIST_INIT(&aiocbe->klist);
1154
1155 suword(&job->_aiocb_private.status, -1);
1156 suword(&job->_aiocb_private.error, 0);
1157 suword(&job->_aiocb_private.kernelinfo, -1);
1158
1159 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1160 if (error) {
1161 suword(&job->_aiocb_private.error, error);
1162 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1163 return error;
1164 }
1165 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1166 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1167 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1168 return EINVAL;
1169 }
1170
1171 /* Save userspace address of the job info. */
1172 aiocbe->uuaiocb = job;
1173
1174 /* Get the opcode. */
1175 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1176 if (type != LIO_NOP)
1177 aiocbe->uaiocb.aio_lio_opcode = type;
1178 opcode = aiocbe->uaiocb.aio_lio_opcode;
1179
1180 /* Get the fd info for process. */
1181 fdp = p->p_fd;
1182
1183 /*
1184 * Range check file descriptor.
1185 */
1186 fd = aiocbe->uaiocb.aio_fildes;
1187 if (fd >= fdp->fd_nfiles) {
1188 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1189 if (type == 0)
1190 suword(&job->_aiocb_private.error, EBADF);
1191 return EBADF;
1192 }
1193
1194 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1195 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1196 0))) {
1197 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1198 if (type == 0)
1199 suword(&job->_aiocb_private.error, EBADF);
1200 return EBADF;
1201 }
1202 fhold(fp);
1203
1204 if (aiocbe->uaiocb.aio_offset == -1LL) {
1205 error = EINVAL;
1206 goto aqueue_fail;
1207 }
1208 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1209 if (error) {
1210 error = EINVAL;
1211 goto aqueue_fail;
1212 }
1213 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1214 if (jobrefid == LONG_MAX)
1215 jobrefid = 1;
1216 else
1217 jobrefid++;
1218
1219 if (opcode == LIO_NOP) {
7b95be2a 1220 fdrop(fp, p->p_thread);
984263bc
MD
1221 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1222 if (type == 0) {
1223 suword(&job->_aiocb_private.error, 0);
1224 suword(&job->_aiocb_private.status, 0);
1225 suword(&job->_aiocb_private.kernelinfo, 0);
1226 }
1227 return 0;
1228 }
1229 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1230 if (type == 0)
1231 suword(&job->_aiocb_private.status, 0);
1232 error = EINVAL;
1233 goto aqueue_fail;
1234 }
1235
1236 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1237 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1238 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1239 }
1240 else {
1241 /*
1242 * This method for requesting kevent-based notification won't
1243 * work on the alpha, since we're passing in a pointer
1244 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1245 * based method instead.
1246 */
1247 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1248 user_opcode == LIO_WRITE)
1249 goto no_kqueue;
1250
1251 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1252 &kev, sizeof(kev));
1253 if (error)
1254 goto aqueue_fail;
1255 }
1256 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1257 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1258 (kq_fp->f_type != DTYPE_KQUEUE)) {
1259 error = EBADF;
1260 goto aqueue_fail;
1261 }
1262 kq = (struct kqueue *)kq_fp->f_data;
1263 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1264 kev.filter = EVFILT_AIO;
1265 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1266 kev.data = (intptr_t)aiocbe;
7b95be2a 1267 error = kqueue_register(kq, &kev, p->p_thread);
984263bc
MD
1268aqueue_fail:
1269 if (error) {
7b95be2a 1270 fdrop(fp, p->p_thread);
984263bc
MD
1271 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1272 if (type == 0)
1273 suword(&job->_aiocb_private.error, error);
1274 goto done;
1275 }
1276no_kqueue:
1277
1278 suword(&job->_aiocb_private.error, EINPROGRESS);
1279 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1280 aiocbe->userproc = p;
1281 aiocbe->jobflags = 0;
1282 aiocbe->lio = lj;
1283 ki = p->p_aioinfo;
1284
1285 if (fp->f_type == DTYPE_SOCKET) {
1286 /*
1287 * Alternate queueing for socket ops: Reach down into the
1288 * descriptor to get the socket data. Then check to see if the
1289 * socket is ready to be read or written (based on the requested
1290 * operation).
