Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_aio.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 1997 John S. Dyson. All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. John S. Dyson's name may not be used to endorse or promote products 12 * derived from this software without specific prior written permission. 13 * 14 * DISCLAIMER: This code isn't warranted to do anything useful. Anything 15 * bad that happens because of using this software isn't the responsibility 16 * of the author. This software is distributed AS-IS. 17 */ 18 19 /* 20 * This file contains support for the POSIX 1003.1B AIO/LIO facility. 21 */ 22 23 #include <sys/cdefs.h> 24 __FBSDID("$FreeBSD$"); 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 | CTLFLAG_MPSAFE, 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) count of jobs */ 252 int lioj_finished_count; /* (a) count of finished jobs */ 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 (*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty); 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 static void aio_biocleanup(struct bio *bp); 311 void aio_init_aioinfo(struct proc *p); 312 static int aio_onceonly(void); 313 static int aio_free_entry(struct kaiocb *job); 314 static void aio_process_rw(struct kaiocb *job); 315 static void aio_process_sync(struct kaiocb *job); 316 static void aio_process_mlock(struct kaiocb *job); 317 static void aio_schedule_fsync(void *context, int pending); 318 static int aio_newproc(int *); 319 int aio_aqueue(struct thread *td, struct aiocb *ujob, 320 struct aioliojob *lio, int type, struct aiocb_ops *ops); 321 static int aio_queue_file(struct file *fp, struct kaiocb *job); 322 static void aio_biowakeup(struct bio *bp); 323 static void aio_proc_rundown(void *arg, struct proc *p); 324 static void aio_proc_rundown_exec(void *arg, struct proc *p, 325 struct image_params *imgp); 326 static int aio_qbio(struct proc *p, struct kaiocb *job); 327 static void aio_daemon(void *param); 328 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job); 329 static bool aio_clear_cancel_function_locked(struct kaiocb *job); 330 static int aio_kick(struct proc *userp); 331 static void aio_kick_nowait(struct proc *userp); 332 static void aio_kick_helper(void *context, int pending); 333 static int filt_aioattach(struct knote *kn); 334 static void filt_aiodetach(struct knote *kn); 335 static int filt_aio(struct knote *kn, long hint); 336 static int filt_lioattach(struct knote *kn); 337 static void filt_liodetach(struct knote *kn); 338 static int filt_lio(struct knote *kn, long hint); 339 340 /* 341 * Zones for: 342 * kaio Per process async io info 343 * aiop async io process data 344 * aiocb async io jobs 345 * aiolio list io jobs 346 */ 347 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone; 348 349 /* kqueue filters for aio */ 350 static struct filterops aio_filtops = { 351 .f_isfd = 0, 352 .f_attach = filt_aioattach, 353 .f_detach = filt_aiodetach, 354 .f_event = filt_aio, 355 }; 356 static struct filterops lio_filtops = { 357 .f_isfd = 0, 358 .f_attach = filt_lioattach, 359 .f_detach = filt_liodetach, 360 .f_event = filt_lio 361 }; 362 363 static eventhandler_tag exit_tag, exec_tag; 364 365 TASKQUEUE_DEFINE_THREAD(aiod_kick); 366 367 /* 368 * Main operations function for use as a kernel module. 369 */ 370 static int 371 aio_modload(struct module *module, int cmd, void *arg) 372 { 373 int error = 0; 374 375 switch (cmd) { 376 case MOD_LOAD: 377 aio_onceonly(); 378 break; 379 case MOD_SHUTDOWN: 380 break; 381 default: 382 error = EOPNOTSUPP; 383 break; 384 } 385 return (error); 386 } 387 388 static moduledata_t aio_mod = { 389 "aio", 390 &aio_modload, 391 NULL 392 }; 393 394 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY); 395 MODULE_VERSION(aio, 1); 396 397 /* 398 * Startup initialization 399 */ 400 static int 401 aio_onceonly(void) 402 { 403 404 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 405 EVENTHANDLER_PRI_ANY); 406 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, 407 NULL, EVENTHANDLER_PRI_ANY); 408 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 409 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 410 TAILQ_INIT(&aio_freeproc); 411 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 412 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 413 TAILQ_INIT(&aio_jobs); 414 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 415 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 416 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 417 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL, 418 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 419 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL, 420 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 421 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 422 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 423 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 424 jobrefid = 1; 425 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO); 426 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 427 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 428 429 return (0); 430 } 431 432 /* 433 * Init the per-process aioinfo structure. The aioinfo limits are set 434 * per-process for user limit (resource) management. 