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("$FreeBSD$");
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/capsicum.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/rwlock.h>
50 #include <sys/sema.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/syscall.h>
54 #include <sys/sysent.h>
55 #include <sys/sysctl.h>
56 #include <sys/sx.h>
57 #include <sys/taskqueue.h>
58 #include <sys/vnode.h>
59 #include <sys/conf.h>
60 #include <sys/event.h>
61 #include <sys/mount.h>
62 #include <geom/geom.h>
63
64 #include <machine/atomic.h>
65
66 #include <vm/vm.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_extern.h>
69 #include <vm/pmap.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_object.h>
72 #include <vm/uma.h>
73 #include <sys/aio.h>
74
75 #include "opt_vfs_aio.h"
76
77 /*
78 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
79 * overflow. (XXX will be removed soon.)
80 */
81 static u_long jobrefid;
82
83 /*
84 * Counter for aio_fsync.
85 */
86 static uint64_t jobseqno;
87
88 #define JOBST_NULL 0
89 #define JOBST_JOBQSOCK 1
90 #define JOBST_JOBQGLOBAL 2
91 #define JOBST_JOBRUNNING 3
92 #define JOBST_JOBFINISHED 4
93 #define JOBST_JOBQBUF 5
94 #define JOBST_JOBQSYNC 6
95
96 #ifndef MAX_AIO_PER_PROC
97 #define MAX_AIO_PER_PROC 32
98 #endif
99
100 #ifndef MAX_AIO_QUEUE_PER_PROC
101 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
102 #endif
103
104 #ifndef MAX_AIO_PROCS
105 #define MAX_AIO_PROCS 32
106 #endif
107
108 #ifndef MAX_AIO_QUEUE
109 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
110 #endif
111
112 #ifndef TARGET_AIO_PROCS
113 #define TARGET_AIO_PROCS 4
114 #endif
115
116 #ifndef MAX_BUF_AIO
117 #define MAX_BUF_AIO 16
118 #endif
119
120 #ifndef AIOD_TIMEOUT_DEFAULT
121 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
122 #endif
123
124 #ifndef AIOD_LIFETIME_DEFAULT
125 #define AIOD_LIFETIME_DEFAULT (30 * hz)
126 #endif
127
128 FEATURE(aio, "Asynchronous I/O");
129
130 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
131
132 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
133
134 static int max_aio_procs = MAX_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
136 CTLFLAG_RW, &max_aio_procs, 0,
137 "Maximum number of kernel threads to use for handling async IO ");
138
139 static int num_aio_procs = 0;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
141 CTLFLAG_RD, &num_aio_procs, 0,
142 "Number of presently active kernel threads for async IO");
143
144 /*
145 * The code will adjust the actual number of AIO processes towards this
146 * number when it gets a chance.
147 */
148 static int target_aio_procs = TARGET_AIO_PROCS;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
150 0, "Preferred number of ready kernel threads for async IO");
151
152 static int max_queue_count = MAX_AIO_QUEUE;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
154 "Maximum number of aio requests to queue, globally");
155
156 static int num_queue_count = 0;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
158 "Number of queued aio requests");
159
160 static int num_buf_aio = 0;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
162 "Number of aio requests presently handled by the buf subsystem");
163
164 /* Number of async I/O thread in the process of being started */
165 /* XXX This should be local to aio_aqueue() */
166 static int num_aio_resv_start = 0;
167
168 static int aiod_timeout;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
170 "Timeout value for synchronous aio operations");
171
172 static int aiod_lifetime;
173 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
174 "Maximum lifetime for idle aiod");
175
176 static int unloadable = 0;
177 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
178 "Allow unload of aio (not recommended)");
179
180
181 static int max_aio_per_proc = MAX_AIO_PER_PROC;
182 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
183 0, "Maximum active aio requests per process (stored in the process)");
184
185 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
186 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
187 &max_aio_queue_per_proc, 0,
188 "Maximum queued aio requests per process (stored in the process)");
189
190 static int max_buf_aio = MAX_BUF_AIO;
191 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
192 "Maximum buf aio requests per process (stored in the process)");
193
194 typedef struct oaiocb {
195 int aio_fildes; /* File descriptor */
196 off_t aio_offset; /* File offset for I/O */
197 volatile void *aio_buf; /* I/O buffer in process space */
198 size_t aio_nbytes; /* Number of bytes for I/O */
199 struct osigevent aio_sigevent; /* Signal to deliver */
200 int aio_lio_opcode; /* LIO opcode */
201 int aio_reqprio; /* Request priority -- ignored */
202 struct __aiocb_private _aiocb_private;
203 } oaiocb_t;
204
205 /*
206 * Below is a key of locks used to protect each member of struct aiocblist
207 * aioliojob and kaioinfo and any backends.
208 *
209 * * - need not protected
210 * a - locked by kaioinfo lock
211 * b - locked by backend lock, the backend lock can be null in some cases,
212 * for example, BIO belongs to this type, in this case, proc lock is
213 * reused.
214 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
215 */
216
217 /*
218 * Current, there is only two backends: BIO and generic file I/O.
219 * socket I/O is served by generic file I/O, this is not a good idea, since
220 * disk file I/O and any other types without O_NONBLOCK flag can block daemon
221 * threads, if there is no thread to serve socket I/O, the socket I/O will be
222 * delayed too long or starved, we should create some threads dedicated to
223 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
224 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
225 * structure is not safe because there is race between userland and aio
226 * daemons.
227 */
228
229 struct aiocblist {
230 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */
231 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */
232 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */
233 int jobflags; /* (a) job flags */
234 int jobstate; /* (b) job state */
235 int inputcharge; /* (*) input blockes */
236 int outputcharge; /* (*) output blockes */
237 struct bio *bp; /* (*) BIO backend BIO pointer */
238 struct buf *pbuf; /* (*) BIO backend buffer pointer */
239 struct vm_page *pages[btoc(MAXPHYS)+1]; /* BIO backend pages */
240 int npages; /* BIO backend number of pages */
241 struct proc *userproc; /* (*) user process */
242 struct ucred *cred; /* (*) active credential when created */
243 struct file *fd_file; /* (*) pointer to file structure */
244 struct aioliojob *lio; /* (*) optional lio job */
245 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */
246 struct knlist klist; /* (a) list of knotes */
247 struct aiocb uaiocb; /* (*) kernel I/O control block */
248 ksiginfo_t ksi; /* (a) realtime signal info */
249 uint64_t seqno; /* (*) job number */
250 int pending; /* (a) number of pending I/O, aio_fsync only */
251 };
252
253 /* jobflags */
254 #define AIOCBLIST_DONE 0x01
255 #define AIOCBLIST_BUFDONE 0x02
256 #define AIOCBLIST_RUNDOWN 0x04
257 #define AIOCBLIST_CHECKSYNC 0x08
258
259 /*
260 * AIO process info
261 */
262 #define AIOP_FREE 0x1 /* proc on free queue */
263
264 struct aiothreadlist {
265 int aiothreadflags; /* (c) AIO proc flags */
266 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */
267 struct thread *aiothread; /* (*) the AIO thread */
268 };
269
270 /*
271 * data-structure for lio signal management
272 */
273 struct aioliojob {
274 int lioj_flags; /* (a) listio flags */
275 int lioj_count; /* (a) listio flags */
276 int lioj_finished_count; /* (a) listio flags */
277 struct sigevent lioj_signal; /* (a) signal on all I/O done */
278 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
279 struct knlist klist; /* (a) list of knotes */
280 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
281 };
282
283 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
284 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
285 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
286
287 /*
288 * per process aio data structure
289 */
290 struct kaioinfo {
291 struct mtx kaio_mtx; /* the lock to protect this struct */
292 int kaio_flags; /* (a) per process kaio flags */
293 int kaio_maxactive_count; /* (*) maximum number of AIOs */
294 int kaio_active_count; /* (c) number of currently used AIOs */
295 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */
296 int kaio_count; /* (a) size of AIO queue */
297 int kaio_ballowed_count; /* (*) maximum number of buffers */
298 int kaio_buffer_count; /* (a) number of physio buffers */
299 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */
300 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */
301 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
302 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */
303 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */
304 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets,
305 * NOT USED YET.
306 */
307 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */
308 struct task kaio_task; /* (*) task to kick aio threads */
309 };
310
311 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
312 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
313 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
314 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
315
316 #define KAIO_RUNDOWN 0x1 /* process is being run down */
317 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
318
319 /*
320 * Operations used to interact with userland aio control blocks.
321 * Different ABIs provide their own operations.
322 */
323 struct aiocb_ops {
324 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
325 long (*fetch_status)(struct aiocb *ujob);
326 long (*fetch_error)(struct aiocb *ujob);
327 int (*store_status)(struct aiocb *ujob, long status);
328 int (*store_error)(struct aiocb *ujob, long error);
329 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
330 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
331 };
332
333 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */
334 static struct sema aio_newproc_sem;
335 static struct mtx aio_job_mtx;
336 static struct mtx aio_sock_mtx;
337 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */
338 static struct unrhdr *aiod_unr;
339
340 void aio_init_aioinfo(struct proc *p);
341 static int aio_onceonly(void);
342 static int aio_free_entry(struct aiocblist *aiocbe);
343 static void aio_process_rw(struct aiocblist *aiocbe);
344 static void aio_process_sync(struct aiocblist *aiocbe);
345 static void aio_process_mlock(struct aiocblist *aiocbe);
346 static int aio_newproc(int *);
347 int aio_aqueue(struct thread *td, struct aiocb *job,
348 struct aioliojob *lio, int type, struct aiocb_ops *ops);
349 static void aio_physwakeup(struct bio *bp);
350 static void aio_proc_rundown(void *arg, struct proc *p);
351 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
352 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
353 static void aio_daemon(void *param);
354 static void aio_swake_cb(struct socket *, struct sockbuf *);
355 static int aio_unload(void);
356 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
357 #define DONE_BUF 1
358 #define DONE_QUEUE 2
359 static int aio_kick(struct proc *userp);
360 static void aio_kick_nowait(struct proc *userp);
361 static void aio_kick_helper(void *context, int pending);
362 static int filt_aioattach(struct knote *kn);
363 static void filt_aiodetach(struct knote *kn);
364 static int filt_aio(struct knote *kn, long hint);
365 static int filt_lioattach(struct knote *kn);
366 static void filt_liodetach(struct knote *kn);
367 static int filt_lio(struct knote *kn, long hint);
368
369 /*
370 * Zones for:
371 * kaio Per process async io info
372 * aiop async io thread data
373 * aiocb async io jobs
374 * aiol list io job pointer - internal to aio_suspend XXX
375 * aiolio list io jobs
376 */
377 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
378
379 /* kqueue filters for aio */
380 static struct filterops aio_filtops = {
381 .f_isfd = 0,
382 .f_attach = filt_aioattach,
383 .f_detach = filt_aiodetach,
384 .f_event = filt_aio,
385 };
386 static struct filterops lio_filtops = {
387 .f_isfd = 0,
388 .f_attach = filt_lioattach,
389 .f_detach = filt_liodetach,
390 .f_event = filt_lio
391 };
392
393 static eventhandler_tag exit_tag, exec_tag;
394
395 TASKQUEUE_DEFINE_THREAD(aiod_bio);
396
397 /*
398 * Main operations function for use as a kernel module.