1291 *
1292 * If it is not ready for io, then queue the aiocbe on the
1293 * socket, and set the flags so we get a call when sbnotify()
1294 * happens.
1295 */
1296 so = (struct socket *)fp->f_data;
e43a034f 1297 crit_enter();
984263bc
MD
1298 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1299 LIO_WRITE) && (!sowriteable(so)))) {
1300 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1301 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1302 if (opcode == LIO_READ)
1303 so->so_rcv.sb_flags |= SB_AIO;
1304 else
1305 so->so_snd.sb_flags |= SB_AIO;
1306 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1307 ki->kaio_queue_count++;
1308 num_queue_count++;
e43a034f 1309 crit_exit();
984263bc
MD
1310 error = 0;
1311 goto done;
1312 }
e43a034f 1313 crit_exit();
984263bc
MD
1314 }
1315
1316 if ((error = aio_qphysio(p, aiocbe)) == 0)
1317 goto done;
1318 if (error > 0) {
1319 suword(&job->_aiocb_private.status, 0);
1320 aiocbe->uaiocb._aiocb_private.error = error;
1321 suword(&job->_aiocb_private.error, error);
1322 goto done;
1323 }
1324
1325 /* No buffer for daemon I/O. */
1326 aiocbe->bp = NULL;
1327
1328 ki->kaio_queue_count++;
1329 if (lj)
1330 lj->lioj_queue_count++;
e43a034f 1331 crit_enter();
984263bc
MD
1332 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1333 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
e43a034f 1334 crit_exit();
984263bc
MD
1335 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1336
1337 num_queue_count++;
1338 error = 0;
1339
1340 /*
1341 * If we don't have a free AIO process, and we are below our quota, then
1342 * start one. Otherwise, depend on the subsequent I/O completions to
d0d91865 1343 * pick-up this job. If we don't successfully create the new process
984263bc
MD
1344 * (thread) due to resource issues, we return an error for now (EAGAIN),
1345 * which is likely not the correct thing to do.
1346 */
e43a034f 1347 crit_enter();
984263bc
MD
1348retryproc:
1349 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1350 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1351 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1352 aiop->aioprocflags &= ~AIOP_FREE;
1353 wakeup(aiop->aioproc);
1354 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1355 ((ki->kaio_active_count + num_aio_resv_start) <
1356 ki->kaio_maxactive_count)) {
1357 num_aio_resv_start++;
1358 if ((error = aio_newproc()) == 0) {
1359 num_aio_resv_start--;
1360 goto retryproc;
1361 }
1362 num_aio_resv_start--;
1363 }
e43a034f 1364 crit_exit();
984263bc
MD
1365done:
1366 return error;
1367}
1368
1369/*
1370 * This routine queues an AIO request, checking for quotas.
1371 */
1372static int
41c20dac 1373aio_aqueue(struct aiocb *job, int type)
984263bc 1374{
7b95be2a 1375 struct proc *p = curproc;
984263bc
MD
1376 struct kaioinfo *ki;
1377
1378 if (p->p_aioinfo == NULL)
1379 aio_init_aioinfo(p);
1380
1381 if (num_queue_count >= max_queue_count)
1382 return EAGAIN;
1383
1384 ki = p->p_aioinfo;
1385 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1386 return EAGAIN;
1387
41c20dac 1388 return _aio_aqueue(job, NULL, type);
984263bc
MD
1389}
1390#endif /* VFS_AIO */
1391
1392/*
1393 * Support the aio_return system call, as a side-effect, kernel resources are
1394 * released.