435 */ 436 void 437 aio_init_aioinfo(struct proc *p) 438 { 439 struct kaioinfo *ki; 440 441 ki = uma_zalloc(kaio_zone, M_WAITOK); 442 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW); 443 ki->kaio_flags = 0; 444 ki->kaio_active_count = 0; 445 ki->kaio_count = 0; 446 ki->kaio_buffer_count = 0; 447 TAILQ_INIT(&ki->kaio_all); 448 TAILQ_INIT(&ki->kaio_done); 449 TAILQ_INIT(&ki->kaio_jobqueue); 450 TAILQ_INIT(&ki->kaio_liojoblist); 451 TAILQ_INIT(&ki->kaio_syncqueue); 452 TAILQ_INIT(&ki->kaio_syncready); 453 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 454 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki); 455 PROC_LOCK(p); 456 if (p->p_aioinfo == NULL) { 457 p->p_aioinfo = ki; 458 PROC_UNLOCK(p); 459 } else { 460 PROC_UNLOCK(p); 461 mtx_destroy(&ki->kaio_mtx); 462 uma_zfree(kaio_zone, ki); 463 } 464 465 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs)) 466 aio_newproc(NULL); 467 } 468 469 static int 470 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext) 471 { 472 struct thread *td; 473 int error; 474 475 error = sigev_findtd(p, sigev, &td); 476 if (error) 477 return (error); 478 if (!KSI_ONQ(ksi)) { 479 ksiginfo_set_sigev(ksi, sigev); 480 ksi->ksi_code = SI_ASYNCIO; 481 ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0; 482 tdsendsignal(p, td, ksi->ksi_signo, ksi); 483 } 484 PROC_UNLOCK(p); 485 return (error); 486 } 487 488 /* 489 * Free a job entry. Wait for completion if it is currently active, but don't 490 * delay forever. If we delay, we return a flag that says that we have to 491 * restart the queue scan. 492 */ 493 static int 494 aio_free_entry(struct kaiocb *job) 495 { 496 struct kaioinfo *ki; 497 struct aioliojob *lj; 498 struct proc *p; 499 500 p = job->userproc; 501 MPASS(curproc == p); 502 ki = p->p_aioinfo; 503 MPASS(ki != NULL); 504 505 AIO_LOCK_ASSERT(ki, MA_OWNED); 506 MPASS(job->jobflags & KAIOCB_FINISHED); 507 508 atomic_subtract_int(&num_queue_count, 1); 509 510 ki->kaio_count--; 511 MPASS(ki->kaio_count >= 0); 512 513 TAILQ_REMOVE(&ki->kaio_done, job, plist); 514 TAILQ_REMOVE(&ki->kaio_all, job, allist); 515 516 lj = job->lio; 517 if (lj) { 518 lj->lioj_count--; 519 lj->lioj_finished_count--; 520 521 if (lj->lioj_count == 0) { 522 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 523 /* lio is going away, we need to destroy any knotes */ 524 knlist_delete(&lj->klist, curthread, 1); 525 PROC_LOCK(p); 526 sigqueue_take(&lj->lioj_ksi); 527 PROC_UNLOCK(p); 528 uma_zfree(aiolio_zone, lj); 529 } 530 } 531 532 /* job is going away, we need to destroy any knotes */ 533 knlist_delete(&job->klist, curthread, 1); 534 PROC_LOCK(p); 535 sigqueue_take(&job->ksi); 536 PROC_UNLOCK(p); 537 538 AIO_UNLOCK(ki); 539 540 /* 541 * The thread argument here is used to find the owning process 542 * and is also passed to fo_close() which may pass it to various 543 * places such as devsw close() routines. Because of that, we 544 * need a thread pointer from the process owning the job that is 545 * persistent and won't disappear out from under us or move to 546 * another process. 547 * 548 * Currently, all the callers of this function call it to remove 549 * a kaiocb from the current process' job list either via a 550 * syscall or due to the current process calling exit() or 551 * execve(). Thus, we know that p == curproc. We also know that 552 * curthread can't exit since we are curthread. 553 * 554 * Therefore, we use curthread as the thread to pass to 555 * knlist_delete(). This does mean that it is possible for the 556 * thread pointer at close time to differ from the thread pointer 557 * at open time, but this is already true of file descriptors in 558 * a multithreaded process. 559 */ 560 if (job->fd_file) 561 fdrop(job->fd_file, curthread); 562 crfree(job->cred); 563 if (job->uiop != &job->uio) 564 free(job->uiop, M_IOV); 565 uma_zfree(aiocb_zone, job); 566 AIO_LOCK(ki); 567 568 return (0); 569 } 570 571 static void 572 aio_proc_rundown_exec(void *arg, struct proc *p, 573 struct image_params *imgp __unused) 574 { 575 aio_proc_rundown(arg, p); 576 } 577 578 static int 579 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job) 580 { 581 aio_cancel_fn_t *func; 582 int cancelled; 583 584 AIO_LOCK_ASSERT(ki, MA_OWNED); 585 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED)) 586 return (0); 587 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0); 588 job->jobflags |= KAIOCB_CANCELLED; 589 590 func = job->cancel_fn; 591 592 /* 593 * If there is no cancel routine, just leave the job marked as 594 * cancelled. The job should be in active use by a caller who 595 * should complete it normally or when it fails to install a 596 * cancel routine. 597 */ 598 if (func == NULL) 599 return (0); 600 601 /* 602 * Set the CANCELLING flag so that aio_complete() will defer 603 * completions of this job. This prevents the job from being 604 * freed out from under the cancel callback. After the 605 * callback any deferred completion (whether from the callback 606 * or any other source) will be completed. 607 */ 608 job->jobflags |= KAIOCB_CANCELLING; 609 AIO_UNLOCK(ki); 610 func(job); 611 AIO_LOCK(ki); 612 job->jobflags &= ~KAIOCB_CANCELLING; 613 if (job->jobflags & KAIOCB_FINISHED) { 614 cancelled = job->uaiocb._aiocb_private.error == ECANCELED; 615 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 616 aio_bio_done_notify(p, job); 617 } else { 618 /* 619 * The cancel callback might have scheduled an 620 * operation to cancel this request, but it is 621 * only counted as cancelled if the request is 622 * cancelled when the callback returns. 