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/11.2/sys/kern/vfs_aio.c 328582 2018-01-30 00:52:22Z jhb $");
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/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sema.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 #include <sys/sx.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
61 #include <sys/conf.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
65
66 #include <machine/atomic.h>
67
68 #include <vm/vm.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 #include <vm/uma.h>
75 #include <sys/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 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC 32
90 #endif
91
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC 256
94 #endif
95
96 #ifndef MAX_AIO_QUEUE
97 #define MAX_AIO_QUEUE 1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
98 #endif
99
100 #ifndef MAX_BUF_AIO
101 #define MAX_BUF_AIO 16
102 #endif
103
104 FEATURE(aio, "Asynchronous I/O");
105 SYSCTL_DECL(_p1003_1b);
106
107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
109
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
111 "Async IO management");
112
113 static int enable_aio_unsafe = 0;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
115 "Permit asynchronous IO on all file types, not just known-safe types");
116
117 static unsigned int unsafe_warningcnt = 1;
118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
119 &unsafe_warningcnt, 0,
120 "Warnings that will be triggered upon failed IO requests on unsafe files");
121
122 static int max_aio_procs = MAX_AIO_PROCS;
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
124 "Maximum number of kernel processes to use for handling async IO ");
125
126 static int num_aio_procs = 0;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
128 "Number of presently active kernel processes for async IO");
129
130 /*
131 * The code will adjust the actual number of AIO processes towards this
132 * number when it gets a chance.
133 */
134 static int target_aio_procs = TARGET_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
136 0,
137 "Preferred number of ready kernel processes for async IO");
138
139 static int max_queue_count = MAX_AIO_QUEUE;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
141 "Maximum number of aio requests to queue, globally");
142
143 static int num_queue_count = 0;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
145 "Number of queued aio requests");
146
147 static int num_buf_aio = 0;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
149 "Number of aio requests presently handled by the buf subsystem");
150
151 static int num_unmapped_aio = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
153 0,
154 "Number of aio requests presently handled by unmapped I/O buffers");
155
156 /* Number of async I/O processes in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
159
160 static int aiod_lifetime;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
162 "Maximum lifetime for idle aiod");
163
164 static int max_aio_per_proc = MAX_AIO_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
166 0,
167 "Maximum active aio requests per process");
168
169 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
171 &max_aio_queue_per_proc, 0,
172 "Maximum queued aio requests per process");
173
174 static int max_buf_aio = MAX_BUF_AIO;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
176 "Maximum buf aio requests per process");
177
178 /*
179 * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
180 * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
181 * vfs.aio.aio_listio_max.
182 */
183 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
184 CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
185 0, "Maximum aio requests for a single lio_listio call");
186
187 #ifdef COMPAT_FREEBSD6
188 typedef struct oaiocb {
189 int aio_fildes; /* File descriptor */
190 off_t aio_offset; /* File offset for I/O */
191 volatile void *aio_buf; /* I/O buffer in process space */
192 size_t aio_nbytes; /* Number of bytes for I/O */
193 struct osigevent aio_sigevent; /* Signal to deliver */
194 int aio_lio_opcode; /* LIO opcode */
195 int aio_reqprio; /* Request priority -- ignored */
196 struct __aiocb_private _aiocb_private;
197 } oaiocb_t;
198 #endif
199
200 /*
201 * Below is a key of locks used to protect each member of struct kaiocb
202 * aioliojob and kaioinfo and any backends.
203 *
204 * * - need not protected
205 * a - locked by kaioinfo lock
206 * b - locked by backend lock, the backend lock can be null in some cases,
207 * for example, BIO belongs to this type, in this case, proc lock is
208 * reused.
209 * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210 */
211
212 /*
213 * If the routine that services an AIO request blocks while running in an
214 * AIO kernel process it can starve other I/O requests. BIO requests
215 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
216 * processes at all. Socket I/O requests use a separate pool of
217 * kprocs and also force non-blocking I/O. Other file I/O requests
218 * use the generic fo_read/fo_write operations which can block. The
219 * fsync and mlock operations can also block while executing. Ideally
220 * none of these requests would block while executing.
221 *
222 * Note that the service routines cannot toggle O_NONBLOCK in the file
223 * structure directly while handling a request due to races with
224 * userland threads.
225 */
226
227 /* jobflags */
228 #define KAIOCB_QUEUEING 0x01
229 #define KAIOCB_CANCELLED 0x02
230 #define KAIOCB_CANCELLING 0x04
231 #define KAIOCB_CHECKSYNC 0x08
232 #define KAIOCB_CLEARED 0x10
233 #define KAIOCB_FINISHED 0x20
234
235 /*
236 * AIO process info
237 */
238 #define AIOP_FREE 0x1 /* proc on free queue */
239
240 struct aioproc {
241 int aioprocflags; /* (c) AIO proc flags */
242 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */
243 struct proc *aioproc; /* (*) the AIO proc */
244 };
245
246 /*
247 * data-structure for lio signal management
248 */
249 struct aioliojob {
250 int lioj_flags; /* (a) listio flags */
251 int lioj_count; /* (a) listio flags */
252 int lioj_finished_count; /* (a) listio flags */
253 struct sigevent lioj_signal; /* (a) signal on all I/O done */
254 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */
255 struct knlist klist; /* (a) list of knotes */
256 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */
257 };
258
259 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
260 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
261 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */
262
263 /*
264 * per process aio data structure
265 */
266 struct kaioinfo {
267 struct mtx kaio_mtx; /* the lock to protect this struct */
268 int kaio_flags; /* (a) per process kaio flags */
269 int kaio_active_count; /* (c) number of currently used AIOs */
270 int kaio_count; /* (a) size of AIO queue */
271 int kaio_buffer_count; /* (a) number of physio buffers */
272 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */
273 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */
274 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */
276 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */
277 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */
278 struct task kaio_task; /* (*) task to kick aio processes */
279 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */
280 };
281
282 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki) (&(ki)->kaio_mtx)
286
287 #define KAIO_RUNDOWN 0x1 /* process is being run down */
288 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */
289
290 /*
291 * Operations used to interact with userland aio control blocks.
292 * Different ABIs provide their own operations.
293 */
294 struct aiocb_ops {
295 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob);
296 long (*fetch_status)(struct aiocb *ujob);
297 long (*fetch_error)(struct aiocb *ujob);
298 int (*store_status)(struct aiocb *ujob, long status);
299 int (*store_error)(struct aiocb *ujob, long error);
300 int (*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
302 };
303
304 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */
308 static struct unrhdr *aiod_unr;
309
310 void aio_init_aioinfo(struct proc *p);
311 static int aio_onceonly(void);
312 static int aio_free_entry(struct kaiocb *job);
313 static void aio_process_rw(struct kaiocb *job);
314 static void aio_process_sync(struct kaiocb *job);
315 static void aio_process_mlock(struct kaiocb *job);
316 static void aio_schedule_fsync(void *context, int pending);
317 static int aio_newproc(int *);
318 int aio_aqueue(struct thread *td, struct aiocb *ujob,
319 struct aioliojob *lio, int type, struct aiocb_ops *ops);
320 static int aio_queue_file(struct file *fp, struct kaiocb *job);
321 static void aio_physwakeup(struct bio *bp);
322 static void aio_proc_rundown(void *arg, struct proc *p);
323 static void aio_proc_rundown_exec(void *arg, struct proc *p,
324 struct image_params *imgp);
325 static int aio_qphysio(struct proc *p, struct kaiocb *job);
326 static void aio_daemon(void *param);
327 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
328 static bool aio_clear_cancel_function_locked(struct kaiocb *job);
329 static int aio_kick(struct proc *userp);
330 static void aio_kick_nowait(struct proc *userp);
331 static void aio_kick_helper(void *context, int pending);
332 static int filt_aioattach(struct knote *kn);
333 static void filt_aiodetach(struct knote *kn);
334 static int filt_aio(struct knote *kn, long hint);
335 static int filt_lioattach(struct knote *kn);
336 static void filt_liodetach(struct knote *kn);
337 static int filt_lio(struct knote *kn, long hint);
338
339 /*
340 * Zones for:
341 * kaio Per process async io info
342 * aiop async io process data
343 * aiocb async io jobs
344 * aiolio list io jobs
345 */
346 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
347
348 /* kqueue filters for aio */
349 static struct filterops aio_filtops = {
350 .f_isfd = 0,
351 .f_attach = filt_aioattach,
352 .f_detach = filt_aiodetach,
353 .f_event = filt_aio,
354 };
355 static struct filterops lio_filtops = {
356 .f_isfd = 0,
357 .f_attach = filt_lioattach,
358 .f_detach = filt_liodetach,
359 .f_event = filt_lio
360 };
361
362 static eventhandler_tag exit_tag, exec_tag;
363
364 TASKQUEUE_DEFINE_THREAD(aiod_kick);
365
366 /*
367 * Main operations function for use as a kernel module.
368 */
369 static int
370 aio_modload(struct module *module, int cmd, void *arg)
371 {
372 int error = 0;
373
374 switch (cmd) {
375 case MOD_LOAD:
376 aio_onceonly();
377 break;
378 case MOD_SHUTDOWN:
379 break;
380 default:
381 error = EOPNOTSUPP;
382 break;
383 }
384 return (error);
385 }
386
387 static moduledata_t aio_mod = {
388 "aio",
389 &aio_modload,
390 NULL
391 };
392
393 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
394 MODULE_VERSION(aio, 1);
395
396 /*
397 * Startup initialization
398 */
399 static int
400 aio_onceonly(void)
401 {
402
403 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
404 EVENTHANDLER_PRI_ANY);
405 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
406 NULL, EVENTHANDLER_PRI_ANY);
407 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
408 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
409 TAILQ_INIT(&aio_freeproc);
410 sema_init(&aio_newproc_sem, 0, "aio_new_proc");
411 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
412 TAILQ_INIT(&aio_jobs);
413 aiod_unr = new_unrhdr(1, INT_MAX, NULL);
414 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
415 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
416 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
417 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
418 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
419 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
420 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
421 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
422 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
423 jobrefid = 1;
424 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
425 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
426 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
427
428 return (0);
429 }
430
431 /*
432 * Init the per-process aioinfo structure. The aioinfo limits are set
433 * per-process for user limit (resource) management.
