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/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_qbio() complete asynchronously 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 bio 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_biowakeup(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_qbio(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
aio_modload(struct module * module,int cmd,void * arg)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
aio_onceonly(void)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
aio_init_aioinfo(struct proc * p)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
aio_sendsig(struct proc * p,struct sigevent * sigev,ksiginfo_t * ksi)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
aio_free_entry(struct kaiocb * job)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
aio_proc_rundown_exec(void * arg,struct proc * p,struct image_params * imgp __unused)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
aio_cancel_job(struct proc * p,struct kaioinfo * ki,struct kaiocb * job)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
aio_proc_rundown(void * arg,struct proc * p)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 *
aio_selectjob(struct aioproc * aiop)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
aio_fsync_vnode(struct thread * td,struct vnode * vp)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-bio version of the operations. The normal
745 * vn operations are used, and this code should work in all instances for every
746 * 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
aio_process_rw(struct kaiocb * job)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
aio_process_sync(struct kaiocb * job)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
aio_process_mlock(struct kaiocb * job)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
aio_bio_done_notify(struct proc * userp,struct kaiocb * job)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
aio_schedule_fsync(void * context,int pending)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
aio_cancel_cleared(struct kaiocb * job)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
aio_clear_cancel_function_locked(struct kaiocb * job)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
aio_clear_cancel_function(struct kaiocb * job)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
aio_set_cancel_function_locked(struct kaiocb * job,aio_cancel_fn_t * func)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
aio_set_cancel_function(struct kaiocb * job,aio_cancel_fn_t * func)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
aio_complete(struct kaiocb * job,long status,int error)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
aio_cancel(struct kaiocb * job)1050 aio_cancel(struct kaiocb *job)
1051 {
1052
1053 aio_complete(job, -1, ECANCELED);
1054 }
1055
1056 void
aio_switch_vmspace(struct kaiocb * job)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
aio_daemon(void * _id)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
aio_newproc(int * start)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 bio 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
aio_qbio(struct proc * p,struct kaiocb * job)1209 aio_qbio(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_biowakeup;
1283 bp->bio_offset = cb->aio_offset;
1284 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1285 bp->bio_dev = dev;
1286 bp->bio_caller1 = (void *)job;
1287
1288 prot = VM_PROT_READ;
1289 if (cb->aio_lio_opcode == LIO_READ)
1290 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
1291 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1292 (vm_offset_t)cb->aio_buf, bp->bio_length, prot, job->pages,
1293 nitems(job->pages));
1294 if (job->npages < 0) {
1295 error = EFAULT;
1296 goto doerror;
1297 }
1298 if (pbuf != NULL) {
1299 pmap_qenter((vm_offset_t)pbuf->b_data,
1300 job->pages, job->npages);
1301 bp->bio_data = pbuf->b_data + poff;
1302 atomic_add_int(&num_buf_aio, 1);
1303 } else {
1304 bp->bio_ma = job->pages;
1305 bp->bio_ma_n = job->npages;
1306 bp->bio_ma_offset = poff;
1307 bp->bio_data = unmapped_buf;
1308 bp->bio_flags |= BIO_UNMAPPED;
1309 atomic_add_int(&num_unmapped_aio, 1);
1310 }
1311
1312 /* Perform transfer. */
1313 csw->d_strategy(bp);
1314 dev_relthread(dev, ref);
1315 return (0);
1316
1317 doerror:
1318 if (pbuf != NULL) {
1319 AIO_LOCK(ki);
1320 ki->kaio_buffer_count--;
1321 AIO_UNLOCK(ki);
1322 relpbuf(pbuf, NULL);
1323 job->pbuf = NULL;
1324 }
1325 g_destroy_bio(bp);
1326 job->bp = NULL;
1327 unref:
1328 dev_relthread(dev, ref);
1329 return (error);
1330 }
1331
1332 #ifdef COMPAT_FREEBSD6
1333 static int
convert_old_sigevent(struct osigevent * osig,struct sigevent * nsig)1334 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1335 {
1336
1337 /*
1338 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1339 * supported by AIO with the old sigevent structure.