623 */ 624 cancelled = 0; 625 } 626 return (cancelled); 627 } 628 629 /* 630 * Rundown the jobs for a given process. 631 */ 632 static void 633 aio_proc_rundown(void *arg, struct proc *p) 634 { 635 struct kaioinfo *ki; 636 struct aioliojob *lj; 637 struct kaiocb *job, *jobn; 638 639 KASSERT(curthread->td_proc == p, 640 ("%s: called on non-curproc", __func__)); 641 ki = p->p_aioinfo; 642 if (ki == NULL) 643 return; 644 645 AIO_LOCK(ki); 646 ki->kaio_flags |= KAIO_RUNDOWN; 647 648 restart: 649 650 /* 651 * Try to cancel all pending requests. This code simulates 652 * aio_cancel on all pending I/O requests. 653 */ 654 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 655 aio_cancel_job(p, ki, job); 656 } 657 658 /* Wait for all running I/O to be finished */ 659 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) { 660 ki->kaio_flags |= KAIO_WAKEUP; 661 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 662 goto restart; 663 } 664 665 /* Free all completed I/O requests. */ 666 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL) 667 aio_free_entry(job); 668 669 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 670 if (lj->lioj_count == 0) { 671 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 672 knlist_delete(&lj->klist, curthread, 1); 673 PROC_LOCK(p); 674 sigqueue_take(&lj->lioj_ksi); 675 PROC_UNLOCK(p); 676 uma_zfree(aiolio_zone, lj); 677 } else { 678 panic("LIO job not cleaned up: C:%d, FC:%d\n", 679 lj->lioj_count, lj->lioj_finished_count); 680 } 681 } 682 AIO_UNLOCK(ki); 683 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task); 684 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task); 685 mtx_destroy(&ki->kaio_mtx); 686 uma_zfree(kaio_zone, ki); 687 p->p_aioinfo = NULL; 688 } 689 690 /* 691 * Select a job to run (called by an AIO daemon). 692 */ 693 static struct kaiocb * 694 aio_selectjob(struct aioproc *aiop) 695 { 696 struct kaiocb *job; 697 struct kaioinfo *ki; 698 struct proc *userp; 699 700 mtx_assert(&aio_job_mtx, MA_OWNED); 701 restart: 702 TAILQ_FOREACH(job, &aio_jobs, list) { 703 userp = job->userproc; 704 ki = userp->p_aioinfo; 705 706 if (ki->kaio_active_count < max_aio_per_proc) { 707 TAILQ_REMOVE(&aio_jobs, job, list); 708 if (!aio_clear_cancel_function(job)) 709 goto restart; 710 711 /* Account for currently active jobs. */ 712 ki->kaio_active_count++; 713 break; 714 } 715 } 716 return (job); 717 } 718 719 /* 720 * Move all data to a permanent storage device. This code 721 * simulates the fsync and fdatasync syscalls. 722 */ 723 static int 724 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op) 725 { 726 struct mount *mp; 727 vm_object_t obj; 728 int error; 729 730 for (;;) { 731 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 732 if (error != 0) 733 break; 734 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 735 obj = vp->v_object; 736 if (obj != NULL) { 737 VM_OBJECT_WLOCK(obj); 738 vm_object_page_clean(obj, 0, 0, 0); 739 VM_OBJECT_WUNLOCK(obj); 740 } 741 if (op == LIO_DSYNC) 742 error = VOP_FDATASYNC(vp, td); 743 else 744 error = VOP_FSYNC(vp, MNT_WAIT, td); 745 746 VOP_UNLOCK(vp); 747 vn_finished_write(mp); 748 if (error != ERELOOKUP) 749 break; 750 } 751 return (error); 752 } 753 754 /* 755 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that 756 * does the I/O request for the non-bio version of the operations. The normal 757 * vn operations are used, and this code should work in all instances for every 758 * type of file, including pipes, sockets, fifos, and regular files. 759 * 760 * XXX I don't think it works well for socket, pipe, and fifo. 761 */ 762 static void 763 aio_process_rw(struct kaiocb *job) 764 { 765 struct ucred *td_savedcred; 766 struct thread *td; 767 struct aiocb *cb; 768 struct file *fp; 769 ssize_t cnt; 770 long msgsnd_st, msgsnd_end; 771 long msgrcv_st, msgrcv_end; 772 long oublock_st, oublock_end; 773 long inblock_st, inblock_end; 774 int error, opcode; 775 776 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ || 777 job->uaiocb.aio_lio_opcode == LIO_READV || 778 job->uaiocb.aio_lio_opcode == LIO_WRITE || 779 job->uaiocb.aio_lio_opcode == LIO_WRITEV, 780 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 781 782 aio_switch_vmspace(job); 783 td = curthread; 784 td_savedcred = td->td_ucred; 785 td->td_ucred = job->cred; 786 job->uiop->uio_td = td; 787 cb = &job->uaiocb; 788 fp = job->fd_file; 789 790 opcode = job->uaiocb.aio_lio_opcode; 791 cnt = job->uiop->uio_resid; 792 793 msgrcv_st = td->td_ru.ru_msgrcv; 794 msgsnd_st = td->td_ru.ru_msgsnd; 795 inblock_st = td->td_ru.ru_inblock; 796 oublock_st = td->td_ru.ru_oublock; 797 798 /* 799 * aio_aqueue() acquires a reference to the file that is 800 * released in aio_free_entry(). 