399 */
400 static int
401 aio_modload(struct module *module, int cmd, void *arg)
402 {
403 int error = 0;
404
405 switch (cmd) {
406 case MOD_LOAD:
407 aio_onceonly();
408 break;
409 case MOD_UNLOAD:
410 error = aio_unload();
411 break;
412 case MOD_SHUTDOWN:
413 break;
414 default:
415 error = EINVAL;
416 break;
417 }
418 return (error);
419 }
420
421 static moduledata_t aio_mod = {
422 "aio",
423 &aio_modload,
424 NULL
425 };
426
427 static struct syscall_helper_data aio_syscalls[] = {
428 SYSCALL_INIT_HELPER(aio_cancel),
429 SYSCALL_INIT_HELPER(aio_error),
430 SYSCALL_INIT_HELPER(aio_fsync),
431 SYSCALL_INIT_HELPER(aio_mlock),
432 SYSCALL_INIT_HELPER(aio_read),
433 SYSCALL_INIT_HELPER(aio_return),
434 SYSCALL_INIT_HELPER(aio_suspend),
435 SYSCALL_INIT_HELPER(aio_waitcomplete),
436 SYSCALL_INIT_HELPER(aio_write),
437 SYSCALL_INIT_HELPER(lio_listio),
438 SYSCALL_INIT_HELPER(oaio_read),
439 SYSCALL_INIT_HELPER(oaio_write),
440 SYSCALL_INIT_HELPER(olio_listio),
441 SYSCALL_INIT_LAST
442 };
443
444 #ifdef COMPAT_FREEBSD32
445 #include <sys/mount.h>
446 #include <sys/socket.h>
447 #include <compat/freebsd32/freebsd32.h>
448 #include <compat/freebsd32/freebsd32_proto.h>
449 #include <compat/freebsd32/freebsd32_signal.h>
450 #include <compat/freebsd32/freebsd32_syscall.h>
451 #include <compat/freebsd32/freebsd32_util.h>
452
453 static struct syscall_helper_data aio32_syscalls[] = {
454 SYSCALL32_INIT_HELPER(freebsd32_aio_return),
455 SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
456 SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
457 SYSCALL32_INIT_HELPER(freebsd32_aio_error),
458 SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
459 SYSCALL32_INIT_HELPER(freebsd32_aio_mlock),
460 SYSCALL32_INIT_HELPER(freebsd32_aio_read),
461 SYSCALL32_INIT_HELPER(freebsd32_aio_write),
462 SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
463 SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
464 SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
465 SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
466 SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
467 SYSCALL_INIT_LAST
468 };
469 #endif
470
471 DECLARE_MODULE(aio, aio_mod,
472 SI_SUB_VFS, SI_ORDER_ANY);
473 MODULE_VERSION(aio, 1);
474
475 /*
476 * Startup initialization
477 */
478 static int
479 aio_onceonly(void)
480 {
481 int error;
482
483 /* XXX: should probably just use so->callback */
484 aio_swake = &aio_swake_cb;
485 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
486 EVENTHANDLER_PRI_ANY);
487 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
488 EVENTHANDLER_PRI_ANY);
489 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
490 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
491 TAILQ_INIT(&aio_freeproc);
492 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
493 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
494 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
495 TAILQ_INIT(&aio_jobs);
496 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
497 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
498 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
499 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
500 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
501 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
502 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
503 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
504 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
505 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
506 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
507 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
508 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
509 jobrefid = 1;
510 async_io_version = _POSIX_VERSION;
511 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
512 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
513 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
514
515 error = syscall_helper_register(aio_syscalls);
516 if (error)
517 return (error);
518 #ifdef COMPAT_FREEBSD32
519 error = syscall32_helper_register(aio32_syscalls);
520 if (error)
521 return (error);
522 #endif
523 return (0);
524 }
525
526 /*
527 * Callback for unload of AIO when used as a module.
528 */
529 static int
530 aio_unload(void)
531 {
532 int error;
533
534 /*
535 * XXX: no unloads by default, it's too dangerous.
536 * perhaps we could do it if locked out callers and then
537 * did an aio_proc_rundown() on each process.
538 *
539 * jhb: aio_proc_rundown() needs to run on curproc though,
540 * so I don't think that would fly.
541 */
542 if (!unloadable)
543 return (EOPNOTSUPP);
544
545 #ifdef COMPAT_FREEBSD32
546 syscall32_helper_unregister(aio32_syscalls);
547 #endif
548 syscall_helper_unregister(aio_syscalls);
549
550 error = kqueue_del_filteropts(EVFILT_AIO);
551 if (error)
552 return error;
553 error = kqueue_del_filteropts(EVFILT_LIO);
554 if (error)
555 return error;
556 async_io_version = 0;
557 aio_swake = NULL;
558 taskqueue_free(taskqueue_aiod_bio);
559 delete_unrhdr(aiod_unr);
560 uma_zdestroy(kaio_zone);
561 uma_zdestroy(aiop_zone);
562 uma_zdestroy(aiocb_zone);
563 uma_zdestroy(aiol_zone);
564 uma_zdestroy(aiolio_zone);
565 EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
566 EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
567 mtx_destroy(&aio_job_mtx);
568 mtx_destroy(&aio_sock_mtx);
569 sema_destroy(&aio_newproc_sem);
570 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
571 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
572 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
573 return (0);
574 }
575
576 /*
577 * Init the per-process aioinfo structure. The aioinfo limits are set
578 * per-process for user limit (resource) management.
579 */
580 void
581 aio_init_aioinfo(struct proc *p)
582 {
583 struct kaioinfo *ki;
584
585 ki = uma_zalloc(kaio_zone, M_WAITOK);
586 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
587 ki->kaio_flags = 0;
588 ki->kaio_maxactive_count = max_aio_per_proc;
589 ki->kaio_active_count = 0;
590 ki->kaio_qallowed_count = max_aio_queue_per_proc;
591 ki->kaio_count = 0;
592 ki->kaio_ballowed_count = max_buf_aio;
593 ki->kaio_buffer_count = 0;
594 TAILQ_INIT(&ki->kaio_all);
595 TAILQ_INIT(&ki->kaio_done);
596 TAILQ_INIT(&ki->kaio_jobqueue);
597 TAILQ_INIT(&ki->kaio_bufqueue);
598 TAILQ_INIT(&ki->kaio_liojoblist);
599 TAILQ_INIT(&ki->kaio_sockqueue);
600 TAILQ_INIT(&ki->kaio_syncqueue);
601 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
602 PROC_LOCK(p);
603 if (p->p_aioinfo == NULL) {
604 p->p_aioinfo = ki;
605 PROC_UNLOCK(p);
606 } else {
607 PROC_UNLOCK(p);
608 mtx_destroy(&ki->kaio_mtx);
609 uma_zfree(kaio_zone, ki);
610 }
611
612 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
613 aio_newproc(NULL);
614 }
615
616 static int
617 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
618 {
619 struct thread *td;
620 int error;
621
622 error = sigev_findtd(p, sigev, &td);
623 if (error)
624 return (error);
625 if (!KSI_ONQ(ksi)) {
626 ksiginfo_set_sigev(ksi, sigev);
627 ksi->ksi_code = SI_ASYNCIO;
628 ksi->ksi_flags |= KSI_EXT | KSI_INS;
629 tdsendsignal(p, td, ksi->ksi_signo, ksi);
630 }
631 PROC_UNLOCK(p);
632 return (error);
633 }
634
635 /*
636 * Free a job entry. Wait for completion if it is currently active, but don't
637 * delay forever. If we delay, we return a flag that says that we have to
638 * restart the queue scan.
639 */
640 static int
641 aio_free_entry(struct aiocblist *aiocbe)
642 {
643 struct kaioinfo *ki;
644 struct aioliojob *lj;
645 struct proc *p;
646
647 p = aiocbe->userproc;
648 MPASS(curproc == p);
649 ki = p->p_aioinfo;
650 MPASS(ki != NULL);
651
652 AIO_LOCK_ASSERT(ki, MA_OWNED);
653 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
654
655 atomic_subtract_int(&num_queue_count, 1);
656
657 ki->kaio_count--;
658 MPASS(ki->kaio_count >= 0);
659
660 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
661 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
662
663 lj = aiocbe->lio;
664 if (lj) {
665 lj->lioj_count--;
666 lj->lioj_finished_count--;
667
668 if (lj->lioj_count == 0) {
669 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
670 /* lio is going away, we need to destroy any knotes */
671 knlist_delete(&lj->klist, curthread, 1);
672 PROC_LOCK(p);
673 sigqueue_take(&lj->lioj_ksi);
674 PROC_UNLOCK(p);
675 uma_zfree(aiolio_zone, lj);
676 }
677 }
678
679 /* aiocbe is going away, we need to destroy any knotes */
680 knlist_delete(&aiocbe->klist, curthread, 1);
681 PROC_LOCK(p);
682 sigqueue_take(&aiocbe->ksi);
683 PROC_UNLOCK(p);
684
685 MPASS(aiocbe->bp == NULL);
686 aiocbe->jobstate = JOBST_NULL;
687 AIO_UNLOCK(ki);
688
689 /*
690 * The thread argument here is used to find the owning process
691 * and is also passed to fo_close() which may pass it to various
692 * places such as devsw close() routines. Because of that, we
693 * need a thread pointer from the process owning the job that is
694 * persistent and won't disappear out from under us or move to
695 * another process.
696 *
697 * Currently, all the callers of this function call it to remove
698 * an aiocblist from the current process' job list either via a
699 * syscall or due to the current process calling exit() or
700 * execve(). Thus, we know that p == curproc. We also know that
701 * curthread can't exit since we are curthread.
702 *
703 * Therefore, we use curthread as the thread to pass to
704 * knlist_delete(). This does mean that it is possible for the
705 * thread pointer at close time to differ from the thread pointer
706 * at open time, but this is already true of file descriptors in
707 * a multithreaded process.
708 */
709 if (aiocbe->fd_file)
710 fdrop(aiocbe->fd_file, curthread);
711 crfree(aiocbe->cred);
712 uma_zfree(aiocb_zone, aiocbe);
713 AIO_LOCK(ki);
714
715 return (0);
716 }
717
718 static void
719 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
720 {
721 aio_proc_rundown(arg, p);
722 }
723
724 /*
725 * Rundown the jobs for a given process.
726 */
727 static void
728 aio_proc_rundown(void *arg, struct proc *p)
729 {
730 struct kaioinfo *ki;
731 struct aioliojob *lj;
732 struct aiocblist *cbe, *cbn;
733 struct file *fp;
734 struct socket *so;
735 int remove;
736
737 KASSERT(curthread->td_proc == p,
738 ("%s: called on non-curproc", __func__));
739 ki = p->p_aioinfo;
740 if (ki == NULL)
741 return;
742
743 AIO_LOCK(ki);
744 ki->kaio_flags |= KAIO_RUNDOWN;
745
746 restart:
747
748 /*
749 * Try to cancel all pending requests. This code simulates
750 * aio_cancel on all pending I/O requests.