1395 */
1396int
41c20dac 1397aio_return(struct aio_return_args *uap)
984263bc
MD
1398{
1399#ifndef VFS_AIO
1400 return ENOSYS;
1401#else
41c20dac 1402 struct proc *p = curproc;
984263bc
MD
1403 long jobref;
1404 struct aiocblist *cb, *ncb;
1405 struct aiocb *ujob;
1406 struct kaioinfo *ki;
1407
1408 ki = p->p_aioinfo;
1409 if (ki == NULL)
1410 return EINVAL;
1411
1412 ujob = uap->aiocbp;
1413
1414 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1415 if (jobref == -1 || jobref == 0)
1416 return EINVAL;
1417
1418 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1419 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1420 jobref) {
1421 if (ujob == cb->uuaiocb) {
c7114eea 1422 uap->sysmsg_result =
984263bc
MD
1423 cb->uaiocb._aiocb_private.status;
1424 } else
c7114eea 1425 uap->sysmsg_result = EFAULT;
984263bc
MD
1426 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1427 p->p_stats->p_ru.ru_oublock +=
1428 cb->outputcharge;
1429 cb->outputcharge = 0;
1430 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1431 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1432 cb->inputcharge = 0;
1433 }
1434 aio_free_entry(cb);
1435 return 0;
1436 }
1437 }
e43a034f 1438 crit_enter();
984263bc
MD
1439 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1440 ncb = TAILQ_NEXT(cb, plist);
1441 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1442 == jobref) {
e43a034f 1443 crit_exit();
984263bc 1444 if (ujob == cb->uuaiocb) {
c7114eea 1445 uap->sysmsg_result =
984263bc
MD
1446 cb->uaiocb._aiocb_private.status;
1447 } else
c7114eea 1448 uap->sysmsg_result = EFAULT;
984263bc
MD
1449 aio_free_entry(cb);
1450 return 0;
1451 }
1452 }
e43a034f 1453 crit_exit();
984263bc
MD
1454
1455 return (EINVAL);
1456#endif /* VFS_AIO */
1457}
1458
1459/*
1460 * Allow a process to wakeup when any of the I/O requests are completed.
1461 */
1462int
41c20dac 1463aio_suspend(struct aio_suspend_args *uap)
984263bc
MD
1464{
1465#ifndef VFS_AIO
1466 return ENOSYS;
1467#else
41c20dac 1468 struct proc *p = curproc;
984263bc
MD
1469 struct timeval atv;
1470 struct timespec ts;
1471 struct aiocb *const *cbptr, *cbp;
1472 struct kaioinfo *ki;
1473 struct aiocblist *cb;
1474 int i;
1475 int njoblist;
831f78e5 1476 int error, timo;
984263bc
MD
1477 long *ijoblist;
1478 struct aiocb **ujoblist;
1479
1480 if (uap->nent > AIO_LISTIO_MAX)
1481 return EINVAL;
1482
1483 timo = 0;
1484 if (uap->timeout) {
1485 /* Get timespec struct. */
1486 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1487 return error;
1488
1489 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1490 return (EINVAL);
1491
1492 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1493 if (itimerfix(&atv))
1494 return (EINVAL);
a94976ad 1495 timo = tvtohz_high(&atv);
984263bc
MD
1496 }
1497
1498 ki = p->p_aioinfo;
1499 if (ki == NULL)
1500 return EAGAIN;
1501
1502 njoblist = 0;
1503 ijoblist = zalloc(aiol_zone);
1504 ujoblist = zalloc(aiol_zone);
1505 cbptr = uap->aiocbp;
1506
1507 for (i = 0; i < uap->nent; i++) {
1508 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1509 if (cbp == 0)
1510 continue;
1511 ujoblist[njoblist] = cbp;
1512 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1513 njoblist++;
1514 }
1515
1516 if (njoblist == 0) {
1517 zfree(aiol_zone, ijoblist);
1518 zfree(aiol_zone, ujoblist);
1519 return 0;
1520 }
1521
1522 error = 0;
1523 for (;;) {
1524 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1525 for (i = 0; i < njoblist; i++) {