1340 */
1341 nsig->sigev_notify = osig->sigev_notify;
1342 switch (nsig->sigev_notify) {
1343 case SIGEV_NONE:
1344 break;
1345 case SIGEV_SIGNAL:
1346 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1347 break;
1348 case SIGEV_KEVENT:
1349 nsig->sigev_notify_kqueue =
1350 osig->__sigev_u.__sigev_notify_kqueue;
1351 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1352 break;
1353 default:
1354 return (EINVAL);
1355 }
1356 return (0);
1357 }
1358
1359 static int
aiocb_copyin_old_sigevent(struct aiocb * ujob,struct aiocb * kjob)1360 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1361 {
1362 struct oaiocb *ojob;
1363 int error;
1364
1365 bzero(kjob, sizeof(struct aiocb));
1366 error = copyin(ujob, kjob, sizeof(struct oaiocb));
1367 if (error)
1368 return (error);
1369 ojob = (struct oaiocb *)kjob;
1370 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1371 }
1372 #endif
1373
1374 static int
aiocb_copyin(struct aiocb * ujob,struct aiocb * kjob)1375 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1376 {
1377
1378 return (copyin(ujob, kjob, sizeof(struct aiocb)));
1379 }
1380
1381 static long
aiocb_fetch_status(struct aiocb * ujob)1382 aiocb_fetch_status(struct aiocb *ujob)
1383 {
1384
1385 return (fuword(&ujob->_aiocb_private.status));
1386 }
1387
1388 static long
aiocb_fetch_error(struct aiocb * ujob)1389 aiocb_fetch_error(struct aiocb *ujob)
1390 {
1391
1392 return (fuword(&ujob->_aiocb_private.error));
1393 }
1394
1395 static int
aiocb_store_status(struct aiocb * ujob,long status)1396 aiocb_store_status(struct aiocb *ujob, long status)
1397 {
1398
1399 return (suword(&ujob->_aiocb_private.status, status));
1400 }
1401
1402 static int
aiocb_store_error(struct aiocb * ujob,long error)1403 aiocb_store_error(struct aiocb *ujob, long error)
1404 {
1405
1406 return (suword(&ujob->_aiocb_private.error, error));
1407 }
1408
1409 static int
aiocb_store_kernelinfo(struct aiocb * ujob,long jobref)1410 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1411 {
1412
1413 return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1414 }
1415
1416 static int
aiocb_store_aiocb(struct aiocb ** ujobp,struct aiocb * ujob)1417 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1418 {
1419
1420 return (suword(ujobp, (long)ujob));
1421 }
1422
1423 static struct aiocb_ops aiocb_ops = {
1424 .copyin = aiocb_copyin,
1425 .fetch_status = aiocb_fetch_status,
1426 .fetch_error = aiocb_fetch_error,
1427 .store_status = aiocb_store_status,
1428 .store_error = aiocb_store_error,
1429 .store_kernelinfo = aiocb_store_kernelinfo,
1430 .store_aiocb = aiocb_store_aiocb,
1431 };
1432
1433 #ifdef COMPAT_FREEBSD6
1434 static struct aiocb_ops aiocb_ops_osigevent = {
1435 .copyin = aiocb_copyin_old_sigevent,
1436 .fetch_status = aiocb_fetch_status,
1437 .fetch_error = aiocb_fetch_error,
1438 .store_status = aiocb_store_status,
1439 .store_error = aiocb_store_error,
1440 .store_kernelinfo = aiocb_store_kernelinfo,
1441 .store_aiocb = aiocb_store_aiocb,
1442 };
1443 #endif
1444
1445 /*
1446 * Queue a new AIO request. Choosing either the threaded or direct bio VCHR
1447 * technique is done in this code.
1448 */
1449 int
aio_aqueue(struct thread * td,struct aiocb * ujob,struct aioliojob * lj,int type,struct aiocb_ops * ops)1450 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1451 int type, struct aiocb_ops *ops)
1452 {
1453 struct proc *p = td->td_proc;
1454 cap_rights_t rights;
1455 struct file *fp;
1456 struct kaiocb *job;
1457 struct kaioinfo *ki;
1458 struct kevent kev;
1459 int opcode;
1460 int error;
1461 int fd, kqfd;
1462 int jid;
1463 u_short evflags;
1464
1465 if (p->p_aioinfo == NULL)
1466 aio_init_aioinfo(p);
1467
1468 ki = p->p_aioinfo;
1469
1470 ops->store_status(ujob, -1);
1471 ops->store_error(ujob, 0);
1472 ops->store_kernelinfo(ujob, -1);
1473
1474 if (num_queue_count >= max_queue_count ||
1475 ki->kaio_count >= max_aio_queue_per_proc) {
1476 ops->store_error(ujob, EAGAIN);
1477 return (EAGAIN);
1478 }
1479
1480 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1481 knlist_init_mtx(&job->klist, AIO_MTX(ki));
1482
1483 error = ops->copyin(ujob, &job->uaiocb);
1484 if (error) {
1485 ops->store_error(ujob, error);
1486 uma_zfree(aiocb_zone, job);
1487 return (error);
1488 }
1489
1490 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1491 uma_zfree(aiocb_zone, job);
1492 return (EINVAL);
1493 }
1494
1495 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1496 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1497 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1498 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1499 ops->store_error(ujob, EINVAL);
1500 uma_zfree(aiocb_zone, job);
1501 return (EINVAL);
1502 }
1503
1504 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1505 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1506 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1507 uma_zfree(aiocb_zone, job);
1508 return (EINVAL);
1509 }
1510
1511 ksiginfo_init(&job->ksi);
1512
1513 /* Save userspace address of the job info. */
1514 job->ujob = ujob;
1515
1516 /* Get the opcode. */
1517 if (type != LIO_NOP)
1518 job->uaiocb.aio_lio_opcode = type;
1519 opcode = job->uaiocb.aio_lio_opcode;
1520
1521 /*
1522 * Validate the opcode and fetch the file object for the specified
1523 * file descriptor.
1524 *
1525 * XXXRW: Moved the opcode validation up here so that we don't
1526 * retrieve a file descriptor without knowing what the capabiltity
1527 * should be.
