FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_aio.c
1 /*-
2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. John S. Dyson's name may not be used to endorse or promote products
10 * derived from this software without specific prior written permission.
11 *
12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
13 * bad that happens because of using this software isn't the responsibility
14 * of the author. This software is distributed AS-IS.
15 */
16
17 /*
18 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
19 */
20
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD: releng/9.0/sys/kern/vfs_aio.c 225617 2011-09-16 13:58:51Z kmacy $");
23
24 #include "opt_compat.h"
25
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
29 #include <sys/bio.h>
30 #include <sys/buf.h>
31 #include <sys/capability.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
39 #include <sys/file.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
45 #include <sys/proc.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/protosw.h>
49 #include <sys/sema.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/syscall.h>
53 #include <sys/sysent.h>
54 #include <sys/sysctl.h>
55 #include <sys/sx.h>
56 #include <sys/taskqueue.h>
57 #include <sys/vnode.h>
58 #include <sys/conf.h>
59 #include <sys/event.h>
60 #include <sys/mount.h>
61
62 #include <machine/atomic.h>
63
64 #include <vm/vm.h>
65 #include <vm/vm_extern.h>
66 #include <vm/pmap.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_object.h>
69 #include <vm/uma.h>
70 #include <sys/aio.h>
71
72 #include "opt_vfs_aio.h"
73
74 /*
75 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
76 * overflow. (XXX will be removed soon.)
77 */
78 static u_long jobrefid;
79
80 /*
81 * Counter for aio_fsync.
82 */
83 static uint64_t jobseqno;
84
85 #define JOBST_NULL 0
86 #define JOBST_JOBQSOCK 1
87 #define JOBST_JOBQGLOBAL 2
88 #define JOBST_JOBRUNNING 3
89 #define JOBST_JOBFINISHED 4
90 #define JOBST_JOBQBUF 5
91 #define JOBST_JOBQSYNC 6
92
93 #ifndef MAX_AIO_PER_PROC
94 #define MAX_AIO_PER_PROC 32
95 #endif
96
97 #ifndef MAX_AIO_QUEUE_PER_PROC
98 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
99 #endif
100
101 #ifndef MAX_AIO_PROCS
102 #define MAX_AIO_PROCS 32
103 #endif
104
105 #ifndef MAX_AIO_QUEUE
106 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
107 #endif
108
109 #ifndef TARGET_AIO_PROCS
110 #define TARGET_AIO_PROCS 4
111 #endif
112
113 #ifndef MAX_BUF_AIO
114 #define MAX_BUF_AIO 16
115 #endif
116
117 #ifndef AIOD_TIMEOUT_DEFAULT
118 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
119 #endif
120
121 #ifndef AIOD_LIFETIME_DEFAULT
122 #define AIOD_LIFETIME_DEFAULT (30 * hz)
123 #endif
124
125 FEATURE(aio, "Asynchronous I/O");
126
127 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
128
129 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
130
131 static int max_aio_procs = MAX_AIO_PROCS;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
133 CTLFLAG_RW, &max_aio_procs, 0,
134 "Maximum number of kernel threads to use for handling async IO ");
135
136 static int num_aio_procs = 0;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
138 CTLFLAG_RD, &num_aio_procs, 0,
139 "Number of presently active kernel threads for async IO");
140
141 /*
142 * The code will adjust the actual number of AIO processes towards this
143 * number when it gets a chance.
144 */
145 static int target_aio_procs = TARGET_AIO_PROCS;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
147 0, "Preferred number of ready kernel threads for async IO");
148
149 static int max_queue_count = MAX_AIO_QUEUE;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
151 "Maximum number of aio requests to queue, globally");
152
153 static int num_queue_count = 0;
154 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
155 "Number of queued aio requests");
156
157 static int num_buf_aio = 0;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
159 "Number of aio requests presently handled by the buf subsystem");
160
161 /* Number of async I/O thread in the process of being started */
162 /* XXX This should be local to aio_aqueue() */
163 static int num_aio_resv_start = 0;
164
165 static int aiod_timeout;
166 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
167 "Timeout value for synchronous aio operations");
168
169 static int aiod_lifetime;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
171 "Maximum lifetime for idle aiod");
172
173 static int unloadable = 0;
174 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
175 "Allow unload of aio (not recommended)");
176
177
178 static int max_aio_per_proc = MAX_AIO_PER_PROC;
179 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
180 0, "Maximum active aio requests per process (stored in the process)");
181
182 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
183 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
184 &max_aio_queue_per_proc, 0,
185 "Maximum queued aio requests per process (stored in the process)");
186
187 static int max_buf_aio = MAX_BUF_AIO;
188 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
189 "Maximum buf aio requests per process (stored in the process)");
190
191 typedef struct oaiocb {
192 int aio_fildes; /* File descriptor */
193 off_t aio_offset; /* File offset for I/O */
194 volatile void *aio_buf; /* I/O buffer in process space */
195 size_t aio_nbytes; /* Number of bytes for I/O */
196 struct osigevent aio_sigevent; /* Signal to deliver */
197 int aio_lio_opcode; /* LIO opcode */
198 int aio_reqprio; /* Request priority -- ignored */
199 struct __aiocb_private _aiocb_private;
200 } oaiocb_t;
201
202 /*
203 * Below is a key of locks used to protect each member of struct aiocblist
204 * aioliojob and kaioinfo and any backends.
205 *
206 * * - need not protected
207 * a - locked by kaioinfo lock
208 * b - locked by backend lock, the backend lock can be null in some cases,
209 * for example, BIO belongs to this type, in this case, proc lock is
210 * reused.
211 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
212 */
213
214 /*
215 * Current, there is only two backends: BIO and generic file I/O.
216 * socket I/O is served by generic file I/O, this is not a good idea, since
217 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
218 * threads, if there is no thread to serve socket I/O, the socket I/O will be
219 * delayed too long or starved, we should create some threads dedicated to
220 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
221 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
222 * structure is not safe because there is race between userland and aio
223 * daemons.
224 */
225
226 struct aiocblist {
227 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
228 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
229 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
230 int jobflags; /* (a) job flags */
231 int jobstate; /* (b) job state */
232 int inputcharge; /* (*) input blockes */
233 int outputcharge; /* (*) output blockes */
234 struct buf *bp; /* (*) private to BIO backend,
235 * buffer pointer
236 */
237 struct proc *userproc; /* (*) user process */
238 struct ucred *cred; /* (*) active credential when created */
239 struct file *fd_file; /* (*) pointer to file structure */
240 struct aioliojob *lio; /* (*) optional lio job */
241 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
242 struct knlist klist; /* (a) list of knotes */
243 struct aiocb uaiocb; /* (*) kernel I/O control block */
244 ksiginfo_t ksi; /* (a) realtime signal info */
245 struct task biotask; /* (*) private to BIO backend */
246 uint64_t seqno; /* (*) job number */
247 int pending; /* (a) number of pending I/O, aio_fsync only */
248 };
249
250 /* jobflags */
251 #define AIOCBLIST_DONE 0x01
252 #define AIOCBLIST_BUFDONE 0x02
253 #define AIOCBLIST_RUNDOWN 0x04
254 #define AIOCBLIST_CHECKSYNC 0x08
255
256 /*
257 * AIO process info
258 */
259 #define AIOP_FREE 0x1 /* proc on free queue */
260
261 struct aiothreadlist {
262 int aiothreadflags; /* (c) AIO proc flags */
263 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
264 struct thread *aiothread; /* (*) the AIO thread */
265 };
266
267 /*
268 * data-structure for lio signal management
269 */
270 struct aioliojob {
271 int lioj_flags; /* (a) listio flags */
272 int lioj_count; /* (a) listio flags */
273 int lioj_finished_count; /* (a) listio flags */
274 struct sigevent lioj_signal; /* (a) signal on all I/O done */
275 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
276 struct knlist klist; /* (a) list of knotes */
277 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
278 };
279
280 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
281 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
282 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
283
284 /*
285 * per process aio data structure
286 */
287 struct kaioinfo {
288 struct mtx kaio_mtx; /* the lock to protect this struct */
289 int kaio_flags; /* (a) per process kaio flags */
290 int kaio_maxactive_count; /* (*) maximum number of AIOs */
291 int kaio_active_count; /* (c) number of currently used AIOs */
292 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
293 int kaio_count; /* (a) size of AIO queue */
294 int kaio_ballowed_count; /* (*) maximum number of buffers */
295 int kaio_buffer_count; /* (a) number of physio buffers */
296 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
297 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
298 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
299 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
300 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
301 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
302 * NOT USED YET.
303 */
304 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
305 struct task kaio_task; /* (*) task to kick aio threads */
306 };
307
308 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
309 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
310 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
311 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
312
313 #define KAIO_RUNDOWN 0x1 /* process is being run down */
314 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
315
316 /*
317 * Operations used to interact with userland aio control blocks.
318 * Different ABIs provide their own operations.
319 */
320 struct aiocb_ops {
321 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
322 long (*fetch_status)(struct aiocb *ujob);
323 long (*fetch_error)(struct aiocb *ujob);
324 int (*store_status)(struct aiocb *ujob, long status);
325 int (*store_error)(struct aiocb *ujob, long error);
326 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
327 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
328 };
329
330 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
331 static struct sema aio_newproc_sem;
332 static struct mtx aio_job_mtx;
333 static struct mtx aio_sock_mtx;
334 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
335 static struct unrhdr *aiod_unr;
336
337 void aio_init_aioinfo(struct proc *p);
338 static int aio_onceonly(void);
339 static int aio_free_entry(struct aiocblist *aiocbe);
340 static void aio_process(struct aiocblist *aiocbe);
341 static int aio_newproc(int *);
342 int aio_aqueue(struct thread *td, struct aiocb *job,
343 struct aioliojob *lio, int type, struct aiocb_ops *ops);
344 static void aio_physwakeup(struct buf *bp);
345 static void aio_proc_rundown(void *arg, struct proc *p);
346 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
347 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
348 static void biohelper(void *, int);
349 static void aio_daemon(void *param);
350 static void aio_swake_cb(struct socket *, struct sockbuf *);
351 static int aio_unload(void);
352 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
353 #define DONE_BUF 1
354 #define DONE_QUEUE 2
355 static int aio_kick(struct proc *userp);
356 static void aio_kick_nowait(struct proc *userp);
357 static void aio_kick_helper(void *context, int pending);
358 static int filt_aioattach(struct knote *kn);
359 static void filt_aiodetach(struct knote *kn);
360 static int filt_aio(struct knote *kn, long hint);
361 static int filt_lioattach(struct knote *kn);
362 static void filt_liodetach(struct knote *kn);
363 static int filt_lio(struct knote *kn, long hint);
364
365 /*
366 * Zones for:
367 * kaio Per process async io info
368 * aiop async io thread data
369 * aiocb async io jobs
370 * aiol list io job pointer - internal to aio_suspend XXX
371 * aiolio list io jobs
372 */
373 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
374
375 /* kqueue filters for aio */
376 static struct filterops aio_filtops = {
377 .f_isfd = 0,
378 .f_attach = filt_aioattach,
379 .f_detach = filt_aiodetach,
380 .f_event = filt_aio,
381 };
382 static struct filterops lio_filtops = {
383 .f_isfd = 0,
384 .f_attach = filt_lioattach,
385 .f_detach = filt_liodetach,
386 .f_event = filt_lio
387 };
388
389 static eventhandler_tag exit_tag, exec_tag;
390
391 TASKQUEUE_DEFINE_THREAD(aiod_bio);
392
393 /*
394 * Main operations function for use as a kernel module.