434 */
435 void
436 aio_init_aioinfo(struct proc *p)
437 {
438 struct kaioinfo *ki;
439
440 ki = uma_zalloc(kaio_zone, M_WAITOK);
441 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
442 ki->kaio_flags = 0;
443 ki->kaio_active_count = 0;
444 ki->kaio_count = 0;
445 ki->kaio_buffer_count = 0;
446 TAILQ_INIT(&ki->kaio_all);
447 TAILQ_INIT(&ki->kaio_done);
448 TAILQ_INIT(&ki->kaio_jobqueue);
449 TAILQ_INIT(&ki->kaio_liojoblist);
450 TAILQ_INIT(&ki->kaio_syncqueue);
451 TAILQ_INIT(&ki->kaio_syncready);
452 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
453 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
454 PROC_LOCK(p);
455 if (p->p_aioinfo == NULL) {
456 p->p_aioinfo = ki;
457 PROC_UNLOCK(p);
458 } else {
459 PROC_UNLOCK(p);
460 mtx_destroy(&ki->kaio_mtx);
461 uma_zfree(kaio_zone, ki);
462 }
463
464 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
465 aio_newproc(NULL);
466 }
467
468 static int
469 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
470 {
471 struct thread *td;
472 int error;
473
474 error = sigev_findtd(p, sigev, &td);
475 if (error)
476 return (error);
477 if (!KSI_ONQ(ksi)) {
478 ksiginfo_set_sigev(ksi, sigev);
479 ksi->ksi_code = SI_ASYNCIO;
480 ksi->ksi_flags |= KSI_EXT | KSI_INS;
481 tdsendsignal(p, td, ksi->ksi_signo, ksi);
482 }
483 PROC_UNLOCK(p);
484 return (error);
485 }
486
487 /*
488 * Free a job entry. Wait for completion if it is currently active, but don't
489 * delay forever. If we delay, we return a flag that says that we have to
490 * restart the queue scan.
491 */
492 static int
493 aio_free_entry(struct kaiocb *job)
494 {
495 struct kaioinfo *ki;
496 struct aioliojob *lj;
497 struct proc *p;
498
499 p = job->userproc;
500 MPASS(curproc == p);
501 ki = p->p_aioinfo;
502 MPASS(ki != NULL);
503
504 AIO_LOCK_ASSERT(ki, MA_OWNED);
505 MPASS(job->jobflags & KAIOCB_FINISHED);
506
507 atomic_subtract_int(&num_queue_count, 1);
508
509 ki->kaio_count--;
510 MPASS(ki->kaio_count >= 0);
511
512 TAILQ_REMOVE(&ki->kaio_done, job, plist);
513 TAILQ_REMOVE(&ki->kaio_all, job, allist);
514
515 lj = job->lio;
516 if (lj) {
517 lj->lioj_count--;
518 lj->lioj_finished_count--;
519
520 if (lj->lioj_count == 0) {
521 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
522 /* lio is going away, we need to destroy any knotes */
523 knlist_delete(&lj->klist, curthread, 1);
524 PROC_LOCK(p);
525 sigqueue_take(&lj->lioj_ksi);
526 PROC_UNLOCK(p);
527 uma_zfree(aiolio_zone, lj);
528 }
529 }
530
531 /* job is going away, we need to destroy any knotes */
532 knlist_delete(&job->klist, curthread, 1);
533 PROC_LOCK(p);
534 sigqueue_take(&job->ksi);
535 PROC_UNLOCK(p);
536
537 AIO_UNLOCK(ki);
538
539 /*
540 * The thread argument here is used to find the owning process
541 * and is also passed to fo_close() which may pass it to various
542 * places such as devsw close() routines. Because of that, we
543 * need a thread pointer from the process owning the job that is
544 * persistent and won't disappear out from under us or move to
545 * another process.
546 *
547 * Currently, all the callers of this function call it to remove
548 * a kaiocb from the current process' job list either via a
549 * syscall or due to the current process calling exit() or
550 * execve(). Thus, we know that p == curproc. We also know that
551 * curthread can't exit since we are curthread.
552 *
553 * Therefore, we use curthread as the thread to pass to
554 * knlist_delete(). This does mean that it is possible for the
555 * thread pointer at close time to differ from the thread pointer
556 * at open time, but this is already true of file descriptors in
557 * a multithreaded process.
558 */
559 if (job->fd_file)
560 fdrop(job->fd_file, curthread);
561 crfree(job->cred);
562 uma_zfree(aiocb_zone, job);
563 AIO_LOCK(ki);
564
565 return (0);
566 }
567
568 static void
569 aio_proc_rundown_exec(void *arg, struct proc *p,
570 struct image_params *imgp __unused)
571 {
572 aio_proc_rundown(arg, p);
573 }
574
575 static int
576 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
577 {
578 aio_cancel_fn_t *func;
579 int cancelled;
580
581 AIO_LOCK_ASSERT(ki, MA_OWNED);
582 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
583 return (0);
584 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
585 job->jobflags |= KAIOCB_CANCELLED;
586
587 func = job->cancel_fn;
588
589 /*
590 * If there is no cancel routine, just leave the job marked as
591 * cancelled. The job should be in active use by a caller who
592 * should complete it normally or when it fails to install a
593 * cancel routine.
594 */
595 if (func == NULL)
596 return (0);
597
598 /*
599 * Set the CANCELLING flag so that aio_complete() will defer
600 * completions of this job. This prevents the job from being
601 * freed out from under the cancel callback. After the
602 * callback any deferred completion (whether from the callback
603 * or any other source) will be completed.
604 */
605 job->jobflags |= KAIOCB_CANCELLING;
606 AIO_UNLOCK(ki);
607 func(job);
608 AIO_LOCK(ki);
609 job->jobflags &= ~KAIOCB_CANCELLING;
610 if (job->jobflags & KAIOCB_FINISHED) {
611 cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
612 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
613 aio_bio_done_notify(p, job);
614 } else {
615 /*
616 * The cancel callback might have scheduled an
617 * operation to cancel this request, but it is
618 * only counted as cancelled if the request is
619 * cancelled when the callback returns.
620 */
621 cancelled = 0;
622 }
623 return (cancelled);
624 }
625
626 /*
627 * Rundown the jobs for a given process.
628 */
629 static void
630 aio_proc_rundown(void *arg, struct proc *p)
631 {
632 struct kaioinfo *ki;
633 struct aioliojob *lj;
634 struct kaiocb *job, *jobn;
635
636 KASSERT(curthread->td_proc == p,
637 ("%s: called on non-curproc", __func__));
638 ki = p->p_aioinfo;
639 if (ki == NULL)
640 return;
641
642 AIO_LOCK(ki);
643 ki->kaio_flags |= KAIO_RUNDOWN;
644
645 restart:
646
647 /*
648 * Try to cancel all pending requests. This code simulates
649 * aio_cancel on all pending I/O requests.
650 */
651 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
652 aio_cancel_job(p, ki, job);
653 }
654
655 /* Wait for all running I/O to be finished */
656 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
657 ki->kaio_flags |= KAIO_WAKEUP;
658 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
659 goto restart;
660 }
661
662 /* Free all completed I/O requests. */
663 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
664 aio_free_entry(job);
665
666 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
667 if (lj->lioj_count == 0) {
668 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
669 knlist_delete(&lj->klist, curthread, 1);
670 PROC_LOCK(p);
671 sigqueue_take(&lj->lioj_ksi);
672 PROC_UNLOCK(p);
673 uma_zfree(aiolio_zone, lj);
674 } else {
675 panic("LIO job not cleaned up: C:%d, FC:%d\n",
676 lj->lioj_count, lj->lioj_finished_count);
677 }
678 }
679 AIO_UNLOCK(ki);
680 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
681 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
682 mtx_destroy(&ki->kaio_mtx);
683 uma_zfree(kaio_zone, ki);
684 p->p_aioinfo = NULL;
685 }
686
687 /*
688 * Select a job to run (called by an AIO daemon).
689 */
690 static struct kaiocb *
691 aio_selectjob(struct aioproc *aiop)
692 {
693 struct kaiocb *job;
694 struct kaioinfo *ki;
695 struct proc *userp;
696
697 mtx_assert(&aio_job_mtx, MA_OWNED);
698 restart:
699 TAILQ_FOREACH(job, &aio_jobs, list) {
700 userp = job->userproc;
701 ki = userp->p_aioinfo;
702
703 if (ki->kaio_active_count < max_aio_per_proc) {
704 TAILQ_REMOVE(&aio_jobs, job, list);
705 if (!aio_clear_cancel_function(job))
706 goto restart;
707
708 /* Account for currently active jobs. */
709 ki->kaio_active_count++;
710 break;
711 }
712 }
713 return (job);
714 }
715
716 /*
717 * Move all data to a permanent storage device. This code
718 * simulates the fsync syscall.
719 */
720 static int
721 aio_fsync_vnode(struct thread *td, struct vnode *vp)
722 {
723 struct mount *mp;
724 int error;
725
726 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
727 goto drop;
728 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
729 if (vp->v_object != NULL) {
730 VM_OBJECT_WLOCK(vp->v_object);
731 vm_object_page_clean(vp->v_object, 0, 0, 0);
732 VM_OBJECT_WUNLOCK(vp->v_object);
733 }
734 error = VOP_FSYNC(vp, MNT_WAIT, td);
735
736 VOP_UNLOCK(vp, 0);
737 vn_finished_write(mp);
738 drop:
739 return (error);
740 }
741
742 /*
743 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that
744 * does the I/O request for the non-physio version of the operations. The
745 * normal vn operations are used, and this code should work in all instances
746 * for every type of file, including pipes, sockets, fifos, and regular files.
747 *
748 * XXX I don't think it works well for socket, pipe, and fifo.
749 */
750 static void
751 aio_process_rw(struct kaiocb *job)
752 {
753 struct ucred *td_savedcred;
754 struct thread *td;
755 struct aiocb *cb;
756 struct file *fp;
757 struct uio auio;
758 struct iovec aiov;
759 ssize_t cnt;
760 long msgsnd_st, msgsnd_end;
761 long msgrcv_st, msgrcv_end;
762 long oublock_st, oublock_end;
763 long inblock_st, inblock_end;
764 int error;
765
766 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
767 job->uaiocb.aio_lio_opcode == LIO_WRITE,
768 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
769
770 aio_switch_vmspace(job);
771 td = curthread;
772 td_savedcred = td->td_ucred;
773 td->td_ucred = job->cred;
774 cb = &job->uaiocb;
775 fp = job->fd_file;
776
777 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
778 aiov.iov_len = cb->aio_nbytes;
779
780 auio.uio_iov = &aiov;
781 auio.uio_iovcnt = 1;
782 auio.uio_offset = cb->aio_offset;
783 auio.uio_resid = cb->aio_nbytes;
784 cnt = cb->aio_nbytes;
785 auio.uio_segflg = UIO_USERSPACE;
786 auio.uio_td = td;
787
788 msgrcv_st = td->td_ru.ru_msgrcv;
789 msgsnd_st = td->td_ru.ru_msgsnd;
790 inblock_st = td->td_ru.ru_inblock;
791 oublock_st = td->td_ru.ru_oublock;
792
793 /*
794 * aio_aqueue() acquires a reference to the file that is
795 * released in aio_free_entry().