1528 */
1529 fd = job->uaiocb.aio_fildes;
1530 switch (opcode) {
1531 case LIO_WRITE:
1532 error = fget_write(td, fd,
1533 cap_rights_init(&rights, CAP_PWRITE), &fp);
1534 break;
1535 case LIO_READ:
1536 error = fget_read(td, fd,
1537 cap_rights_init(&rights, CAP_PREAD), &fp);
1538 break;
1539 case LIO_SYNC:
1540 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1541 break;
1542 case LIO_MLOCK:
1543 fp = NULL;
1544 break;
1545 case LIO_NOP:
1546 error = fget(td, fd, cap_rights_init(&rights), &fp);
1547 break;
1548 default:
1549 error = EINVAL;
1550 }
1551 if (error) {
1552 uma_zfree(aiocb_zone, job);
1553 ops->store_error(ujob, error);
1554 return (error);
1555 }
1556
1557 if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1558 error = EINVAL;
1559 goto aqueue_fail;
1560 }
1561
1562 if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1563 job->uaiocb.aio_offset < 0 &&
1564 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1565 error = EINVAL;
1566 goto aqueue_fail;
1567 }
1568
1569 job->fd_file = fp;
1570
1571 mtx_lock(&aio_job_mtx);
1572 jid = jobrefid++;
1573 job->seqno = jobseqno++;
1574 mtx_unlock(&aio_job_mtx);
1575 error = ops->store_kernelinfo(ujob, jid);
1576 if (error) {
1577 error = EINVAL;
1578 goto aqueue_fail;
1579 }
1580 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1581
1582 if (opcode == LIO_NOP) {
1583 fdrop(fp, td);
1584 uma_zfree(aiocb_zone, job);
1585 return (0);
1586 }
1587
1588 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1589 goto no_kqueue;
1590 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1591 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1592 error = EINVAL;
1593 goto aqueue_fail;
1594 }
1595 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1596 memset(&kev, 0, sizeof(kev));
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
aio_cancel_daemon_job(struct kaiocb * job)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
aio_schedule(struct kaiocb * job,aio_handle_fn_t * func)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
aio_cancel_sync(struct kaiocb * job)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
aio_queue_file(struct file * fp,struct kaiocb * job)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_qbio(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
aio_kick_nowait(struct proc * userp)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
aio_kick(struct proc * userp)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
aio_kick_helper(void * context,int pending)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
kern_aio_return(struct thread * td,struct aiocb * ujob,struct aiocb_ops * ops)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
sys_aio_return(struct thread * td,struct aio_return_args * uap)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
kern_aio_suspend(struct thread * td,int njoblist,struct aiocb ** ujoblist,struct timespec * ts)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
sys_aio_suspend(struct thread * td,struct aio_suspend_args * uap)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-bio aio operations not currently in progress.
1957 */
1958 int
sys_aio_cancel(struct thread * td,struct aio_cancel_args * uap)1959 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1960 {
1961 struct proc *p = td->td_proc;
1962 struct kaioinfo *ki;
1963 struct kaiocb *job, *jobn;
1964 struct file *fp;
1965 cap_rights_t rights;
1966 int error;
1967 int cancelled = 0;
1968 int notcancelled = 0;
1969 struct vnode *vp;
1970
1971 /* Lookup file object. */
1972 error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1973 if (error)
1974 return (error);
1975
1976 ki = p->p_aioinfo;
1977 if (ki == NULL)
1978 goto done;
1979
1980 if (fp->f_type == DTYPE_VNODE) {
1981 vp = fp->f_vnode;
1982 if (vn_isdisk(vp, &error)) {
1983 fdrop(fp, td);
1984 td->td_retval[0] = AIO_NOTCANCELED;
1985 return (0);
1986 }
1987 }
1988
1989 AIO_LOCK(ki);
1990 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1991 if ((uap->fd == job->uaiocb.aio_fildes) &&
1992 ((uap->aiocbp == NULL) ||
1993 (uap->aiocbp == job->ujob))) {
1994 if (aio_cancel_job(p, ki, job)) {
1995 cancelled++;
1996 } else {
1997 notcancelled++;
1998 }
1999 if (uap->aiocbp != NULL)
2000 break;
2001 }
2002 }
2003 AIO_UNLOCK(ki);
2004
2005 done:
2006 fdrop(fp, td);
2007
2008 if (uap->aiocbp != NULL) {
2009 if (cancelled) {
2010 td->td_retval[0] = AIO_CANCELED;
2011 return (0);
2012 }
2013 }
2014
2015 if (notcancelled) {
2016 td->td_retval[0] = AIO_NOTCANCELED;
2017 return (0);
2018 }
2019
2020 if (cancelled) {
2021 td->td_retval[0] = AIO_CANCELED;
2022 return (0);
2023 }
2024
2025 td->td_retval[0] = AIO_ALLDONE;
2026
2027 return (0);
2028 }
2029
2030 /*
2031 * aio_error is implemented in the kernel level for compatibility purposes
2032 * only. For a user mode async implementation, it would be best to do it in
2033 * a userland subroutine.
2034 */
2035 static int
kern_aio_error(struct thread * td,struct aiocb * ujob,struct aiocb_ops * ops)2036 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2037 {
2038 struct proc *p = td->td_proc;
2039 struct kaiocb *job;
2040 struct kaioinfo *ki;
2041 int status;
2042
2043 ki = p->p_aioinfo;
2044 if (ki == NULL) {
2045 td->td_retval[0] = EINVAL;
2046 return (0);
2047 }
2048
2049 AIO_LOCK(ki);
2050 TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2051 if (job->ujob == ujob) {
2052 if (job->jobflags & KAIOCB_FINISHED)
2053 td->td_retval[0] =
2054 job->uaiocb._aiocb_private.error;
2055 else
2056 td->td_retval[0] = EINPROGRESS;
2057 AIO_UNLOCK(ki);
2058 return (0);
2059 }
2060 }
2061 AIO_UNLOCK(ki);
2062
2063 /*
2064 * Hack for failure of aio_aqueue.