801 */ 802 if (opcode == LIO_READ || opcode == LIO_READV) { 803 if (job->uiop->uio_resid == 0) 804 error = 0; 805 else 806 error = fo_read(fp, job->uiop, fp->f_cred, FOF_OFFSET, 807 td); 808 } else { 809 if (fp->f_type == DTYPE_VNODE) 810 bwillwrite(); 811 error = fo_write(fp, job->uiop, fp->f_cred, FOF_OFFSET, td); 812 } 813 msgrcv_end = td->td_ru.ru_msgrcv; 814 msgsnd_end = td->td_ru.ru_msgsnd; 815 inblock_end = td->td_ru.ru_inblock; 816 oublock_end = td->td_ru.ru_oublock; 817 818 job->msgrcv = msgrcv_end - msgrcv_st; 819 job->msgsnd = msgsnd_end - msgsnd_st; 820 job->inblock = inblock_end - inblock_st; 821 job->outblock = oublock_end - oublock_st; 822 823 if (error != 0 && job->uiop->uio_resid != cnt) { 824 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 825 error = 0; 826 if (error == EPIPE && (opcode & LIO_WRITE)) { 827 PROC_LOCK(job->userproc); 828 kern_psignal(job->userproc, SIGPIPE); 829 PROC_UNLOCK(job->userproc); 830 } 831 } 832 833 cnt -= job->uiop->uio_resid; 834 td->td_ucred = td_savedcred; 835 if (error) 836 aio_complete(job, -1, error); 837 else 838 aio_complete(job, cnt, 0); 839 } 840 841 static void 842 aio_process_sync(struct kaiocb *job) 843 { 844 struct thread *td = curthread; 845 struct ucred *td_savedcred = td->td_ucred; 846 struct file *fp = job->fd_file; 847 int error = 0; 848 849 KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC, 850 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 851 852 td->td_ucred = job->cred; 853 if (fp->f_vnode != NULL) { 854 error = aio_fsync_vnode(td, fp->f_vnode, 855 job->uaiocb.aio_lio_opcode); 856 } 857 td->td_ucred = td_savedcred; 858 if (error) 859 aio_complete(job, -1, error); 860 else 861 aio_complete(job, 0, 0); 862 } 863 864 static void 865 aio_process_mlock(struct kaiocb *job) 866 { 867 struct aiocb *cb = &job->uaiocb; 868 int error; 869 870 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK, 871 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 872 873 aio_switch_vmspace(job); 874 error = kern_mlock(job->userproc, job->cred, 875 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes); 876 aio_complete(job, error != 0 ? -1 : 0, error); 877 } 878 879 static void 880 aio_bio_done_notify(struct proc *userp, struct kaiocb *job) 881 { 882 struct aioliojob *lj; 883 struct kaioinfo *ki; 884 struct kaiocb *sjob, *sjobn; 885 int lj_done; 886 bool schedule_fsync; 887 888 ki = userp->p_aioinfo; 889 AIO_LOCK_ASSERT(ki, MA_OWNED); 890 lj = job->lio; 891 lj_done = 0; 892 if (lj) { 893 lj->lioj_finished_count++; 894 if (lj->lioj_count == lj->lioj_finished_count) 895 lj_done = 1; 896 } 897 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist); 898 MPASS(job->jobflags & KAIOCB_FINISHED); 899 900 if (ki->kaio_flags & KAIO_RUNDOWN) 901 goto notification_done; 902 903 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 904 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 905 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true); 906 907 KNOTE_LOCKED(&job->klist, 1); 908 909 if (lj_done) { 910 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 911 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 912 KNOTE_LOCKED(&lj->klist, 1); 913 } 914 if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED)) 915 == LIOJ_SIGNAL && 916 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 917 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 918 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi, 919 true); 920 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 921 } 922 } 923 924 notification_done: 925 if (job->jobflags & KAIOCB_CHECKSYNC) { 926 schedule_fsync = false; 927 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) { 928 if (job->fd_file != sjob->fd_file || 929 job->seqno >= sjob->seqno) 930 continue; 931 if (--sjob->pending > 0) 932 continue; 933 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list); 934 if (!aio_clear_cancel_function_locked(sjob)) 935 continue; 936 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list); 937 schedule_fsync = true; 938 } 939 if (schedule_fsync) 940 taskqueue_enqueue(taskqueue_aiod_kick, 941 &ki->kaio_sync_task); 942 } 943 if (ki->kaio_flags & KAIO_WAKEUP) { 944 ki->kaio_flags &= ~KAIO_WAKEUP; 945 wakeup(&userp->p_aioinfo); 946 } 947 } 948 949 static void 950 aio_schedule_fsync(void *context, int pending) 951 { 952 struct kaioinfo *ki; 953 struct kaiocb *job; 954 955 ki = context; 956 AIO_LOCK(ki); 957 while (!TAILQ_EMPTY(&ki->kaio_syncready)) { 958 job = TAILQ_FIRST(&ki->kaio_syncready); 959 TAILQ_REMOVE(&ki->kaio_syncready, job, list); 960 AIO_UNLOCK(ki); 961 aio_schedule(job, aio_process_sync); 962 AIO_LOCK(ki); 963 } 964 AIO_UNLOCK(ki); 965 } 966 967 bool 968 aio_cancel_cleared(struct kaiocb *job) 969 { 970 971 /* 972 * The caller should hold the same queue lock held when 973 * aio_clear_cancel_function() was called and set this flag 974 * ensuring this check sees an up-to-date value. However, 975 * there is no way to assert that. 976 */ 977 return ((job->jobflags & KAIOCB_CLEARED) != 0); 978 } 979 980 static bool 981 aio_clear_cancel_function_locked(struct kaiocb *job) 982 { 983 984 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 985 MPASS(job->cancel_fn != NULL); 986 if (job->jobflags & KAIOCB_CANCELLING) { 987 job->jobflags |= KAIOCB_CLEARED; 988 return (false); 989 } 990 job->cancel_fn = NULL; 991 return (true); 992 } 993 994 bool 995 aio_clear_cancel_function(struct kaiocb *job) 996 { 997 struct kaioinfo *ki; 998 bool ret; 999 1000 ki = job->userproc->p_aioinfo; 1001 AIO_LOCK(ki); 1002 ret = aio_clear_cancel_function_locked(job); 1003 AIO_UNLOCK(ki); 1004 return (ret); 1005 } 1006 1007 static bool 1008 