751 */
752 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
753 remove = 0;
754 mtx_lock(&aio_job_mtx);
755 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
756 TAILQ_REMOVE(&aio_jobs, cbe, list);
757 remove = 1;
758 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
759 fp = cbe->fd_file;
760 MPASS(fp->f_type == DTYPE_SOCKET);
761 so = fp->f_data;
762 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
763 remove = 1;
764 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
765 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
766 remove = 1;
767 }
768 mtx_unlock(&aio_job_mtx);
769
770 if (remove) {
771 cbe->jobstate = JOBST_JOBFINISHED;
772 cbe->uaiocb._aiocb_private.status = -1;
773 cbe->uaiocb._aiocb_private.error = ECANCELED;
774 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
775 aio_bio_done_notify(p, cbe, DONE_QUEUE);
776 }
777 }
778
779 /* Wait for all running I/O to be finished */
780 if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
781 TAILQ_FIRST(&ki->kaio_jobqueue)) {
782 ki->kaio_flags |= KAIO_WAKEUP;
783 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
784 goto restart;
785 }
786
787 /* Free all completed I/O requests. */
788 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
789 aio_free_entry(cbe);
790
791 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
792 if (lj->lioj_count == 0) {
793 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
794 knlist_delete(&lj->klist, curthread, 1);
795 PROC_LOCK(p);
796 sigqueue_take(&lj->lioj_ksi);
797 PROC_UNLOCK(p);
798 uma_zfree(aiolio_zone, lj);
799 } else {
800 panic("LIO job not cleaned up: C:%d, FC:%d\n",
801 lj->lioj_count, lj->lioj_finished_count);
802 }
803 }
804 AIO_UNLOCK(ki);
805 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
806 mtx_destroy(&ki->kaio_mtx);
807 uma_zfree(kaio_zone, ki);
808 p->p_aioinfo = NULL;
809 }
810
811 /*
812 * Select a job to run (called by an AIO daemon).
813 */
814 static struct aiocblist *
815 aio_selectjob(struct aiothreadlist *aiop)
816 {
817 struct aiocblist *aiocbe;
818 struct kaioinfo *ki;
819 struct proc *userp;
820
821 mtx_assert(&aio_job_mtx, MA_OWNED);
822 TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
823 userp = aiocbe->userproc;
824 ki = userp->p_aioinfo;
825
826 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
827 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
828 /* Account for currently active jobs. */
829 ki->kaio_active_count++;
830 aiocbe->jobstate = JOBST_JOBRUNNING;
831 break;
832 }
833 }
834 return (aiocbe);
835 }
836
837 /*
838 * Move all data to a permanent storage device, this code
839 * simulates fsync syscall.
840 */
841 static int
842 aio_fsync_vnode(struct thread *td, struct vnode *vp)
843 {
844 struct mount *mp;
845 int error;
846
847 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
848 goto drop;
849 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
850 if (vp->v_object != NULL) {
851 VM_OBJECT_WLOCK(vp->v_object);
852 vm_object_page_clean(vp->v_object, 0, 0, 0);
853 VM_OBJECT_WUNLOCK(vp->v_object);
854 }
855 error = VOP_FSYNC(vp, MNT_WAIT, td);
856
857 VOP_UNLOCK(vp, 0);
858 vn_finished_write(mp);
859 drop:
860 return (error);
861 }
862
863 /*
864 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
865 * does the I/O request for the non-physio version of the operations. The
866 * normal vn operations are used, and this code should work in all instances
867 * for every type of file, including pipes, sockets, fifos, and regular files.
868 *
869 * XXX I don't think it works well for socket, pipe, and fifo.
870 */
871 static void
872 aio_process_rw(struct aiocblist *aiocbe)
873 {
874 struct ucred *td_savedcred;
875 struct thread *td;
876 struct aiocb *cb;
877 struct file *fp;
878 struct socket *so;
879 struct uio auio;
880 struct iovec aiov;
881 int cnt;
882 int error;
883 int oublock_st, oublock_end;
884 int inblock_st, inblock_end;
885
886 KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_READ ||
887 aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE,
888 ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
889
890 td = curthread;
891 td_savedcred = td->td_ucred;
892 td->td_ucred = aiocbe->cred;
893 cb = &aiocbe->uaiocb;
894 fp = aiocbe->fd_file;
895
896 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
897 aiov.iov_len = cb->aio_nbytes;
898
899 auio.uio_iov = &aiov;
900 auio.uio_iovcnt = 1;
901 auio.uio_offset = cb->aio_offset;
902 auio.uio_resid = cb->aio_nbytes;
903 cnt = cb->aio_nbytes;
904 auio.uio_segflg = UIO_USERSPACE;
905 auio.uio_td = td;
906
907 inblock_st = td->td_ru.ru_inblock;
908 oublock_st = td->td_ru.ru_oublock;
909 /*
910 * aio_aqueue() acquires a reference to the file that is
911 * released in aio_free_entry().
912 */
913 if (cb->aio_lio_opcode == LIO_READ) {
914 auio.uio_rw = UIO_READ;
915 if (auio.uio_resid == 0)
916 error = 0;
917 else
918 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
919 } else {
920 if (fp->f_type == DTYPE_VNODE)
921 bwillwrite();
922 auio.uio_rw = UIO_WRITE;
923 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
924 }
925 inblock_end = td->td_ru.ru_inblock;
926 oublock_end = td->td_ru.ru_oublock;
927
928 aiocbe->inputcharge = inblock_end - inblock_st;
929 aiocbe->outputcharge = oublock_end - oublock_st;
930
931 if ((error) && (auio.uio_resid != cnt)) {
932 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
933 error = 0;
934 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
935 int sigpipe = 1;
936 if (fp->f_type == DTYPE_SOCKET) {
937 so = fp->f_data;
938 if (so->so_options & SO_NOSIGPIPE)
939 sigpipe = 0;
940 }
941 if (sigpipe) {
942 PROC_LOCK(aiocbe->userproc);
943 kern_psignal(aiocbe->userproc, SIGPIPE);
944 PROC_UNLOCK(aiocbe->userproc);
945 }
946 }
947 }
948
949 cnt -= auio.uio_resid;
950 cb->_aiocb_private.error = error;
951 cb->_aiocb_private.status = cnt;
952 td->td_ucred = td_savedcred;
953 }
954
955 static void
956 aio_process_sync(struct aiocblist *aiocbe)
957 {
958 struct thread *td = curthread;
959 struct ucred *td_savedcred = td->td_ucred;
960 struct aiocb *cb = &aiocbe->uaiocb;
961 struct file *fp = aiocbe->fd_file;
962 int error = 0;
963
964 KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_SYNC,
965 ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
966
967 td->td_ucred = aiocbe->cred;
968 if (fp->f_vnode != NULL)
969 error = aio_fsync_vnode(td, fp->f_vnode);
970 cb->_aiocb_private.error = error;
971 cb->_aiocb_private.status = 0;
972 td->td_ucred = td_savedcred;
973 }
974
975 static void
976 aio_process_mlock(struct aiocblist *aiocbe)
977 {
978 struct aiocb *cb = &aiocbe->uaiocb;
979 int error;
980
981 KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_MLOCK,
982 ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
983
984 error = vm_mlock(aiocbe->userproc, aiocbe->cred,
985 __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
986 cb->_aiocb_private.error = error;
987 cb->_aiocb_private.status = 0;
988 }
989
990 static void
991 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
992 {
993 struct aioliojob *lj;
994 struct kaioinfo *ki;
995 struct aiocblist *scb, *scbn;
996 int lj_done;
997
998 ki = userp->p_aioinfo;
999 AIO_LOCK_ASSERT(ki, MA_OWNED);
1000 lj = aiocbe->lio;
1001 lj_done = 0;
1002 if (lj) {
1003 lj->lioj_finished_count++;
1004 if (lj->lioj_count == lj->lioj_finished_count)
1005 lj_done = 1;
1006 }
1007 if (type == DONE_QUEUE) {
1008 aiocbe->jobflags |= AIOCBLIST_DONE;
1009 } else {
1010 aiocbe->jobflags |= AIOCBLIST_BUFDONE;
1011 }
1012 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
1013 aiocbe->jobstate = JOBST_JOBFINISHED;
1014
1015 if (ki->kaio_flags & KAIO_RUNDOWN)
1016 goto notification_done;
1017
1018 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1019 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
1020 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
1021
1022 KNOTE_LOCKED(&aiocbe->klist, 1);
1023
1024 if (lj_done) {
1025 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
1026 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
1027 KNOTE_LOCKED(&lj->klist, 1);
1028 }
1029 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
1030 == LIOJ_SIGNAL
1031 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
1032 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
1033 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
1034 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1035 }
1036 }
1037
1038 notification_done:
1039 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1040 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1041 if (aiocbe->fd_file == scb->fd_file &&
1042 aiocbe->seqno < scb->seqno) {
1043 if (--scb->pending == 0) {
1044 mtx_lock(&aio_job_mtx);
1045 scb->jobstate = JOBST_JOBQGLOBAL;
1046 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1047 TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1048 aio_kick_nowait(userp);
1049 mtx_unlock(&aio_job_mtx);
1050 }
1051 }
1052 }
1053 }
1054 if (ki->kaio_flags & KAIO_WAKEUP) {
1055 ki->kaio_flags &= ~KAIO_WAKEUP;
1056 wakeup(&userp->p_aioinfo);
1057 }
1058 }
1059
1060 /*
1061 * The AIO daemon, most of the actual work is done in aio_process_*,
1062 * but the setup (and address space mgmt) is done in this routine.
1063 */
1064 static void
1065 aio_daemon(void *_id)
1066 {
1067 struct aiocblist *aiocbe;
1068 struct aiothreadlist *aiop;
1069 struct kaioinfo *ki;
1070 struct proc *curcp, *mycp, *userp;
1071 struct vmspace *myvm, *tmpvm;
1072 struct thread *td = curthread;
1073 int id = (intptr_t)_id;
1074
1075 /*
1076 * Local copies of curproc (cp) and vmspace (myvm)
1077 */
1078 mycp = td->td_proc;
1079 myvm = mycp->p_vmspace;
1080
1081 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1082
1083 /*
1084 * Allocate and ready the aio control info. There is one aiop structure
1085 * per daemon.
1086 */
1087 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1088 aiop->aiothread = td;
1089 aiop->aiothreadflags = 0;
1090
1091 /* The daemon resides in its own pgrp. */
1092 sys_setsid(td, NULL);
1093
1094 /*
1095 * Wakeup parent process. (Parent sleeps to keep from blasting away
1096 * and creating too many daemons.)
1097 */
1098 sema_post(&aio_newproc_sem);
1099
1100 mtx_lock(&aio_job_mtx);
1101 for (;;) {
1102 /*
1103 * curcp is the current daemon process context.
1104 * userp is the current user process context.
1105 */
1106 curcp = mycp;
1107
1108 /*
1109 * Take daemon off of free queue
1110 */
1111 if (aiop->aiothreadflags & AIOP_FREE) {
1112 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1113 aiop->aiothreadflags &= ~AIOP_FREE;
1114 }
1115
1116 /*
1117 * Check for jobs.
1118 */
1119 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1120 mtx_unlock(&aio_job_mtx);
1121 userp = aiocbe->userproc;
1122
1123 /*
1124 * Connect to process address space for user program.
1125 */
1126 if (userp != curcp) {
1127 /*
1128 * Save the current address space that we are
1129 * connected to.