2065 */
2066 status = ops->fetch_status(ujob);
2067 if (status == -1) {
2068 td->td_retval[0] = ops->fetch_error(ujob);
2069 return (0);
2070 }
2071
2072 td->td_retval[0] = EINVAL;
2073 return (0);
2074 }
2075
2076 int
sys_aio_error(struct thread * td,struct aio_error_args * uap)2077 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2078 {
2079
2080 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2081 }
2082
2083 /* syscall - asynchronous read from a file (REALTIME) */
2084 #ifdef COMPAT_FREEBSD6
2085 int
freebsd6_aio_read(struct thread * td,struct freebsd6_aio_read_args * uap)2086 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2087 {
2088
2089 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2090 &aiocb_ops_osigevent));
2091 }
2092 #endif
2093
2094 int
sys_aio_read(struct thread * td,struct aio_read_args * uap)2095 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2096 {
2097
2098 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2099 }
2100
2101 /* syscall - asynchronous write to a file (REALTIME) */
2102 #ifdef COMPAT_FREEBSD6
2103 int
freebsd6_aio_write(struct thread * td,struct freebsd6_aio_write_args * uap)2104 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2105 {
2106
2107 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2108 &aiocb_ops_osigevent));
2109 }
2110 #endif
2111
2112 int
sys_aio_write(struct thread * td,struct aio_write_args * uap)2113 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2114 {
2115
2116 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2117 }
2118
2119 int
sys_aio_mlock(struct thread * td,struct aio_mlock_args * uap)2120 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2121 {
2122
2123 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2124 }
2125
2126 static int
kern_lio_listio(struct thread * td,int mode,struct aiocb * const * uacb_list,struct aiocb ** acb_list,int nent,struct sigevent * sig,struct aiocb_ops * ops)2127 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2128 struct aiocb **acb_list, int nent, struct sigevent *sig,
2129 struct aiocb_ops *ops)
2130 {
2131 struct proc *p = td->td_proc;
2132 struct aiocb *job;
2133 struct kaioinfo *ki;
2134 struct aioliojob *lj;
2135 struct kevent kev;
2136 int error;
2137 int nagain, nerror;
2138 int i;
2139
2140 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2141 return (EINVAL);
2142
2143 if (nent < 0 || nent > max_aio_queue_per_proc)
2144 return (EINVAL);
2145
2146 if (p->p_aioinfo == NULL)
2147 aio_init_aioinfo(p);
2148
2149 ki = p->p_aioinfo;
2150
2151 lj = uma_zalloc(aiolio_zone, M_WAITOK);
2152 lj->lioj_flags = 0;
2153 lj->lioj_count = 0;
2154 lj->lioj_finished_count = 0;
2155 knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2156 ksiginfo_init(&lj->lioj_ksi);
2157
2158 /*
2159 * Setup signal.
2160 */
2161 if (sig && (mode == LIO_NOWAIT)) {
2162 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2163 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2164 /* Assume only new style KEVENT */
2165 memset(&kev, 0, sizeof(kev));
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 nagain = 0;
2208 nerror = 0;
2209 for (i = 0; i < nent; i++) {
2210 job = acb_list[i];
2211 if (job != NULL) {
2212 error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2213 if (error == EAGAIN)
2214 nagain++;
2215 else if (error != 0)
2216 nerror++;
2217 }
2218 }
2219
2220 error = 0;
2221 AIO_LOCK(ki);
2222 if (mode == LIO_WAIT) {
2223 while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2224 ki->kaio_flags |= KAIO_WAKEUP;
2225 error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2226 PRIBIO | PCATCH, "aiospn", 0);
2227 if (error == ERESTART)
2228 error = EINTR;
2229 if (error)
2230 break;
2231 }
2232 } else {
2233 if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2234 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2235 lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2236 KNOTE_LOCKED(&lj->klist, 1);
2237 }
2238 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2239 == LIOJ_SIGNAL
2240 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2241 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2242 aio_sendsig(p, &lj->lioj_signal,
2243 &lj->lioj_ksi);
2244 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2245 }
2246 }
2247 }
2248 lj->lioj_count--;
2249 if (lj->lioj_count == 0) {
2250 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2251 knlist_delete(&lj->klist, curthread, 1);
2252 PROC_LOCK(p);
2253 sigqueue_take(&lj->lioj_ksi);
2254 PROC_UNLOCK(p);
2255 AIO_UNLOCK(ki);
2256 uma_zfree(aiolio_zone, lj);
2257 } else
2258 AIO_UNLOCK(ki);
2259
2260 if (nerror)
2261 return (EIO);
2262 else if (nagain)
2263 return (EAGAIN);
2264 else
2265 return (error);
2266 }
2267
2268 /* syscall - list directed I/O (REALTIME) */
2269 #ifdef COMPAT_FREEBSD6
2270 int
freebsd6_lio_listio(struct thread * td,struct freebsd6_lio_listio_args * uap)2271 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2272 {
2273 struct aiocb **acb_list;
2274 struct sigevent *sigp, sig;
2275 struct osigevent osig;
2276 int error, nent;
2277
2278 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2279 return (EINVAL);
2280
2281 nent = uap->nent;
2282 if (nent < 0 || nent > max_aio_queue_per_proc)
2283 return (EINVAL);
2284
2285 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2286 error = copyin(uap->sig, &osig, sizeof(osig));
2287 if (error)
2288 return (error);
2289 error = convert_old_sigevent(&osig, &sig);
2290 if (error)
2291 return (error);
2292 sigp = &sig;
2293 } else
2294 sigp = NULL;
2295
2296 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2297 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2298 if (error == 0)