796 */
797 if (cb->aio_lio_opcode == LIO_READ) {
798 auio.uio_rw = UIO_READ;
799 if (auio.uio_resid == 0)
800 error = 0;
801 else
802 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
803 } else {
804 if (fp->f_type == DTYPE_VNODE)
805 bwillwrite();
806 auio.uio_rw = UIO_WRITE;
807 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
808 }
809 msgrcv_end = td->td_ru.ru_msgrcv;
810 msgsnd_end = td->td_ru.ru_msgsnd;
811 inblock_end = td->td_ru.ru_inblock;
812 oublock_end = td->td_ru.ru_oublock;
813
814 job->msgrcv = msgrcv_end - msgrcv_st;
815 job->msgsnd = msgsnd_end - msgsnd_st;
816 job->inblock = inblock_end - inblock_st;
817 job->outblock = oublock_end - oublock_st;
818
819 if ((error) && (auio.uio_resid != cnt)) {
820 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
821 error = 0;
822 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
823 PROC_LOCK(job->userproc);
824 kern_psignal(job->userproc, SIGPIPE);
825 PROC_UNLOCK(job->userproc);
826 }
827 }
828
829 cnt -= auio.uio_resid;
830 td->td_ucred = td_savedcred;
831 if (error)
832 aio_complete(job, -1, error);
833 else
834 aio_complete(job, cnt, 0);
835 }
836
837 static void
838 aio_process_sync(struct kaiocb *job)
839 {
840 struct thread *td = curthread;
841 struct ucred *td_savedcred = td->td_ucred;
842 struct file *fp = job->fd_file;
843 int error = 0;
844
845 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
846 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
847
848 td->td_ucred = job->cred;
849 if (fp->f_vnode != NULL)
850 error = aio_fsync_vnode(td, fp->f_vnode);
851 td->td_ucred = td_savedcred;
852 if (error)
853 aio_complete(job, -1, error);
854 else
855 aio_complete(job, 0, 0);
856 }
857
858 static void
859 aio_process_mlock(struct kaiocb *job)
860 {
861 struct aiocb *cb = &job->uaiocb;
862 int error;
863
864 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
865 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
866
867 aio_switch_vmspace(job);
868 error = kern_mlock(job->userproc, job->cred,
869 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
870 aio_complete(job, error != 0 ? -1 : 0, error);
871 }
872
873 static void
874 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
875 {
876 struct aioliojob *lj;
877 struct kaioinfo *ki;
878 struct kaiocb *sjob, *sjobn;
879 int lj_done;
880 bool schedule_fsync;
881
882 ki = userp->p_aioinfo;
883 AIO_LOCK_ASSERT(ki, MA_OWNED);
884 lj = job->lio;
885 lj_done = 0;
886 if (lj) {
887 lj->lioj_finished_count++;
888 if (lj->lioj_count == lj->lioj_finished_count)
889 lj_done = 1;
890 }
891 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
892 MPASS(job->jobflags & KAIOCB_FINISHED);
893
894 if (ki->kaio_flags & KAIO_RUNDOWN)
895 goto notification_done;
896
897 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
898 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
899 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
900
901 KNOTE_LOCKED(&job->klist, 1);
902
903 if (lj_done) {
904 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
905 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
906 KNOTE_LOCKED(&lj->klist, 1);
907 }
908 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
909 == LIOJ_SIGNAL
910 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
911 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
912 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
913 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
914 }
915 }
916
917 notification_done:
918 if (job->jobflags & KAIOCB_CHECKSYNC) {
919 schedule_fsync = false;
920 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
921 if (job->fd_file != sjob->fd_file ||
922 job->seqno >= sjob->seqno)
923 continue;
924 if (--sjob->pending > 0)
925 continue;
926 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
927 if (!aio_clear_cancel_function_locked(sjob))
928 continue;
929 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
930 schedule_fsync = true;
931 }
932 if (schedule_fsync)
933 taskqueue_enqueue(taskqueue_aiod_kick,
934 &ki->kaio_sync_task);
935 }
936 if (ki->kaio_flags & KAIO_WAKEUP) {
937 ki->kaio_flags &= ~KAIO_WAKEUP;
938 wakeup(&userp->p_aioinfo);
939 }
940 }
941
942 static void
943 aio_schedule_fsync(void *context, int pending)
944 {
945 struct kaioinfo *ki;
946 struct kaiocb *job;
947
948 ki = context;
949 AIO_LOCK(ki);
950 while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
951 job = TAILQ_FIRST(&ki->kaio_syncready);
952 TAILQ_REMOVE(&ki->kaio_syncready, job, list);
953 AIO_UNLOCK(ki);
954 aio_schedule(job, aio_process_sync);
955 AIO_LOCK(ki);
956 }
957 AIO_UNLOCK(ki);
958 }
959
960 bool
961 aio_cancel_cleared(struct kaiocb *job)
962 {
963 struct kaioinfo *ki;
964
965 /*
966 * The caller should hold the same queue lock held when
967 * aio_clear_cancel_function() was called and set this flag
968 * ensuring this check sees an up-to-date value. However,
969 * there is no way to assert that.
970 */
971 ki = job->userproc->p_aioinfo;
972 return ((job->jobflags & KAIOCB_CLEARED) != 0);
973 }
974
975 static bool
976 aio_clear_cancel_function_locked(struct kaiocb *job)
977 {
978
979 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
980 MPASS(job->cancel_fn != NULL);
981 if (job->jobflags & KAIOCB_CANCELLING) {
982 job->jobflags |= KAIOCB_CLEARED;
983 return (false);
984 }
985 job->cancel_fn = NULL;
986 return (true);
987 }
988
989 bool
990 aio_clear_cancel_function(struct kaiocb *job)
991 {
992 struct kaioinfo *ki;
993 bool ret;
994
995 ki = job->userproc->p_aioinfo;
996 AIO_LOCK(ki);
997 ret = aio_clear_cancel_function_locked(job);
998 AIO_UNLOCK(ki);
999 return (ret);
1000 }
1001
1002 static bool
1003 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1004 {
1005
1006 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1007 if (job->jobflags & KAIOCB_CANCELLED)
1008 return (false);
1009 job->cancel_fn = func;
1010 return (true);
1011 }
1012
1013 bool
1014 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1015 {
1016 struct kaioinfo *ki;
1017 bool ret;
1018
1019 ki = job->userproc->p_aioinfo;
1020 AIO_LOCK(ki);
1021 ret = aio_set_cancel_function_locked(job, func);
1022 AIO_UNLOCK(ki);
1023 return (ret);
1024 }
1025
1026 void
1027 aio_complete(struct kaiocb *job, long status, int error)
1028 {
1029 struct kaioinfo *ki;
1030 struct proc *userp;
1031
1032 job->uaiocb._aiocb_private.error = error;
1033 job->uaiocb._aiocb_private.status = status;
1034
1035 userp = job->userproc;
1036 ki = userp->p_aioinfo;
1037
1038 AIO_LOCK(ki);
1039 KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1040 ("duplicate aio_complete"));
1041 job->jobflags |= KAIOCB_FINISHED;
1042 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1043 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1044 aio_bio_done_notify(userp, job);
1045 }
1046 AIO_UNLOCK(ki);
1047 }
1048
1049 void
1050 aio_cancel(struct kaiocb *job)
1051 {
1052
1053 aio_complete(job, -1, ECANCELED);
1054 }
1055
1056 void
1057 aio_switch_vmspace(struct kaiocb *job)
1058 {
1059
1060 vmspace_switch_aio(job->userproc->p_vmspace);
1061 }
1062
1063 /*
1064 * The AIO daemon, most of the actual work is done in aio_process_*,
1065 * but the setup (and address space mgmt) is done in this routine.
1066 */
1067 static void
1068 aio_daemon(void *_id)
1069 {
1070 struct kaiocb *job;
1071 struct aioproc *aiop;
1072 struct kaioinfo *ki;
1073 struct proc *p;
1074 struct vmspace *myvm;
1075 struct thread *td = curthread;
1076 int id = (intptr_t)_id;
1077
1078 /*
1079 * Grab an extra reference on the daemon's vmspace so that it
1080 * doesn't get freed by jobs that switch to a different
1081 * vmspace.
1082 */
1083 p = td->td_proc;
1084 myvm = vmspace_acquire_ref(p);
1085
1086 KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1087
1088 /*
1089 * Allocate and ready the aio control info. There is one aiop structure
1090 * per daemon.
1091 */
1092 aiop = uma_zalloc(aiop_zone, M_WAITOK);
1093 aiop->aioproc = p;
1094 aiop->aioprocflags = 0;
1095
1096 /*
1097 * Wakeup parent process. (Parent sleeps to keep from blasting away
1098 * and creating too many daemons.)
1099 */
1100 sema_post(&aio_newproc_sem);
1101
1102 mtx_lock(&aio_job_mtx);
1103 for (;;) {
1104 /*
1105 * Take daemon off of free queue
1106 */
1107 if (aiop->aioprocflags & AIOP_FREE) {
1108 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1109 aiop->aioprocflags &= ~AIOP_FREE;
1110 }
1111
1112 /*
1113 * Check for jobs.
1114 */
1115 while ((job = aio_selectjob(aiop)) != NULL) {
1116 mtx_unlock(&aio_job_mtx);
1117
1118 ki = job->userproc->p_aioinfo;
1119 job->handle_fn(job);
1120
1121 mtx_lock(&aio_job_mtx);
1122 /* Decrement the active job count. */
1123 ki->kaio_active_count--;
1124 }
1125
1126 /*
1127 * Disconnect from user address space.
1128 */
1129 if (p->p_vmspace != myvm) {
1130 mtx_unlock(&aio_job_mtx);
1131 vmspace_switch_aio(myvm);
1132 mtx_lock(&aio_job_mtx);
1133 /*
1134 * We have to restart to avoid race, we only sleep if
1135 * no job can be selected.
1136 */
1137 continue;
1138 }
1139
1140 mtx_assert(&aio_job_mtx, MA_OWNED);
1141
1142 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1143 aiop->aioprocflags |= AIOP_FREE;
1144
1145 /*
1146 * If daemon is inactive for a long time, allow it to exit,
1147 * thereby freeing resources.