395 */
396 static int
397 aio_modload(struct module *module, int cmd, void *arg)
398 {
399 int error = 0;
400
401 switch (cmd) {
402 case MOD_LOAD:
403 aio_onceonly();
404 break;
405 case MOD_UNLOAD:
406 error = aio_unload();
407 break;
408 case MOD_SHUTDOWN:
409 break;
410 default:
411 error = EINVAL;
412 break;
413 }
414 return (error);
415 }
416
417 static moduledata_t aio_mod = {
418 "aio",
419 &aio_modload,
420 NULL
421 };
422
423 static struct syscall_helper_data aio_syscalls[] = {
424 SYSCALL_INIT_HELPER(aio_cancel),
425 SYSCALL_INIT_HELPER(aio_error),
426 SYSCALL_INIT_HELPER(aio_fsync),
427 SYSCALL_INIT_HELPER(aio_read),
428 SYSCALL_INIT_HELPER(aio_return),
429 SYSCALL_INIT_HELPER(aio_suspend),
430 SYSCALL_INIT_HELPER(aio_waitcomplete),
431 SYSCALL_INIT_HELPER(aio_write),
432 SYSCALL_INIT_HELPER(lio_listio),
433 SYSCALL_INIT_HELPER(oaio_read),
434 SYSCALL_INIT_HELPER(oaio_write),
435 SYSCALL_INIT_HELPER(olio_listio),
436 SYSCALL_INIT_LAST
437 };
438
439 #ifdef COMPAT_FREEBSD32
440 #include <sys/mount.h>
441 #include <sys/socket.h>
442 #include <compat/freebsd32/freebsd32.h>
443 #include <compat/freebsd32/freebsd32_proto.h>
444 #include <compat/freebsd32/freebsd32_signal.h>
445 #include <compat/freebsd32/freebsd32_syscall.h>
446 #include <compat/freebsd32/freebsd32_util.h>
447
448 static struct syscall_helper_data aio32_syscalls[] = {
449 SYSCALL32_INIT_HELPER(freebsd32_aio_return),
450 SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
451 SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
452 SYSCALL32_INIT_HELPER(freebsd32_aio_error),
453 SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
454 SYSCALL32_INIT_HELPER(freebsd32_aio_read),
455 SYSCALL32_INIT_HELPER(freebsd32_aio_write),
456 SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
457 SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
458 SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
459 SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
460 SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
461 SYSCALL_INIT_LAST
462 };
463 #endif
464
465 DECLARE_MODULE(aio, aio_mod,
466 SI_SUB_VFS, SI_ORDER_ANY);
467 MODULE_VERSION(aio, 1);
468
469 /*
470 * Startup initialization
471 */
472 static int
473 aio_onceonly(void)
474 {
475 int error;
476
477 /* XXX: should probably just use so->callback */
478 aio_swake = &aio_swake_cb;
479 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
480 EVENTHANDLER_PRI_ANY);
481 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
482 EVENTHANDLER_PRI_ANY);
483 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
484 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
485 TAILQ_INIT(&aio_freeproc);
486 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
487 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
488 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
489 TAILQ_INIT(&aio_jobs);
490 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
491 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
492 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
493 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
494 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
495 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
496 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
497 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
498 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
499 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
500 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
501 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
502 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
503 jobrefid = 1;
504 async_io_version = _POSIX_VERSION;
505 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
506 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
507 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
508
509 error = syscall_helper_register(aio_syscalls);
510 if (error)
511 return (error);
512 #ifdef COMPAT_FREEBSD32
513 error = syscall32_helper_register(aio32_syscalls);
514 if (error)
515 return (error);
516 #endif
517 return (0);
518 }
519
520 /*
521 * Callback for unload of AIO when used as a module.
522 */
523 static int
524 aio_unload(void)
525 {
526 int error;
527
528 /*
529 * XXX: no unloads by default, it's too dangerous.
530 * perhaps we could do it if locked out callers and then
531 * did an aio_proc_rundown() on each process.
532 *
533 * jhb: aio_proc_rundown() needs to run on curproc though,
534 * so I don't think that would fly.
535 */
536 if (!unloadable)
537 return (EOPNOTSUPP);
538
539 #ifdef COMPAT_FREEBSD32
540 syscall32_helper_unregister(aio32_syscalls);
541 #endif
542 syscall_helper_unregister(aio_syscalls);
543
544 error = kqueue_del_filteropts(EVFILT_AIO);
545 if (error)
546 return error;
547 error = kqueue_del_filteropts(EVFILT_LIO);
548 if (error)
549 return error;
550 async_io_version = 0;
551 aio_swake = NULL;
552 taskqueue_free(taskqueue_aiod_bio);
553 delete_unrhdr(aiod_unr);
554 uma_zdestroy(kaio_zone);
555 uma_zdestroy(aiop_zone);
556 uma_zdestroy(aiocb_zone);
557 uma_zdestroy(aiol_zone);
558 uma_zdestroy(aiolio_zone);
559 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
560 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
561 mtx_destroy(&aio_job_mtx);
562 mtx_destroy(&aio_sock_mtx);
563 sema_destroy(&aio_newproc_sem);
564 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
565 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
566 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
567 return (0);
568 }
569
570 /*
571 * Init the per-process aioinfo structure. The aioinfo limits are set
572 * per-process for user limit (resource) management.
573 */
574 void
575 aio_init_aioinfo(struct proc *p)
576 {
577 struct kaioinfo *ki;
578
579 ki = uma_zalloc(kaio_zone, M_WAITOK);
580 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
581 ki->kaio_flags = 0;
582 ki->kaio_maxactive_count = max_aio_per_proc;
583 ki->kaio_active_count = 0;
584 ki->kaio_qallowed_count = max_aio_queue_per_proc;
585 ki->kaio_count = 0;
586 ki->kaio_ballowed_count = max_buf_aio;
587 ki->kaio_buffer_count = 0;
588 TAILQ_INIT(&ki->kaio_all);
589 TAILQ_INIT(&ki->kaio_done);
590 TAILQ_INIT(&ki->kaio_jobqueue);
591 TAILQ_INIT(&ki->kaio_bufqueue);
592 TAILQ_INIT(&ki->kaio_liojoblist);
593 TAILQ_INIT(&ki->kaio_sockqueue);
594 TAILQ_INIT(&ki->kaio_syncqueue);
595 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
596 PROC_LOCK(p);
597 if (p->p_aioinfo == NULL) {
598 p->p_aioinfo = ki;
599 PROC_UNLOCK(p);
600 } else {
601 PROC_UNLOCK(p);
602 mtx_destroy(&ki->kaio_mtx);
603 uma_zfree(kaio_zone, ki);
604 }
605
606 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
607 aio_newproc(NULL);
608 }
609
610 static int
611 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
612 {
613 struct thread *td;
614 int error;
615
616 error = sigev_findtd(p, sigev, &td);
617 if (error)
618 return (error);
619 if (!KSI_ONQ(ksi)) {
620 ksiginfo_set_sigev(ksi, sigev);
621 ksi->ksi_code = SI_ASYNCIO;
622 ksi->ksi_flags |= KSI_EXT | KSI_INS;
623 tdsendsignal(p, td, ksi->ksi_signo, ksi);
624 }
625 PROC_UNLOCK(p);
626 return (error);
627 }
628
629 /*
630 * Free a job entry. Wait for completion if it is currently active, but don't
631 * delay forever. If we delay, we return a flag that says that we have to
632 * restart the queue scan.
633 */
634 static int
635 aio_free_entry(struct aiocblist *aiocbe)
636 {
637 struct kaioinfo *ki;
638 struct aioliojob *lj;
639 struct proc *p;
640
641 p = aiocbe->userproc;
642 MPASS(curproc == p);
643 ki = p->p_aioinfo;
644 MPASS(ki != NULL);
645
646 AIO_LOCK_ASSERT(ki, MA_OWNED);
647 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
648
649 atomic_subtract_int(&num_queue_count, 1);
650
651 ki->kaio_count--;
652 MPASS(ki->kaio_count >= 0);
653
654 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
655 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
656
657 lj = aiocbe->lio;
658 if (lj) {
659 lj->lioj_count--;
660 lj->lioj_finished_count--;
661
662 if (lj->lioj_count == 0) {
663 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
664 /* lio is going away, we need to destroy any knotes */
665 knlist_delete(&lj->klist, curthread, 1);
666 PROC_LOCK(p);
667 sigqueue_take(&lj->lioj_ksi);
668 PROC_UNLOCK(p);
669 uma_zfree(aiolio_zone, lj);
670 }
671 }
672
673 /* aiocbe is going away, we need to destroy any knotes */
674 knlist_delete(&aiocbe->klist, curthread, 1);
675 PROC_LOCK(p);
676 sigqueue_take(&aiocbe->ksi);
677 PROC_UNLOCK(p);
678
679 MPASS(aiocbe->bp == NULL);
680 aiocbe->jobstate = JOBST_NULL;
681 AIO_UNLOCK(ki);
682
683 /*
684 * The thread argument here is used to find the owning process
685 * and is also passed to fo_close() which may pass it to various
686 * places such as devsw close() routines. Because of that, we
687 * need a thread pointer from the process owning the job that is
688 * persistent and won't disappear out from under us or move to
689 * another process.
690 *
691 * Currently, all the callers of this function call it to remove
692 * an aiocblist from the current process' job list either via a
693 * syscall or due to the current process calling exit() or
694 * execve(). Thus, we know that p == curproc. We also know that
695 * curthread can't exit since we are curthread.
696 *
697 * Therefore, we use curthread as the thread to pass to
698 * knlist_delete(). This does mean that it is possible for the
699 * thread pointer at close time to differ from the thread pointer
700 * at open time, but this is already true of file descriptors in
701 * a multithreaded process.
702 */
703 fdrop(aiocbe->fd_file, curthread);
704 crfree(aiocbe->cred);
705 uma_zfree(aiocb_zone, aiocbe);
706 AIO_LOCK(ki);
707
708 return (0);
709 }
710
711 static void
712 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
713 {
714 aio_proc_rundown(arg, p);
715 }
716
717 /*
718 * Rundown the jobs for a given process.