2299 error = kern_lio_listio(td, uap->mode,
2300 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2301 &aiocb_ops_osigevent);
2302 free(acb_list, M_LIO);
2303 return (error);
2304 }
2305 #endif
2306
2307 /* syscall - list directed I/O (REALTIME) */
2308 int
sys_lio_listio(struct thread * td,struct lio_listio_args * uap)2309 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2310 {
2311 struct aiocb **acb_list;
2312 struct sigevent *sigp, sig;
2313 int error, nent;
2314
2315 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2316 return (EINVAL);
2317
2318 nent = uap->nent;
2319 if (nent < 0 || nent > max_aio_queue_per_proc)
2320 return (EINVAL);
2321
2322 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2323 error = copyin(uap->sig, &sig, sizeof(sig));
2324 if (error)
2325 return (error);
2326 sigp = &sig;
2327 } else
2328 sigp = NULL;
2329
2330 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2331 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2332 if (error == 0)
2333 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2334 nent, sigp, &aiocb_ops);
2335 free(acb_list, M_LIO);
2336 return (error);
2337 }
2338
2339 static void
aio_biowakeup(struct bio * bp)2340 aio_biowakeup(struct bio *bp)
2341 {
2342 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2343 struct proc *userp;
2344 struct kaioinfo *ki;
2345 size_t nbytes;
2346 int error, nblks;
2347
2348 /* Release mapping into kernel space. */
2349 userp = job->userproc;
2350 ki = userp->p_aioinfo;
2351 if (job->pbuf) {
2352 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2353 relpbuf(job->pbuf, NULL);
2354 job->pbuf = NULL;
2355 atomic_subtract_int(&num_buf_aio, 1);
2356 AIO_LOCK(ki);
2357 ki->kaio_buffer_count--;
2358 AIO_UNLOCK(ki);
2359 } else
2360 atomic_subtract_int(&num_unmapped_aio, 1);
2361 vm_page_unhold_pages(job->pages, job->npages);
2362
2363 bp = job->bp;
2364 job->bp = NULL;
2365 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2366 error = 0;
2367 if (bp->bio_flags & BIO_ERROR)
2368 error = bp->bio_error;
2369 nblks = btodb(nbytes);
2370 if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2371 job->outblock += nblks;
2372 else
2373 job->inblock += nblks;
2374
2375 if (error)
2376 aio_complete(job, -1, error);
2377 else
2378 aio_complete(job, nbytes, 0);
2379
2380 g_destroy_bio(bp);
2381 }
2382
2383 /* syscall - wait for the next completion of an aio request */
2384 static int
kern_aio_waitcomplete(struct thread * td,struct aiocb ** ujobp,struct timespec * ts,struct aiocb_ops * ops)2385 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2386 struct timespec *ts, struct aiocb_ops *ops)
2387 {
2388 struct proc *p = td->td_proc;
2389 struct timeval atv;
2390 struct kaioinfo *ki;
2391 struct kaiocb *job;
2392 struct aiocb *ujob;
2393 long error, status;
2394 int timo;
2395
2396 ops->store_aiocb(ujobp, NULL);
2397
2398 if (ts == NULL) {
2399 timo = 0;
2400 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2401 timo = -1;
2402 } else {
2403 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2404 return (EINVAL);
2405
2406 TIMESPEC_TO_TIMEVAL(&atv, ts);
2407 if (itimerfix(&atv))
2408 return (EINVAL);
2409 timo = tvtohz(&atv);
2410 }
2411
2412 if (p->p_aioinfo == NULL)
2413 aio_init_aioinfo(p);
2414 ki = p->p_aioinfo;
2415
2416 error = 0;
2417 job = NULL;
2418 AIO_LOCK(ki);
2419 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2420 if (timo == -1) {
2421 error = EWOULDBLOCK;
2422 break;
2423 }
2424 ki->kaio_flags |= KAIO_WAKEUP;
2425 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2426 "aiowc", timo);
2427 if (timo && error == ERESTART)
2428 error = EINTR;
2429 if (error)
2430 break;
2431 }
2432
2433 if (job != NULL) {
2434 MPASS(job->jobflags & KAIOCB_FINISHED);
2435 ujob = job->ujob;
2436 status = job->uaiocb._aiocb_private.status;
2437 error = job->uaiocb._aiocb_private.error;
2438 td->td_retval[0] = status;
2439 td->td_ru.ru_oublock += job->outblock;
2440 td->td_ru.ru_inblock += job->inblock;
2441 td->td_ru.ru_msgsnd += job->msgsnd;
2442 td->td_ru.ru_msgrcv += job->msgrcv;
2443 aio_free_entry(job);
2444 AIO_UNLOCK(ki);
2445 ops->store_aiocb(ujobp, ujob);
2446 ops->store_error(ujob, error);
2447 ops->store_status(ujob, status);
2448 } else
2449 AIO_UNLOCK(ki);
2450
2451 return (error);
2452 }
2453
2454 int
sys_aio_waitcomplete(struct thread * td,struct aio_waitcomplete_args * uap)2455 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2456 {
2457 struct timespec ts, *tsp;
2458 int error;
2459
2460 if (uap->timeout) {
2461 /* Get timespec struct. */
2462 error = copyin(uap->timeout, &ts, sizeof(ts));
2463 if (error)
2464 return (error);
2465 tsp = &ts;
2466 } else
2467 tsp = NULL;
2468
2469 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2470 }
2471
2472 static int
kern_aio_fsync(struct thread * td,int op,struct aiocb * ujob,struct aiocb_ops * ops)2473 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2474 struct aiocb_ops *ops)
2475 {
2476
2477 if (op != O_SYNC) /* XXX lack of O_DSYNC */
2478 return (EINVAL);
2479 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2480 }
2481
2482 int
sys_aio_fsync(struct thread * td,struct aio_fsync_args * uap)2483 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2484 {
2485
2486 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2487 }
2488
2489 /* kqueue attach function */
2490 static int
filt_aioattach(struct knote * kn)2491 filt_aioattach(struct knote *kn)
2492 {
2493 struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2494
2495 /*
2496 * The job pointer must be validated before using it, so
2497 * registration is restricted to the kernel; the user cannot
2498 * set EV_FLAG1.