1130 */
1131 tmpvm = mycp->p_vmspace;
1132
1133 /*
1134 * Point to the new user address space, and
1135 * refer to it.
1136 */
1137 mycp->p_vmspace = userp->p_vmspace;
1138 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1139
1140 /* Activate the new mapping. */
1141 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1142
1143 /*
1144 * If the old address space wasn't the daemons
1145 * own address space, then we need to remove the
1146 * daemon's reference from the other process
1147 * that it was acting on behalf of.
1148 */
1149 if (tmpvm != myvm) {
1150 vmspace_free(tmpvm);
1151 }
1152 curcp = userp;
1153 }
1154
1155 ki = userp->p_aioinfo;
1156
1157 /* Do the I/O function. */
1158 switch(aiocbe->uaiocb.aio_lio_opcode) {
1159 case LIO_READ:
1160 case LIO_WRITE:
1161 aio_process_rw(aiocbe);
1162 break;
1163 case LIO_SYNC:
1164 aio_process_sync(aiocbe);
1165 break;
1166 case LIO_MLOCK:
1167 aio_process_mlock(aiocbe);
1168 break;
1169 }
1170
1171 mtx_lock(&aio_job_mtx);
1172 /* Decrement the active job count. */
1173 ki->kaio_active_count--;
1174 mtx_unlock(&aio_job_mtx);
1175
1176 AIO_LOCK(ki);
1177 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1178 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1179 AIO_UNLOCK(ki);
1180
1181 mtx_lock(&aio_job_mtx);
1182 }
1183
1184 /*
1185 * Disconnect from user address space.
1186 */
1187 if (curcp != mycp) {
1188
1189 mtx_unlock(&aio_job_mtx);
1190
1191 /* Get the user address space to disconnect from. */
1192 tmpvm = mycp->p_vmspace;
1193
1194 /* Get original address space for daemon. */
1195 mycp->p_vmspace = myvm;
1196
1197 /* Activate the daemon's address space. */
1198 pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1199 #ifdef DIAGNOSTIC
1200 if (tmpvm == myvm) {
1201 printf("AIOD: vmspace problem -- %d\n",
1202 mycp->p_pid);
1203 }
1204 #endif
1205 /* Remove our vmspace reference. */
1206 vmspace_free(tmpvm);
1207
1208 curcp = mycp;
1209
1210 mtx_lock(&aio_job_mtx);
1211 /*
1212 * We have to restart to avoid race, we only sleep if
1213 * no job can be selected, that should be
1214 * curcp == mycp.
1215 */
1216 continue;
1217 }
1218
1219 mtx_assert(&aio_job_mtx, MA_OWNED);
1220
1221 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1222 aiop->aiothreadflags |= AIOP_FREE;
1223
1224 /*
1225 * If daemon is inactive for a long time, allow it to exit,
1226 * thereby freeing resources.
1227 */
1228 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1229 aiod_lifetime)) {
1230 if (TAILQ_EMPTY(&aio_jobs)) {
1231 if ((aiop->aiothreadflags & AIOP_FREE) &&
1232 (num_aio_procs > target_aio_procs)) {
1233 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1234 num_aio_procs--;
1235 mtx_unlock(&aio_job_mtx);
1236 uma_zfree(aiop_zone, aiop);
1237 free_unr(aiod_unr, id);
1238 #ifdef DIAGNOSTIC
1239 if (mycp->p_vmspace->vm_refcnt <= 1) {
1240 printf("AIOD: bad vm refcnt for"
1241 " exiting daemon: %d\n",
1242 mycp->p_vmspace->vm_refcnt);
1243 }
1244 #endif
1245 kproc_exit(0);
1246 }
1247 }
1248 }
1249 }
1250 mtx_unlock(&aio_job_mtx);
1251 panic("shouldn't be here\n");
1252 }
1253
1254 /*
1255 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1256 * AIO daemon modifies its environment itself.
1257 */
1258 static int
1259 aio_newproc(int *start)
1260 {
1261 int error;
1262 struct proc *p;
1263 int id;
1264
1265 id = alloc_unr(aiod_unr);
1266 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1267 RFNOWAIT, 0, "aiod%d", id);
1268 if (error == 0) {
1269 /*
1270 * Wait until daemon is started.
1271 */
1272 sema_wait(&aio_newproc_sem);
1273 mtx_lock(&aio_job_mtx);
1274 num_aio_procs++;
1275 if (start != NULL)
1276 (*start)--;
1277 mtx_unlock(&aio_job_mtx);
1278 } else {
1279 free_unr(aiod_unr, id);
1280 }
1281 return (error);
1282 }
1283
1284 /*
1285 * Try the high-performance, low-overhead physio method for eligible
1286 * VCHR devices. This method doesn't use an aio helper thread, and
1287 * thus has very low overhead.
1288 *
1289 * Assumes that the caller, aio_aqueue(), has incremented the file
1290 * structure's reference count, preventing its deallocation for the
1291 * duration of this call.
1292 */
1293 static int
1294 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1295 {
1296 struct aiocb *cb;
1297 struct file *fp;
1298 struct bio *bp;
1299 struct buf *pbuf;
1300 struct vnode *vp;
1301 struct cdevsw *csw;
1302 struct cdev *dev;
1303 struct kaioinfo *ki;
1304 struct aioliojob *lj;
1305 int error, ref, unmap, poff;
1306 vm_prot_t prot;
1307
1308 cb = &aiocbe->uaiocb;
1309 fp = aiocbe->fd_file;
1310
1311 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1312 return (-1);
1313
1314 vp = fp->f_vnode;
1315 if (vp->v_type != VCHR)
1316 return (-1);
1317 if (vp->v_bufobj.bo_bsize == 0)
1318 return (-1);
1319 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1320 return (-1);
1321
1322 ref = 0;
1323 csw = devvn_refthread(vp, &dev, &ref);
1324 if (csw == NULL)
1325 return (ENXIO);
1326
1327 if ((csw->d_flags & D_DISK) == 0) {
1328 error = -1;
1329 goto unref;
1330 }
1331 if (cb->aio_nbytes > dev->si_iosize_max) {
1332 error = -1;
1333 goto unref;
1334 }
1335
1336 ki = p->p_aioinfo;
1337 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1338 unmap = ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed);
1339 if (unmap) {
1340 if (cb->aio_nbytes > MAXPHYS) {
1341 error = -1;
1342 goto unref;
1343 }
1344 } else {
1345 if (cb->aio_nbytes > MAXPHYS - poff) {
1346 error = -1;
1347 goto unref;
1348 }
1349 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1350 error = -1;
1351 goto unref;
1352 }
1353 }
1354 aiocbe->bp = bp = g_alloc_bio();
1355 if (!unmap) {
1356 aiocbe->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1357 BUF_KERNPROC(pbuf);
1358 } else
1359 pbuf = NULL;
1360
1361 AIO_LOCK(ki);
1362 ki->kaio_count++;
1363 if (!unmap)
1364 ki->kaio_buffer_count++;
1365 lj = aiocbe->lio;
1366 if (lj)
1367 lj->lioj_count++;
1368 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1369 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1370 aiocbe->jobstate = JOBST_JOBQBUF;
1371 cb->_aiocb_private.status = cb->aio_nbytes;
1372 AIO_UNLOCK(ki);
1373
1374 bp->bio_length = cb->aio_nbytes;
1375 bp->bio_bcount = cb->aio_nbytes;
1376 bp->bio_done = aio_physwakeup;
1377 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1378 bp->bio_offset = cb->aio_offset;
1379 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1380 bp->bio_dev = dev;
1381 bp->bio_caller1 = (void *)aiocbe;
1382
1383 prot = VM_PROT_READ;
1384 if (cb->aio_lio_opcode == LIO_READ)
1385 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1386 if ((aiocbe->npages = vm_fault_quick_hold_pages(
1387 &curproc->p_vmspace->vm_map,
1388 (vm_offset_t)bp->bio_data, bp->bio_length, prot, aiocbe->pages,
1389 sizeof(aiocbe->pages)/sizeof(aiocbe->pages[0]))) < 0) {
1390 error = EFAULT;
1391 goto doerror;
1392 }
1393 if (!unmap) {
1394 pmap_qenter((vm_offset_t)pbuf->b_data,
1395 aiocbe->pages, aiocbe->npages);
1396 bp->bio_data = pbuf->b_data + poff;
1397 } else {
1398 bp->bio_ma = aiocbe->pages;
1399 bp->bio_ma_n = aiocbe->npages;
1400 bp->bio_ma_offset = poff;
1401 bp->bio_data = unmapped_buf;
1402 bp->bio_flags |= BIO_UNMAPPED;
1403 }
1404
1405 atomic_add_int(&num_queue_count, 1);
1406 if (!unmap)
1407 atomic_add_int(&num_buf_aio, 1);
1408
1409 /* Perform transfer. */
1410 csw->d_strategy(bp);
1411 dev_relthread(dev, ref);
1412 return (0);
1413
1414 doerror:
1415 AIO_LOCK(ki);
1416 aiocbe->jobstate = JOBST_NULL;
1417 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1418 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
1419 ki->kaio_count--;
1420 if (!unmap)
1421 ki->kaio_buffer_count--;
1422 if (lj)
1423 lj->lioj_count--;
1424 AIO_UNLOCK(ki);
1425 if (pbuf) {
1426 relpbuf(pbuf, NULL);
1427 aiocbe->pbuf = NULL;
1428 }
1429 g_destroy_bio(bp);
1430 aiocbe->bp = NULL;
1431 unref:
1432 dev_relthread(dev, ref);
1433 return (error);
1434 }
1435
1436 /*
1437 * Wake up aio requests that may be serviceable now.
1438 */
1439 static void
1440 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1441 {
1442 struct aiocblist *cb, *cbn;
1443 int opcode;
1444
1445 SOCKBUF_LOCK_ASSERT(sb);
1446 if (sb == &so->so_snd)
1447 opcode = LIO_WRITE;
1448 else
1449 opcode = LIO_READ;
1450
1451 sb->sb_flags &= ~SB_AIO;
1452 mtx_lock(&aio_job_mtx);
1453 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1454 if (opcode == cb->uaiocb.aio_lio_opcode) {
1455 if (cb->jobstate != JOBST_JOBQSOCK)
1456 panic("invalid queue value");
1457 /* XXX
1458 * We don't have actual sockets backend yet,
1459 * so we simply move the requests to the generic
1460 * file I/O backend.
1461 */
1462 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1463 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1464 aio_kick_nowait(cb->userproc);
1465 }
1466 }
1467 mtx_unlock(&aio_job_mtx);
1468 }
1469
1470 static int
1471 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1472 {
1473
1474 /*
1475 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1476 * supported by AIO with the old sigevent structure.