1148 */
1149 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1150 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1151 (aiop->aioprocflags & AIOP_FREE) &&
1152 num_aio_procs > target_aio_procs)
1153 break;
1154 }
1155 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1156 num_aio_procs--;
1157 mtx_unlock(&aio_job_mtx);
1158 uma_zfree(aiop_zone, aiop);
1159 free_unr(aiod_unr, id);
1160 vmspace_free(myvm);
1161
1162 KASSERT(p->p_vmspace == myvm,
1163 ("AIOD: bad vmspace for exiting daemon"));
1164 KASSERT(myvm->vm_refcnt > 1,
1165 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1166 kproc_exit(0);
1167 }
1168
1169 /*
1170 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1171 * AIO daemon modifies its environment itself.
1172 */
1173 static int
1174 aio_newproc(int *start)
1175 {
1176 int error;
1177 struct proc *p;
1178 int id;
1179
1180 id = alloc_unr(aiod_unr);
1181 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1182 RFNOWAIT, 0, "aiod%d", id);
1183 if (error == 0) {
1184 /*
1185 * Wait until daemon is started.
1186 */
1187 sema_wait(&aio_newproc_sem);
1188 mtx_lock(&aio_job_mtx);
1189 num_aio_procs++;
1190 if (start != NULL)
1191 (*start)--;
1192 mtx_unlock(&aio_job_mtx);
1193 } else {
1194 free_unr(aiod_unr, id);
1195 }
1196 return (error);
1197 }
1198
1199 /*
1200 * Try the high-performance, low-overhead physio method for eligible
1201 * VCHR devices. This method doesn't use an aio helper thread, and
1202 * thus has very low overhead.
1203 *
1204 * Assumes that the caller, aio_aqueue(), has incremented the file
1205 * structure's reference count, preventing its deallocation for the
1206 * duration of this call.
1207 */
1208 static int
1209 aio_qphysio(struct proc *p, struct kaiocb *job)
1210 {
1211 struct aiocb *cb;
1212 struct file *fp;
1213 struct bio *bp;
1214 struct buf *pbuf;
1215 struct vnode *vp;
1216 struct cdevsw *csw;
1217 struct cdev *dev;
1218 struct kaioinfo *ki;
1219 int error, ref, poff;
1220 vm_prot_t prot;
1221
1222 cb = &job->uaiocb;
1223 fp = job->fd_file;
1224
1225 if (!(cb->aio_lio_opcode == LIO_WRITE ||
1226 cb->aio_lio_opcode == LIO_READ))
1227 return (-1);
1228 if (fp == NULL || fp->f_type != DTYPE_VNODE)
1229 return (-1);
1230
1231 vp = fp->f_vnode;
1232 if (vp->v_type != VCHR)
1233 return (-1);
1234 if (vp->v_bufobj.bo_bsize == 0)
1235 return (-1);
1236 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1237 return (-1);
1238
1239 ref = 0;
1240 csw = devvn_refthread(vp, &dev, &ref);
1241 if (csw == NULL)
1242 return (ENXIO);
1243
1244 if ((csw->d_flags & D_DISK) == 0) {
1245 error = -1;
1246 goto unref;
1247 }
1248 if (cb->aio_nbytes > dev->si_iosize_max) {
1249 error = -1;
1250 goto unref;
1251 }
1252
1253 ki = p->p_aioinfo;
1254 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1255 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1256 if (cb->aio_nbytes > MAXPHYS) {
1257 error = -1;
1258 goto unref;
1259 }
1260
1261 pbuf = NULL;
1262 } else {
1263 if (cb->aio_nbytes > MAXPHYS - poff) {
1264 error = -1;
1265 goto unref;
1266 }
1267 if (ki->kaio_buffer_count >= max_buf_aio) {
1268 error = EAGAIN;
1269 goto unref;
1270 }
1271
1272 job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1273 BUF_KERNPROC(pbuf);
1274 AIO_LOCK(ki);
1275 ki->kaio_buffer_count++;
1276 AIO_UNLOCK(ki);
1277 }
1278 job->bp = bp = g_alloc_bio();
1279
1280 bp->bio_length = cb->aio_nbytes;
1281 bp->bio_bcount = cb->aio_nbytes;
1282 bp->bio_done = aio_physwakeup;
1283 bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1284 bp->bio_offset = cb->aio_offset;
1285 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1286 bp->bio_dev = dev;
1287 bp->bio_caller1 = (void *)job;
1288
1289 prot = VM_PROT_READ;
1290 if (cb->aio_lio_opcode == LIO_READ)
1291 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1292 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1293 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1294 nitems(job->pages));
1295 if (job->npages < 0) {
1296 error = EFAULT;
1297 goto doerror;
1298 }
1299 if (pbuf != NULL) {
1300 pmap_qenter((vm_offset_t)pbuf->b_data,
1301 job->pages, job->npages);
1302 bp->bio_data = pbuf->b_data + poff;
1303 atomic_add_int(&num_buf_aio, 1);
1304 } else {
1305 bp->bio_ma = job->pages;
1306 bp->bio_ma_n = job->npages;
1307 bp->bio_ma_offset = poff;
1308 bp->bio_data = unmapped_buf;
1309 bp->bio_flags |= BIO_UNMAPPED;
1310 atomic_add_int(&num_unmapped_aio, 1);
1311 }
1312
1313 /* Perform transfer. */
1314 csw->d_strategy(bp);
1315 dev_relthread(dev, ref);
1316 return (0);
1317
1318 doerror:
1319 if (pbuf != NULL) {
1320 AIO_LOCK(ki);
1321 ki->kaio_buffer_count--;
1322 AIO_UNLOCK(ki);
1323 relpbuf(pbuf, NULL);
1324 job->pbuf = NULL;
1325 }
1326 g_destroy_bio(bp);
1327 job->bp = NULL;
1328 unref:
1329 dev_relthread(dev, ref);
1330 return (error);
1331 }
1332
1333 #ifdef COMPAT_FREEBSD6
1334 static int
1335 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1336 {
1337
1338 /*
1339 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1340 * supported by AIO with the old sigevent structure.
1341 */
1342 nsig->sigev_notify = osig->sigev_notify;
1343 switch (nsig->sigev_notify) {
1344 case SIGEV_NONE:
1345 break;
1346 case SIGEV_SIGNAL:
1347 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1348 break;
1349 case SIGEV_KEVENT:
1350 nsig->sigev_notify_kqueue =
1351 osig->__sigev_u.__sigev_notify_kqueue;
1352 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1353 break;
1354 default:
1355 return (EINVAL);
1356 }
1357 return (0);
1358 }
1359
1360 static int
1361 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1362 {
1363 struct oaiocb *ojob;
1364 int error;
1365
1366 bzero(kjob, sizeof(struct aiocb));
1367 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1368 if (error)
1369 return (error);
1370 ojob = (struct oaiocb *)kjob;
1371 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1372 }
1373 #endif
1374
1375 static int
1376 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1377 {
1378
1379 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1380 }
1381
1382 static long
1383 aiocb_fetch_status(struct aiocb *ujob)
1384 {
1385
1386 return (fuword(&ujob->_aiocb_private.status));
1387 }
1388
1389 static long
1390 aiocb_fetch_error(struct aiocb *ujob)
1391 {
1392
1393 return (fuword(&ujob->_aiocb_private.error));
1394 }
1395
1396 static int
1397 aiocb_store_status(struct aiocb *ujob, long status)
1398 {
1399
1400 return (suword(&ujob->_aiocb_private.status, status));
1401 }
1402
1403 static int
1404 aiocb_store_error(struct aiocb *ujob, long error)
1405 {
1406
1407 return (suword(&ujob->_aiocb_private.error, error));
1408 }
1409
1410 static int
1411 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1412 {
1413
1414 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1415 }
1416
1417 static int
1418 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1419 {
1420
1421 return (suword(ujobp, (long)ujob));
1422 }
1423
1424 static struct aiocb_ops aiocb_ops = {
1425 .copyin = aiocb_copyin,
1426 .fetch_status = aiocb_fetch_status,
1427 .fetch_error = aiocb_fetch_error,
1428 .store_status = aiocb_store_status,
1429 .store_error = aiocb_store_error,
1430 .store_kernelinfo = aiocb_store_kernelinfo,
1431 .store_aiocb = aiocb_store_aiocb,
1432 };
1433
1434 #ifdef COMPAT_FREEBSD6
1435 static struct aiocb_ops aiocb_ops_osigevent = {
1436 .copyin = aiocb_copyin_old_sigevent,
1437 .fetch_status = aiocb_fetch_status,
1438 .fetch_error = aiocb_fetch_error,
1439 .store_status = aiocb_store_status,
1440 .store_error = aiocb_store_error,
1441 .store_kernelinfo = aiocb_store_kernelinfo,
1442 .store_aiocb = aiocb_store_aiocb,
1443 };
1444 #endif
1445
1446 /*
1447 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1448 * technique is done in this code.
1449 */
1450 int
1451 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1452 int type, struct aiocb_ops *ops)
1453 {
1454 struct proc *p = td->td_proc;
1455 cap_rights_t rights;
1456 struct file *fp;
1457 struct kaiocb *job;
1458 struct kaioinfo *ki;
1459 struct kevent kev;
1460 int opcode;
1461 int error;
1462 int fd, kqfd;
1463 int jid;
1464 u_short evflags;
1465
1466 if (p->p_aioinfo == NULL)
1467 aio_init_aioinfo(p);
1468
1469 ki = p->p_aioinfo;
1470
1471 ops->store_status(ujob, -1);
1472 ops->store_error(ujob, 0);
1473 ops->store_kernelinfo(ujob, -1);
1474
1475 if (num_queue_count >= max_queue_count ||
1476 ki->kaio_count >= max_aio_queue_per_proc) {
1477 ops->store_error(ujob, EAGAIN);
1478 return (EAGAIN);
1479 }
1480
1481 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1482 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1483
1484 error = ops->copyin(ujob, &job->uaiocb);
1485 if (error) {
1486 ops->store_error(ujob, error);
1487 uma_zfree(aiocb_zone, job);
1488 return (error);
1489 }
1490
1491 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1492 uma_zfree(aiocb_zone, job);
1493 return (EINVAL);
1494 }
1495
1496 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1497 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1498 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1499 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1500 ops->store_error(ujob, EINVAL);
1501 uma_zfree(aiocb_zone, job);
1502 return (EINVAL);
1503 }
1504
1505 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1506 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1507 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1508 uma_zfree(aiocb_zone, job);
1509 return (EINVAL);
1510 }
1511
1512 ksiginfo_init(&job->ksi);
1513
1514 /* Save userspace address of the job info. */
1515 job->ujob = ujob;
1516
1517 /* Get the opcode. */
1518 if (type != LIO_NOP)
1519 job->uaiocb.aio_lio_opcode = type;
1520 opcode = job->uaiocb.aio_lio_opcode;
1521
1522 /*
1523 * Validate the opcode and fetch the file object for the specified
1524 * file descriptor.