2499 */
2500 if ((kn->kn_flags & EV_FLAG1) == 0)
2501 return (EPERM);
2502 kn->kn_ptr.p_aio = job;
2503 kn->kn_flags &= ~EV_FLAG1;
2504
2505 knlist_add(&job->klist, kn, 0);
2506
2507 return (0);
2508 }
2509
2510 /* kqueue detach function */
2511 static void
filt_aiodetach(struct knote * kn)2512 filt_aiodetach(struct knote *kn)
2513 {
2514 struct knlist *knl;
2515
2516 knl = &kn->kn_ptr.p_aio->klist;
2517 knl->kl_lock(knl->kl_lockarg);
2518 if (!knlist_empty(knl))
2519 knlist_remove(knl, kn, 1);
2520 knl->kl_unlock(knl->kl_lockarg);
2521 }
2522
2523 /* kqueue filter function */
2524 /*ARGSUSED*/
2525 static int
filt_aio(struct knote * kn,long hint)2526 filt_aio(struct knote *kn, long hint)
2527 {
2528 struct kaiocb *job = kn->kn_ptr.p_aio;
2529
2530 kn->kn_data = job->uaiocb._aiocb_private.error;
2531 if (!(job->jobflags & KAIOCB_FINISHED))
2532 return (0);
2533 kn->kn_flags |= EV_EOF;
2534 return (1);
2535 }
2536
2537 /* kqueue attach function */
2538 static int
filt_lioattach(struct knote * kn)2539 filt_lioattach(struct knote *kn)
2540 {
2541 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2542
2543 /*
2544 * The aioliojob pointer must be validated before using it, so
2545 * registration is restricted to the kernel; the user cannot
2546 * set EV_FLAG1.
2547 */
2548 if ((kn->kn_flags & EV_FLAG1) == 0)
2549 return (EPERM);
2550 kn->kn_ptr.p_lio = lj;
2551 kn->kn_flags &= ~EV_FLAG1;
2552
2553 knlist_add(&lj->klist, kn, 0);
2554
2555 return (0);
2556 }
2557
2558 /* kqueue detach function */
2559 static void
filt_liodetach(struct knote * kn)2560 filt_liodetach(struct knote *kn)
2561 {
2562 struct knlist *knl;
2563
2564 knl = &kn->kn_ptr.p_lio->klist;
2565 knl->kl_lock(knl->kl_lockarg);
2566 if (!knlist_empty(knl))
2567 knlist_remove(knl, kn, 1);
2568 knl->kl_unlock(knl->kl_lockarg);
2569 }
2570
2571 /* kqueue filter function */
2572 /*ARGSUSED*/
2573 static int
filt_lio(struct knote * kn,long hint)2574 filt_lio(struct knote *kn, long hint)
2575 {
2576 struct aioliojob * lj = kn->kn_ptr.p_lio;
2577
2578 return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2579 }
2580
2581 #ifdef COMPAT_FREEBSD32
2582 #include <sys/mount.h>
2583 #include <sys/socket.h>
2584 #include <compat/freebsd32/freebsd32.h>
2585 #include <compat/freebsd32/freebsd32_proto.h>
2586 #include <compat/freebsd32/freebsd32_signal.h>
2587 #include <compat/freebsd32/freebsd32_syscall.h>
2588 #include <compat/freebsd32/freebsd32_util.h>
2589
2590 struct __aiocb_private32 {
2591 int32_t status;
2592 int32_t error;
2593 uint32_t kernelinfo;
2594 };
2595
2596 #ifdef COMPAT_FREEBSD6
2597 typedef struct oaiocb32 {
2598 int aio_fildes; /* File descriptor */
2599 uint64_t aio_offset __packed; /* File offset for I/O */
2600 uint32_t aio_buf; /* I/O buffer in process space */
2601 uint32_t aio_nbytes; /* Number of bytes for I/O */
2602 struct osigevent32 aio_sigevent; /* Signal to deliver */
2603 int aio_lio_opcode; /* LIO opcode */
2604 int aio_reqprio; /* Request priority -- ignored */
2605 struct __aiocb_private32 _aiocb_private;
2606 } oaiocb32_t;
2607 #endif
2608
2609 typedef struct aiocb32 {
2610 int32_t aio_fildes; /* File descriptor */
2611 uint64_t aio_offset __packed; /* File offset for I/O */
2612 uint32_t aio_buf; /* I/O buffer in process space */
2613 uint32_t aio_nbytes; /* Number of bytes for I/O */
2614 int __spare__[2];
2615 uint32_t __spare2__;
2616 int aio_lio_opcode; /* LIO opcode */
2617 int aio_reqprio; /* Request priority -- ignored */
2618 struct __aiocb_private32 _aiocb_private;
2619 struct sigevent32 aio_sigevent; /* Signal to deliver */
2620 } aiocb32_t;
2621
2622 #ifdef COMPAT_FREEBSD6
2623 static int
convert_old_sigevent32(struct osigevent32 * osig,struct sigevent * nsig)2624 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2625 {
2626
2627 /*
2628 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2629 * supported by AIO with the old sigevent structure.