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func) 1009 { 1010 1011 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 1012 if (job->jobflags & KAIOCB_CANCELLED) 1013 return (false); 1014 job->cancel_fn = func; 1015 return (true); 1016 } 1017 1018 bool 1019 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func) 1020 { 1021 struct kaioinfo *ki; 1022 bool ret; 1023 1024 ki = job->userproc->p_aioinfo; 1025 AIO_LOCK(ki); 1026 ret = aio_set_cancel_function_locked(job, func); 1027 AIO_UNLOCK(ki); 1028 return (ret); 1029 } 1030 1031 void 1032 aio_complete(struct kaiocb *job, long status, int error) 1033 { 1034 struct kaioinfo *ki; 1035 struct proc *userp; 1036 1037 job->uaiocb._aiocb_private.error = error; 1038 job->uaiocb._aiocb_private.status = status; 1039 1040 userp = job->userproc; 1041 ki = userp->p_aioinfo; 1042 1043 AIO_LOCK(ki); 1044 KASSERT(!(job->jobflags & KAIOCB_FINISHED), 1045 ("duplicate aio_complete")); 1046 job->jobflags |= KAIOCB_FINISHED; 1047 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) { 1048 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 1049 aio_bio_done_notify(userp, job); 1050 } 1051 AIO_UNLOCK(ki); 1052 } 1053 1054 void 1055 aio_cancel(struct kaiocb *job) 1056 { 1057 1058 aio_complete(job, -1, ECANCELED); 1059 } 1060 1061 void 1062 aio_switch_vmspace(struct kaiocb *job) 1063 { 1064 1065 vmspace_switch_aio(job->userproc->p_vmspace); 1066 } 1067 1068 /* 1069 * The AIO daemon, most of the actual work is done in aio_process_*, 1070 * but the setup (and address space mgmt) is done in this routine. 1071 */ 1072 static void 1073 aio_daemon(void *_id) 1074 { 1075 struct kaiocb *job; 1076 struct aioproc *aiop; 1077 struct kaioinfo *ki; 1078 struct proc *p; 1079 struct vmspace *myvm; 1080 struct thread *td = curthread; 1081 int id = (intptr_t)_id; 1082 1083 /* 1084 * Grab an extra reference on the daemon's vmspace so that it 1085 * doesn't get freed by jobs that switch to a different 1086 * vmspace. 1087 */ 1088 p = td->td_proc; 1089 myvm = vmspace_acquire_ref(p); 1090 1091 KASSERT(p->p_textvp == NULL, ("kthread has a textvp")); 1092 1093 /* 1094 * Allocate and ready the aio control info. There is one aiop structure 1095 * per daemon. 1096 */ 1097 aiop = uma_zalloc(aiop_zone, M_WAITOK); 1098 aiop->aioproc = p; 1099 aiop->aioprocflags = 0; 1100 1101 /* 1102 * Wakeup parent process. (Parent sleeps to keep from blasting away 1103 * and creating too many daemons.) 1104 */ 1105 sema_post(&aio_newproc_sem); 1106 1107 mtx_lock(&aio_job_mtx); 1108 for (;;) { 1109 /* 1110 * Take daemon off of free queue 1111 */ 1112 if (aiop->aioprocflags & AIOP_FREE) { 1113 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1114 aiop->aioprocflags &= ~AIOP_FREE; 1115 } 1116 1117 /* 1118 * Check for jobs. 1119 */ 1120 while ((job = aio_selectjob(aiop)) != NULL) { 1121 mtx_unlock(&aio_job_mtx); 1122 1123 ki = job->userproc->p_aioinfo; 1124 job->handle_fn(job); 1125 1126 mtx_lock(&aio_job_mtx); 1127 /* Decrement the active job count. */ 1128 ki->kaio_active_count--; 1129 } 1130 1131 /* 1132 * Disconnect from user address space. 1133 */ 1134 if (p->p_vmspace != myvm) { 1135 mtx_unlock(&aio_job_mtx); 1136 vmspace_switch_aio(myvm); 1137 mtx_lock(&aio_job_mtx); 1138 /* 1139 * We have to restart to avoid race, we only sleep if 1140 * no job can be selected. 1141 */ 1142 continue; 1143 } 1144 1145 mtx_assert(&aio_job_mtx, MA_OWNED); 1146 1147 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1148 aiop->aioprocflags |= AIOP_FREE; 1149 1150 /* 1151 * If daemon is inactive for a long time, allow it to exit, 1152 * thereby freeing resources. 1153 */ 1154 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy", 1155 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) && 1156 (aiop->aioprocflags & AIOP_FREE) && 1157 num_aio_procs > target_aio_procs) 1158 break; 1159 } 1160 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1161 num_aio_procs--; 1162 mtx_unlock(&aio_job_mtx); 1163 uma_zfree(aiop_zone, aiop); 1164 free_unr(aiod_unr, id); 1165 vmspace_free(myvm); 1166 1167 KASSERT(p->p_vmspace == myvm, 1168 ("AIOD: bad vmspace for exiting daemon")); 1169 KASSERT(refcount_load(&myvm->vm_refcnt) > 1, 1170 ("AIOD: bad vm refcnt for exiting daemon: %d", 1171 refcount_load(&myvm->vm_refcnt))); 1172 kproc_exit(0); 1173 } 1174 1175 /* 1176 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1177 * AIO daemon modifies its environment itself. 1178 */ 1179 static int 1180 aio_newproc(int *start) 1181 { 1182 int error; 1183 struct proc *p; 1184 int id; 1185 1186 id = alloc_unr(aiod_unr); 1187 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p, 1188 RFNOWAIT, 0, "aiod%d", id); 1189 if (error == 0) { 1190 /* 1191 * Wait until daemon is started. 1192 */ 1193 sema_wait(&aio_newproc_sem); 1194 mtx_lock(&aio_job_mtx); 1195 num_aio_procs++; 1196 if (start != NULL) 1197 (*start)--; 1198 mtx_unlock(&aio_job_mtx); 1199 } else { 1200 free_unr(aiod_unr, id); 1201 } 1202 return (error); 1203 } 1204 1205 /* 1206 * Try the high-performance, low-overhead bio method for eligible 1207 * VCHR devices. This method doesn't use an aio helper thread, and 1208 * thus has very low overhead. 