719 */
720 static void
721 aio_proc_rundown(void *arg, struct proc *p)
722 {
723 struct kaioinfo *ki;
724 struct aioliojob *lj;
725 struct aiocblist *cbe, *cbn;
726 struct file *fp;
727 struct socket *so;
728 int remove;
729
730 KASSERT(curthread->td_proc == p,
731 ("%s: called on non-curproc", __func__));
732 ki = p->p_aioinfo;
733 if (ki == NULL)
734 return;
735
736 AIO_LOCK(ki);
737 ki->kaio_flags |= KAIO_RUNDOWN;
738
739 restart:
740
741 /*
742 * Try to cancel all pending requests. This code simulates
743 * aio_cancel on all pending I/O requests.
744 */
745 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
746 remove = 0;
747 mtx_lock(&aio_job_mtx);
748 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
749 TAILQ_REMOVE(&aio_jobs, cbe, list);
750 remove = 1;
751 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
752 fp = cbe->fd_file;
753 MPASS(fp->f_type == DTYPE_SOCKET);
754 so = fp->f_data;
755 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
756 remove = 1;
757 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
758 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
759 remove = 1;
760 }
761 mtx_unlock(&aio_job_mtx);
762
763 if (remove) {
764 cbe->jobstate = JOBST_JOBFINISHED;
765 cbe->uaiocb._aiocb_private.status = -1;
766 cbe->uaiocb._aiocb_private.error = ECANCELED;
767 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
768 aio_bio_done_notify(p, cbe, DONE_QUEUE);
769 }
770 }
771
772 /* Wait for all running I/O to be finished */
773 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
774 TAILQ_FIRST(&ki->kaio_jobqueue)) {
775 ki->kaio_flags |= KAIO_WAKEUP;
776 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
777 goto restart;
778 }
779
780 /* Free all completed I/O requests. */
781 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
782 aio_free_entry(cbe);
783
784 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
785 if (lj->lioj_count == 0) {
786 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
787 knlist_delete(&lj->klist, curthread, 1);
788 PROC_LOCK(p);
789 sigqueue_take(&lj->lioj_ksi);
790 PROC_UNLOCK(p);
791 uma_zfree(aiolio_zone, lj);
792 } else {
793 panic("LIO job not cleaned up: C:%d, FC:%d\n",
794 lj->lioj_count, lj->lioj_finished_count);
795 }
796 }
797 AIO_UNLOCK(ki);
798 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
799 mtx_destroy(&ki->kaio_mtx);
800 uma_zfree(kaio_zone, ki);
801 p->p_aioinfo = NULL;
802 }
803
804 /*
805 * Select a job to run (called by an AIO daemon).
806 */
807 static struct aiocblist *
808 aio_selectjob(struct aiothreadlist *aiop)
809 {
810 struct aiocblist *aiocbe;
811 struct kaioinfo *ki;
812 struct proc *userp;
813
814 mtx_assert(&aio_job_mtx, MA_OWNED);
815 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
816 userp = aiocbe->userproc;
817 ki = userp->p_aioinfo;
818
819 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
820 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
821 /* Account for currently active jobs. */
822 ki->kaio_active_count++;
823 aiocbe->jobstate = JOBST_JOBRUNNING;
824 break;
825 }
826 }
827 return (aiocbe);
828 }
829
830 /*
831 * Move all data to a permanent storage device, this code
832 * simulates fsync syscall.
833 */
834 static int
835 aio_fsync_vnode(struct thread *td, struct vnode *vp)
836 {
837 struct mount *mp;
838 int vfslocked;
839 int error;
840
841 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
842 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
843 goto drop;
844 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
845 if (vp->v_object != NULL) {
846 VM_OBJECT_LOCK(vp->v_object);
847 vm_object_page_clean(vp->v_object, 0, 0, 0);
848 VM_OBJECT_UNLOCK(vp->v_object);
849 }
850 error = VOP_FSYNC(vp, MNT_WAIT, td);
851
852 VOP_UNLOCK(vp, 0);
853 vn_finished_write(mp);
854 drop:
855 VFS_UNLOCK_GIANT(vfslocked);
856 return (error);
857 }
858
859 /*
860 * The AIO processing activity. This is the code that does the I/O request for
861 * the non-physio version of the operations. The normal vn operations are used,
862 * and this code should work in all instances for every type of file, including
863 * pipes, sockets, fifos, and regular files.
864 *
865 * XXX I don't think it works well for socket, pipe, and fifo.
866 */
867 static void
868 aio_process(struct aiocblist *aiocbe)
869 {
870 struct ucred *td_savedcred;
871 struct thread *td;
872 struct aiocb *cb;
873 struct file *fp;
874 struct socket *so;
875 struct uio auio;
876 struct iovec aiov;
877 int cnt;
878 int error;
879 int oublock_st, oublock_end;
880 int inblock_st, inblock_end;
881
882 td = curthread;
883 td_savedcred = td->td_ucred;
884 td->td_ucred = aiocbe->cred;
885 cb = &aiocbe->uaiocb;
886 fp = aiocbe->fd_file;
887
888 if (cb->aio_lio_opcode == LIO_SYNC) {
889 error = 0;
890 cnt = 0;
891 if (fp->f_vnode != NULL)
892 error = aio_fsync_vnode(td, fp->f_vnode);
893 cb->_aiocb_private.error = error;
894 cb->_aiocb_private.status = 0;
895 td->td_ucred = td_savedcred;
896 return;
897 }
898
899 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
900 aiov.iov_len = cb->aio_nbytes;
901
902 auio.uio_iov = &aiov;
903 auio.uio_iovcnt = 1;
904 auio.uio_offset = cb->aio_offset;
905 auio.uio_resid = cb->aio_nbytes;
906 cnt = cb->aio_nbytes;
907 auio.uio_segflg = UIO_USERSPACE;
908 auio.uio_td = td;
909
910 inblock_st = td->td_ru.ru_inblock;
911 oublock_st = td->td_ru.ru_oublock;
912 /*
913 * aio_aqueue() acquires a reference to the file that is
914 * released in aio_free_entry().
915 */
916 if (cb->aio_lio_opcode == LIO_READ) {
917 auio.uio_rw = UIO_READ;
918 if (auio.uio_resid == 0)
919 error = 0;
920 else
921 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
922 } else {
923 if (fp->f_type == DTYPE_VNODE)
924 bwillwrite();
925 auio.uio_rw = UIO_WRITE;
926 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
927 }
928 inblock_end = td->td_ru.ru_inblock;
929 oublock_end = td->td_ru.ru_oublock;
930
931 aiocbe->inputcharge = inblock_end - inblock_st;
932 aiocbe->outputcharge = oublock_end - oublock_st;
933
934 if ((error) && (auio.uio_resid != cnt)) {
935 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
936 error = 0;
937 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
938 int sigpipe = 1;
939 if (fp->f_type == DTYPE_SOCKET) {
940 so = fp->f_data;
941 if (so->so_options & SO_NOSIGPIPE)
942 sigpipe = 0;
943 }
944 if (sigpipe) {
945 PROC_LOCK(aiocbe->userproc);
946 kern_psignal(aiocbe->userproc, SIGPIPE);
947 PROC_UNLOCK(aiocbe->userproc);
948 }
949 }
950 }
951
952 cnt -= auio.uio_resid;
953 cb->_aiocb_private.error = error;
954 cb->_aiocb_private.status = cnt;
955 td->td_ucred = td_savedcred;
956 }
957
958 static void
959 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
960 {
961 struct aioliojob *lj;
962 struct kaioinfo *ki;
963 struct aiocblist *scb, *scbn;
964 int lj_done;
965
966 ki = userp->p_aioinfo;
967 AIO_LOCK_ASSERT(ki, MA_OWNED);
968 lj = aiocbe->lio;
969 lj_done = 0;
970 if (lj) {
971 lj->lioj_finished_count++;
972 if (lj->lioj_count == lj->lioj_finished_count)
973 lj_done = 1;
974 }
975 if (type == DONE_QUEUE) {
976 aiocbe->jobflags |= AIOCBLIST_DONE;
977 } else {
978 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
979 }
980 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
981 aiocbe->jobstate = JOBST_JOBFINISHED;
982
983 if (ki->kaio_flags & KAIO_RUNDOWN)
984 goto notification_done;
985
986 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
987 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
988 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
989
990 KNOTE_LOCKED(&aiocbe->klist, 1);
991
992 if (lj_done) {
993 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
994 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
995 KNOTE_LOCKED(&lj->klist, 1);
996 }
997 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
998 == LIOJ_SIGNAL
999 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
1000 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
1001 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
1002 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1003 }
1004 }
1005
1006 notification_done:
1007 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1008 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1009 if (aiocbe->fd_file == scb->fd_file &&
1010 aiocbe->seqno < scb->seqno) {
1011 if (--scb->pending == 0) {
1012 mtx_lock(&aio_job_mtx);
1013 scb->jobstate = JOBST_JOBQGLOBAL;
1014 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1015 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1016 aio_kick_nowait(userp);
1017 mtx_unlock(&aio_job_mtx);
1018 }
1019 }
1020 }
1021 }
1022 if (ki->kaio_flags & KAIO_WAKEUP) {
1023 ki->kaio_flags &= ~KAIO_WAKEUP;
1024 wakeup(&userp->p_aioinfo);
1025 }
1026 }
1027
1028 /*
1029 * The AIO daemon, most of the actual work is done in aio_process,
1030 * but the setup (and address space mgmt) is done in this routine.
1031 */
1032 static void
1033 aio_daemon(void *_id)
1034 {
1035 struct aiocblist *aiocbe;
1036 struct aiothreadlist *aiop;
1037 struct kaioinfo *ki;
1038 struct proc *curcp, *mycp, *userp;
1039 struct vmspace *myvm, *tmpvm;
1040 struct thread *td = curthread;
1041 int id = (intptr_t)_id;
1042
1043 /*
1044 * Local copies of curproc (cp) and vmspace (myvm)
1045 */
1046 mycp = td->td_proc;
1047 myvm = mycp->p_vmspace;
1048
1049 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1050
1051 /*
1052 * Allocate and ready the aio control info. There is one aiop structure
1053 * per daemon.
1054 */
1055 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1056 aiop->aiothread = td;
1057 aiop->aiothreadflags = 0;
1058
1059 /* The daemon resides in its own pgrp. */
1060 sys_setsid(td, NULL);
1061
1062 /*
1063 * Wakeup parent process. (Parent sleeps to keep from blasting away
1064 * and creating too many daemons.)
1065 */
1066 sema_post(&aio_newproc_sem);
1067
1068 mtx_lock(&aio_job_mtx);
1069 for (;;) {
1070 /*
1071 * curcp is the current daemon process context.
1072 * userp is the current user process context.
1073 */
1074 curcp = mycp;
1075
1076 /*
1077 * Take daemon off of free queue
1078 */
1079 if (aiop->aiothreadflags & AIOP_FREE) {
1080 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1081 aiop->aiothreadflags &= ~AIOP_FREE;
1082 }
1083
1084 /*
1085 * Check for jobs.