1477 */
1478 nsig->sigev_notify = osig->sigev_notify;
1479 switch (nsig->sigev_notify) {
1480 case SIGEV_NONE:
1481 break;
1482 case SIGEV_SIGNAL:
1483 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1484 break;
1485 case SIGEV_KEVENT:
1486 nsig->sigev_notify_kqueue =
1487 osig->__sigev_u.__sigev_notify_kqueue;
1488 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1489 break;
1490 default:
1491 return (EINVAL);
1492 }
1493 return (0);
1494 }
1495
1496 static int
1497 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1498 {
1499 struct oaiocb *ojob;
1500 int error;
1501
1502 bzero(kjob, sizeof(struct aiocb));
1503 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1504 if (error)
1505 return (error);
1506 ojob = (struct oaiocb *)kjob;
1507 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1508 }
1509
1510 static int
1511 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1512 {
1513
1514 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1515 }
1516
1517 static long
1518 aiocb_fetch_status(struct aiocb *ujob)
1519 {
1520
1521 return (fuword(&ujob->_aiocb_private.status));
1522 }
1523
1524 static long
1525 aiocb_fetch_error(struct aiocb *ujob)
1526 {
1527
1528 return (fuword(&ujob->_aiocb_private.error));
1529 }
1530
1531 static int
1532 aiocb_store_status(struct aiocb *ujob, long status)
1533 {
1534
1535 return (suword(&ujob->_aiocb_private.status, status));
1536 }
1537
1538 static int
1539 aiocb_store_error(struct aiocb *ujob, long error)
1540 {
1541
1542 return (suword(&ujob->_aiocb_private.error, error));
1543 }
1544
1545 static int
1546 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1547 {
1548
1549 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1550 }
1551
1552 static int
1553 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1554 {
1555
1556 return (suword(ujobp, (long)ujob));
1557 }
1558
1559 static struct aiocb_ops aiocb_ops = {
1560 .copyin = aiocb_copyin,
1561 .fetch_status = aiocb_fetch_status,
1562 .fetch_error = aiocb_fetch_error,
1563 .store_status = aiocb_store_status,
1564 .store_error = aiocb_store_error,
1565 .store_kernelinfo = aiocb_store_kernelinfo,
1566 .store_aiocb = aiocb_store_aiocb,
1567 };
1568
1569 static struct aiocb_ops aiocb_ops_osigevent = {
1570 .copyin = aiocb_copyin_old_sigevent,
1571 .fetch_status = aiocb_fetch_status,
1572 .fetch_error = aiocb_fetch_error,
1573 .store_status = aiocb_store_status,
1574 .store_error = aiocb_store_error,
1575 .store_kernelinfo = aiocb_store_kernelinfo,
1576 .store_aiocb = aiocb_store_aiocb,
1577 };
1578
1579 /*
1580 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1581 * technique is done in this code.
1582 */
1583 int
1584 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1585 int type, struct aiocb_ops *ops)
1586 {
1587 struct proc *p = td->td_proc;
1588 cap_rights_t rights;
1589 struct file *fp;
1590 struct socket *so;
1591 struct aiocblist *aiocbe, *cb;
1592 struct kaioinfo *ki;
1593 struct kevent kev;
1594 struct sockbuf *sb;
1595 int opcode;
1596 int error;
1597 int fd, kqfd;
1598 int jid;
1599 u_short evflags;
1600
1601 if (p->p_aioinfo == NULL)
1602 aio_init_aioinfo(p);
1603
1604 ki = p->p_aioinfo;
1605
1606 ops->store_status(job, -1);
1607 ops->store_error(job, 0);
1608 ops->store_kernelinfo(job, -1);
1609
1610 if (num_queue_count >= max_queue_count ||
1611 ki->kaio_count >= ki->kaio_qallowed_count) {
1612 ops->store_error(job, EAGAIN);
1613 return (EAGAIN);
1614 }
1615
1616 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1617 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1618
1619 error = ops->copyin(job, &aiocbe->uaiocb);
1620 if (error) {
1621 ops->store_error(job, error);
1622 uma_zfree(aiocb_zone, aiocbe);
1623 return (error);
1624 }
1625
1626 /* XXX: aio_nbytes is later casted to signed types. */
1627 if (aiocbe->uaiocb.aio_nbytes > INT_MAX) {
1628 uma_zfree(aiocb_zone, aiocbe);
1629 return (EINVAL);
1630 }
1631
1632 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1633 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1634 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1635 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1636 ops->store_error(job, EINVAL);
1637 uma_zfree(aiocb_zone, aiocbe);
1638 return (EINVAL);
1639 }
1640
1641 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1642 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1643 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1644 uma_zfree(aiocb_zone, aiocbe);
1645 return (EINVAL);
1646 }
1647
1648 ksiginfo_init(&aiocbe->ksi);
1649
1650 /* Save userspace address of the job info. */
1651 aiocbe->uuaiocb = job;
1652
1653 /* Get the opcode. */
1654 if (type != LIO_NOP)
1655 aiocbe->uaiocb.aio_lio_opcode = type;
1656 opcode = aiocbe->uaiocb.aio_lio_opcode;
1657
1658 /*
1659 * Validate the opcode and fetch the file object for the specified
1660 * file descriptor.
1661 *
1662 * XXXRW: Moved the opcode validation up here so that we don't
1663 * retrieve a file descriptor without knowing what the capabiltity
1664 * should be.
1665 */
1666 fd = aiocbe->uaiocb.aio_fildes;
1667 switch (opcode) {
1668 case LIO_WRITE:
1669 error = fget_write(td, fd,
1670 cap_rights_init(&rights, CAP_PWRITE), &fp);
1671 break;
1672 case LIO_READ:
1673 error = fget_read(td, fd,
1674 cap_rights_init(&rights, CAP_PREAD), &fp);
1675 break;
1676 case LIO_SYNC:
1677 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1678 break;
1679 case LIO_MLOCK:
1680 fp = NULL;
1681 break;
1682 case LIO_NOP:
1683 error = fget(td, fd, cap_rights_init(&rights), &fp);
1684 break;
1685 default:
1686 error = EINVAL;
1687 }
1688 if (error) {
1689 uma_zfree(aiocb_zone, aiocbe);
1690 ops->store_error(job, error);
1691 return (error);
1692 }
1693
1694 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1695 error = EINVAL;
1696 goto aqueue_fail;
1697 }
1698
1699 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1700 error = EINVAL;
1701 goto aqueue_fail;
1702 }
1703
1704 aiocbe->fd_file = fp;
1705
1706 mtx_lock(&aio_job_mtx);
1707 jid = jobrefid++;
1708 aiocbe->seqno = jobseqno++;
1709 mtx_unlock(&aio_job_mtx);
1710 error = ops->store_kernelinfo(job, jid);
1711 if (error) {
1712 error = EINVAL;
1713 goto aqueue_fail;
1714 }
1715 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1716
1717 if (opcode == LIO_NOP) {
1718 fdrop(fp, td);
1719 uma_zfree(aiocb_zone, aiocbe);
1720 return (0);
1721 }
1722
1723 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1724 goto no_kqueue;
1725 evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1726 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1727 error = EINVAL;
1728 goto aqueue_fail;
1729 }
1730 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1731 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1732 kev.filter = EVFILT_AIO;
1733 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1734 kev.data = (intptr_t)aiocbe;
1735 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1736 error = kqfd_register(kqfd, &kev, td, 1);
1737 aqueue_fail:
1738 if (error) {
1739 if (fp)
1740 fdrop(fp, td);
1741 uma_zfree(aiocb_zone, aiocbe);
1742 ops->store_error(job, error);
1743 goto done;
1744 }
1745 no_kqueue:
1746
1747 ops->store_error(job, EINPROGRESS);
1748 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1749 aiocbe->userproc = p;
1750 aiocbe->cred = crhold(td->td_ucred);
1751 aiocbe->jobflags = 0;
1752 aiocbe->lio = lj;
1753
1754 if (opcode == LIO_SYNC)
1755 goto queueit;
1756
1757 if (fp && fp->f_type == DTYPE_SOCKET) {
1758 /*
1759 * Alternate queueing for socket ops: Reach down into the
1760 * descriptor to get the socket data. Then check to see if the
1761 * socket is ready to be read or written (based on the requested
1762 * operation).
1763 *
1764 * If it is not ready for io, then queue the aiocbe on the
1765 * socket, and set the flags so we get a call when sbnotify()
1766 * happens.
1767 *
1768 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1769 * and unlock the snd sockbuf for no reason.
1770 */
1771 so = fp->f_data;
1772 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1773 SOCKBUF_LOCK(sb);
1774 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1775 LIO_WRITE) && (!sowriteable(so)))) {
1776 sb->sb_flags |= SB_AIO;
1777
1778 mtx_lock(&aio_job_mtx);
1779 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1780 mtx_unlock(&aio_job_mtx);
1781
1782 AIO_LOCK(ki);
1783 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1784 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1785 aiocbe->jobstate = JOBST_JOBQSOCK;
1786 ki->kaio_count++;
1787 if (lj)
1788 lj->lioj_count++;
1789 AIO_UNLOCK(ki);
1790 SOCKBUF_UNLOCK(sb);
1791 atomic_add_int(&num_queue_count, 1);
1792 error = 0;
1793 goto done;
1794 }
1795 SOCKBUF_UNLOCK(sb);
1796 }
1797
1798 if ((error = aio_qphysio(p, aiocbe)) == 0)
1799 goto done;
1800 #if 0
1801 if (error > 0) {
1802 aiocbe->uaiocb._aiocb_private.error = error;
1803 ops->store_error(job, error);
1804 goto done;
1805 }
1806 #endif
1807 queueit:
1808 atomic_add_int(&num_queue_count, 1);
1809
1810 AIO_LOCK(ki);
1811 ki->kaio_count++;
1812 if (lj)
1813 lj->lioj_count++;
1814 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1815 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1816 if (opcode == LIO_SYNC) {
1817 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1818 if (cb->fd_file == aiocbe->fd_file &&
1819 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1820 cb->seqno < aiocbe->seqno) {
1821 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1822 aiocbe->pending++;
1823 }
1824 }
1825 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1826 if (cb->fd_file == aiocbe->fd_file &&
1827 cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1828 cb->seqno < aiocbe->seqno) {
1829 cb->jobflags |= AIOCBLIST_CHECKSYNC;
1830 aiocbe->pending++;
1831 }
1832 }
1833 if (aiocbe->pending != 0) {
1834 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1835 aiocbe->jobstate = JOBST_JOBQSYNC;
1836 AIO_UNLOCK(ki);
1837 goto done;
1838 }
1839 }
1840 mtx_lock(&aio_job_mtx);
1841 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1842 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1843 aio_kick_nowait(p);
1844 mtx_unlock(&aio_job_mtx);
1845 AIO_UNLOCK(ki);
1846 error = 0;
1847 done:
1848 return (error);
1849 }
1850
1851 static void
1852 aio_kick_nowait(struct proc *userp)
1853 {
1854 struct kaioinfo *ki = userp->p_aioinfo;
1855 struct aiothreadlist *aiop;
1856
1857 mtx_assert(&aio_job_mtx, MA_OWNED);
1858 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1859 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1860 aiop->aiothreadflags &= ~AIOP_FREE;
1861 wakeup(aiop->aiothread);
1862 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1863 ((ki->kaio_active_count + num_aio_resv_start) <
1864 ki->kaio_maxactive_count)) {
1865 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1866 }
1867 }
1868
1869 static int
1870 aio_kick(struct proc *userp)
1871 {
1872 struct kaioinfo *ki = userp->p_aioinfo;
1873 struct aiothreadlist *aiop;
1874 int error, ret = 0;
1875
1876 mtx_assert(&aio_job_mtx, MA_OWNED);
1877 retryproc:
1878 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1879 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1880 aiop->aiothreadflags &= ~AIOP_FREE;
1881 wakeup(aiop->aiothread);
1882 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1883 ((ki->kaio_active_count + num_aio_resv_start) <
1884 ki->kaio_maxactive_count)) {
1885 num_aio_resv_start++;
1886 mtx_unlock(&aio_job_mtx);
1887 error = aio_newproc(&num_aio_resv_start);
1888 mtx_lock(&aio_job_mtx);
1889 if (error) {
1890 num_aio_resv_start--;
1891 goto retryproc;
1892 }
1893 } else {
1894 ret = -1;
1895 }
1896 return (ret);
1897 }
1898
1899 static void
1900 aio_kick_helper(void *context, int pending)
1901 {
1902 struct proc *userp = context;
1903
1904 mtx_lock(&aio_job_mtx);
1905 while (--pending >= 0) {
1906 if (aio_kick(userp))
1907 break;
1908 }
1909 mtx_unlock(&aio_job_mtx);
1910 }
1911
1912 /*
1913 * Support the aio_return system call, as a side-effect, kernel resources are
1914 * released.