1526 if (((intptr_t)
1527 cb->uaiocb._aiocb_private.kernelinfo) ==
1528 ijoblist[i]) {
1529 if (ujoblist[i] != cb->uuaiocb)
1530 error = EINVAL;
1531 zfree(aiol_zone, ijoblist);
1532 zfree(aiol_zone, ujoblist);
1533 return error;
1534 }
1535 }
1536 }
1537
e43a034f 1538 crit_enter();
984263bc
MD
1539 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1540 TAILQ_NEXT(cb, plist)) {
1541 for (i = 0; i < njoblist; i++) {
1542 if (((intptr_t)
1543 cb->uaiocb._aiocb_private.kernelinfo) ==
1544 ijoblist[i]) {
e43a034f 1545 crit_exit();
984263bc
MD
1546 if (ujoblist[i] != cb->uuaiocb)
1547 error = EINVAL;
1548 zfree(aiol_zone, ijoblist);
1549 zfree(aiol_zone, ujoblist);
1550 return error;
1551 }
1552 }
1553 }
1554
1555 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 1556 error = tsleep(p, PCATCH, "aiospn", timo);
e43a034f 1557 crit_exit();
984263bc
MD
1558
1559 if (error == ERESTART || error == EINTR) {
1560 zfree(aiol_zone, ijoblist);
1561 zfree(aiol_zone, ujoblist);
1562 return EINTR;
1563 } else if (error == EWOULDBLOCK) {
1564 zfree(aiol_zone, ijoblist);
1565 zfree(aiol_zone, ujoblist);
1566 return EAGAIN;
1567 }
1568 }
1569
1570/* NOTREACHED */
1571 return EINVAL;
1572#endif /* VFS_AIO */
1573}
1574
1575/*
1576 * aio_cancel cancels any non-physio aio operations not currently in
1577 * progress.
1578 */
1579int
41c20dac 1580aio_cancel(struct aio_cancel_args *uap)
984263bc
MD
1581{
1582#ifndef VFS_AIO
1583 return ENOSYS;
1584#else
41c20dac 1585 struct proc *p = curproc;
984263bc
MD
1586 struct kaioinfo *ki;
1587 struct aiocblist *cbe, *cbn;
1588 struct file *fp;
1589 struct filedesc *fdp;
1590 struct socket *so;
1591 struct proc *po;
e43a034f 1592 int error;
984263bc
MD
1593 int cancelled=0;
1594 int notcancelled=0;
1595 struct vnode *vp;
1596
1597 fdp = p->p_fd;
1598 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1599 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1600 return (EBADF);
1601
1602 if (fp->f_type == DTYPE_VNODE) {
1603 vp = (struct vnode *)fp->f_data;
1604
1605 if (vn_isdisk(vp,&error)) {
c7114eea 1606 uap->sysmsg_result = AIO_NOTCANCELED;
984263bc
MD
1607 return 0;
1608 }
1609 } else if (fp->f_type == DTYPE_SOCKET) {
1610 so = (struct socket *)fp->f_data;
1611
e43a034f 1612 crit_enter();
984263bc
MD
1613
1614 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1615 cbn = TAILQ_NEXT(cbe, list);
1616 if ((uap->aiocbp == NULL) ||
1617 (uap->aiocbp == cbe->uuaiocb) ) {
1618 po = cbe->userproc;
1619 ki = po->p_aioinfo;
1620 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1621 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1622 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1623 if (ki->kaio_flags & KAIO_WAKEUP) {
1624 wakeup(po);
1625 }
1626 cbe->jobstate = JOBST_JOBFINISHED;
1627 cbe->uaiocb._aiocb_private.status=-1;
1628 cbe->uaiocb._aiocb_private.error=ECANCELED;
1629 cancelled++;
1630/* XXX cancelled, knote? */
1631 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1632 SIGEV_SIGNAL)
1633 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1634 if (uap->aiocbp)
1635 break;
1636 }
1637 }
e43a034f 1638 crit_exit();
984263bc
MD
1639
1640 if ((cancelled) && (uap->aiocbp)) {
c7114eea 1641 uap->sysmsg_result = AIO_CANCELED;
984263bc
MD
1642 return 0;
1643 }
1644 }
1645 ki=p->p_aioinfo;
1646 if (ki == NULL)
1647 goto done;
e43a034f 1648 crit_enter();
984263bc
MD
1649