1525 *
1526 * XXXRW: Moved the opcode validation up here so that we don't
1527 * retrieve a file descriptor without knowing what the capabiltity
1528 * should be.
1529 */
1530 fd = job->uaiocb.aio_fildes;
1531 switch (opcode) {
1532 case LIO_WRITE:
1533 error = fget_write(td, fd,
1534 cap_rights_init(&rights, CAP_PWRITE), &fp);
1535 break;
1536 case LIO_READ:
1537 error = fget_read(td, fd,
1538 cap_rights_init(&rights, CAP_PREAD), &fp);
1539 break;
1540 case LIO_SYNC:
1541 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1542 break;
1543 case LIO_MLOCK:
1544 fp = NULL;
1545 break;
1546 case LIO_NOP:
1547 error = fget(td, fd, cap_rights_init(&rights), &fp);
1548 break;
1549 default:
1550 error = EINVAL;
1551 }
1552 if (error) {
1553 uma_zfree(aiocb_zone, job);
1554 ops->store_error(ujob, error);
1555 return (error);
1556 }
1557
1558 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1559 error = EINVAL;
1560 goto aqueue_fail;
1561 }
1562
1563 if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1564 job->uaiocb.aio_offset < 0 &&
1565 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1566 error = EINVAL;
1567 goto aqueue_fail;
1568 }
1569
1570 job->fd_file = fp;
1571
1572 mtx_lock(&aio_job_mtx);
1573 jid = jobrefid++;
1574 job->seqno = jobseqno++;
1575 mtx_unlock(&aio_job_mtx);
1576 error = ops->store_kernelinfo(ujob, jid);
1577 if (error) {
1578 error = EINVAL;
1579 goto aqueue_fail;
1580 }
1581 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1582
1583 if (opcode == LIO_NOP) {
1584 fdrop(fp, td);
1585 uma_zfree(aiocb_zone, job);
1586 return (0);
1587 }
1588
1589 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1590 goto no_kqueue;
1591 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1592 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1593 error = EINVAL;
1594 goto aqueue_fail;
1595 }
1596 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1597 kev.ident = (uintptr_t)job->ujob;
1598 kev.filter = EVFILT_AIO;
1599 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1600 kev.data = (intptr_t)job;
1601 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1602 error = kqfd_register(kqfd, &kev, td, 1);
1603 if (error)
1604 goto aqueue_fail;
1605
1606 no_kqueue:
1607
1608 ops->store_error(ujob, EINPROGRESS);
1609 job->uaiocb._aiocb_private.error = EINPROGRESS;
1610 job->userproc = p;
1611 job->cred = crhold(td->td_ucred);
1612 job->jobflags = KAIOCB_QUEUEING;
1613 job->lio = lj;
1614
1615 if (opcode == LIO_MLOCK) {
1616 aio_schedule(job, aio_process_mlock);
1617 error = 0;
1618 } else if (fp->f_ops->fo_aio_queue == NULL)
1619 error = aio_queue_file(fp, job);
1620 else
1621 error = fo_aio_queue(fp, job);
1622 if (error)
1623 goto aqueue_fail;
1624
1625 AIO_LOCK(ki);
1626 job->jobflags &= ~KAIOCB_QUEUEING;
1627 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1628 ki->kaio_count++;
1629 if (lj)
1630 lj->lioj_count++;
1631 atomic_add_int(&num_queue_count, 1);
1632 if (job->jobflags & KAIOCB_FINISHED) {
1633 /*
1634 * The queue callback completed the request synchronously.
1635 * The bulk of the completion is deferred in that case
1636 * until this point.
1637 */
1638 aio_bio_done_notify(p, job);
1639 } else
1640 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1641 AIO_UNLOCK(ki);
1642 return (0);
1643
1644 aqueue_fail:
1645 knlist_delete(&job->klist, curthread, 0);
1646 if (fp)
1647 fdrop(fp, td);
1648 uma_zfree(aiocb_zone, job);
1649 ops->store_error(ujob, error);
1650 return (error);
1651 }
1652
1653 static void
1654 aio_cancel_daemon_job(struct kaiocb *job)
1655 {
1656
1657 mtx_lock(&aio_job_mtx);
1658 if (!aio_cancel_cleared(job))
1659 TAILQ_REMOVE(&aio_jobs, job, list);
1660 mtx_unlock(&aio_job_mtx);
1661 aio_cancel(job);
1662 }
1663
1664 void
1665 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1666 {
1667
1668 mtx_lock(&aio_job_mtx);
1669 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1670 mtx_unlock(&aio_job_mtx);
1671 aio_cancel(job);
1672 return;
1673 }
1674 job->handle_fn = func;
1675 TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1676 aio_kick_nowait(job->userproc);
1677 mtx_unlock(&aio_job_mtx);
1678 }
1679
1680 static void
1681 aio_cancel_sync(struct kaiocb *job)
1682 {
1683 struct kaioinfo *ki;
1684
1685 ki = job->userproc->p_aioinfo;
1686 AIO_LOCK(ki);
1687 if (!aio_cancel_cleared(job))
1688 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1689 AIO_UNLOCK(ki);
1690 aio_cancel(job);
1691 }
1692
1693 int
1694 aio_queue_file(struct file *fp, struct kaiocb *job)
1695 {
1696 struct aioliojob *lj;
1697 struct kaioinfo *ki;
1698 struct kaiocb *job2;
1699 struct vnode *vp;
1700 struct mount *mp;
1701 int error;
1702 bool safe;
1703
1704 lj = job->lio;
1705 ki = job->userproc->p_aioinfo;
1706 error = aio_qphysio(job->userproc, job);
1707 if (error >= 0)
1708 return (error);
1709 safe = false;
1710 if (fp->f_type == DTYPE_VNODE) {
1711 vp = fp->f_vnode;
1712 if (vp->v_type == VREG || vp->v_type == VDIR) {
1713 mp = fp->f_vnode->v_mount;
1714 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1715 safe = true;
1716 }
1717 }
1718 if (!(safe || enable_aio_unsafe)) {
1719 counted_warning(&unsafe_warningcnt,
1720 "is attempting to use unsafe AIO requests");
1721 return (EOPNOTSUPP);
1722 }
1723
1724 switch (job->uaiocb.aio_lio_opcode) {
1725 case LIO_READ:
1726 case LIO_WRITE:
1727 aio_schedule(job, aio_process_rw);
1728 error = 0;
1729 break;
1730 case LIO_SYNC:
1731 AIO_LOCK(ki);
1732 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1733 if (job2->fd_file == job->fd_file &&
1734 job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1735 job2->seqno < job->seqno) {
1736 job2->jobflags |= KAIOCB_CHECKSYNC;
1737 job->pending++;
1738 }
1739 }
1740 if (job->pending != 0) {
1741 if (!aio_set_cancel_function_locked(job,
1742 aio_cancel_sync)) {
1743 AIO_UNLOCK(ki);
1744 aio_cancel(job);
1745 return (0);
1746 }
1747 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1748 AIO_UNLOCK(ki);
1749 return (0);
1750 }
1751 AIO_UNLOCK(ki);
1752 aio_schedule(job, aio_process_sync);
1753 error = 0;
1754 break;
1755 default:
1756 error = EINVAL;
1757 }
1758 return (error);
1759 }
1760
1761 static void
1762 aio_kick_nowait(struct proc *userp)
1763 {
1764 struct kaioinfo *ki = userp->p_aioinfo;
1765 struct aioproc *aiop;
1766
1767 mtx_assert(&aio_job_mtx, MA_OWNED);
1768 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1769 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1770 aiop->aioprocflags &= ~AIOP_FREE;
1771 wakeup(aiop->aioproc);
1772 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1773 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1774 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1775 }
1776 }
1777
1778 static int
1779 aio_kick(struct proc *userp)
1780 {
1781 struct kaioinfo *ki = userp->p_aioinfo;
1782 struct aioproc *aiop;
1783 int error, ret = 0;
1784
1785 mtx_assert(&aio_job_mtx, MA_OWNED);
1786 retryproc:
1787 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1788 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1789 aiop->aioprocflags &= ~AIOP_FREE;
1790 wakeup(aiop->aioproc);
1791 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1792 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1793 num_aio_resv_start++;
1794 mtx_unlock(&aio_job_mtx);
1795 error = aio_newproc(&num_aio_resv_start);
1796 mtx_lock(&aio_job_mtx);
1797 if (error) {
1798 num_aio_resv_start--;
1799 goto retryproc;
1800 }
1801 } else {
1802 ret = -1;
1803 }
1804 return (ret);
1805 }
1806
1807 static void
1808 aio_kick_helper(void *context, int pending)
1809 {
1810 struct proc *userp = context;
1811
1812 mtx_lock(&aio_job_mtx);
1813 while (--pending >= 0) {
1814 if (aio_kick(userp))
1815 break;
1816 }
1817 mtx_unlock(&aio_job_mtx);
1818 }
1819
1820 /*
1821 * Support the aio_return system call, as a side-effect, kernel resources are
1822 * released.
1823 */
1824 static int
1825 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1826 {
1827 struct proc *p = td->td_proc;
1828 struct kaiocb *job;
1829 struct kaioinfo *ki;
1830 long status, error;
1831
1832 ki = p->p_aioinfo;
1833 if (ki == NULL)
1834 return (EINVAL);
1835 AIO_LOCK(ki);
1836 TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1837 if (job->ujob == ujob)
1838 break;
1839 }
1840 if (job != NULL) {
1841 MPASS(job->jobflags & KAIOCB_FINISHED);
1842 status = job->uaiocb._aiocb_private.status;
1843 error = job->uaiocb._aiocb_private.error;
1844 td->td_retval[0] = status;
1845 td->td_ru.ru_oublock += job->outblock;
1846 td->td_ru.ru_inblock += job->inblock;
1847 td->td_ru.ru_msgsnd += job->msgsnd;
1848 td->td_ru.ru_msgrcv += job->msgrcv;
1849 aio_free_entry(job);
1850 AIO_UNLOCK(ki);
1851 ops->store_error(ujob, error);
1852 ops->store_status(ujob, status);
1853 } else {
1854 error = EINVAL;
1855 AIO_UNLOCK(ki);
1856 }
1857 return (error);
1858 }
1859
1860 int
1861 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1862 {
1863
1864 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1865 }
1866
1867 /*
1868 * Allow a process to wakeup when any of the I/O requests are completed.