1209 * 1210 * Assumes that the caller, aio_aqueue(), has incremented the file 1211 * structure's reference count, preventing its deallocation for the 1212 * duration of this call. 1213 */ 1214 static int 1215 aio_qbio(struct proc *p, struct kaiocb *job) 1216 { 1217 struct aiocb *cb; 1218 struct file *fp; 1219 struct buf *pbuf; 1220 struct vnode *vp; 1221 struct cdevsw *csw; 1222 struct cdev *dev; 1223 struct kaioinfo *ki; 1224 struct bio **bios = NULL; 1225 off_t offset; 1226 int bio_cmd, error, i, iovcnt, opcode, poff, ref; 1227 vm_prot_t prot; 1228 bool use_unmapped; 1229 1230 cb = &job->uaiocb; 1231 fp = job->fd_file; 1232 opcode = cb->aio_lio_opcode; 1233 1234 if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV || 1235 opcode == LIO_READ || opcode == LIO_READV)) 1236 return (-1); 1237 if (fp == NULL || fp->f_type != DTYPE_VNODE) 1238 return (-1); 1239 1240 vp = fp->f_vnode; 1241 if (vp->v_type != VCHR) 1242 return (-1); 1243 if (vp->v_bufobj.bo_bsize == 0) 1244 return (-1); 1245 1246 bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ; 1247 iovcnt = job->uiop->uio_iovcnt; 1248 if (iovcnt > max_buf_aio) 1249 return (-1); 1250 for (i = 0; i < iovcnt; i++) { 1251 if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0) 1252 return (-1); 1253 if (job->uiop->uio_iov[i].iov_len > maxphys) { 1254 error = -1; 1255 return (-1); 1256 } 1257 } 1258 offset = cb->aio_offset; 1259 1260 ref = 0; 1261 csw = devvn_refthread(vp, &dev, &ref); 1262 if (csw == NULL) 1263 return (ENXIO); 1264 1265 if ((csw->d_flags & D_DISK) == 0) { 1266 error = -1; 1267 goto unref; 1268 } 1269 if (job->uiop->uio_resid > dev->si_iosize_max) { 1270 error = -1; 1271 goto unref; 1272 } 1273 1274 ki = p->p_aioinfo; 1275 job->error = 0; 1276 1277 use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed; 1278 if (!use_unmapped) { 1279 AIO_LOCK(ki); 1280 if (ki->kaio_buffer_count + iovcnt > max_buf_aio) { 1281 AIO_UNLOCK(ki); 1282 error = EAGAIN; 1283 goto unref; 1284 } 1285 ki->kaio_buffer_count += iovcnt; 1286 AIO_UNLOCK(ki); 1287 } 1288 1289 bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK); 1290 atomic_store_int(&job->nbio, iovcnt); 1291 for (i = 0; i < iovcnt; i++) { 1292 struct vm_page** pages; 1293 struct bio *bp; 1294 void *buf; 1295 size_t nbytes; 1296 int npages; 1297 1298 buf = job->uiop->uio_iov[i].iov_base; 1299 nbytes = job->uiop->uio_iov[i].iov_len; 1300 1301 bios[i] = g_alloc_bio(); 1302 bp = bios[i]; 1303 1304 poff = (vm_offset_t)buf & PAGE_MASK; 1305 if (use_unmapped) { 1306 pbuf = NULL; 1307 pages = malloc(sizeof(vm_page_t) * (atop(round_page( 1308 nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO); 1309 } else { 1310 pbuf = uma_zalloc(pbuf_zone, M_WAITOK); 1311 BUF_KERNPROC(pbuf); 1312 pages = pbuf->b_pages; 1313 } 1314 1315 bp->bio_length = nbytes; 1316 bp->bio_bcount = nbytes; 1317 bp->bio_done = aio_biowakeup; 1318 bp->bio_offset = offset; 1319 bp->bio_cmd = bio_cmd; 1320 bp->bio_dev = dev; 1321 bp->bio_caller1 = job; 1322 bp->bio_caller2 = pbuf; 1323 1324 prot = VM_PROT_READ; 1325 if (opcode == LIO_READ || opcode == LIO_READV) 1326 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 1327 npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 1328 (vm_offset_t)buf, bp->bio_length, prot, pages, 1329 atop(maxphys) + 1); 1330 if (npages < 0) { 1331 if (pbuf != NULL) 1332 uma_zfree(pbuf_zone, pbuf); 1333 else 1334 free(pages, M_TEMP); 1335 error = EFAULT; 1336 g_destroy_bio(bp); 1337 i--; 1338 goto destroy_bios; 1339 } 1340 if (pbuf != NULL) { 1341 pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages); 1342 bp->bio_data = pbuf->b_data + poff; 1343 pbuf->b_npages = npages; 1344 atomic_add_int(&num_buf_aio, 1); 1345 } else { 1346 bp->bio_ma = pages; 1347 bp->bio_ma_n = npages; 1348 bp->bio_ma_offset = poff; 1349 bp->bio_data = unmapped_buf; 1350 bp->bio_flags |= BIO_UNMAPPED; 1351 atomic_add_int(&num_unmapped_aio, 1); 1352 } 1353 1354 offset += nbytes; 1355 } 1356 1357 /* Perform transfer. */ 1358 for (i = 0; i < iovcnt; i++) 1359 csw->d_strategy(bios[i]); 1360 free(bios, M_TEMP); 1361 1362 dev_relthread(dev, ref); 1363 return (0); 1364 1365 destroy_bios: 1366 for (; i >= 0; i--) 1367 aio_biocleanup(bios[i]); 1368 free(bios, M_TEMP); 1369 unref: 1370 dev_relthread(dev, ref); 1371 return (error); 1372 } 1373 1374 #ifdef COMPAT_FREEBSD6 1375 static int 1376 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig) 1377 { 1378 1379 /* 1380 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 1381 * supported by AIO with the old sigevent structure. 1382 */ 1383 nsig->sigev_notify = osig->sigev_notify; 1384 switch (nsig->sigev_notify) { 1385 case SIGEV_NONE: 1386 break; 1387 case SIGEV_SIGNAL: 1388 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 1389 break; 1390 case SIGEV_KEVENT: 1391 nsig->sigev_notify_kqueue = 1392 osig->__sigev_u.