2630 */
2631 CP(*osig, *nsig, sigev_notify);
2632 switch (nsig->sigev_notify) {
2633 case SIGEV_NONE:
2634 break;
2635 case SIGEV_SIGNAL:
2636 nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2637 break;
2638 case SIGEV_KEVENT:
2639 nsig->sigev_notify_kqueue =
2640 osig->__sigev_u.__sigev_notify_kqueue;
2641 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2642 break;
2643 default:
2644 return (EINVAL);
2645 }
2646 return (0);
2647 }
2648
2649 static int
aiocb32_copyin_old_sigevent(struct aiocb * ujob,struct aiocb * kjob)2650 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2651 {
2652 struct oaiocb32 job32;
2653 int error;
2654
2655 bzero(kjob, sizeof(struct aiocb));
2656 error = copyin(ujob, &job32, sizeof(job32));
2657 if (error)
2658 return (error);
2659
2660 CP(job32, *kjob, aio_fildes);
2661 CP(job32, *kjob, aio_offset);
2662 PTRIN_CP(job32, *kjob, aio_buf);
2663 CP(job32, *kjob, aio_nbytes);
2664 CP(job32, *kjob, aio_lio_opcode);
2665 CP(job32, *kjob, aio_reqprio);
2666 CP(job32, *kjob, _aiocb_private.status);
2667 CP(job32, *kjob, _aiocb_private.error);
2668 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2669 return (convert_old_sigevent32(&job32.aio_sigevent,
2670 &kjob->aio_sigevent));
2671 }
2672 #endif
2673
2674 static int
aiocb32_copyin(struct aiocb * ujob,struct aiocb * kjob)2675 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2676 {
2677 struct aiocb32 job32;
2678 int error;
2679
2680 error = copyin(ujob, &job32, sizeof(job32));
2681 if (error)
2682 return (error);
2683 CP(job32, *kjob, aio_fildes);
2684 CP(job32, *kjob, aio_offset);
2685 PTRIN_CP(job32, *kjob, aio_buf);
2686 CP(job32, *kjob, aio_nbytes);
2687 CP(job32, *kjob, aio_lio_opcode);
2688 CP(job32, *kjob, aio_reqprio);
2689 CP(job32, *kjob, _aiocb_private.status);
2690 CP(job32, *kjob, _aiocb_private.error);
2691 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2692 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2693 }
2694
2695 static long
aiocb32_fetch_status(struct aiocb * ujob)2696 aiocb32_fetch_status(struct aiocb *ujob)
2697 {
2698 struct aiocb32 *ujob32;
2699
2700 ujob32 = (struct aiocb32 *)ujob;
2701 return (fuword32(&ujob32->_aiocb_private.status));
2702 }
2703
2704 static long
aiocb32_fetch_error(struct aiocb * ujob)2705 aiocb32_fetch_error(struct aiocb *ujob)
2706 {
2707 struct aiocb32 *ujob32;
2708
2709 ujob32 = (struct aiocb32 *)ujob;
2710 return (fuword32(&ujob32->_aiocb_private.error));
2711 }
2712
2713 static int
aiocb32_store_status(struct aiocb * ujob,long status)2714 aiocb32_store_status(struct aiocb *ujob, long status)
2715 {
2716 struct aiocb32 *ujob32;
2717
2718 ujob32 = (struct aiocb32 *)ujob;
2719 return (suword32(&ujob32->_aiocb_private.status, status));
2720 }
2721
2722 static int
aiocb32_store_error(struct aiocb * ujob,long error)2723 aiocb32_store_error(struct aiocb *ujob, long error)
2724 {
2725 struct aiocb32 *ujob32;
2726
2727 ujob32 = (struct aiocb32 *)ujob;
2728 return (suword32(&ujob32->_aiocb_private.error, error));
2729 }
2730
2731 static int
aiocb32_store_kernelinfo(struct aiocb * ujob,long jobref)2732 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2733 {
2734 struct aiocb32 *ujob32;
2735
2736 ujob32 = (struct aiocb32 *)ujob;
2737 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2738 }
2739
2740 static int
aiocb32_store_aiocb(struct aiocb ** ujobp,struct aiocb * ujob)2741 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2742 {
2743
2744 return (suword32(ujobp, (long)ujob));
2745 }
2746
2747 static struct aiocb_ops aiocb32_ops = {
2748 .copyin = aiocb32_copyin,
2749 .fetch_status = aiocb32_fetch_status,
2750 .fetch_error = aiocb32_fetch_error,
2751 .store_status = aiocb32_store_status,
2752 .store_error = aiocb32_store_error,
2753 .store_kernelinfo = aiocb32_store_kernelinfo,
2754 .store_aiocb = aiocb32_store_aiocb,
2755 };
2756
2757 #ifdef COMPAT_FREEBSD6
2758 static struct aiocb_ops aiocb32_ops_osigevent = {
2759 .copyin = aiocb32_copyin_old_sigevent,
2760 .fetch_status = aiocb32_fetch_status,
2761 .fetch_error = aiocb32_fetch_error,
2762 .store_status = aiocb32_store_status,
2763 .store_error = aiocb32_store_error,
2764 .store_kernelinfo = aiocb32_store_kernelinfo,
2765 .store_aiocb = aiocb32_store_aiocb,
2766 };
2767 #endif
2768
2769 int
freebsd32_aio_return(struct thread * td,struct freebsd32_aio_return_args * uap)2770 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2771 {
2772
2773 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2774 }
2775
2776 int
freebsd32_aio_suspend(struct thread * td,struct freebsd32_aio_suspend_args * uap)2777 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2778 {
2779 struct timespec32 ts32;
2780 struct timespec ts, *tsp;
2781 struct aiocb **ujoblist;
2782 uint32_t *ujoblist32;
2783 int error, i;
2784
2785 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2786 return (EINVAL);
2787
2788 if (uap->timeout) {
2789 /* Get timespec struct. */
2790 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2791 return (error);
2792 CP(ts32, ts, tv_sec);
2793 CP(ts32, ts, tv_nsec);
2794 tsp = &ts;
2795 } else
2796 tsp = NULL;
2797
2798 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2799 ujoblist32 = (uint32_t *)ujoblist;
2800 error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2801 sizeof(ujoblist32[0]));
2802 if (error == 0) {
2803 for (i = uap->nent - 1; i >= 0; i--)
2804 ujoblist[i] = PTRIN(ujoblist32[i]);
2805