1915 */
1916 static int
1917 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1918 {
1919 struct proc *p = td->td_proc;
1920 struct aiocblist *cb;
1921 struct kaioinfo *ki;
1922 int status, error;
1923
1924 ki = p->p_aioinfo;
1925 if (ki == NULL)
1926 return (EINVAL);
1927 AIO_LOCK(ki);
1928 TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1929 if (cb->uuaiocb == uaiocb)
1930 break;
1931 }
1932 if (cb != NULL) {
1933 MPASS(cb->jobstate == JOBST_JOBFINISHED);
1934 status = cb->uaiocb._aiocb_private.status;
1935 error = cb->uaiocb._aiocb_private.error;
1936 td->td_retval[0] = status;
1937 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1938 td->td_ru.ru_oublock += cb->outputcharge;
1939 cb->outputcharge = 0;
1940 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1941 td->td_ru.ru_inblock += cb->inputcharge;
1942 cb->inputcharge = 0;
1943 }
1944 aio_free_entry(cb);
1945 AIO_UNLOCK(ki);
1946 ops->store_error(uaiocb, error);
1947 ops->store_status(uaiocb, status);
1948 } else {
1949 error = EINVAL;
1950 AIO_UNLOCK(ki);
1951 }
1952 return (error);
1953 }
1954
1955 int
1956 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1957 {
1958
1959 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1960 }
1961
1962 /*
1963 * Allow a process to wakeup when any of the I/O requests are completed.
1964 */
1965 static int
1966 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1967 struct timespec *ts)
1968 {
1969 struct proc *p = td->td_proc;
1970 struct timeval atv;
1971 struct kaioinfo *ki;
1972 struct aiocblist *cb, *cbfirst;
1973 int error, i, timo;
1974
1975 timo = 0;
1976 if (ts) {
1977 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1978 return (EINVAL);
1979
1980 TIMESPEC_TO_TIMEVAL(&atv, ts);
1981 if (itimerfix(&atv))
1982 return (EINVAL);
1983 timo = tvtohz(&atv);
1984 }
1985
1986 ki = p->p_aioinfo;
1987 if (ki == NULL)
1988 return (EAGAIN);
1989
1990 if (njoblist == 0)
1991 return (0);
1992
1993 AIO_LOCK(ki);
1994 for (;;) {
1995 cbfirst = NULL;
1996 error = 0;
1997 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1998 for (i = 0; i < njoblist; i++) {
1999 if (cb->uuaiocb == ujoblist[i]) {
2000 if (cbfirst == NULL)
2001 cbfirst = cb;
2002 if (cb->jobstate == JOBST_JOBFINISHED)
2003 goto RETURN;
2004 }
2005 }
2006 }
2007 /* All tasks were finished. */
2008 if (cbfirst == NULL)
2009 break;
2010
2011 ki->kaio_flags |= KAIO_WAKEUP;
2012 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2013 "aiospn", timo);
2014 if (error == ERESTART)
2015 error = EINTR;
2016 if (error)
2017 break;
2018 }
2019 RETURN:
2020 AIO_UNLOCK(ki);
2021 return (error);
2022 }
2023
2024 int
2025 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2026 {
2027 struct timespec ts, *tsp;
2028 struct aiocb **ujoblist;
2029 int error;
2030
2031 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2032 return (EINVAL);
2033
2034 if (uap->timeout) {
2035 /* Get timespec struct. */
2036 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2037 return (error);
2038 tsp = &ts;
2039 } else
2040 tsp = NULL;
2041
2042 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2043 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2044 if (error == 0)
2045 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2046 uma_zfree(aiol_zone, ujoblist);
2047 return (error);
2048 }
2049
2050 /*
2051 * aio_cancel cancels any non-physio aio operations not currently in
2052 * progress.
2053 */
2054 int
2055 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2056 {
2057 struct proc *p = td->td_proc;
2058 struct kaioinfo *ki;
2059 struct aiocblist *cbe, *cbn;
2060 struct file *fp;
2061 struct socket *so;
2062 int error;
2063 int remove;
2064 int cancelled = 0;
2065 int notcancelled = 0;
2066 struct vnode *vp;
2067
2068 /* Lookup file object. */
2069 error = fget(td, uap->fd, NULL, &fp);
2070 if (error)
2071 return (error);
2072
2073 ki = p->p_aioinfo;
2074 if (ki == NULL)
2075 goto done;
2076
2077 if (fp->f_type == DTYPE_VNODE) {
2078 vp = fp->f_vnode;
2079 if (vn_isdisk(vp, &error)) {
2080 fdrop(fp, td);
2081 td->td_retval[0] = AIO_NOTCANCELED;
2082 return (0);
2083 }
2084 }
2085
2086 AIO_LOCK(ki);
2087 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
2088 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
2089 ((uap->aiocbp == NULL) ||
2090 (uap->aiocbp == cbe->uuaiocb))) {
2091 remove = 0;
2092
2093 mtx_lock(&aio_job_mtx);
2094 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
2095 TAILQ_REMOVE(&aio_jobs, cbe, list);
2096 remove = 1;
2097 } else if (cbe->jobstate == JOBST_JOBQSOCK) {
2098 MPASS(fp->f_type == DTYPE_SOCKET);
2099 so = fp->f_data;
2100 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2101 remove = 1;
2102 } else if (cbe->jobstate == JOBST_JOBQSYNC) {
2103 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2104 remove = 1;
2105 }
2106 mtx_unlock(&aio_job_mtx);
2107
2108 if (remove) {
2109 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2110 cbe->uaiocb._aiocb_private.status = -1;
2111 cbe->uaiocb._aiocb_private.error = ECANCELED;
2112 aio_bio_done_notify(p, cbe, DONE_QUEUE);
2113 cancelled++;
2114 } else {
2115 notcancelled++;
2116 }
2117 if (uap->aiocbp != NULL)
2118 break;
2119 }
2120 }
2121 AIO_UNLOCK(ki);
2122
2123 done:
2124 fdrop(fp, td);
2125
2126 if (uap->aiocbp != NULL) {
2127 if (cancelled) {
2128 td->td_retval[0] = AIO_CANCELED;
2129 return (0);
2130 }
2131 }
2132
2133 if (notcancelled) {
2134 td->td_retval[0] = AIO_NOTCANCELED;
2135 return (0);
2136 }
2137
2138 if (cancelled) {
2139 td->td_retval[0] = AIO_CANCELED;
2140 return (0);
2141 }
2142
2143 td->td_retval[0] = AIO_ALLDONE;
2144
2145 return (0);
2146 }
2147
2148 /*
2149 * aio_error is implemented in the kernel level for compatibility purposes
2150 * only. For a user mode async implementation, it would be best to do it in
2151 * a userland subroutine.
2152 */
2153 static int
2154 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2155 {
2156 struct proc *p = td->td_proc;
2157 struct aiocblist *cb;
2158 struct kaioinfo *ki;
2159 int status;
2160
2161 ki = p->p_aioinfo;
2162 if (ki == NULL) {
2163 td->td_retval[0] = EINVAL;
2164 return (0);
2165 }
2166
2167 AIO_LOCK(ki);
2168 TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2169 if (cb->uuaiocb == aiocbp) {
2170 if (cb->jobstate == JOBST_JOBFINISHED)
2171 td->td_retval[0] =
2172 cb->uaiocb._aiocb_private.error;
2173 else
2174 td->td_retval[0] = EINPROGRESS;
2175 AIO_UNLOCK(ki);
2176 return (0);
2177 }
2178 }
2179 AIO_UNLOCK(ki);
2180
2181 /*
2182 * Hack for failure of aio_aqueue.
2183 */
2184 status = ops->fetch_status(aiocbp);
2185 if (status == -1) {
2186 td->td_retval[0] = ops->fetch_error(aiocbp);
2187 return (0);
2188 }
2189
2190 td->td_retval[0] = EINVAL;
2191 return (0);
2192 }
2193
2194 int
2195 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2196 {
2197
2198 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2199 }
2200
2201 /* syscall - asynchronous read from a file (REALTIME) */
2202 int
2203 sys_oaio_read(struct thread *td, struct oaio_read_args *uap)
2204 {
2205
2206 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2207 &aiocb_ops_osigevent));
2208 }
2209
2210 int
2211 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2212 {
2213
2214 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2215 }
2216
2217 /* syscall - asynchronous write to a file (REALTIME) */
2218 int
2219 sys_oaio_write(struct thread *td, struct oaio_write_args *uap)
2220 {
2221
2222 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2223 &aiocb_ops_osigevent));
2224 }
2225
2226 int
2227 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2228 {
2229
2230 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2231 }
2232
2233 int
2234 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2235 {
2236
2237 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2238 }
2239
2240 static int
2241 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2242 struct aiocb **acb_list, int nent, struct sigevent *sig,
2243 struct aiocb_ops *ops)
2244 {
2245 struct proc *p = td->td_proc;
2246 struct aiocb *iocb;
2247 struct kaioinfo *ki;
2248 struct aioliojob *lj;
2249 struct kevent kev;
2250 int error;
2251 int nerror;
2252 int i;
2253
2254 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2255 return (EINVAL);
2256
2257 if (nent < 0 || nent > AIO_LISTIO_MAX)
2258 return (EINVAL);
2259
2260 if (p->p_aioinfo == NULL)
2261 aio_init_aioinfo(p);
2262
2263 ki = p->p_aioinfo;
2264
2265 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2266 lj->lioj_flags = 0;
2267 lj->lioj_count = 0;
2268 lj->lioj_finished_count = 0;
2269 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2270 ksiginfo_init(&lj->lioj_ksi);
2271
2272 /*
2273 * Setup signal.