1650 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1651 cbn = TAILQ_NEXT(cbe, plist);
1652
1653 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1654 ((uap->aiocbp == NULL ) ||
1655 (uap->aiocbp == cbe->uuaiocb))) {
1656
1657 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1658 TAILQ_REMOVE(&aio_jobs, cbe, list);
1659 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1660 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1661 plist);
1662 cancelled++;
1663 ki->kaio_queue_finished_count++;
1664 cbe->jobstate = JOBST_JOBFINISHED;
1665 cbe->uaiocb._aiocb_private.status = -1;
1666 cbe->uaiocb._aiocb_private.error = ECANCELED;
1667/* XXX cancelled, knote? */
1668 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1669 SIGEV_SIGNAL)
1670 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1671 } else {
1672 notcancelled++;
1673 }
1674 }
1675 }
e43a034f 1676 crit_exit();
984263bc
MD
1677done:
1678 if (notcancelled) {
c7114eea 1679 uap->sysmsg_result = AIO_NOTCANCELED;
984263bc
MD
1680 return 0;
1681 }
1682 if (cancelled) {
c7114eea 1683 uap->sysmsg_result = AIO_CANCELED;
984263bc
MD
1684 return 0;
1685 }
c7114eea 1686 uap->sysmsg_result = AIO_ALLDONE;
984263bc
MD
1687
1688 return 0;
1689#endif /* VFS_AIO */
1690}
1691
1692/*
1693 * aio_error is implemented in the kernel level for compatibility purposes only.
1694 * For a user mode async implementation, it would be best to do it in a userland
1695 * subroutine.
1696 */
1697int
41c20dac 1698aio_error(struct aio_error_args *uap)
984263bc
MD
1699{
1700#ifndef VFS_AIO
1701 return ENOSYS;
1702#else
41c20dac 1703 struct proc *p = curproc;
984263bc
MD
1704 struct aiocblist *cb;
1705 struct kaioinfo *ki;
1706 long jobref;
1707
1708 ki = p->p_aioinfo;
1709 if (ki == NULL)
1710 return EINVAL;
1711
1712 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1713 if ((jobref == -1) || (jobref == 0))
1714 return EINVAL;
1715
1716 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1717 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1718 jobref) {
c7114eea 1719 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
984263bc
MD
1720 return 0;
1721 }
1722 }
1723
e43a034f 1724 crit_enter();
984263bc
MD
1725
1726 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1727 plist)) {
1728 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1729 jobref) {
c7114eea 1730 uap->sysmsg_result = EINPROGRESS;
e43a034f 1731 crit_exit();
984263bc
MD
1732 return 0;
1733 }
1734 }
1735
1736 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1737 plist)) {
1738 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1739 jobref) {
c7114eea 1740 uap->sysmsg_result = EINPROGRESS;
e43a034f 1741 crit_exit();
984263bc
MD
1742 return 0;
1743 }
1744 }
e43a034f 1745 crit_exit();
984263bc 1746
e43a034f 1747 crit_enter();
984263bc
MD
1748 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1749 plist)) {
1750 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1751 jobref) {
c7114eea 1752 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
e43a034f 1753 crit_exit();
984263bc
MD
1754 return 0;
1755 }
1756 }
1757
1758 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1759 plist)) {
1760 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1761 jobref) {
c7114eea 1762 uap->sysmsg_result = EINPROGRESS;
e43a034f 1763 crit_exit();
984263bc
MD
1764 return 0;
1765 }
1766 }
e43a034f 1767 crit_exit();
984263bc
MD
1768
1769#if (0)
1770 /*
1771 * Hack for lio.