2806 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2807 }
2808 free(ujoblist, M_AIOS);
2809 return (error);
2810 }
2811
2812 int
freebsd32_aio_error(struct thread * td,struct freebsd32_aio_error_args * uap)2813 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2814 {
2815
2816 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2817 }
2818
2819 #ifdef COMPAT_FREEBSD6
2820 int
freebsd6_freebsd32_aio_read(struct thread * td,struct freebsd6_freebsd32_aio_read_args * uap)2821 freebsd6_freebsd32_aio_read(struct thread *td,
2822 struct freebsd6_freebsd32_aio_read_args *uap)
2823 {
2824
2825 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2826 &aiocb32_ops_osigevent));
2827 }
2828 #endif
2829
2830 int
freebsd32_aio_read(struct thread * td,struct freebsd32_aio_read_args * uap)2831 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2832 {
2833
2834 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2835 &aiocb32_ops));
2836 }
2837
2838 #ifdef COMPAT_FREEBSD6
2839 int
freebsd6_freebsd32_aio_write(struct thread * td,struct freebsd6_freebsd32_aio_write_args * uap)2840 freebsd6_freebsd32_aio_write(struct thread *td,
2841 struct freebsd6_freebsd32_aio_write_args *uap)
2842 {
2843
2844 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2845 &aiocb32_ops_osigevent));
2846 }
2847 #endif
2848
2849 int
freebsd32_aio_write(struct thread * td,struct freebsd32_aio_write_args * uap)2850 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2851 {
2852
2853 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2854 &aiocb32_ops));
2855 }
2856
2857 int
freebsd32_aio_mlock(struct thread * td,struct freebsd32_aio_mlock_args * uap)2858 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2859 {
2860
2861 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2862 &aiocb32_ops));
2863 }
2864
2865 int
freebsd32_aio_waitcomplete(struct thread * td,struct freebsd32_aio_waitcomplete_args * uap)2866 freebsd32_aio_waitcomplete(struct thread *td,
2867 struct freebsd32_aio_waitcomplete_args *uap)
2868 {
2869 struct timespec32 ts32;
2870 struct timespec ts, *tsp;
2871 int error;
2872
2873 if (uap->timeout) {
2874 /* Get timespec struct. */
2875 error = copyin(uap->timeout, &ts32, sizeof(ts32));
2876 if (error)
2877 return (error);
2878 CP(ts32, ts, tv_sec);
2879 CP(ts32, ts, tv_nsec);
2880 tsp = &ts;
2881 } else
2882 tsp = NULL;
2883
2884 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2885 &aiocb32_ops));
2886 }
2887
2888 int
freebsd32_aio_fsync(struct thread * td,struct freebsd32_aio_fsync_args * uap)2889 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2890 {
2891
2892 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2893 &aiocb32_ops));
2894 }
2895
2896 #ifdef COMPAT_FREEBSD6
2897 int
freebsd6_freebsd32_lio_listio(struct thread * td,struct freebsd6_freebsd32_lio_listio_args * uap)2898 freebsd6_freebsd32_lio_listio(struct thread *td,
2899 struct freebsd6_freebsd32_lio_listio_args *uap)
2900 {
2901 struct aiocb **acb_list;
2902 struct sigevent *sigp, sig;
2903 struct osigevent32 osig;
2904 uint32_t *acb_list32;
2905 int error, i, nent;
2906
2907 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2908 return (EINVAL);
2909
2910 nent = uap->nent;
2911 if (nent < 0 || nent > max_aio_queue_per_proc)
2912 return (EINVAL);
2913
2914 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2915 error = copyin(uap->sig, &osig, sizeof(osig));
2916 if (error)
2917 return (error);
2918 error = convert_old_sigevent32(&osig, &sig);
2919 if (error)
2920 return (error);
2921 sigp = &sig;
2922 } else
2923 sigp = NULL;
2924
2925 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2926 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2927 if (error) {
2928 free(acb_list32, M_LIO);
2929 return (error);
2930 }
2931 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2932 for (i = 0; i < nent; i++)
2933 acb_list[i] = PTRIN(acb_list32[i]);
2934 free(acb_list32, M_LIO);
2935
2936 error = kern_lio_listio(td, uap->mode,
2937 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2938 &aiocb32_ops_osigevent);
2939 free(acb_list, M_LIO);
2940 return (error);
2941 }
2942 #endif
2943
2944 int
freebsd32_lio_listio(struct thread * td,struct freebsd32_lio_listio_args * uap)2945 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2946 {
2947 struct aiocb **acb_list;
2948 struct sigevent *sigp, sig;
2949 struct sigevent32 sig32;
2950 uint32_t *acb_list32;
2951 int error, i, nent;
2952
2953 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2954 return (EINVAL);
2955
2956 nent = uap->nent;
2957 if (nent < 0 || nent > max_aio_queue_per_proc)
2958 return (EINVAL);
2959
2960 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2961 error = copyin(uap->sig, &sig32, sizeof(sig32));
2962 if (error)
2963 return (error);
2964 error = convert_sigevent32(&sig32, &sig);
2965 if (error)
2966 return (error);
2967 sigp = &sig;
2968 } else
2969 sigp = NULL;
2970
2971 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2972 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2973 if (error) {
2974 free(acb_list32, M_LIO);
2975 return (error);
2976 }
2977 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2978 for (i = 0; i < nent; i++)
2979 acb_list[i] = PTRIN(acb_list32[i]);
2980 free(acb_list32, M_LIO);
2981
2982 error = kern_lio_listio(td, uap->mode,
2983 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2984 &aiocb32_ops);
2985 free(acb_list, M_LIO);
2986 return (error);
2987 }
2988
2989 #endif
2990