1086 */
1087 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1088 mtx_unlock(&aio_job_mtx);
1089 userp = aiocbe->userproc;
1090
1091 /*
1092 * Connect to process address space for user program.
1093 */
1094 if (userp != curcp) {
1095 /*
1096 * Save the current address space that we are
1097 * connected to.
1098 */
1099 tmpvm = mycp->p_vmspace;
1100
1101 /*
1102 * Point to the new user address space, and
1103 * refer to it.
1104 */
1105 mycp->p_vmspace = userp->p_vmspace;
1106 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1107
1108 /* Activate the new mapping. */
1109 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1110
1111 /*
1112 * If the old address space wasn't the daemons
1113 * own address space, then we need to remove the
1114 * daemon's reference from the other process
1115 * that it was acting on behalf of.
1116 */
1117 if (tmpvm != myvm) {
1118 vmspace_free(tmpvm);
1119 }
1120 curcp = userp;
1121 }
1122
1123 ki = userp->p_aioinfo;
1124
1125 /* Do the I/O function. */
1126 aio_process(aiocbe);
1127
1128 mtx_lock(&aio_job_mtx);
1129 /* Decrement the active job count. */
1130 ki->kaio_active_count--;
1131 mtx_unlock(&aio_job_mtx);
1132
1133 AIO_LOCK(ki);
1134 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1135 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1136 AIO_UNLOCK(ki);
1137
1138 mtx_lock(&aio_job_mtx);
1139 }
1140
1141 /*
1142 * Disconnect from user address space.
1143 */
1144 if (curcp != mycp) {
1145
1146 mtx_unlock(&aio_job_mtx);
1147
1148 /* Get the user address space to disconnect from. */
1149 tmpvm = mycp->p_vmspace;
1150
1151 /* Get original address space for daemon. */
1152 mycp->p_vmspace = myvm;
1153
1154 /* Activate the daemon's address space. */
1155 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1156 #ifdef DIAGNOSTIC
1157 if (tmpvm == myvm) {
1158 printf("AIOD: vmspace problem -- %d\n",
1159 mycp->p_pid);
1160 }
1161 #endif
1162 /* Remove our vmspace reference. */
1163 vmspace_free(tmpvm);
1164
1165 curcp = mycp;
1166
1167 mtx_lock(&aio_job_mtx);
1168 /*
1169 * We have to restart to avoid race, we only sleep if
1170 * no job can be selected, that should be
1171 * curcp == mycp.
1172 */
1173 continue;
1174 }
1175
1176 mtx_assert(&aio_job_mtx, MA_OWNED);
1177
1178 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1179 aiop->aiothreadflags |= AIOP_FREE;
1180
1181 /*
1182 * If daemon is inactive for a long time, allow it to exit,
1183 * thereby freeing resources.
1184 */
1185 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1186 aiod_lifetime)) {
1187 if (TAILQ_EMPTY(&aio_jobs)) {
1188 if ((aiop->aiothreadflags & AIOP_FREE) &&
1189 (num_aio_procs > target_aio_procs)) {
1190 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1191 num_aio_procs--;
1192 mtx_unlock(&aio_job_mtx);
1193 uma_zfree(aiop_zone, aiop);
1194 free_unr(aiod_unr, id);
1195 #ifdef DIAGNOSTIC
1196 if (mycp->p_vmspace->vm_refcnt <= 1) {
1197 printf("AIOD: bad vm refcnt for"
1198 " exiting daemon: %d\n",
1199 mycp->p_vmspace->vm_refcnt);
1200 }
1201 #endif
1202 kproc_exit(0);
1203 }
1204 }
1205 }
1206 }
1207 mtx_unlock(&aio_job_mtx);
1208 panic("shouldn't be here\n");
1209 }
1210
1211 /*
1212 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1213 * AIO daemon modifies its environment itself.
1214 */
1215 static int
1216 aio_newproc(int *start)
1217 {
1218 int error;
1219 struct proc *p;
1220 int id;
1221
1222 id = alloc_unr(aiod_unr);
1223 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1224 RFNOWAIT, 0, "aiod%d", id);
1225 if (error == 0) {
1226 /*
1227 * Wait until daemon is started.
1228 */
1229 sema_wait(&aio_newproc_sem);
1230 mtx_lock(&aio_job_mtx);
1231 num_aio_procs++;
1232 if (start != NULL)
1233 (*start)--;
1234 mtx_unlock(&aio_job_mtx);
1235 } else {
1236 free_unr(aiod_unr, id);
1237 }
1238 return (error);
1239 }
1240
1241 /*
1242 * Try the high-performance, low-overhead physio method for eligible
1243 * VCHR devices. This method doesn't use an aio helper thread, and
1244 * thus has very low overhead.
1245 *
1246 * Assumes that the caller, aio_aqueue(), has incremented the file
1247 * structure's reference count, preventing its deallocation for the
1248 * duration of this call.
1249 */
1250 static int
1251 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1252 {
1253 struct aiocb *cb;
1254 struct file *fp;
1255 struct buf *bp;
1256 struct vnode *vp;
1257 struct kaioinfo *ki;
1258 struct aioliojob *lj;
1259 int error;
1260
1261 cb = &aiocbe->uaiocb;
1262 fp = aiocbe->fd_file;
1263
1264 if (fp->f_type != DTYPE_VNODE)
1265 return (-1);
1266
1267 vp = fp->f_vnode;
1268
1269 /*
1270 * If its not a disk, we don't want to return a positive error.
1271 * It causes the aio code to not fall through to try the thread
1272 * way when you're talking to a regular file.
1273 */
1274 if (!vn_isdisk(vp, &error)) {
1275 if (error == ENOTBLK)
1276 return (-1);
1277 else
1278 return (error);
1279 }
1280
1281 if (vp->v_bufobj.bo_bsize == 0)
1282 return (-1);
1283
1284 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1285 return (-1);
1286
1287 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1288 return (-1);
1289
1290 if (cb->aio_nbytes >
1291 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1292 return (-1);
1293
1294 ki = p->p_aioinfo;
1295 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1296 return (-1);
1297
1298 /* Create and build a buffer header for a transfer. */
1299 bp = (struct buf *)getpbuf(NULL);
1300 BUF_KERNPROC(bp);
1301
1302 AIO_LOCK(ki);
1303 ki->kaio_count++;
1304 ki->kaio_buffer_count++;
1305 lj = aiocbe->lio;
1306 if (lj)
1307 lj->lioj_count++;
1308 AIO_UNLOCK(ki);
1309
1310 /*
1311 * Get a copy of the kva from the physical buffer.
1312 */
1313 error = 0;
1314
1315 bp->b_bcount = cb->aio_nbytes;
1316 bp->b_bufsize = cb->aio_nbytes;
1317 bp->b_iodone = aio_physwakeup;
1318 bp->b_saveaddr = bp->b_data;
1319 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1320 bp->b_offset = cb->aio_offset;
1321 bp->b_iooffset = cb->aio_offset;
1322 bp->b_blkno = btodb(cb->aio_offset);
1323 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1324
1325 /*
1326 * Bring buffer into kernel space.
1327 */
1328 if (vmapbuf(bp) < 0) {
1329 error = EFAULT;
1330 goto doerror;
1331 }
1332
1333 AIO_LOCK(ki);
1334 aiocbe->bp = bp;
1335 bp->b_caller1 = (void *)aiocbe;
1336 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1337 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1338 aiocbe->jobstate = JOBST_JOBQBUF;
1339 cb->_aiocb_private.status = cb->aio_nbytes;
1340 AIO_UNLOCK(ki);
1341
1342 atomic_add_int(&num_queue_count, 1);
1343 atomic_add_int(&num_buf_aio, 1);
1344
1345 bp->b_error = 0;
1346
1347 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1348
1349 /* Perform transfer. */
1350 dev_strategy(vp->v_rdev, bp);
1351 return (0);
1352
1353 doerror:
1354 AIO_LOCK(ki);
1355 ki->kaio_count--;
1356 ki->kaio_buffer_count--;
1357 if (lj)
1358 lj->lioj_count--;
1359 aiocbe->bp = NULL;
1360 AIO_UNLOCK(ki);
1361 relpbuf(bp, NULL);
1362 return (error);
1363 }
1364
1365 /*
1366 * Wake up aio requests that may be serviceable now.
1367 */
1368 static void
1369 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1370 {
1371 struct aiocblist *cb, *cbn;
1372 int opcode;
1373
1374 SOCKBUF_LOCK_ASSERT(sb);
1375 if (sb == &so->so_snd)
1376 opcode = LIO_WRITE;
1377 else
1378 opcode = LIO_READ;
1379
1380 sb->sb_flags &= ~SB_AIO;
1381 mtx_lock(&aio_job_mtx);
1382 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1383 if (opcode == cb->uaiocb.aio_lio_opcode) {
1384 if (cb->jobstate != JOBST_JOBQSOCK)
1385 panic("invalid queue value");
1386 /* XXX
1387 * We don't have actual sockets backend yet,
1388 * so we simply move the requests to the generic
1389 * file I/O backend.
1390 */
1391 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1392 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1393 aio_kick_nowait(cb->userproc);
1394 }
1395 }
1396 mtx_unlock(&aio_job_mtx);
1397 }
1398
1399 static int
1400 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1401 {
1402
1403 /*
1404 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1405 * supported by AIO with the old sigevent structure.
1406 */
1407 nsig->sigev_notify = osig->sigev_notify;
1408 switch (nsig->sigev_notify) {
1409 case SIGEV_NONE:
1410 break;
1411 case SIGEV_SIGNAL:
1412 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1413 break;
1414 case SIGEV_KEVENT:
1415 nsig->sigev_notify_kqueue =
1416 osig->__sigev_u.__sigev_notify_kqueue;
1417 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1418 break;
1419 default:
1420 return (EINVAL);
1421 }
1422 return (0);
1423 }
1424
1425 static int
1426 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1427 {
1428 struct oaiocb *ojob;
1429 int error;
1430
1431 bzero(kjob, sizeof(struct aiocb));
1432 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1433 if (error)
1434 return (error);
1435 ojob = (struct oaiocb *)kjob;
1436 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1437 }
1438
1439 static int
1440 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1441 {
1442
1443 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1444 }
1445
1446 static long
1447 aiocb_fetch_status(struct aiocb *ujob)
1448 {
1449
1450 return (fuword(&ujob->_aiocb_private.status));
1451 }
1452
1453 static long
1454 aiocb_fetch_error(struct aiocb *ujob)
1455 {
1456
1457 return (fuword(&ujob->_aiocb_private.error));
1458 }
1459
1460 static int
1461 aiocb_store_status(struct aiocb *ujob, long status)
1462 {
1463
1464 return (suword(&ujob->_aiocb_private.status, status));
1465 }
1466
1467 static int
1468 aiocb_store_error(struct aiocb *ujob, long error)
1469 {
1470
1471 return (suword(&ujob->_aiocb_private.error, error));
1472 }
1473
1474 static int
1475 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1476 {
1477
1478 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1479 }
1480
1481 static int
1482 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1483 {
1484
1485 return (suword(ujobp, (long)ujob));
1486 }
1487
1488 static struct aiocb_ops aiocb_ops = {
1489 .copyin = aiocb_copyin,
1490 .fetch_status = aiocb_fetch_status,
1491 .fetch_error = aiocb_fetch_error,
1492 .store_status = aiocb_store_status,
1493 .store_error = aiocb_store_error,
1494 .store_kernelinfo = aiocb_store_kernelinfo,
1495 .store_aiocb = aiocb_store_aiocb,
1496 };
1497
1498 static struct aiocb_ops aiocb_ops_osigevent = {
1499 .copyin = aiocb_copyin_old_sigevent,
1500 .fetch_status = aiocb_fetch_status,
1501 .fetch_error = aiocb_fetch_error,
1502 .store_status = aiocb_store_status,
1503 .store_error = aiocb_store_error,
1504 .store_kernelinfo = aiocb_store_kernelinfo,
1505 .store_aiocb = aiocb_store_aiocb,
1506 };
1507
1508 /*
1509 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1510 * technique is done in this code.