1772 */
1773 status = fuword(&uap->aiocbp->_aiocb_private.status);
1774 if (status == -1)
1775 return fuword(&uap->aiocbp->_aiocb_private.error);
1776#endif
1777 return EINVAL;
1778#endif /* VFS_AIO */
1779}
1780
1781/* syscall - asynchronous read from a file (REALTIME) */
1782int
41c20dac 1783aio_read(struct aio_read_args *uap)
984263bc
MD
1784{
1785#ifndef VFS_AIO
1786 return ENOSYS;
1787#else
41c20dac 1788 return aio_aqueue(uap->aiocbp, LIO_READ);
984263bc
MD
1789#endif /* VFS_AIO */
1790}
1791
1792/* syscall - asynchronous write to a file (REALTIME) */
1793int
41c20dac 1794aio_write(struct aio_write_args *uap)
984263bc
MD
1795{
1796#ifndef VFS_AIO
1797 return ENOSYS;
1798#else
41c20dac 1799 return aio_aqueue(uap->aiocbp, LIO_WRITE);
984263bc
MD
1800#endif /* VFS_AIO */
1801}
1802
1803/* syscall - XXX undocumented */
1804int
41c20dac 1805lio_listio(struct lio_listio_args *uap)
984263bc
MD
1806{
1807#ifndef VFS_AIO
1808 return ENOSYS;
1809#else
41c20dac 1810 struct proc *p = curproc;
984263bc
MD
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;
984263bc
MD
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 if (p->p_aioinfo == NULL)
1828 aio_init_aioinfo(p);
1829
1830 if ((nent + num_queue_count) > max_queue_count)
1831 return EAGAIN;
1832
1833 ki = p->p_aioinfo;
1834 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1835 return EAGAIN;
1836
1837 lj = zalloc(aiolio_zone);
1838 if (!lj)
1839 return EAGAIN;
1840
1841 lj->lioj_flags = 0;
1842 lj->lioj_buffer_count = 0;
1843 lj->lioj_buffer_finished_count = 0;
1844 lj->lioj_queue_count = 0;
1845 lj->lioj_queue_finished_count = 0;
1846 lj->lioj_ki = ki;
1847
1848 /*
1849 * Setup signal.
1850 */
1851 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1852 error = copyin(uap->sig, &lj->lioj_signal,
1853 sizeof(lj->lioj_signal));
1854 if (error) {
1855 zfree(aiolio_zone, lj);
1856 return error;
1857 }
1858 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1859 zfree(aiolio_zone, lj);
1860 return EINVAL;
1861 }
1862 lj->lioj_flags |= LIOJ_SIGNAL;
1863 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1864 } else
1865 lj->lioj_flags &= ~LIOJ_SIGNAL;
1866
1867 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1868 /*
1869 * Get pointers to the list of I/O requests.
1870 */
1871 nerror = 0;
1872 nentqueued = 0;
1873 cbptr = uap->acb_list;
1874 for (i = 0; i < uap->nent; i++) {
1875 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1876 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
41c20dac 1877 error = _aio_aqueue(iocb, lj, 0);
984263bc
MD
1878 if (error == 0)
1879 nentqueued++;
1880 else
1881 nerror++;
1882 }
1883 }
1884
1885 /*
1886 * If we haven't queued any, then just return error.
1887 */
1888 if (nentqueued == 0)
1889 return 0;
1890
1891 /*
1892 * Calculate the appropriate error return.
1893 */
1894 runningcode = 0;
1895 if (nerror)
1896 runningcode = EIO;
1897
1898 if (uap->mode == LIO_WAIT) {
1899 int command, found, jobref;
1900
1901 for (;;) {
1902 found = 0;
1903 for (i = 0; i < uap->nent; i++) {
1904 /*
1905 * Fetch address of the control buf pointer in
1906 * user space.
1907 */
1908 iocb = (struct aiocb *)
1909 (intptr_t)fuword(&cbptr[i]);
1910 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1911 == 0))
1912 continue;
1913
1914 /*
1915 * Fetch the associated command from user space.