2274 */
2275 if (sig && (mode == LIO_NOWAIT)) {
2276 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2277 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2278 /* Assume only new style KEVENT */
2279 kev.filter = EVFILT_LIO;
2280 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2281 kev.ident = (uintptr_t)uacb_list; /* something unique */
2282 kev.data = (intptr_t)lj;
2283 /* pass user defined sigval data */
2284 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2285 error = kqfd_register(
2286 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2287 if (error) {
2288 uma_zfree(aiolio_zone, lj);
2289 return (error);
2290 }
2291 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2292 ;
2293 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2294 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2295 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2296 uma_zfree(aiolio_zone, lj);
2297 return EINVAL;
2298 }
2299 lj->lioj_flags |= LIOJ_SIGNAL;
2300 } else {
2301 uma_zfree(aiolio_zone, lj);
2302 return EINVAL;
2303 }
2304 }
2305
2306 AIO_LOCK(ki);
2307 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2308 /*
2309 * Add extra aiocb count to avoid the lio to be freed
2310 * by other threads doing aio_waitcomplete or aio_return,
2311 * and prevent event from being sent until we have queued
2312 * all tasks.
2313 */
2314 lj->lioj_count = 1;
2315 AIO_UNLOCK(ki);
2316
2317 /*
2318 * Get pointers to the list of I/O requests.
2319 */
2320 nerror = 0;
2321 for (i = 0; i < nent; i++) {
2322 iocb = acb_list[i];
2323 if (iocb != NULL) {
2324 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2325 if (error != 0)
2326 nerror++;
2327 }
2328 }
2329
2330 error = 0;
2331 AIO_LOCK(ki);
2332 if (mode == LIO_WAIT) {
2333 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2334 ki->kaio_flags |= KAIO_WAKEUP;
2335 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2336 PRIBIO | PCATCH, "aiospn", 0);
2337 if (error == ERESTART)
2338 error = EINTR;
2339 if (error)
2340 break;
2341 }
2342 } else {
2343 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2344 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2345 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2346 KNOTE_LOCKED(&lj->klist, 1);
2347 }
2348 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2349 == LIOJ_SIGNAL
2350 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2351 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2352 aio_sendsig(p, &lj->lioj_signal,
2353 &lj->lioj_ksi);
2354 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2355 }
2356 }
2357 }
2358 lj->lioj_count--;
2359 if (lj->lioj_count == 0) {
2360 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2361 knlist_delete(&lj->klist, curthread, 1);
2362 PROC_LOCK(p);
2363 sigqueue_take(&lj->lioj_ksi);
2364 PROC_UNLOCK(p);
2365 AIO_UNLOCK(ki);
2366 uma_zfree(aiolio_zone, lj);
2367 } else
2368 AIO_UNLOCK(ki);
2369
2370 if (nerror)
2371 return (EIO);
2372 return (error);
2373 }
2374
2375 /* syscall - list directed I/O (REALTIME) */
2376 int
2377 sys_olio_listio(struct thread *td, struct olio_listio_args *uap)
2378 {
2379 struct aiocb **acb_list;
2380 struct sigevent *sigp, sig;
2381 struct osigevent osig;
2382 int error, nent;
2383
2384 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2385 return (EINVAL);
2386
2387 nent = uap->nent;
2388 if (nent < 0 || nent > AIO_LISTIO_MAX)
2389 return (EINVAL);
2390
2391 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2392 error = copyin(uap->sig, &osig, sizeof(osig));
2393 if (error)
2394 return (error);
2395 error = convert_old_sigevent(&osig, &sig);
2396 if (error)
2397 return (error);
2398 sigp = &sig;
2399 } else
2400 sigp = NULL;
2401
2402 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2403 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2404 if (error == 0)
2405 error = kern_lio_listio(td, uap->mode,
2406 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2407 &aiocb_ops_osigevent);
2408 free(acb_list, M_LIO);
2409 return (error);
2410 }
2411
2412 /* syscall - list directed I/O (REALTIME) */
2413 int
2414 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2415 {
2416 struct aiocb **acb_list;
2417 struct sigevent *sigp, sig;
2418 int error, nent;
2419
2420 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2421 return (EINVAL);
2422
2423 nent = uap->nent;
2424 if (nent < 0 || nent > AIO_LISTIO_MAX)
2425 return (EINVAL);
2426
2427 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2428 error = copyin(uap->sig, &sig, sizeof(sig));
2429 if (error)
2430 return (error);
2431 sigp = &sig;
2432 } else
2433 sigp = NULL;
2434
2435 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2436 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2437 if (error == 0)
2438 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2439 nent, sigp, &aiocb_ops);
2440 free(acb_list, M_LIO);
2441 return (error);
2442 }
2443
2444 static void
2445 aio_physwakeup(struct bio *bp)
2446 {
2447 struct aiocblist *aiocbe = (struct aiocblist *)bp->bio_caller1;
2448 struct proc *userp;
2449 struct kaioinfo *ki;
2450 int nblks;
2451
2452 /* Release mapping into kernel space. */
2453 if (aiocbe->pbuf) {
2454 pmap_qremove((vm_offset_t)aiocbe->pbuf->b_data, aiocbe->npages);
2455 relpbuf(aiocbe->pbuf, NULL);
2456 aiocbe->pbuf = NULL;
2457 atomic_subtract_int(&num_buf_aio, 1);
2458 }
2459 vm_page_unhold_pages(aiocbe->pages, aiocbe->npages);
2460
2461 bp = aiocbe->bp;
2462 aiocbe->bp = NULL;
2463 userp = aiocbe->userproc;
2464 ki = userp->p_aioinfo;
2465 AIO_LOCK(ki);
2466 aiocbe->uaiocb._aiocb_private.status -= bp->bio_resid;
2467 aiocbe->uaiocb._aiocb_private.error = 0;
2468 if (bp->bio_flags & BIO_ERROR)
2469 aiocbe->uaiocb._aiocb_private.error = bp->bio_error;
2470 nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2471 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2472 aiocbe->outputcharge += nblks;
2473 else
2474 aiocbe->inputcharge += nblks;
2475 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2476 ki->kaio_buffer_count--;
2477 aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2478 AIO_UNLOCK(ki);
2479
2480 g_destroy_bio(bp);
2481 }
2482
2483 /* syscall - wait for the next completion of an aio request */
2484 static int
2485 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2486 struct timespec *ts, struct aiocb_ops *ops)
2487 {
2488 struct proc *p = td->td_proc;
2489 struct timeval atv;
2490 struct kaioinfo *ki;
2491 struct aiocblist *cb;
2492 struct aiocb *uuaiocb;
2493 int error, status, timo;
2494
2495 ops->store_aiocb(aiocbp, NULL);
2496
2497 timo = 0;
2498 if (ts) {
2499 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2500 return (EINVAL);
2501
2502 TIMESPEC_TO_TIMEVAL(&atv, ts);
2503 if (itimerfix(&atv))
2504 return (EINVAL);
2505 timo = tvtohz(&atv);
2506 }
2507
2508 if (p->p_aioinfo == NULL)
2509 aio_init_aioinfo(p);
2510 ki = p->p_aioinfo;
2511
2512 error = 0;
2513 cb = NULL;
2514 AIO_LOCK(ki);
2515 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2516 ki->kaio_flags |= KAIO_WAKEUP;
2517 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2518 "aiowc", timo);
2519 if (timo && error == ERESTART)
2520 error = EINTR;
2521 if (error)
2522 break;
2523 }
2524
2525 if (cb != NULL) {
2526 MPASS(cb->jobstate == JOBST_JOBFINISHED);
2527 uuaiocb = cb->uuaiocb;
2528 status = cb->uaiocb._aiocb_private.status;
2529 error = cb->uaiocb._aiocb_private.error;
2530 td->td_retval[0] = status;
2531 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2532 td->td_ru.ru_oublock += cb->outputcharge;
2533 cb->outputcharge = 0;
2534 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2535 td->td_ru.ru_inblock += cb->inputcharge;
2536 cb->inputcharge = 0;
2537 }
2538 aio_free_entry(cb);
2539 AIO_UNLOCK(ki);
2540 ops->store_aiocb(aiocbp, uuaiocb);
2541 ops->store_error(uuaiocb, error);
2542 ops->store_status(uuaiocb, status);
2543 } else
2544 AIO_UNLOCK(ki);
2545
2546 return (error);
2547 }
2548
2549 int
2550 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2551 {
2552 struct timespec ts, *tsp;
2553 int error;
2554
2555 if (uap->timeout) {
2556 /* Get timespec struct. */
2557 error = copyin(uap->timeout, &ts, sizeof(ts));
2558 if (error)
2559 return (error);
2560 tsp = &ts;
2561 } else
2562 tsp = NULL;
2563
2564 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2565 }
2566
2567 static int
2568 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2569 struct aiocb_ops *ops)
2570 {
2571
2572 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2573 return (EINVAL);
2574 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2575 }
2576
2577 int
2578 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2579 {
2580
2581 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2582 }
2583
2584 /* kqueue attach function */
2585 static int
2586 filt_aioattach(struct knote *kn)
2587 {
2588 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2589
2590 /*
2591 * The aiocbe pointer must be validated before using it, so
2592 * registration is restricted to the kernel; the user cannot
2593 * set EV_FLAG1.
2594 */
2595 if ((kn->kn_flags & EV_FLAG1) == 0)
2596 return (EPERM);
2597 kn->kn_ptr.p_aio = aiocbe;
2598 kn->kn_flags &= ~EV_FLAG1;
2599
2600 knlist_add(&aiocbe->klist, kn, 0);
2601
2602 return (0);
2603 }
2604
2605 /* kqueue detach function */
2606 static void
2607 filt_aiodetach(struct knote *kn)
2608 {
2609 struct knlist *knl;
2610
2611 knl = &kn->kn_ptr.p_aio->klist;
2612 knl->kl_lock(knl->kl_lockarg);
2613 if (!knlist_empty(knl))
2614 knlist_remove(knl, kn, 1);
2615 knl->kl_unlock(knl->kl_lockarg);
2616 }
2617
2618 /* kqueue filter function */
2619 /*ARGSUSED*/
2620 static int
2621 filt_aio(struct knote *kn, long hint)
2622 {
2623 struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2624
2625 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2626 if (aiocbe->jobstate != JOBST_JOBFINISHED)
2627 return (0);
2628 kn->kn_flags |= EV_EOF;
2629 return (1);
2630 }
2631
2632 /* kqueue attach function */
2633 static int
2634 filt_lioattach(struct knote *kn)
2635 {
2636 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2637
2638 /*
2639 * The aioliojob pointer must be validated before using it, so
2640 * registration is restricted to the kernel; the user cannot
2641 * set EV_FLAG1.