1511 */
1512 int
1513 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1514 int type, struct aiocb_ops *ops)
1515 {
1516 struct proc *p = td->td_proc;
1517 struct file *fp;
1518 struct socket *so;
1519 struct aiocblist *aiocbe, *cb;
1520 struct kaioinfo *ki;
1521 struct kevent kev;
1522 struct sockbuf *sb;
1523 int opcode;
1524 int error;
1525 int fd, kqfd;
1526 int jid;
1527
1528 if (p->p_aioinfo == NULL)
1529 aio_init_aioinfo(p);
1530
1531 ki = p->p_aioinfo;
1532
1533 ops->store_status(job, -1);
1534 ops->store_error(job, 0);
1535 ops->store_kernelinfo(job, -1);
1536
1537 if (num_queue_count >= max_queue_count ||
1538 ki->kaio_count >= ki->kaio_qallowed_count) {
1539 ops->store_error(job, EAGAIN);
1540 return (EAGAIN);
1541 }
1542
1543 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1544 aiocbe->inputcharge = 0;
1545 aiocbe->outputcharge = 0;
1546 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1547
1548 error = ops->copyin(job, &aiocbe->uaiocb);
1549 if (error) {
1550 ops->store_error(job, error);
1551 uma_zfree(aiocb_zone, aiocbe);
1552 return (error);
1553 }
1554
1555 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1556 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1557 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1558 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1559 ops->store_error(job, EINVAL);
1560 uma_zfree(aiocb_zone, aiocbe);
1561 return (EINVAL);
1562 }
1563
1564 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1565 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1566 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1567 uma_zfree(aiocb_zone, aiocbe);
1568 return (EINVAL);
1569 }
1570
1571 ksiginfo_init(&aiocbe->ksi);
1572
1573 /* Save userspace address of the job info. */
1574 aiocbe->uuaiocb = job;
1575
1576 /* Get the opcode. */
1577 if (type != LIO_NOP)
1578 aiocbe->uaiocb.aio_lio_opcode = type;
1579 opcode = aiocbe->uaiocb.aio_lio_opcode;
1580
1581 /*
1582 * Validate the opcode and fetch the file object for the specified
1583 * file descriptor.
1584 *
1585 * XXXRW: Moved the opcode validation up here so that we don't
1586 * retrieve a file descriptor without knowing what the capabiltity
1587 * should be.
1588 */
1589 fd = aiocbe->uaiocb.aio_fildes;
1590 switch (opcode) {
1591 case LIO_WRITE:
1592 error = fget_write(td, fd, CAP_WRITE | CAP_SEEK, &fp);
1593 break;
1594 case LIO_READ:
1595 error = fget_read(td, fd, CAP_READ | CAP_SEEK, &fp);
1596 break;
1597 case LIO_SYNC:
1598 error = fget(td, fd, CAP_FSYNC, &fp);
1599 break;
1600 case LIO_NOP:
1601 error = fget(td, fd, 0, &fp);
1602 break;
1603 default:
1604 error = EINVAL;
1605 }
1606 if (error) {
1607 uma_zfree(aiocb_zone, aiocbe);
1608 ops->store_error(job, error);
1609 return (error);
1610 }
1611
1612 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1613 error = EINVAL;
1614 goto aqueue_fail;
1615 }
1616
1617 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1618 error = EINVAL;
1619 goto aqueue_fail;
1620 }
1621
1622 aiocbe->fd_file = fp;
1623
1624 mtx_lock(&aio_job_mtx);
1625 jid = jobrefid++;
1626 aiocbe->seqno = jobseqno++;
1627 mtx_unlock(&aio_job_mtx);
1628 error = ops->store_kernelinfo(job, jid);
1629 if (error) {
1630 error = EINVAL;
1631 goto aqueue_fail;
1632 }
1633 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1634
1635 if (opcode == LIO_NOP) {
1636 fdrop(fp, td);
1637 uma_zfree(aiocb_zone, aiocbe);
1638 return (0);
1639 }
1640
1641 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1642 goto no_kqueue;
1643 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1644 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1645 kev.filter = EVFILT_AIO;
1646 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1647 kev.data = (intptr_t)aiocbe;
1648 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1649 error = kqfd_register(kqfd, &kev, td, 1);
1650 aqueue_fail:
1651 if (error) {
1652 fdrop(fp, td);
1653 uma_zfree(aiocb_zone, aiocbe);
1654 ops->store_error(job, error);
1655 goto done;
1656 }
1657 no_kqueue:
1658
1659 ops->store_error(job, EINPROGRESS);
1660 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1661 aiocbe->userproc = p;
1662 aiocbe->cred = crhold(td->td_ucred);
1663 aiocbe->jobflags = 0;
1664 aiocbe->lio = lj;
1665
1666 if (opcode == LIO_SYNC)
1667 goto queueit;
1668
1669 if (fp->f_type == DTYPE_SOCKET) {
1670 /*
1671 * Alternate queueing for socket ops: Reach down into the
1672 * descriptor to get the socket data. Then check to see if the
1673 * socket is ready to be read or written (based on the requested
1674 * operation).
1675 *
1676 * If it is not ready for io, then queue the aiocbe on the
1677 * socket, and set the flags so we get a call when sbnotify()
1678 * happens.
1679 *
1680 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1681 * and unlock the snd sockbuf for no reason.
1682 */
1683 so = fp->f_data;
1684 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1685 SOCKBUF_LOCK(sb);
1686 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1687 LIO_WRITE) && (!sowriteable(so)))) {
1688 sb->sb_flags |= SB_AIO;
1689
1690 mtx_lock(&aio_job_mtx);
1691 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1692 mtx_unlock(&aio_job_mtx);
1693
1694 AIO_LOCK(ki);
1695 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1696 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1697 aiocbe->jobstate = JOBST_JOBQSOCK;
1698 ki->kaio_count++;
1699 if (lj)
1700 lj->lioj_count++;
1701 AIO_UNLOCK(ki);
1702 SOCKBUF_UNLOCK(sb);
1703 atomic_add_int(&num_queue_count, 1);
1704 error = 0;
1705 goto done;
1706 }
1707 SOCKBUF_UNLOCK(sb);
1708 }
1709
1710 if ((error = aio_qphysio(p, aiocbe)) == 0)
1711 goto done;
1712 #if 0
1713 if (error > 0) {
1714 aiocbe->uaiocb._aiocb_private.error = error;
1715 ops->store_error(job, error);
1716 goto done;
1717 }
1718 #endif
1719 queueit:
1720 /* No buffer for daemon I/O. */
1721 aiocbe->bp = NULL;
1722 atomic_add_int(&num_queue_count, 1);
1723
1724 AIO_LOCK(ki);
1725 ki->kaio_count++;
1726 if (lj)
1727 lj->lioj_count++;
1728 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1729 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1730 if (opcode == LIO_SYNC) {
1731 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1732 if (cb->fd_file == aiocbe->fd_file &&
1733 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1734 cb->seqno < aiocbe->seqno) {
1735 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1736 aiocbe->pending++;
1737 }
1738 }
1739 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1740 if (cb->fd_file == aiocbe->fd_file &&
1741 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1742 cb->seqno < aiocbe->seqno) {
1743 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1744 aiocbe->pending++;
1745 }
1746 }
1747 if (aiocbe->pending != 0) {
1748 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1749 aiocbe->jobstate = JOBST_JOBQSYNC;
1750 AIO_UNLOCK(ki);
1751 goto done;
1752 }
1753 }
1754 mtx_lock(&aio_job_mtx);
1755 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1756 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1757 aio_kick_nowait(p);
1758 mtx_unlock(&aio_job_mtx);
1759 AIO_UNLOCK(ki);
1760 error = 0;
1761 done:
1762 return (error);
1763 }
1764
1765 static void
1766 aio_kick_nowait(struct proc *userp)
1767 {
1768 struct kaioinfo *ki = userp->p_aioinfo;
1769 struct aiothreadlist *aiop;
1770
1771 mtx_assert(&aio_job_mtx, MA_OWNED);
1772 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1773 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1774 aiop->aiothreadflags &= ~AIOP_FREE;
1775 wakeup(aiop->aiothread);
1776 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1777 ((ki->kaio_active_count + num_aio_resv_start) <
1778 ki->kaio_maxactive_count)) {
1779 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1780 }
1781 }
1782
1783 static int
1784 aio_kick(struct proc *userp)
1785 {
1786 struct kaioinfo *ki = userp->p_aioinfo;
1787 struct aiothreadlist *aiop;
1788 int error, ret = 0;
1789
1790 mtx_assert(&aio_job_mtx, MA_OWNED);
1791 retryproc:
1792 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1793 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1794 aiop->aiothreadflags &= ~AIOP_FREE;
1795 wakeup(aiop->aiothread);
1796 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1797 ((ki->kaio_active_count + num_aio_resv_start) <
1798 ki->kaio_maxactive_count)) {
1799 num_aio_resv_start++;
1800 mtx_unlock(&aio_job_mtx);
1801 error = aio_newproc(&num_aio_resv_start);
1802 mtx_lock(&aio_job_mtx);
1803 if (error) {
1804 num_aio_resv_start--;
1805 goto retryproc;
1806 }
1807 } else {
1808 ret = -1;
1809 }
1810 return (ret);
1811 }
1812
1813 static void
1814 aio_kick_helper(void *context, int pending)
1815 {
1816 struct proc *userp = context;
1817
1818 mtx_lock(&aio_job_mtx);
1819 while (--pending >= 0) {
1820 if (aio_kick(userp))
1821 break;
1822 }
1823 mtx_unlock(&aio_job_mtx);
1824 }
1825
1826 /*
1827 * Support the aio_return system call, as a side-effect, kernel resources are
1828 * released.