1916 */
1917 command = fuword(&iocb->aio_lio_opcode);
1918 if (command == LIO_NOP) {
1919 found++;
1920 continue;
1921 }
1922
1923 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1924
1925 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1926 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1927 == jobref) {
1928 if (cb->uaiocb.aio_lio_opcode
1929 == LIO_WRITE) {
1930 p->p_stats->p_ru.ru_oublock
1931 +=
1932 cb->outputcharge;
1933 cb->outputcharge = 0;
1934 } else if (cb->uaiocb.aio_lio_opcode
1935 == LIO_READ) {
1936 p->p_stats->p_ru.ru_inblock
1937 += cb->inputcharge;
1938 cb->inputcharge = 0;
1939 }
1940 found++;
1941 break;
1942 }
1943 }
1944
e43a034f 1945 crit_enter();
984263bc
MD
1946 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1947 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1948 == jobref) {
1949 found++;
1950 break;
1951 }
1952 }
e43a034f 1953 crit_exit();
984263bc
MD
1954 }
1955
1956 /*
1957 * If all I/Os have been disposed of, then we can
1958 * return.
1959 */
1960 if (found == nentqueued)
1961 return runningcode;
1962
1963 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 1964 error = tsleep(p, PCATCH, "aiospn", 0);
984263bc
MD
1965
1966 if (error == EINTR)
1967 return EINTR;
1968 else if (error == EWOULDBLOCK)
1969 return EAGAIN;
1970 }
1971 }
1972
1973 return runningcode;
1974#endif /* VFS_AIO */
1975}
1976
1977#ifdef VFS_AIO
1978/*
1979 * This is a weird hack so that we can post a signal. It is safe to do so from
1980 * a timeout routine, but *not* from an interrupt routine.
1981 */
1982static void
1983process_signal(void *aioj)
1984{
1985 struct aiocblist *aiocbe = aioj;
1986 struct aio_liojob *lj = aiocbe->lio;
1987 struct aiocb *cb = &aiocbe->uaiocb;
1988
1989 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
1990 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
1991 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1992 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1993 }
1994
1995 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
1996 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
1997}
1998
1999/*
2000 * Interrupt handler for physio, performs the necessary process wakeups, and
2001 * signals.
2002 */
2003static void
2004aio_physwakeup(struct buf *bp)
2005{
2006 struct aiocblist *aiocbe;
2007 struct proc *p;
2008 struct kaioinfo *ki;
2009 struct aio_liojob *lj;
2010
2011 wakeup(bp);
2012
2013 aiocbe = (struct aiocblist *)bp->b_spc;
2014 if (aiocbe) {
2015 p = bp->b_caller1;
2016
2017 aiocbe->jobstate = JOBST_JOBBFINISHED;
2018 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2019 aiocbe->uaiocb._aiocb_private.error = 0;
2020 aiocbe->jobflags |= AIOCBLIST_DONE;
2021
2022 if (bp->b_flags & B_ERROR)
2023 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2024
2025 lj = aiocbe->lio;
2026 if (lj) {
2027 lj->lioj_buffer_finished_count++;
2028
2029 /*
2030 * wakeup/signal if all of the interrupt jobs are done.
2031 */
2032 if (lj->lioj_buffer_finished_count ==
2033 lj->lioj_buffer_count) {
2034 /*
2035 * Post a signal if it is called for.