1869 */
1870 static int
1871 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1872 struct timespec *ts)
1873 {
1874 struct proc *p = td->td_proc;
1875 struct timeval atv;
1876 struct kaioinfo *ki;
1877 struct kaiocb *firstjob, *job;
1878 int error, i, timo;
1879
1880 timo = 0;
1881 if (ts) {
1882 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1883 return (EINVAL);
1884
1885 TIMESPEC_TO_TIMEVAL(&atv, ts);
1886 if (itimerfix(&atv))
1887 return (EINVAL);
1888 timo = tvtohz(&atv);
1889 }
1890
1891 ki = p->p_aioinfo;
1892 if (ki == NULL)
1893 return (EAGAIN);
1894
1895 if (njoblist == 0)
1896 return (0);
1897
1898 AIO_LOCK(ki);
1899 for (;;) {
1900 firstjob = NULL;
1901 error = 0;
1902 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1903 for (i = 0; i < njoblist; i++) {
1904 if (job->ujob == ujoblist[i]) {
1905 if (firstjob == NULL)
1906 firstjob = job;
1907 if (job->jobflags & KAIOCB_FINISHED)
1908 goto RETURN;
1909 }
1910 }
1911 }
1912 /* All tasks were finished. */
1913 if (firstjob == NULL)
1914 break;
1915
1916 ki->kaio_flags |= KAIO_WAKEUP;
1917 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1918 "aiospn", timo);
1919 if (error == ERESTART)
1920 error = EINTR;
1921 if (error)
1922 break;
1923 }
1924 RETURN:
1925 AIO_UNLOCK(ki);
1926 return (error);
1927 }
1928
1929 int
1930 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1931 {
1932 struct timespec ts, *tsp;
1933 struct aiocb **ujoblist;
1934 int error;
1935
1936 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1937 return (EINVAL);
1938
1939 if (uap->timeout) {
1940 /* Get timespec struct. */
1941 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1942 return (error);
1943 tsp = &ts;
1944 } else
1945 tsp = NULL;
1946
1947 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
1948 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1949 if (error == 0)
1950 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1951 free(ujoblist, M_AIOS);
1952 return (error);
1953 }
1954
1955 /*
1956 * aio_cancel cancels any non-physio aio operations not currently in
1957 * progress.
1958 */
1959 int
1960 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1961 {
1962 struct proc *p = td->td_proc;
1963 struct kaioinfo *ki;
1964 struct kaiocb *job, *jobn;
1965 struct file *fp;
1966 cap_rights_t rights;
1967 int error;
1968 int cancelled = 0;
1969 int notcancelled = 0;
1970 struct vnode *vp;
1971
1972 /* Lookup file object. */
1973 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1974 if (error)
1975 return (error);
1976
1977 ki = p->p_aioinfo;
1978 if (ki == NULL)
1979 goto done;
1980
1981 if (fp->f_type == DTYPE_VNODE) {
1982 vp = fp->f_vnode;
1983 if (vn_isdisk(vp, &error)) {
1984 fdrop(fp, td);
1985 td->td_retval[0] = AIO_NOTCANCELED;
1986 return (0);
1987 }
1988 }
1989
1990 AIO_LOCK(ki);
1991 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1992 if ((uap->fd == job->uaiocb.aio_fildes) &&
1993 ((uap->aiocbp == NULL) ||
1994 (uap->aiocbp == job->ujob))) {
1995 if (aio_cancel_job(p, ki, job)) {
1996 cancelled++;
1997 } else {
1998 notcancelled++;
1999 }
2000 if (uap->aiocbp != NULL)
2001 break;
2002 }
2003 }
2004 AIO_UNLOCK(ki);
2005
2006 done:
2007 fdrop(fp, td);
2008
2009 if (uap->aiocbp != NULL) {
2010 if (cancelled) {
2011 td->td_retval[0] = AIO_CANCELED;
2012 return (0);
2013 }
2014 }
2015
2016 if (notcancelled) {
2017 td->td_retval[0] = AIO_NOTCANCELED;
2018 return (0);
2019 }
2020
2021 if (cancelled) {
2022 td->td_retval[0] = AIO_CANCELED;
2023 return (0);
2024 }
2025
2026 td->td_retval[0] = AIO_ALLDONE;
2027
2028 return (0);
2029 }
2030
2031 /*
2032 * aio_error is implemented in the kernel level for compatibility purposes
2033 * only. For a user mode async implementation, it would be best to do it in
2034 * a userland subroutine.
2035 */
2036 static int
2037 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2038 {
2039 struct proc *p = td->td_proc;
2040 struct kaiocb *job;
2041 struct kaioinfo *ki;
2042 int status;
2043
2044 ki = p->p_aioinfo;
2045 if (ki == NULL) {
2046 td->td_retval[0] = EINVAL;
2047 return (0);
2048 }
2049
2050 AIO_LOCK(ki);
2051 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2052 if (job->ujob == ujob) {
2053 if (job->jobflags & KAIOCB_FINISHED)
2054 td->td_retval[0] =
2055 job->uaiocb._aiocb_private.error;
2056 else
2057 td->td_retval[0] = EINPROGRESS;
2058 AIO_UNLOCK(ki);
2059 return (0);
2060 }
2061 }
2062 AIO_UNLOCK(ki);
2063
2064 /*
2065 * Hack for failure of aio_aqueue.
2066 */
2067 status = ops->fetch_status(ujob);
2068 if (status == -1) {
2069 td->td_retval[0] = ops->fetch_error(ujob);
2070 return (0);
2071 }
2072
2073 td->td_retval[0] = EINVAL;
2074 return (0);
2075 }
2076
2077 int
2078 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2079 {
2080
2081 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2082 }
2083
2084 /* syscall - asynchronous read from a file (REALTIME) */
2085 #ifdef COMPAT_FREEBSD6
2086 int
2087 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2088 {
2089
2090 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2091 &aiocb_ops_osigevent));
2092 }
2093 #endif
2094
2095 int
2096 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2097 {
2098
2099 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2100 }
2101
2102 /* syscall - asynchronous write to a file (REALTIME) */
2103 #ifdef COMPAT_FREEBSD6
2104 int
2105 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2106 {
2107
2108 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2109 &aiocb_ops_osigevent));
2110 }
2111 #endif
2112
2113 int
2114 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2115 {
2116
2117 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2118 }
2119
2120 int
2121 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2122 {
2123
2124 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2125 }
2126
2127 static int
2128 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2129 struct aiocb **acb_list, int nent, struct sigevent *sig,
2130 struct aiocb_ops *ops)
2131 {
2132 struct proc *p = td->td_proc;
2133 struct aiocb *job;
2134 struct kaioinfo *ki;
2135 struct aioliojob *lj;
2136 struct kevent kev;
2137 int error;
2138 int nerror;
2139 int i;
2140
2141 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2142 return (EINVAL);
2143
2144 if (nent < 0 || nent > max_aio_queue_per_proc)
2145 return (EINVAL);
2146
2147 if (p->p_aioinfo == NULL)
2148 aio_init_aioinfo(p);
2149
2150 ki = p->p_aioinfo;
2151
2152 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2153 lj->lioj_flags = 0;
2154 lj->lioj_count = 0;
2155 lj->lioj_finished_count = 0;
2156 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2157 ksiginfo_init(&lj->lioj_ksi);
2158
2159 /*
2160 * Setup signal.
2161 */
2162 if (sig && (mode == LIO_NOWAIT)) {
2163 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2164 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2165 /* Assume only new style KEVENT */
2166 kev.filter = EVFILT_LIO;
2167 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2168 kev.ident = (uintptr_t)uacb_list; /* something unique */
2169 kev.data = (intptr_t)lj;
2170 /* pass user defined sigval data */
2171 kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2172 error = kqfd_register(
2173 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2174 if (error) {
2175 uma_zfree(aiolio_zone, lj);
2176 return (error);
2177 }
2178 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2179 ;
2180 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2181 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2182 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2183 uma_zfree(aiolio_zone, lj);
2184 return EINVAL;
2185 }
2186 lj->lioj_flags |= LIOJ_SIGNAL;
2187 } else {
2188 uma_zfree(aiolio_zone, lj);
2189 return EINVAL;
2190 }
2191 }
2192
2193 AIO_LOCK(ki);
2194 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2195 /*
2196 * Add extra aiocb count to avoid the lio to be freed
2197 * by other threads doing aio_waitcomplete or aio_return,
2198 * and prevent event from being sent until we have queued
2199 * all tasks.
2200 */
2201 lj->lioj_count = 1;
2202 AIO_UNLOCK(ki);
2203
2204 /*
2205 * Get pointers to the list of I/O requests.