1829 */
1830 static int
1831 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1832 {
1833 struct proc *p = td->td_proc;
1834 struct aiocblist *cb;
1835 struct kaioinfo *ki;
1836 int status, error;
1837
1838 ki = p->p_aioinfo;
1839 if (ki == NULL)
1840 return (EINVAL);
1841 AIO_LOCK(ki);
1842 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1843 if (cb->uuaiocb == uaiocb)
1844 break;
1845 }
1846 if (cb != NULL) {
1847 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1848 status = cb->uaiocb._aiocb_private.status;
1849 error = cb->uaiocb._aiocb_private.error;
1850 td->td_retval[0] = status;
1851 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1852 td->td_ru.ru_oublock += cb->outputcharge;
1853 cb->outputcharge = 0;
1854 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1855 td->td_ru.ru_inblock += cb->inputcharge;
1856 cb->inputcharge = 0;
1857 }
1858 aio_free_entry(cb);
1859 AIO_UNLOCK(ki);
1860 ops->store_error(uaiocb, error);
1861 ops->store_status(uaiocb, status);
1862 } else {
1863 error = EINVAL;
1864 AIO_UNLOCK(ki);
1865 }
1866 return (error);
1867 }
1868
1869 int
1870 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1871 {
1872
1873 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1874 }
1875
1876 /*
1877 * Allow a process to wakeup when any of the I/O requests are completed.
1878 */
1879 static int
1880 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1881 struct timespec *ts)
1882 {
1883 struct proc *p = td->td_proc;
1884 struct timeval atv;
1885 struct kaioinfo *ki;
1886 struct aiocblist *cb, *cbfirst;
1887 int error, i, timo;
1888
1889 timo = 0;
1890 if (ts) {
1891 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1892 return (EINVAL);
1893
1894 TIMESPEC_TO_TIMEVAL(&atv, ts);
1895 if (itimerfix(&atv))
1896 return (EINVAL);
1897 timo = tvtohz(&atv);
1898 }
1899
1900 ki = p->p_aioinfo;
1901 if (ki == NULL)
1902 return (EAGAIN);
1903
1904 if (njoblist == 0)
1905 return (0);
1906
1907 AIO_LOCK(ki);
1908 for (;;) {
1909 cbfirst = NULL;
1910 error = 0;
1911 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1912 for (i = 0; i < njoblist; i++) {
1913 if (cb->uuaiocb == ujoblist[i]) {
1914 if (cbfirst == NULL)
1915 cbfirst = cb;
1916 if (cb->jobstate == JOBST_JOBFINISHED)
1917 goto RETURN;
1918 }
1919 }
1920 }
1921 /* All tasks were finished. */
1922 if (cbfirst == NULL)
1923 break;
1924
1925 ki->kaio_flags |= KAIO_WAKEUP;
1926 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1927 "aiospn", timo);
1928 if (error == ERESTART)
1929 error = EINTR;
1930 if (error)
1931 break;
1932 }
1933 RETURN:
1934 AIO_UNLOCK(ki);
1935 return (error);
1936 }
1937
1938 int
1939 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1940 {
1941 struct timespec ts, *tsp;
1942 struct aiocb **ujoblist;
1943 int error;
1944
1945 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1946 return (EINVAL);
1947
1948 if (uap->timeout) {
1949 /* Get timespec struct. */
1950 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1951 return (error);
1952 tsp = &ts;
1953 } else
1954 tsp = NULL;
1955
1956 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1957 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1958 if (error == 0)
1959 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1960 uma_zfree(aiol_zone, ujoblist);
1961 return (error);
1962 }
1963
1964 /*
1965 * aio_cancel cancels any non-physio aio operations not currently in
1966 * progress.
1967 */
1968 int
1969 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1970 {
1971 struct proc *p = td->td_proc;
1972 struct kaioinfo *ki;
1973 struct aiocblist *cbe, *cbn;
1974 struct file *fp;
1975 struct socket *so;
1976 int error;
1977 int remove;
1978 int cancelled = 0;
1979 int notcancelled = 0;
1980 struct vnode *vp;
1981
1982 /* Lookup file object. */
1983 error = fget(td, uap->fd, 0, &fp);
1984 if (error)
1985 return (error);
1986
1987 ki = p->p_aioinfo;
1988 if (ki == NULL)
1989 goto done;
1990
1991 if (fp->f_type == DTYPE_VNODE) {
1992 vp = fp->f_vnode;
1993 if (vn_isdisk(vp, &error)) {
1994 fdrop(fp, td);
1995 td->td_retval[0] = AIO_NOTCANCELED;
1996 return (0);
1997 }
1998 }
1999
2000 AIO_LOCK(ki);
2001 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
2002 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
2003 ((uap->aiocbp == NULL) ||
2004 (uap->aiocbp == cbe->uuaiocb))) {
2005 remove = 0;
2006
2007 mtx_lock(&aio_job_mtx);
2008 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
2009 TAILQ_REMOVE(&aio_jobs, cbe, list);
2010 remove = 1;
2011 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
2012 MPASS(fp->f_type == DTYPE_SOCKET);
2013 so = fp->f_data;
2014 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2015 remove = 1;
2016 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
2017 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2018 remove = 1;
2019 }
2020 mtx_unlock(&aio_job_mtx);
2021
2022 if (remove) {
2023 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2024 cbe->uaiocb._aiocb_private.status = -1;
2025 cbe->uaiocb._aiocb_private.error = ECANCELED;
2026 aio_bio_done_notify(p, cbe, DONE_QUEUE);
2027 cancelled++;
2028 } else {
2029 notcancelled++;
2030 }
2031 if (uap->aiocbp != NULL)
2032 break;
2033 }
2034 }
2035 AIO_UNLOCK(ki);
2036
2037 done:
2038 fdrop(fp, td);
2039
2040 if (uap->aiocbp != NULL) {
2041 if (cancelled) {
2042 td->td_retval[0] = AIO_CANCELED;
2043 return (0);
2044 }
2045 }
2046
2047 if (notcancelled) {
2048 td->td_retval[0] = AIO_NOTCANCELED;
2049 return (0);
2050 }
2051
2052 if (cancelled) {
2053 td->td_retval[0] = AIO_CANCELED;
2054 return (0);
2055 }
2056
2057 td->td_retval[0] = AIO_ALLDONE;
2058
2059 return (0);
2060 }
2061
2062 /*
2063 * aio_error is implemented in the kernel level for compatibility purposes
2064 * only. For a user mode async implementation, it would be best to do it in
2065 * a userland subroutine.
2066 */
2067 static int
2068 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2069 {
2070 struct proc *p = td->td_proc;
2071 struct aiocblist *cb;
2072 struct kaioinfo *ki;
2073 int status;
2074
2075 ki = p->p_aioinfo;
2076 if (ki == NULL) {
2077 td->td_retval[0] = EINVAL;
2078 return (0);
2079 }
2080
2081 AIO_LOCK(ki);
2082 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2083 if (cb->uuaiocb == aiocbp) {
2084 if (cb->jobstate == JOBST_JOBFINISHED)
2085 td->td_retval[0] =
2086 cb->uaiocb._aiocb_private.error;
2087 else
2088 td->td_retval[0] = EINPROGRESS;
2089 AIO_UNLOCK(ki);
2090 return (0);
2091 }
2092 }
2093 AIO_UNLOCK(ki);
2094
2095 /*
2096 * Hack for failure of aio_aqueue.
2097 */
2098 status = ops->fetch_status(aiocbp);
2099 if (status == -1) {
2100 td->td_retval[0] = ops->fetch_error(aiocbp);
2101 return (0);
2102 }
2103
2104 td->td_retval[0] = EINVAL;
2105 return (0);
2106 }
2107
2108 int
2109 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2110 {
2111
2112 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2113 }
2114
2115 /* syscall - asynchronous read from a file (REALTIME) */
2116 int
2117 sys_oaio_read(struct thread *td, struct oaio_read_args *uap)
2118 {
2119
2120 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2121 &aiocb_ops_osigevent));
2122 }
2123
2124 int
2125 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2126 {
2127
2128 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2129 }
2130
2131 /* syscall - asynchronous write to a file (REALTIME) */
2132 int
2133 sys_oaio_write(struct thread *td, struct oaio_write_args *uap)
2134 {
2135
2136 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2137 &aiocb_ops_osigevent));
2138 }
2139
2140 int
2141 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2142 {
2143
2144 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2145 }
2146
2147 static int
2148 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2149 struct aiocb **acb_list, int nent, struct sigevent *sig,
2150 struct aiocb_ops *ops)
2151 {
2152 struct proc *p = td->td_proc;
2153 struct aiocb *iocb;
2154 struct kaioinfo *ki;
2155 struct aioliojob *lj;
2156 struct kevent kev;
2157 int error;
2158 int nerror;
2159 int i;
2160
2161 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2162 return (EINVAL);
2163
2164 if (nent < 0 || nent > AIO_LISTIO_MAX)
2165 return (EINVAL);
2166
2167 if (p->p_aioinfo == NULL)
2168 aio_init_aioinfo(p);
2169
2170 ki = p->p_aioinfo;
2171
2172 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2173 lj->lioj_flags = 0;
2174 lj->lioj_count = 0;
2175 lj->lioj_finished_count = 0;
2176 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2177 ksiginfo_init(&lj->lioj_ksi);
2178
2179 /*
2180 * Setup signal.
2181 */
2182 if (sig && (mode == LIO_NOWAIT)) {
2183 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2184 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2185 /* Assume only new style KEVENT */
2186 kev.filter = EVFILT_LIO;
2187 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2188 kev.ident = (uintptr_t)uacb_list; /* something unique */
2189 kev.data = (intptr_t)lj;
2190 /* pass user defined sigval data */
2191 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2192 error = kqfd_register(
2193 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2194 if (error) {
2195 uma_zfree(aiolio_zone, lj);
2196 return (error);
2197 }
2198 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2199 ;
2200 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2201 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2202 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2203 uma_zfree(aiolio_zone, lj);
2204 return EINVAL;
2205 }
2206 lj->lioj_flags |= LIOJ_SIGNAL;
2207 } else {
2208 uma_zfree(aiolio_zone, lj);
2209 return EINVAL;
2210 }
2211 }
2212
2213 AIO_LOCK(ki);
2214 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2215 /*
2216 * Add extra aiocb count to avoid the lio to be freed
2217 * by other threads doing aio_waitcomplete or aio_return,
2218 * and prevent event from being sent until we have queued
2219 * all tasks.
2220 */
2221 lj->lioj_count = 1;
2222 AIO_UNLOCK(ki);
2223
2224 /*
2225 * Get pointers to the list of I/O requests.