2036 */
2037 if ((lj->lioj_flags &
2038 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2039 LIOJ_SIGNAL) {
2040 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
a5eb27b6
MD
2041 callout_reset(&aiocbe->timeout, 0,
2042 process_signal, aiocbe);
984263bc
MD
2043 }
2044 }
2045 }
2046
2047 ki = p->p_aioinfo;
2048 if (ki) {
2049 ki->kaio_buffer_finished_count++;
2050 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2051 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2052 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2053
2054 KNOTE(&aiocbe->klist, 0);
2055 /* Do the wakeup. */
2056 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2057 ki->kaio_flags &= ~KAIO_WAKEUP;
2058 wakeup(p);
2059 }
2060 }
2061
a5eb27b6
MD
2062 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2063 callout_reset(&aiocbe->timeout, 0,
2064 process_signal, aiocbe);
2065 }
984263bc
MD
2066 }
2067}
2068#endif /* VFS_AIO */
2069
2070/* syscall - wait for the next completion of an aio request */
2071int
41c20dac 2072aio_waitcomplete(struct aio_waitcomplete_args *uap)
984263bc
MD
2073{
2074#ifndef VFS_AIO
2075 return ENOSYS;
2076#else
41c20dac 2077 struct proc *p = curproc;
984263bc
MD
2078 struct timeval atv;
2079 struct timespec ts;
2080 struct kaioinfo *ki;
2081 struct aiocblist *cb = NULL;
831f78e5 2082 int error, timo;
984263bc
MD
2083
2084 suword(uap->aiocbp, (int)NULL);
2085
2086 timo = 0;
2087 if (uap->timeout) {
2088 /* Get timespec struct. */
2089 error = copyin(uap->timeout, &ts, sizeof(ts));
2090 if (error)
2091 return error;
2092
2093 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2094 return (EINVAL);
2095
2096 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2097 if (itimerfix(&atv))
2098 return (EINVAL);
a94976ad 2099 timo = tvtohz_high(&atv);
984263bc
MD
2100 }
2101
2102 ki = p->p_aioinfo;
2103 if (ki == NULL)
2104 return EAGAIN;
2105
2106 for (;;) {
2107 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2108 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
c7114eea 2109 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
984263bc
MD
2110 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2111 p->p_stats->p_ru.ru_oublock +=
2112 cb->outputcharge;
2113 cb->outputcharge = 0;
2114 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2115 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2116 cb->inputcharge = 0;
2117 }
2118 aio_free_entry(cb);
2119 return cb->uaiocb._aiocb_private.error;
2120 }
2121
e43a034f 2122 crit_enter();
984263bc 2123 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
e43a034f 2124 crit_exit();
984263bc 2125 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
c7114eea 2126 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
984263bc
MD
2127 aio_free_entry(cb);
2128 return cb->uaiocb._aiocb_private.error;
2129 }
2130
2131 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 2132 error = tsleep(p, PCATCH, "aiowc", timo);
e43a034f 2133 crit_exit();
984263bc
MD
2134
2135 if (error == ERESTART)
2136 return EINTR;
2137 else if (error < 0)
2138 return error;
2139 else if (error == EINTR)
2140 return EINTR;
2141 else if (error == EWOULDBLOCK)
2142 return EAGAIN;
2143 }
2144#endif /* VFS_AIO */
2145}
2146
2147#ifndef VFS_AIO
2148static int
2149filt_aioattach(struct knote *kn)
2150{
2151
2152 return (ENXIO);
2153}
2154
2155struct filterops aio_filtops =
2156 { 0, filt_aioattach, NULL, NULL };
2157
2158#else
2159/* kqueue attach function */
2160static int
2161filt_aioattach(struct knote *kn)
2162{
2163 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2164
2165 /*
2166 * The aiocbe pointer must be validated before using it, so
2167 * registration is restricted to the kernel; the user cannot
2168 * set EV_FLAG1.
2169 */
2170 if ((kn->kn_flags & EV_FLAG1) == 0)
2171 return (EPERM);
2172 kn->kn_flags &= ~EV_FLAG1;
2173
2174 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2175
2176 return (0);
2177}
2178
2179/* kqueue detach function */
2180static void
2181filt_aiodetach(struct knote *kn)
2182{
2183 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2184
2185 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2186}
2187
2188/* kqueue filter function */
2189/*ARGSUSED*/
2190static int
2191filt_aio(struct knote *kn, long hint)
2192{
2193 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2194
2195 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2196 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2197 aiocbe->jobstate != JOBST_JOBBFINISHED)
2198 return (0);
2199 kn->kn_flags |= EV_EOF;
2200 return (1);
2201}
2202
2203struct filterops aio_filtops =
2204 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2205#endif /* VFS_AIO */