2642 */
2643 if ((kn->kn_flags & EV_FLAG1) == 0)
2644 return (EPERM);
2645 kn->kn_ptr.p_lio = lj;
2646 kn->kn_flags &= ~EV_FLAG1;
2647
2648 knlist_add(&lj->klist, kn, 0);
2649
2650 return (0);
2651 }
2652
2653 /* kqueue detach function */
2654 static void
2655 filt_liodetach(struct knote *kn)
2656 {
2657 struct knlist *knl;
2658
2659 knl = &kn->kn_ptr.p_lio->klist;
2660 knl->kl_lock(knl->kl_lockarg);
2661 if (!knlist_empty(knl))
2662 knlist_remove(knl, kn, 1);
2663 knl->kl_unlock(knl->kl_lockarg);
2664 }
2665
2666 /* kqueue filter function */
2667 /*ARGSUSED*/
2668 static int
2669 filt_lio(struct knote *kn, long hint)
2670 {
2671 struct aioliojob * lj = kn->kn_ptr.p_lio;
2672
2673 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2674 }
2675
2676 #ifdef COMPAT_FREEBSD32
2677
2678 struct __aiocb_private32 {
2679 int32_t status;
2680 int32_t error;
2681 uint32_t kernelinfo;
2682 };
2683
2684 typedef struct oaiocb32 {
2685 int aio_fildes; /* File descriptor */
2686 uint64_t aio_offset __packed; /* File offset for I/O */
2687 uint32_t aio_buf; /* I/O buffer in process space */
2688 uint32_t aio_nbytes; /* Number of bytes for I/O */
2689 struct osigevent32 aio_sigevent; /* Signal to deliver */
2690 int aio_lio_opcode; /* LIO opcode */
2691 int aio_reqprio; /* Request priority -- ignored */
2692 struct __aiocb_private32 _aiocb_private;
2693 } oaiocb32_t;
2694
2695 typedef struct aiocb32 {
2696 int32_t aio_fildes; /* File descriptor */
2697 uint64_t aio_offset __packed; /* File offset for I/O */
2698 uint32_t aio_buf; /* I/O buffer in process space */
2699 uint32_t aio_nbytes; /* Number of bytes for I/O */
2700 int __spare__[2];
2701 uint32_t __spare2__;
2702 int aio_lio_opcode; /* LIO opcode */
2703 int aio_reqprio; /* Request priority -- ignored */
2704 struct __aiocb_private32 _aiocb_private;
2705 struct sigevent32 aio_sigevent; /* Signal to deliver */
2706 } aiocb32_t;
2707
2708 static int
2709 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2710 {
2711
2712 /*
2713 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2714 * supported by AIO with the old sigevent structure.
2715 */
2716 CP(*osig, *nsig, sigev_notify);
2717 switch (nsig->sigev_notify) {
2718 case SIGEV_NONE:
2719 break;
2720 case SIGEV_SIGNAL:
2721 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2722 break;
2723 case SIGEV_KEVENT:
2724 nsig->sigev_notify_kqueue =
2725 osig->__sigev_u.__sigev_notify_kqueue;
2726 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2727 break;
2728 default:
2729 return (EINVAL);
2730 }
2731 return (0);
2732 }
2733
2734 static int
2735 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2736 {
2737 struct oaiocb32 job32;
2738 int error;
2739
2740 bzero(kjob, sizeof(struct aiocb));
2741 error = copyin(ujob, &job32, sizeof(job32));
2742 if (error)
2743 return (error);
2744
2745 CP(job32, *kjob, aio_fildes);
2746 CP(job32, *kjob, aio_offset);
2747 PTRIN_CP(job32, *kjob, aio_buf);
2748 CP(job32, *kjob, aio_nbytes);
2749 CP(job32, *kjob, aio_lio_opcode);
2750 CP(job32, *kjob, aio_reqprio);
2751 CP(job32, *kjob, _aiocb_private.status);
2752 CP(job32, *kjob, _aiocb_private.error);
2753 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2754 return (convert_old_sigevent32(&job32.aio_sigevent,
2755 &kjob->aio_sigevent));
2756 }
2757
2758 static int
2759 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2760 {
2761 struct aiocb32 job32;
2762 int error;
2763
2764 error = copyin(ujob, &job32, sizeof(job32));
2765 if (error)
2766 return (error);
2767 CP(job32, *kjob, aio_fildes);
2768 CP(job32, *kjob, aio_offset);
2769 PTRIN_CP(job32, *kjob, aio_buf);
2770 CP(job32, *kjob, aio_nbytes);
2771 CP(job32, *kjob, aio_lio_opcode);
2772 CP(job32, *kjob, aio_reqprio);
2773 CP(job32, *kjob, _aiocb_private.status);
2774 CP(job32, *kjob, _aiocb_private.error);
2775 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2776 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2777 }
2778
2779 static long
2780 aiocb32_fetch_status(struct aiocb *ujob)
2781 {
2782 struct aiocb32 *ujob32;
2783
2784 ujob32 = (struct aiocb32 *)ujob;
2785 return (fuword32(&ujob32->_aiocb_private.status));
2786 }
2787
2788 static long
2789 aiocb32_fetch_error(struct aiocb *ujob)
2790 {
2791 struct aiocb32 *ujob32;
2792
2793 ujob32 = (struct aiocb32 *)ujob;
2794 return (fuword32(&ujob32->_aiocb_private.error));
2795 }
2796
2797 static int
2798 aiocb32_store_status(struct aiocb *ujob, long status)
2799 {
2800 struct aiocb32 *ujob32;
2801
2802 ujob32 = (struct aiocb32 *)ujob;
2803 return (suword32(&ujob32->_aiocb_private.status, status));
2804 }
2805
2806 static int
2807 aiocb32_store_error(struct aiocb *ujob, long error)
2808 {
2809 struct aiocb32 *ujob32;
2810
2811 ujob32 = (struct aiocb32 *)ujob;
2812 return (suword32(&ujob32->_aiocb_private.error, error));
2813 }
2814
2815 static int
2816 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2817 {
2818 struct aiocb32 *ujob32;
2819
2820 ujob32 = (struct aiocb32 *)ujob;
2821 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2822 }
2823
2824 static int
2825 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2826 {
2827
2828 return (suword32(ujobp, (long)ujob));
2829 }
2830
2831 static struct aiocb_ops aiocb32_ops = {
2832 .copyin = aiocb32_copyin,
2833 .fetch_status = aiocb32_fetch_status,
2834 .fetch_error = aiocb32_fetch_error,
2835 .store_status = aiocb32_store_status,
2836 .store_error = aiocb32_store_error,
2837 .store_kernelinfo = aiocb32_store_kernelinfo,
2838 .store_aiocb = aiocb32_store_aiocb,
2839 };
2840
2841 static struct aiocb_ops aiocb32_ops_osigevent = {
2842 .copyin = aiocb32_copyin_old_sigevent,
2843 .fetch_status = aiocb32_fetch_status,
2844 .fetch_error = aiocb32_fetch_error,
2845 .store_status = aiocb32_store_status,
2846 .store_error = aiocb32_store_error,
2847 .store_kernelinfo = aiocb32_store_kernelinfo,
2848 .store_aiocb = aiocb32_store_aiocb,
2849 };
2850
2851 int
2852 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2853 {
2854
2855 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2856 }
2857
2858 int
2859 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2860 {
2861 struct timespec32 ts32;
2862 struct timespec ts, *tsp;
2863 struct aiocb **ujoblist;
2864 uint32_t *ujoblist32;
2865 int error, i;
2866
2867 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2868 return (EINVAL);
2869
2870 if (uap->timeout) {
2871 /* Get timespec struct. */
2872 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2873 return (error);
2874 CP(ts32, ts, tv_sec);
2875 CP(ts32, ts, tv_nsec);
2876 tsp = &ts;
2877 } else
2878 tsp = NULL;
2879
2880 ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2881 ujoblist32 = (uint32_t *)ujoblist;
2882 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2883 sizeof(ujoblist32[0]));
2884 if (error == 0) {
2885 for (i = uap->nent; i > 0; i--)
2886 ujoblist[i] = PTRIN(ujoblist32[i]);
2887
2888 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2889 }
2890 uma_zfree(aiol_zone, ujoblist);
2891 return (error);
2892 }
2893
2894 int
2895 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2896 {
2897
2898 return (sys_aio_cancel(td, (struct aio_cancel_args *)uap));
2899 }
2900
2901 int
2902 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2903 {
2904
2905 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2906 }
2907
2908 int
2909 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2910 {
2911
2912 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2913 &aiocb32_ops_osigevent));
2914 }
2915
2916 int
2917 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2918 {
2919
2920 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2921 &aiocb32_ops));
2922 }
2923
2924 int
2925 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2926 {
2927
2928 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2929 &aiocb32_ops_osigevent));
2930 }
2931
2932 int
2933 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2934 {
2935
2936 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2937 &aiocb32_ops));
2938 }
2939
2940 int
2941 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2942 {
2943
2944 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2945 &aiocb32_ops));
2946 }
2947
2948 int
2949 freebsd32_aio_waitcomplete(struct thread *td,
2950 struct freebsd32_aio_waitcomplete_args *uap)
2951 {
2952 struct timespec32 ts32;
2953 struct timespec ts, *tsp;
2954 int error;
2955
2956 if (uap->timeout) {
2957 /* Get timespec struct. */
2958 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2959 if (error)
2960 return (error);
2961 CP(ts32, ts, tv_sec);
2962 CP(ts32, ts, tv_nsec);
2963 tsp = &ts;
2964 } else
2965 tsp = NULL;
2966
2967 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2968 &aiocb32_ops));
2969 }
2970
2971 int
2972 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2973 {
2974
2975 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2976 &aiocb32_ops));
2977 }
2978
2979 int
2980 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2981 {
2982 struct aiocb **acb_list;
2983 struct sigevent *sigp, sig;
2984 struct osigevent32 osig;
2985 uint32_t *acb_list32;
2986 int error, i, nent;
2987
2988 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2989 return (EINVAL);
2990
2991 nent = uap->nent;
2992 if (nent < 0 || nent > AIO_LISTIO_MAX)
2993 return (EINVAL);
2994
2995 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2996 error = copyin(uap->sig, &osig, sizeof(osig));
2997 if (error)
2998 return (error);
2999 error = convert_old_sigevent32(&osig, &sig);
3000 if (error)
3001 return (error);
3002 sigp = &sig;
3003 } else
3004 sigp = NULL;
3005
3006 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3007 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3008 if (error) {
3009 free(acb_list32, M_LIO);
3010 return (error);
3011 }
3012 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3013 for (i = 0; i < nent; i++)
3014 acb_list[i] = PTRIN(acb_list32[i]);
3015 free(acb_list32, M_LIO);
3016
3017 error = kern_lio_listio(td, uap->mode,
3018 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3019 &aiocb32_ops_osigevent);
3020 free(acb_list, M_LIO);
3021 return (error);
3022 }
3023
3024 int
3025 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3026 {
3027 struct aiocb **acb_list;
3028 struct sigevent *sigp, sig;
3029 struct sigevent32 sig32;
3030 uint32_t *acb_list32;
3031 int error, i, nent;
3032
3033 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3034 return (EINVAL);
3035
3036 nent = uap->nent;
3037 if (nent < 0 || nent > AIO_LISTIO_MAX)
3038 return (EINVAL);
3039
3040 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3041 error = copyin(uap->sig, &sig32, sizeof(sig32));
3042 if (error)
3043 return (error);
3044 error = convert_sigevent32(&sig32, &sig);
3045 if (error)
3046 return (error);
3047 sigp = &sig;
3048 } else
3049 sigp = NULL;
3050
3051 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3052 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3053 if (error) {
3054 free(acb_list32, M_LIO);
3055 return (error);
3056 }
3057 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3058 for (i = 0; i < nent; i++)
3059 acb_list[i] = PTRIN(acb_list32[i]);
3060 free(acb_list32, M_LIO);
3061
3062 error = kern_lio_listio(td, uap->mode,
3063 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3064 &aiocb32_ops);
3065 free(acb_list, M_LIO);
3066 return (error);
3067 }
3068
3069 #endif
Cache object: ea11f2ee454c827c4b4b525e62cd516f
|