2226 */
2227 nerror = 0;
2228 for (i = 0; i < nent; i++) {
2229 iocb = acb_list[i];
2230 if (iocb != NULL) {
2231 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2232 if (error != 0)
2233 nerror++;
2234 }
2235 }
2236
2237 error = 0;
2238 AIO_LOCK(ki);
2239 if (mode == LIO_WAIT) {
2240 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2241 ki->kaio_flags |= KAIO_WAKEUP;
2242 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2243 PRIBIO | PCATCH, "aiospn", 0);
2244 if (error == ERESTART)
2245 error = EINTR;
2246 if (error)
2247 break;
2248 }
2249 } else {
2250 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2251 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2252 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2253 KNOTE_LOCKED(&lj->klist, 1);
2254 }
2255 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2256 == LIOJ_SIGNAL
2257 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2258 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2259 aio_sendsig(p, &lj->lioj_signal,
2260 &lj->lioj_ksi);
2261 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2262 }
2263 }
2264 }
2265 lj->lioj_count--;
2266 if (lj->lioj_count == 0) {
2267 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2268 knlist_delete(&lj->klist, curthread, 1);
2269 PROC_LOCK(p);
2270 sigqueue_take(&lj->lioj_ksi);
2271 PROC_UNLOCK(p);
2272 AIO_UNLOCK(ki);
2273 uma_zfree(aiolio_zone, lj);
2274 } else
2275 AIO_UNLOCK(ki);
2276
2277 if (nerror)
2278 return (EIO);
2279 return (error);
2280 }
2281
2282 /* syscall - list directed I/O (REALTIME) */
2283 int
2284 sys_olio_listio(struct thread *td, struct olio_listio_args *uap)
2285 {
2286 struct aiocb **acb_list;
2287 struct sigevent *sigp, sig;
2288 struct osigevent osig;
2289 int error, nent;
2290
2291 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2292 return (EINVAL);
2293
2294 nent = uap->nent;
2295 if (nent < 0 || nent > AIO_LISTIO_MAX)
2296 return (EINVAL);
2297
2298 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2299 error = copyin(uap->sig, &osig, sizeof(osig));
2300 if (error)
2301 return (error);
2302 error = convert_old_sigevent(&osig, &sig);
2303 if (error)
2304 return (error);
2305 sigp = &sig;
2306 } else
2307 sigp = NULL;
2308
2309 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2310 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2311 if (error == 0)
2312 error = kern_lio_listio(td, uap->mode,
2313 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2314 &aiocb_ops_osigevent);
2315 free(acb_list, M_LIO);
2316 return (error);
2317 }
2318
2319 /* syscall - list directed I/O (REALTIME) */
2320 int
2321 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2322 {
2323 struct aiocb **acb_list;
2324 struct sigevent *sigp, sig;
2325 int error, nent;
2326
2327 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2328 return (EINVAL);
2329
2330 nent = uap->nent;
2331 if (nent < 0 || nent > AIO_LISTIO_MAX)
2332 return (EINVAL);
2333
2334 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2335 error = copyin(uap->sig, &sig, sizeof(sig));
2336 if (error)
2337 return (error);
2338 sigp = &sig;
2339 } else
2340 sigp = NULL;
2341
2342 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2343 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2344 if (error == 0)
2345 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2346 nent, sigp, &aiocb_ops);
2347 free(acb_list, M_LIO);
2348 return (error);
2349 }
2350
2351 /*
2352 * Called from interrupt thread for physio, we should return as fast
2353 * as possible, so we schedule a biohelper task.
2354 */
2355 static void
2356 aio_physwakeup(struct buf *bp)
2357 {
2358 struct aiocblist *aiocbe;
2359
2360 aiocbe = (struct aiocblist *)bp->b_caller1;
2361 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2362 }
2363
2364 /*
2365 * Task routine to perform heavy tasks, process wakeup, and signals.
2366 */
2367 static void
2368 biohelper(void *context, int pending)
2369 {
2370 struct aiocblist *aiocbe = context;
2371 struct buf *bp;
2372 struct proc *userp;
2373 struct kaioinfo *ki;
2374 int nblks;
2375
2376 bp = aiocbe->bp;
2377 userp = aiocbe->userproc;
2378 ki = userp->p_aioinfo;
2379 AIO_LOCK(ki);
2380 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2381 aiocbe->uaiocb._aiocb_private.error = 0;
2382 if (bp->b_ioflags & BIO_ERROR)
2383 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2384 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2385 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2386 aiocbe->outputcharge += nblks;
2387 else
2388 aiocbe->inputcharge += nblks;
2389 aiocbe->bp = NULL;
2390 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2391 ki->kaio_buffer_count--;
2392 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2393 AIO_UNLOCK(ki);
2394
2395 /* Release mapping into kernel space. */
2396 vunmapbuf(bp);
2397 relpbuf(bp, NULL);
2398 atomic_subtract_int(&num_buf_aio, 1);
2399 }
2400
2401 /* syscall - wait for the next completion of an aio request */
2402 static int
2403 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2404 struct timespec *ts, struct aiocb_ops *ops)
2405 {
2406 struct proc *p = td->td_proc;
2407 struct timeval atv;
2408 struct kaioinfo *ki;
2409 struct aiocblist *cb;
2410 struct aiocb *uuaiocb;
2411 int error, status, timo;
2412
2413 ops->store_aiocb(aiocbp, NULL);
2414
2415 timo = 0;
2416 if (ts) {
2417 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2418 return (EINVAL);
2419
2420 TIMESPEC_TO_TIMEVAL(&atv, ts);
2421 if (itimerfix(&atv))
2422 return (EINVAL);
2423 timo = tvtohz(&atv);
2424 }
2425
2426 if (p->p_aioinfo == NULL)
2427 aio_init_aioinfo(p);
2428 ki = p->p_aioinfo;
2429
2430 error = 0;
2431 cb = NULL;
2432 AIO_LOCK(ki);
2433 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2434 ki->kaio_flags |= KAIO_WAKEUP;
2435 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2436 "aiowc", timo);
2437 if (timo && error == ERESTART)
2438 error = EINTR;
2439 if (error)
2440 break;
2441 }
2442
2443 if (cb != NULL) {
2444 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2445 uuaiocb = cb->uuaiocb;
2446 status = cb->uaiocb._aiocb_private.status;
2447 error = cb->uaiocb._aiocb_private.error;
2448 td->td_retval[0] = status;
2449 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2450 td->td_ru.ru_oublock += cb->outputcharge;
2451 cb->outputcharge = 0;
2452 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2453 td->td_ru.ru_inblock += cb->inputcharge;
2454 cb->inputcharge = 0;
2455 }
2456 aio_free_entry(cb);
2457 AIO_UNLOCK(ki);
2458 ops->store_aiocb(aiocbp, uuaiocb);
2459 ops->store_error(uuaiocb, error);
2460 ops->store_status(uuaiocb, status);
2461 } else
2462 AIO_UNLOCK(ki);
2463
2464 return (error);
2465 }
2466
2467 int
2468 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2469 {
2470 struct timespec ts, *tsp;
2471 int error;
2472
2473 if (uap->timeout) {
2474 /* Get timespec struct. */
2475 error = copyin(uap->timeout, &ts, sizeof(ts));
2476 if (error)
2477 return (error);
2478 tsp = &ts;
2479 } else
2480 tsp = NULL;
2481
2482 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2483 }
2484
2485 static int
2486 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2487 struct aiocb_ops *ops)
2488 {
2489 struct proc *p = td->td_proc;
2490 struct kaioinfo *ki;
2491
2492 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2493 return (EINVAL);
2494 ki = p->p_aioinfo;
2495 if (ki == NULL)
2496 aio_init_aioinfo(p);
2497 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2498 }
2499
2500 int
2501 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2502 {
2503
2504 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2505 }
2506
2507 /* kqueue attach function */
2508 static int
2509 filt_aioattach(struct knote *kn)
2510 {
2511 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2512
2513 /*
2514 * The aiocbe pointer must be validated before using it, so
2515 * registration is restricted to the kernel; the user cannot
2516 * set EV_FLAG1.
2517 */
2518 if ((kn->kn_flags & EV_FLAG1) == 0)
2519 return (EPERM);
2520 kn->kn_ptr.p_aio = aiocbe;
2521 kn->kn_flags &= ~EV_FLAG1;
2522
2523 knlist_add(&aiocbe->klist, kn, 0);
2524
2525 return (0);
2526 }
2527
2528 /* kqueue detach function */
2529 static void
2530 filt_aiodetach(struct knote *kn)
2531 {
2532 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2533
2534 if (!knlist_empty(&aiocbe->klist))
2535 knlist_remove(&aiocbe->klist, kn, 0);
2536 }
2537
2538 /* kqueue filter function */
2539 /*ARGSUSED*/
2540 static int
2541 filt_aio(struct knote *kn, long hint)
2542 {
2543 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2544
2545 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2546 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2547 return (0);
2548 kn->kn_flags |= EV_EOF;
2549 return (1);
2550 }
2551
2552 /* kqueue attach function */
2553 static int
2554 filt_lioattach(struct knote *kn)
2555 {
2556 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2557
2558 /*
2559 * The aioliojob pointer must be validated before using it, so
2560 * registration is restricted to the kernel; the user cannot
2561 * set EV_FLAG1.
2562 */
2563 if ((kn->kn_flags & EV_FLAG1) == 0)
2564 return (EPERM);
2565 kn->kn_ptr.p_lio = lj;
2566 kn->kn_flags &= ~EV_FLAG1;
2567
2568 knlist_add(&lj->klist, kn, 0);
2569
2570 return (0);
2571 }
2572
2573 /* kqueue detach function */
2574 static void
2575 filt_liodetach(struct knote *kn)
2576 {
2577 struct aioliojob * lj = kn->kn_ptr.p_lio;
2578
2579 if (!knlist_empty(&lj->klist))
2580 knlist_remove(&lj->klist, kn, 0);
2581 }
2582
2583 /* kqueue filter function */
2584 /*ARGSUSED*/
2585 static int
2586 filt_lio(struct knote *kn, long hint)
2587 {
2588 struct aioliojob * lj = kn->kn_ptr.p_lio;
2589
2590 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2591 }
2592
2593 #ifdef COMPAT_FREEBSD32
2594
2595 struct __aiocb_private32 {
2596 int32_t status;
2597 int32_t error;
2598 uint32_t kernelinfo;
2599 };
2600
2601 typedef struct oaiocb32 {
2602 int aio_fildes; /* File descriptor */
2603 uint64_t aio_offset __packed; /* File offset for I/O */
2604 uint32_t aio_buf; /* I/O buffer in process space */
2605 uint32_t aio_nbytes; /* Number of bytes for I/O */
2606 struct osigevent32 aio_sigevent; /* Signal to deliver */
2607 int aio_lio_opcode; /* LIO opcode */
2608 int aio_reqprio; /* Request priority -- ignored */
2609 struct __aiocb_private32 _aiocb_private;
2610 } oaiocb32_t;
2611
2612 typedef struct aiocb32 {
2613 int32_t aio_fildes; /* File descriptor */
2614 uint64_t aio_offset __packed; /* File offset for I/O */
2615 uint32_t aio_buf; /* I/O buffer in process space */
2616 uint32_t aio_nbytes; /* Number of bytes for I/O */
2617 int __spare__[2];
2618 uint32_t __spare2__;
2619 int aio_lio_opcode; /* LIO opcode */
2620 int aio_reqprio; /* Request priority -- ignored */
2621 struct __aiocb_private32 _aiocb_private;
2622 struct sigevent32 aio_sigevent; /* Signal to deliver */
2623 } aiocb32_t;
2624
2625 static int
2626 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2627 {
2628
2629 /*
2630 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2631 * supported by AIO with the old sigevent structure.