2206 */
2207 nerror = 0;
2208 for (i = 0; i < nent; i++) {
2209 job = acb_list[i];
2210 if (job != NULL) {
2211 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2212 if (error != 0)
2213 nerror++;
2214 }
2215 }
2216
2217 error = 0;
2218 AIO_LOCK(ki);
2219 if (mode == LIO_WAIT) {
2220 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2221 ki->kaio_flags |= KAIO_WAKEUP;
2222 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2223 PRIBIO | PCATCH, "aiospn", 0);
2224 if (error == ERESTART)
2225 error = EINTR;
2226 if (error)
2227 break;
2228 }
2229 } else {
2230 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2231 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2232 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2233 KNOTE_LOCKED(&lj->klist, 1);
2234 }
2235 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2236 == LIOJ_SIGNAL
2237 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2238 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2239 aio_sendsig(p, &lj->lioj_signal,
2240 &lj->lioj_ksi);
2241 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2242 }
2243 }
2244 }
2245 lj->lioj_count--;
2246 if (lj->lioj_count == 0) {
2247 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2248 knlist_delete(&lj->klist, curthread, 1);
2249 PROC_LOCK(p);
2250 sigqueue_take(&lj->lioj_ksi);
2251 PROC_UNLOCK(p);
2252 AIO_UNLOCK(ki);
2253 uma_zfree(aiolio_zone, lj);
2254 } else
2255 AIO_UNLOCK(ki);
2256
2257 if (nerror)
2258 return (EIO);
2259 return (error);
2260 }
2261
2262 /* syscall - list directed I/O (REALTIME) */
2263 #ifdef COMPAT_FREEBSD6
2264 int
2265 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2266 {
2267 struct aiocb **acb_list;
2268 struct sigevent *sigp, sig;
2269 struct osigevent osig;
2270 int error, nent;
2271
2272 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2273 return (EINVAL);
2274
2275 nent = uap->nent;
2276 if (nent < 0 || nent > max_aio_queue_per_proc)
2277 return (EINVAL);
2278
2279 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2280 error = copyin(uap->sig, &osig, sizeof(osig));
2281 if (error)
2282 return (error);
2283 error = convert_old_sigevent(&osig, &sig);
2284 if (error)
2285 return (error);
2286 sigp = &sig;
2287 } else
2288 sigp = NULL;
2289
2290 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2291 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2292 if (error == 0)
2293 error = kern_lio_listio(td, uap->mode,
2294 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2295 &aiocb_ops_osigevent);
2296 free(acb_list, M_LIO);
2297 return (error);
2298 }
2299 #endif
2300
2301 /* syscall - list directed I/O (REALTIME) */
2302 int
2303 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2304 {
2305 struct aiocb **acb_list;
2306 struct sigevent *sigp, sig;
2307 int error, nent;
2308
2309 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2310 return (EINVAL);
2311
2312 nent = uap->nent;
2313 if (nent < 0 || nent > max_aio_queue_per_proc)
2314 return (EINVAL);
2315
2316 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2317 error = copyin(uap->sig, &sig, sizeof(sig));
2318 if (error)
2319 return (error);
2320 sigp = &sig;
2321 } else
2322 sigp = NULL;
2323
2324 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2325 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2326 if (error == 0)
2327 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2328 nent, sigp, &aiocb_ops);
2329 free(acb_list, M_LIO);
2330 return (error);
2331 }
2332
2333 static void
2334 aio_physwakeup(struct bio *bp)
2335 {
2336 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2337 struct proc *userp;
2338 struct kaioinfo *ki;
2339 size_t nbytes;
2340 int error, nblks;
2341
2342 /* Release mapping into kernel space. */
2343 userp = job->userproc;
2344 ki = userp->p_aioinfo;
2345 if (job->pbuf) {
2346 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2347 relpbuf(job->pbuf, NULL);
2348 job->pbuf = NULL;
2349 atomic_subtract_int(&num_buf_aio, 1);
2350 AIO_LOCK(ki);
2351 ki->kaio_buffer_count--;
2352 AIO_UNLOCK(ki);
2353 } else
2354 atomic_subtract_int(&num_unmapped_aio, 1);
2355 vm_page_unhold_pages(job->pages, job->npages);
2356
2357 bp = job->bp;
2358 job->bp = NULL;
2359 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2360 error = 0;
2361 if (bp->bio_flags & BIO_ERROR)
2362 error = bp->bio_error;
2363 nblks = btodb(nbytes);
2364 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2365 job->outblock += nblks;
2366 else
2367 job->inblock += nblks;
2368
2369 if (error)
2370 aio_complete(job, -1, error);
2371 else
2372 aio_complete(job, nbytes, 0);
2373
2374 g_destroy_bio(bp);
2375 }
2376
2377 /* syscall - wait for the next completion of an aio request */
2378 static int
2379 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2380 struct timespec *ts, struct aiocb_ops *ops)
2381 {
2382 struct proc *p = td->td_proc;
2383 struct timeval atv;
2384 struct kaioinfo *ki;
2385 struct kaiocb *job;
2386 struct aiocb *ujob;
2387 long error, status;
2388 int timo;
2389
2390 ops->store_aiocb(ujobp, NULL);
2391
2392 if (ts == NULL) {
2393 timo = 0;
2394 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2395 timo = -1;
2396 } else {
2397 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2398 return (EINVAL);
2399
2400 TIMESPEC_TO_TIMEVAL(&atv, ts);
2401 if (itimerfix(&atv))
2402 return (EINVAL);
2403 timo = tvtohz(&atv);
2404 }
2405
2406 if (p->p_aioinfo == NULL)
2407 aio_init_aioinfo(p);
2408 ki = p->p_aioinfo;
2409
2410 error = 0;
2411 job = NULL;
2412 AIO_LOCK(ki);
2413 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2414 if (timo == -1) {
2415 error = EWOULDBLOCK;
2416 break;
2417 }
2418 ki->kaio_flags |= KAIO_WAKEUP;
2419 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2420 "aiowc", timo);
2421 if (timo && error == ERESTART)
2422 error = EINTR;
2423 if (error)
2424 break;
2425 }
2426
2427 if (job != NULL) {
2428 MPASS(job->jobflags & KAIOCB_FINISHED);
2429 ujob = job->ujob;
2430 status = job->uaiocb._aiocb_private.status;
2431 error = job->uaiocb._aiocb_private.error;
2432 td->td_retval[0] = status;
2433 td->td_ru.ru_oublock += job->outblock;
2434 td->td_ru.ru_inblock += job->inblock;
2435 td->td_ru.ru_msgsnd += job->msgsnd;
2436 td->td_ru.ru_msgrcv += job->msgrcv;
2437 aio_free_entry(job);
2438 AIO_UNLOCK(ki);
2439 ops->store_aiocb(ujobp, ujob);
2440 ops->store_error(ujob, error);
2441 ops->store_status(ujob, status);
2442 } else
2443 AIO_UNLOCK(ki);
2444
2445 return (error);
2446 }
2447
2448 int
2449 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2450 {
2451 struct timespec ts, *tsp;
2452 int error;
2453
2454 if (uap->timeout) {
2455 /* Get timespec struct. */
2456 error = copyin(uap->timeout, &ts, sizeof(ts));
2457 if (error)
2458 return (error);
2459 tsp = &ts;
2460 } else
2461 tsp = NULL;
2462
2463 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2464 }
2465
2466 static int
2467 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2468 struct aiocb_ops *ops)
2469 {
2470
2471 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2472 return (EINVAL);
2473 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2474 }
2475
2476 int
2477 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2478 {
2479
2480 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2481 }
2482
2483 /* kqueue attach function */
2484 static int
2485 filt_aioattach(struct knote *kn)
2486 {
2487 struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2488
2489 /*
2490 * The job pointer must be validated before using it, so
2491 * registration is restricted to the kernel; the user cannot
2492 * set EV_FLAG1.
2493 */
2494 if ((kn->kn_flags & EV_FLAG1) == 0)
2495 return (EPERM);
2496 kn->kn_ptr.p_aio = job;
2497 kn->kn_flags &= ~EV_FLAG1;
2498
2499 knlist_add(&job->klist, kn, 0);
2500
2501 return (0);
2502 }
2503
2504 /* kqueue detach function */
2505 static void
2506 filt_aiodetach(struct knote *kn)
2507 {
2508 struct knlist *knl;
2509
2510 knl = &kn->kn_ptr.p_aio->klist;
2511 knl->kl_lock(knl->kl_lockarg);
2512 if (!knlist_empty(knl))
2513 knlist_remove(knl, kn, 1);
2514 knl->kl_unlock(knl->kl_lockarg);
2515 }
2516
2517 /* kqueue filter function */
2518 /*ARGSUSED*/
2519 static int
2520 filt_aio(struct knote *kn, long hint)
2521 {
2522 struct kaiocb *job = kn->kn_ptr.p_aio;
2523
2524 kn->kn_data = job->uaiocb._aiocb_private.error;
2525 if (!(job->jobflags & KAIOCB_FINISHED))
2526 return (0);
2527 kn->kn_flags |= EV_EOF;
2528 return (1);
2529 }
2530
2531 /* kqueue attach function */
2532 static int
2533 filt_lioattach(struct knote *kn)
2534 {
2535 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2536
2537 /*
2538 * The aioliojob pointer must be validated before using it, so
2539 * registration is restricted to the kernel; the user cannot
2540 * set EV_FLAG1.
2541 */
2542 if ((kn->kn_flags & EV_FLAG1) == 0)
2543 return (EPERM);
2544 kn->kn_ptr.p_lio = lj;
2545 kn->kn_flags &= ~EV_FLAG1;
2546
2547 knlist_add(&lj->klist, kn, 0);
2548
2549 return (0);
2550 }
2551
2552 /* kqueue detach function */
2553 static void
2554 filt_liodetach(struct knote *kn)
2555 {
2556 struct knlist *knl;
2557
2558 knl = &kn->kn_ptr.p_lio->klist;
2559 knl->kl_lock(knl->kl_lockarg);
2560 if (!knlist_empty(knl))
2561 knlist_remove(knl, kn, 1);
2562 knl->kl_unlock(knl->kl_lockarg);
2563 }
2564
2565 /* kqueue filter function */
2566 /*ARGSUSED*/
2567 static int
2568 filt_lio(struct knote *kn, long hint)
2569 {
2570 struct aioliojob * lj = kn->kn_ptr.p_lio;
2571
2572 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2573 }
2574
2575 #ifdef COMPAT_FREEBSD32
2576 #include <sys/mount.h>
2577 #include <sys/socket.h>
2578 #include <compat/freebsd32/freebsd32.h>
2579 #include <compat/freebsd32/freebsd32_proto.h>
2580 #include <compat/freebsd32/freebsd32_signal.h>
2581 #include <compat/freebsd32/freebsd32_syscall.h>
2582 #include <compat/freebsd32/freebsd32_util.h>
2583
2584 struct __aiocb_private32 {
2585 int32_t status;
2586 int32_t error;
2587 uint32_t kernelinfo;
2588 };
2589
2590 #ifdef COMPAT_FREEBSD6
2591 typedef struct oaiocb32 {
2592 int aio_fildes; /* File descriptor */
2593 uint64_t aio_offset __packed; /* File offset for I/O */
2594 uint32_t aio_buf; /* I/O buffer in process space */
2595 uint32_t aio_nbytes; /* Number of bytes for I/O */
2596 struct osigevent32 aio_sigevent; /* Signal to deliver */
2597 int aio_lio_opcode; /* LIO opcode */
2598 int aio_reqprio; /* Request priority -- ignored */
2599 struct __aiocb_private32 _aiocb_private;
2600 } oaiocb32_t;
2601 #endif
2602
2603 typedef struct aiocb32 {
2604 int32_t aio_fildes; /* File descriptor */
2605 uint64_t aio_offset __packed; /* File offset for I/O */
2606 uint32_t aio_buf; /* I/O buffer in process space */
2607 uint32_t aio_nbytes; /* Number of bytes for I/O */
2608 int __spare__[2];
2609 uint32_t __spare2__;
2610 int aio_lio_opcode; /* LIO opcode */
2611 int aio_reqprio; /* Request priority -- ignored */
2612 struct __aiocb_private32 _aiocb_private;
2613 struct sigevent32 aio_sigevent; /* Signal to deliver */
2614 } aiocb32_t;
2615
2616 #ifdef COMPAT_FREEBSD6
2617 static int
2618 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2619 {
2620
2621 /*
2622 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2623 * supported by AIO with the old sigevent structure.