2632 */
2633 CP(*osig, *nsig, sigev_notify);
2634 switch (nsig->sigev_notify) {
2635 case SIGEV_NONE:
2636 break;
2637 case SIGEV_SIGNAL:
2638 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2639 break;
2640 case SIGEV_KEVENT:
2641 nsig->sigev_notify_kqueue =
2642 osig->__sigev_u.__sigev_notify_kqueue;
2643 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2644 break;
2645 default:
2646 return (EINVAL);
2647 }
2648 return (0);
2649 }
2650
2651 static int
2652 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2653 {
2654 struct oaiocb32 job32;
2655 int error;
2656
2657 bzero(kjob, sizeof(struct aiocb));
2658 error = copyin(ujob, &job32, sizeof(job32));
2659 if (error)
2660 return (error);
2661
2662 CP(job32, *kjob, aio_fildes);
2663 CP(job32, *kjob, aio_offset);
2664 PTRIN_CP(job32, *kjob, aio_buf);
2665 CP(job32, *kjob, aio_nbytes);
2666 CP(job32, *kjob, aio_lio_opcode);
2667 CP(job32, *kjob, aio_reqprio);
2668 CP(job32, *kjob, _aiocb_private.status);
2669 CP(job32, *kjob, _aiocb_private.error);
2670 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2671 return (convert_old_sigevent32(&job32.aio_sigevent,
2672 &kjob->aio_sigevent));
2673 }
2674
2675 static int
2676 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig)
2677 {
2678
2679 CP(*sig32, *sig, sigev_notify);
2680 switch (sig->sigev_notify) {
2681 case SIGEV_NONE:
2682 break;
2683 case SIGEV_THREAD_ID:
2684 CP(*sig32, *sig, sigev_notify_thread_id);
2685 /* FALLTHROUGH */
2686 case SIGEV_SIGNAL:
2687 CP(*sig32, *sig, sigev_signo);
2688 break;
2689 case SIGEV_KEVENT:
2690 CP(*sig32, *sig, sigev_notify_kqueue);
2691 PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr);
2692 break;
2693 default:
2694 return (EINVAL);
2695 }
2696 return (0);
2697 }
2698
2699 static int
2700 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2701 {
2702 struct aiocb32 job32;
2703 int error;
2704
2705 error = copyin(ujob, &job32, sizeof(job32));
2706 if (error)
2707 return (error);
2708 CP(job32, *kjob, aio_fildes);
2709 CP(job32, *kjob, aio_offset);
2710 PTRIN_CP(job32, *kjob, aio_buf);
2711 CP(job32, *kjob, aio_nbytes);
2712 CP(job32, *kjob, aio_lio_opcode);
2713 CP(job32, *kjob, aio_reqprio);
2714 CP(job32, *kjob, _aiocb_private.status);
2715 CP(job32, *kjob, _aiocb_private.error);
2716 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2717 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2718 }
2719
2720 static long
2721 aiocb32_fetch_status(struct aiocb *ujob)
2722 {
2723 struct aiocb32 *ujob32;
2724
2725 ujob32 = (struct aiocb32 *)ujob;
2726 return (fuword32(&ujob32->_aiocb_private.status));
2727 }
2728
2729 static long
2730 aiocb32_fetch_error(struct aiocb *ujob)
2731 {
2732 struct aiocb32 *ujob32;
2733
2734 ujob32 = (struct aiocb32 *)ujob;
2735 return (fuword32(&ujob32->_aiocb_private.error));
2736 }
2737
2738 static int
2739 aiocb32_store_status(struct aiocb *ujob, long status)
2740 {
2741 struct aiocb32 *ujob32;
2742
2743 ujob32 = (struct aiocb32 *)ujob;
2744 return (suword32(&ujob32->_aiocb_private.status, status));
2745 }
2746
2747 static int
2748 aiocb32_store_error(struct aiocb *ujob, long error)
2749 {
2750 struct aiocb32 *ujob32;
2751
2752 ujob32 = (struct aiocb32 *)ujob;
2753 return (suword32(&ujob32->_aiocb_private.error, error));
2754 }
2755
2756 static int
2757 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2758 {
2759 struct aiocb32 *ujob32;
2760
2761 ujob32 = (struct aiocb32 *)ujob;
2762 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2763 }
2764
2765 static int
2766 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2767 {
2768
2769 return (suword32(ujobp, (long)ujob));
2770 }
2771
2772 static struct aiocb_ops aiocb32_ops = {
2773 .copyin = aiocb32_copyin,
2774 .fetch_status = aiocb32_fetch_status,
2775 .fetch_error = aiocb32_fetch_error,
2776 .store_status = aiocb32_store_status,
2777 .store_error = aiocb32_store_error,
2778 .store_kernelinfo = aiocb32_store_kernelinfo,
2779 .store_aiocb = aiocb32_store_aiocb,
2780 };
2781
2782 static struct aiocb_ops aiocb32_ops_osigevent = {
2783 .copyin = aiocb32_copyin_old_sigevent,
2784 .fetch_status = aiocb32_fetch_status,
2785 .fetch_error = aiocb32_fetch_error,
2786 .store_status = aiocb32_store_status,
2787 .store_error = aiocb32_store_error,
2788 .store_kernelinfo = aiocb32_store_kernelinfo,
2789 .store_aiocb = aiocb32_store_aiocb,
2790 };
2791
2792 int
2793 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2794 {
2795
2796 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2797 }
2798
2799 int
2800 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2801 {
2802 struct timespec32 ts32;
2803 struct timespec ts, *tsp;
2804 struct aiocb **ujoblist;
2805 uint32_t *ujoblist32;
2806 int error, i;
2807
2808 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2809 return (EINVAL);
2810
2811 if (uap->timeout) {
2812 /* Get timespec struct. */
2813 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2814 return (error);
2815 CP(ts32, ts, tv_sec);
2816 CP(ts32, ts, tv_nsec);
2817 tsp = &ts;
2818 } else
2819 tsp = NULL;
2820
2821 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2822 ujoblist32 = (uint32_t *)ujoblist;
2823 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2824 sizeof(ujoblist32[0]));
2825 if (error == 0) {
2826 for (i = uap->nent; i > 0; i--)
2827 ujoblist[i] = PTRIN(ujoblist32[i]);
2828
2829 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2830 }
2831 uma_zfree(aiol_zone, ujoblist);
2832 return (error);
2833 }
2834
2835 int
2836 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2837 {
2838
2839 return (sys_aio_cancel(td, (struct aio_cancel_args *)uap));
2840 }
2841
2842 int
2843 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2844 {
2845
2846 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2847 }
2848
2849 int
2850 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2851 {
2852
2853 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2854 &aiocb32_ops_osigevent));
2855 }
2856
2857 int
2858 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2859 {
2860
2861 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2862 &aiocb32_ops));
2863 }
2864
2865 int
2866 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2867 {
2868
2869 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2870 &aiocb32_ops_osigevent));
2871 }
2872
2873 int
2874 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2875 {
2876
2877 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2878 &aiocb32_ops));
2879 }
2880
2881 int
2882 freebsd32_aio_waitcomplete(struct thread *td,
2883 struct freebsd32_aio_waitcomplete_args *uap)
2884 {
2885 struct timespec32 ts32;
2886 struct timespec ts, *tsp;
2887 int error;
2888
2889 if (uap->timeout) {
2890 /* Get timespec struct. */
2891 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2892 if (error)
2893 return (error);
2894 CP(ts32, ts, tv_sec);
2895 CP(ts32, ts, tv_nsec);
2896 tsp = &ts;
2897 } else
2898 tsp = NULL;
2899
2900 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2901 &aiocb32_ops));
2902 }
2903
2904 int
2905 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2906 {
2907
2908 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2909 &aiocb32_ops));
2910 }
2911
2912 int
2913 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2914 {
2915 struct aiocb **acb_list;
2916 struct sigevent *sigp, sig;
2917 struct osigevent32 osig;
2918 uint32_t *acb_list32;
2919 int error, i, nent;
2920
2921 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2922 return (EINVAL);
2923
2924 nent = uap->nent;
2925 if (nent < 0 || nent > AIO_LISTIO_MAX)
2926 return (EINVAL);
2927
2928 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2929 error = copyin(uap->sig, &osig, sizeof(osig));
2930 if (error)
2931 return (error);
2932 error = convert_old_sigevent32(&osig, &sig);
2933 if (error)
2934 return (error);
2935 sigp = &sig;
2936 } else
2937 sigp = NULL;
2938
2939 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2940 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2941 if (error) {
2942 free(acb_list32, M_LIO);
2943 return (error);
2944 }
2945 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2946 for (i = 0; i < nent; i++)
2947 acb_list[i] = PTRIN(acb_list32[i]);
2948 free(acb_list32, M_LIO);
2949
2950 error = kern_lio_listio(td, uap->mode,
2951 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2952 &aiocb32_ops_osigevent);
2953 free(acb_list, M_LIO);
2954 return (error);
2955 }
2956
2957 int
2958 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2959 {
2960 struct aiocb **acb_list;
2961 struct sigevent *sigp, sig;
2962 struct sigevent32 sig32;
2963 uint32_t *acb_list32;
2964 int error, i, nent;
2965
2966 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2967 return (EINVAL);
2968
2969 nent = uap->nent;
2970 if (nent < 0 || nent > AIO_LISTIO_MAX)
2971 return (EINVAL);
2972
2973 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2974 error = copyin(uap->sig, &sig32, sizeof(sig32));
2975 if (error)
2976 return (error);
2977 error = convert_sigevent32(&sig32, &sig);
2978 if (error)
2979 return (error);
2980 sigp = &sig;
2981 } else
2982 sigp = NULL;
2983
2984 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2985 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2986 if (error) {
2987 free(acb_list32, M_LIO);
2988 return (error);
2989 }
2990 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2991 for (i = 0; i < nent; i++)
2992 acb_list[i] = PTRIN(acb_list32[i]);
2993 free(acb_list32, M_LIO);
2994
2995 error = kern_lio_listio(td, uap->mode,
2996 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2997 &aiocb32_ops);
2998 free(acb_list, M_LIO);
2999 return (error);
3000 }
3001
3002 #endif
Cache object: ae0301e049ca1fde1c588eb8e8431a29
|