2624 */
2625 CP(*osig, *nsig, sigev_notify);
2626 switch (nsig->sigev_notify) {
2627 case SIGEV_NONE:
2628 break;
2629 case SIGEV_SIGNAL:
2630 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2631 break;
2632 case SIGEV_KEVENT:
2633 nsig->sigev_notify_kqueue =
2634 osig->__sigev_u.__sigev_notify_kqueue;
2635 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2636 break;
2637 default:
2638 return (EINVAL);
2639 }
2640 return (0);
2641 }
2642
2643 static int
2644 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2645 {
2646 struct oaiocb32 job32;
2647 int error;
2648
2649 bzero(kjob, sizeof(struct aiocb));
2650 error = copyin(ujob, &job32, sizeof(job32));
2651 if (error)
2652 return (error);
2653
2654 CP(job32, *kjob, aio_fildes);
2655 CP(job32, *kjob, aio_offset);
2656 PTRIN_CP(job32, *kjob, aio_buf);
2657 CP(job32, *kjob, aio_nbytes);
2658 CP(job32, *kjob, aio_lio_opcode);
2659 CP(job32, *kjob, aio_reqprio);
2660 CP(job32, *kjob, _aiocb_private.status);
2661 CP(job32, *kjob, _aiocb_private.error);
2662 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2663 return (convert_old_sigevent32(&job32.aio_sigevent,
2664 &kjob->aio_sigevent));
2665 }
2666 #endif
2667
2668 static int
2669 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2670 {
2671 struct aiocb32 job32;
2672 int error;
2673
2674 error = copyin(ujob, &job32, sizeof(job32));
2675 if (error)
2676 return (error);
2677 CP(job32, *kjob, aio_fildes);
2678 CP(job32, *kjob, aio_offset);
2679 PTRIN_CP(job32, *kjob, aio_buf);
2680 CP(job32, *kjob, aio_nbytes);
2681 CP(job32, *kjob, aio_lio_opcode);
2682 CP(job32, *kjob, aio_reqprio);
2683 CP(job32, *kjob, _aiocb_private.status);
2684 CP(job32, *kjob, _aiocb_private.error);
2685 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2686 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2687 }
2688
2689 static long
2690 aiocb32_fetch_status(struct aiocb *ujob)
2691 {
2692 struct aiocb32 *ujob32;
2693
2694 ujob32 = (struct aiocb32 *)ujob;
2695 return (fuword32(&ujob32->_aiocb_private.status));
2696 }
2697
2698 static long
2699 aiocb32_fetch_error(struct aiocb *ujob)
2700 {
2701 struct aiocb32 *ujob32;
2702
2703 ujob32 = (struct aiocb32 *)ujob;
2704 return (fuword32(&ujob32->_aiocb_private.error));
2705 }
2706
2707 static int
2708 aiocb32_store_status(struct aiocb *ujob, long status)
2709 {
2710 struct aiocb32 *ujob32;
2711
2712 ujob32 = (struct aiocb32 *)ujob;
2713 return (suword32(&ujob32->_aiocb_private.status, status));
2714 }
2715
2716 static int
2717 aiocb32_store_error(struct aiocb *ujob, long error)
2718 {
2719 struct aiocb32 *ujob32;
2720
2721 ujob32 = (struct aiocb32 *)ujob;
2722 return (suword32(&ujob32->_aiocb_private.error, error));
2723 }
2724
2725 static int
2726 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2727 {
2728 struct aiocb32 *ujob32;
2729
2730 ujob32 = (struct aiocb32 *)ujob;
2731 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2732 }
2733
2734 static int
2735 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2736 {
2737
2738 return (suword32(ujobp, (long)ujob));
2739 }
2740
2741 static struct aiocb_ops aiocb32_ops = {
2742 .copyin = aiocb32_copyin,
2743 .fetch_status = aiocb32_fetch_status,
2744 .fetch_error = aiocb32_fetch_error,
2745 .store_status = aiocb32_store_status,
2746 .store_error = aiocb32_store_error,
2747 .store_kernelinfo = aiocb32_store_kernelinfo,
2748 .store_aiocb = aiocb32_store_aiocb,
2749 };
2750
2751 #ifdef COMPAT_FREEBSD6
2752 static struct aiocb_ops aiocb32_ops_osigevent = {
2753 .copyin = aiocb32_copyin_old_sigevent,
2754 .fetch_status = aiocb32_fetch_status,
2755 .fetch_error = aiocb32_fetch_error,
2756 .store_status = aiocb32_store_status,
2757 .store_error = aiocb32_store_error,
2758 .store_kernelinfo = aiocb32_store_kernelinfo,
2759 .store_aiocb = aiocb32_store_aiocb,
2760 };
2761 #endif
2762
2763 int
2764 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2765 {
2766
2767 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2768 }
2769
2770 int
2771 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2772 {
2773 struct timespec32 ts32;
2774 struct timespec ts, *tsp;
2775 struct aiocb **ujoblist;
2776 uint32_t *ujoblist32;
2777 int error, i;
2778
2779 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2780 return (EINVAL);
2781
2782 if (uap->timeout) {
2783 /* Get timespec struct. */
2784 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2785 return (error);
2786 CP(ts32, ts, tv_sec);
2787 CP(ts32, ts, tv_nsec);
2788 tsp = &ts;
2789 } else
2790 tsp = NULL;
2791
2792 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2793 ujoblist32 = (uint32_t *)ujoblist;
2794 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2795 sizeof(ujoblist32[0]));
2796 if (error == 0) {
2797 for (i = uap->nent - 1; i >= 0; i--)
2798 ujoblist[i] = PTRIN(ujoblist32[i]);
2799
2800 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2801 }
2802 free(ujoblist, M_AIOS);
2803 return (error);
2804 }
2805
2806 int
2807 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2808 {
2809
2810 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2811 }
2812
2813 #ifdef COMPAT_FREEBSD6
2814 int
2815 freebsd6_freebsd32_aio_read(struct thread *td,
2816 struct freebsd6_freebsd32_aio_read_args *uap)
2817 {
2818
2819 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2820 &aiocb32_ops_osigevent));
2821 }
2822 #endif
2823
2824 int
2825 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2826 {
2827
2828 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2829 &aiocb32_ops));
2830 }
2831
2832 #ifdef COMPAT_FREEBSD6
2833 int
2834 freebsd6_freebsd32_aio_write(struct thread *td,
2835 struct freebsd6_freebsd32_aio_write_args *uap)
2836 {
2837
2838 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2839 &aiocb32_ops_osigevent));
2840 }
2841 #endif
2842
2843 int
2844 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2845 {
2846
2847 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2848 &aiocb32_ops));
2849 }
2850
2851 int
2852 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2853 {
2854
2855 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2856 &aiocb32_ops));
2857 }
2858
2859 int
2860 freebsd32_aio_waitcomplete(struct thread *td,
2861 struct freebsd32_aio_waitcomplete_args *uap)
2862 {
2863 struct timespec32 ts32;
2864 struct timespec ts, *tsp;
2865 int error;
2866
2867 if (uap->timeout) {
2868 /* Get timespec struct. */
2869 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2870 if (error)
2871 return (error);
2872 CP(ts32, ts, tv_sec);
2873 CP(ts32, ts, tv_nsec);
2874 tsp = &ts;
2875 } else
2876 tsp = NULL;
2877
2878 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2879 &aiocb32_ops));
2880 }
2881
2882 int
2883 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2884 {
2885
2886 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2887 &aiocb32_ops));
2888 }
2889
2890 #ifdef COMPAT_FREEBSD6
2891 int
2892 freebsd6_freebsd32_lio_listio(struct thread *td,
2893 struct freebsd6_freebsd32_lio_listio_args *uap)
2894 {
2895 struct aiocb **acb_list;
2896 struct sigevent *sigp, sig;
2897 struct osigevent32 osig;
2898 uint32_t *acb_list32;
2899 int error, i, nent;
2900
2901 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2902 return (EINVAL);
2903
2904 nent = uap->nent;
2905 if (nent < 0 || nent > max_aio_queue_per_proc)
2906 return (EINVAL);
2907
2908 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2909 error = copyin(uap->sig, &osig, sizeof(osig));
2910 if (error)
2911 return (error);
2912 error = convert_old_sigevent32(&osig, &sig);
2913 if (error)
2914 return (error);
2915 sigp = &sig;
2916 } else
2917 sigp = NULL;
2918
2919 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2920 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2921 if (error) {
2922 free(acb_list32, M_LIO);
2923 return (error);
2924 }
2925 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2926 for (i = 0; i < nent; i++)
2927 acb_list[i] = PTRIN(acb_list32[i]);
2928 free(acb_list32, M_LIO);
2929
2930 error = kern_lio_listio(td, uap->mode,
2931 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2932 &aiocb32_ops_osigevent);
2933 free(acb_list, M_LIO);
2934 return (error);
2935 }
2936 #endif
2937
2938 int
2939 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2940 {
2941 struct aiocb **acb_list;
2942 struct sigevent *sigp, sig;
2943 struct sigevent32 sig32;
2944 uint32_t *acb_list32;
2945 int error, i, nent;
2946
2947 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2948 return (EINVAL);
2949
2950 nent = uap->nent;
2951 if (nent < 0 || nent > max_aio_queue_per_proc)
2952 return (EINVAL);
2953
2954 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2955 error = copyin(uap->sig, &sig32, sizeof(sig32));
2956 if (error)
2957 return (error);
2958 error = convert_sigevent32(&sig32, &sig);
2959 if (error)
2960 return (error);
2961 sigp = &sig;
2962 } else
2963 sigp = NULL;
2964
2965 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2966 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2967 if (error) {
2968 free(acb_list32, M_LIO);
2969 return (error);
2970 }
2971 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2972 for (i = 0; i < nent; i++)
2973 acb_list[i] = PTRIN(acb_list32[i]);
2974 free(acb_list32, M_LIO);
2975
2976 error = kern_lio_listio(td, uap->mode,
2977 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2978 &aiocb32_ops);
2979 free(acb_list, M_LIO);
2980 return (error);
2981 }
2982
2983 #endif
Cache object: 443dcdd909bdeebfe3d4e260a4d7c5dc
|