__sigev_notify_kqueue; 1393 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr; 1394 break; 1395 default: 1396 return (EINVAL); 1397 } 1398 return (0); 1399 } 1400 1401 static int 1402 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob, 1403 int type __unused) 1404 { 1405 struct oaiocb *ojob; 1406 struct aiocb *kcb = &kjob->uaiocb; 1407 int error; 1408 1409 bzero(kcb, sizeof(struct aiocb)); 1410 error = copyin(ujob, kcb, sizeof(struct oaiocb)); 1411 if (error) 1412 return (error); 1413 /* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */ 1414 ojob = (struct oaiocb *)kcb; 1415 return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent)); 1416 } 1417 #endif 1418 1419 static int 1420 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type) 1421 { 1422 struct aiocb *kcb = &kjob->uaiocb; 1423 int error; 1424 1425 error = copyin(ujob, kcb, sizeof(struct aiocb)); 1426 if (error) 1427 return (error); 1428 if (type & LIO_VECTORED) { 1429 /* malloc a uio and copy in the iovec */ 1430 error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov), 1431 kcb->aio_iovcnt, &kjob->uiop); 1432 } 1433 1434 return (error); 1435 } 1436 1437 static long 1438 aiocb_fetch_status(struct aiocb *ujob) 1439 { 1440 1441 return (fuword(&ujob->_aiocb_private.status)); 1442 } 1443 1444 static long 1445 aiocb_fetch_error(struct aiocb *ujob) 1446 { 1447 1448 return (fuword(&ujob->_aiocb_private.error)); 1449 } 1450 1451 static int 1452 aiocb_store_status(struct aiocb *ujob, long status) 1453 { 1454 1455 return (suword(&ujob->_aiocb_private.status, status)); 1456 } 1457 1458 static int 1459 aiocb_store_error(struct aiocb *ujob, long error) 1460 { 1461 1462 return (suword(&ujob->_aiocb_private.error, error)); 1463 } 1464 1465 static int 1466 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref) 1467 { 1468 1469 return (suword(&ujob->_aiocb_private.kernelinfo, jobref)); 1470 } 1471 1472 static int 1473 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 1474 { 1475 1476 return (suword(ujobp, (long)ujob)); 1477 } 1478 1479 static struct aiocb_ops aiocb_ops = { 1480 .aio_copyin = aiocb_copyin, 1481 .fetch_status = aiocb_fetch_status, 1482 .fetch_error = aiocb_fetch_error, 1483 .store_status = aiocb_store_status, 1484 .store_error = aiocb_store_error, 1485 .store_kernelinfo = aiocb_store_kernelinfo, 1486 .store_aiocb = aiocb_store_aiocb, 1487 }; 1488 1489 #ifdef COMPAT_FREEBSD6 1490 static struct aiocb_ops aiocb_ops_osigevent = { 1491 .aio_copyin = aiocb_copyin_old_sigevent, 1492 .fetch_status = aiocb_fetch_status, 1493 .fetch_error = aiocb_fetch_error, 1494 .store_status = aiocb_store_status, 1495 .store_error = aiocb_store_error, 1496 .store_kernelinfo = aiocb_store_kernelinfo, 1497 .store_aiocb = aiocb_store_aiocb, 1498 }; 1499 #endif 1500 1501 /* 1502 * Queue a new AIO request. Choosing either the threaded or direct bio VCHR 1503 * technique is done in this code. 1504 */ 1505 int 1506 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj, 1507 int type, struct aiocb_ops *ops) 1508 { 1509 struct proc *p = td->td_proc; 1510 struct file *fp = NULL; 1511 struct kaiocb *job; 1512 struct kaioinfo *ki; 1513 struct kevent kev; 1514 int opcode; 1515 int error; 1516 int fd, kqfd; 1517 int jid; 1518 u_short evflags; 1519 1520 if (p->p_aioinfo == NULL) 1521 aio_init_aioinfo(p); 1522 1523 ki = p->p_aioinfo; 1524 1525 ops->store_status(ujob, -1); 1526 ops->store_error(ujob, 0); 1527 ops->store_kernelinfo(ujob, -1); 1528 1529 if (num_queue_count >= max_queue_count || 1530 ki->kaio_count >= max_aio_queue_per_proc) { 1531 error = EAGAIN; 1532 goto err1; 1533 } 1534 1535 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1536 knlist_init_mtx(&job->klist, AIO_MTX(ki)); 1537 1538 error = ops->aio_copyin(ujob, job, type); 1539 if (error) 1540 goto err2; 1541 1542 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) { 1543 error = EINVAL; 1544 goto err2; 1545 } 1546 1547 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1548 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1549 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1550 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1551 error = EINVAL; 1552 goto err2; 1553 } 1554 1555 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1556 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1557 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) { 1558 error = EINVAL; 1559 goto err2; 1560 } 1561 1562 /* Get the opcode. */ 1563 if (type == LIO_NOP) { 1564 switch (job->uaiocb.aio_lio_opcode) { 1565 case LIO_WRITE: 1566 case LIO_NOP: 1567 case LIO_READ: 1568 opcode = job->uaiocb.aio_lio_opcode; 1569 break; 1570 default: 1571 error = EINVAL; 1572 goto err2; 1573 } 1574 } else 1575 opcode = job->uaiocb.aio_lio_opcode = type; 1576 1577 ksiginfo_init(&job->ksi); 1578 1579 /* Save userspace address of the job info. */ 1580 job->ujob = ujob; 1581 1582 /* 1583 * Validate the opcode and fetch the file object for the specified 1584 * file descriptor. 1585 * 1586 * XXXRW: Moved the opcode validation up here so that we don't 1587 * retrieve a file descriptor without knowing what the capabiltity 1588 * should be. 1589 */ 1590 fd = job->uaiocb.aio_fildes; 1591 switch (opcode) { 1592 case LIO_WRITE: 1593 case LIO_WRITEV: 1594 error = fget_write(td, fd, &cap_pwrite_rights, &fp); 1595 break; 1596 case LIO_READ: 1597 case LIO_READV: 1598 error = fget_read(td, fd, &cap_pread_rights, &fp); 1599 break; 1600 case LIO_SYNC: 1601 case LIO_DSYNC: 1602 error = fget(td, fd, &cap_fsync_rights, &fp); 1603 break; 1604 case LIO_MLOCK: 1605 break; 1606 case LIO_NOP: 1607 error = fget(td, fd, &cap_no_rights, &fp); 1608 break; 1609 default: 1610 error = EINVAL; 1611 } 1612 if (error) 1613 goto err3; 1614 1615 if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) { 1616 error = EINVAL; 1617 goto err3; 1618 } 1619 1620 if ((opcode == LIO_READ || opcode == LIO_READV || 1621 opcode == LIO_WRITE || opcode == LIO_WRITEV) && 1622 job->uaiocb.aio_offset < 0 && 1623 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) { 1624 error = EINVAL; 1625 goto err3; 1626 } 1627 1628 if (fp != NULL && fp->f_ops == &path_fileops) { 1629 error = EBADF; 1630 goto err3; 1631 } 1632 1633 job->fd_file = fp; 1634 1635 mtx_lock(&aio_job_mtx); 1636 jid = jobrefid++; 1637 job->seqno = jobseqno++; 1638 mtx_unlock(&aio_job_mtx); 1639 error = ops->store_kernelinfo(ujob, jid); 1640 if (error) { 1641 error = EINVAL; 1642 goto err3; 1643 } 1644 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid; 1645 1646 if (opcode == LIO_NOP) { 1647 fdrop(fp, td); 1648 MPASS(job->uiop == &job->uio || job->uiop == NULL); 1649 uma_zfree(aiocb_zone, job); 1650 return (0); 1651 } 1652 1653 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT) 1654 goto no_kqueue; 1655 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags; 1656 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) { 1657 error = EINVAL; 1658 goto err3; 1659 } 1660 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue; 1661 memset(&kev, 0, sizeof(kev)); 1662 kev.ident = (uintptr_t)job->ujob; 1663 kev.filter = EVFILT_AIO; 1664 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags; 1665 kev.data = (intptr_t)job; 1666 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1667 error = kqfd_register(kqfd, &kev, td, M_WAITOK); 1668 if (error) 1669 goto err3; 1670 1671 no_kqueue: 1672 1673 ops->store_error(ujob, EINPROGRESS); 1674 job->uaiocb._aiocb_private.error = EINPROGRESS; 1675 job->userproc = p; 1676 job->cred = crhold(td->td_ucred); 1677 job->jobflags = KAIOCB_QUEUEING; 1678 job->lio = lj; 1679 1680 if (opcode & LIO_VECTORED) { 1681 /* Use the uio copied in by aio_copyin */ 1682 MPASS(job->uiop != &job->uio && job->uiop != NULL); 1683 } else { 1684 /* Setup the inline uio */ 1685 job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf; 1686 job->iov[0].iov_len = job->uaiocb.aio_nbytes; 1687 job->uio.uio_iov = job->iov; 1688 job->uio.uio_iovcnt = 1; 1689 job->uio.uio_resid = job->uaiocb.aio_nbytes; 1690 job->uio.uio_segflg = UIO_USERSPACE; 1691 job->uiop = &job->uio; 1692 } 1693 switch (opcode & (LIO_READ | LIO_WRITE)) { 1694 case LIO_READ: 1695 job->uiop->uio_rw = UIO_READ; 1696 break; 1697 case LIO_WRITE: 1698 job->uiop->uio_rw = UIO_WRITE; 1699 break; 1700 } 1701 job->uiop->uio_offset = job->uaiocb.aio_offset; 1702 job->uiop->uio_td = td; 1703 1704 if (opcode == LIO_MLOCK) { 1705 aio_schedule(job, aio_process_mlock); 1706 error = 0; 1707 } else if (fp->f_ops->fo_aio_queue == NULL) 1708 error = aio_queue_file(fp, job); 1709 else 1710 error = fo_aio_queue(fp, job); 1711 if (error) 1712 goto err4; 1713 1714 AIO_LOCK(ki); 1715 job->jobflags &= ~KAIOCB_QUEUEING; 1716 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist); 1717 ki->kaio_count++; 1718 if (lj) 1719 lj->lioj_count++; 1720 atomic_add_int(&num_queue_count, 1); 1721 if (job->jobflags & KAIOCB_FINISHED) { 1722 /* 1723 * The queue callback completed the request synchronously. 1724 * The bulk of the completion is deferred in that case 1725 * until this point. 1726 */ 1727 aio_bio_done_notify(p, job); 1728 } else 1729 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist); 1730 AIO_UNLOCK(ki); 1731 return (0); 1732 1733 err4: 1734 crfree(job->cred); 1735 err3: 1736 if (fp) 1737 fdrop(fp, td); 1738 knlist_delete(&job->klist, curthread, 0); 1739 err2: 1740 if (job->uiop != &job->uio) 1741 free(job->uiop, M_IOV); 1742 uma_zfree(aiocb_zone, job); 1743 err1: 1744 ops->store_error(ujob, error); 1745 return (error); 1746 } 1747 1748 static void 1749 aio_cancel_daemon_job(struct kaiocb *job) 1750 { 1751 1752 mtx_lock(&aio_job_mtx); 1753 if (!aio_cancel_cleared(job)) 1754 TAILQ_REMOVE(&aio_jobs, job, list); 1755 mtx_unlock(&aio_job_mtx); 1756 aio_cancel(job); 1757 } 1758 1759 void 1760 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func) 1761 { 1762 1763 mtx_lock(&aio_job_mtx); 1764 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) { 1765 mtx_unlock(&aio_job_mtx); 1766 aio_cancel(job); 1767 return; 1768 } 1769 job->handle_fn = func; 1770 TAILQ_INSERT_TAIL(&aio_jobs, job, list); 1771 aio_kick_nowait(job->userproc); 1772 mtx_unlock(&aio_job_mtx); 1773 } 1774 1775 static void 1776 aio_cancel_sync(struct kaiocb *job) 1777 { 1778 struct kaioinfo *ki; 1779 1780 ki = job->userproc->p_aioinfo; 1781 AIO_LOCK(ki); 1782 if (!aio_cancel_cleared(job)) 1783 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list); 1784 AIO_UNLOCK(ki); 1785 aio_cancel(job); 1786 } 1787 1788 int 1789 aio_queue_file(struct file *fp, struct kaiocb *job) 1790 { 1791 struct kaioinfo *ki; 1792 struct kaiocb *job2; 1793 struct vnode *vp; 1794 struct mount *mp; 1795 int error; 1796 bool safe; 1797 1798 ki = job->userproc->p_aioinfo; 1799 error = aio_qbio(job->userproc, job); 1800 if (error >= 0) 1801 return (error); 1802 safe = false; 1803 if (fp->f_type == DTYPE_VNODE) { 1804 vp = fp->f_vnode; 1805 if (vp->v_type == VREG || vp->v_type == VDIR) { 1806 mp = fp->f_vnode->v_mount; 1807 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0) 1808 safe = true; 1809 } 1810 } 1811 if (!(safe || enable_aio_unsafe)) { 1812 counted_warning(&unsafe_warningcnt, 1813 "is attempting to use unsafe AIO requests"); 1814 return (EOPNOTSUPP); 1815 } 1816 1817 if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) { 1818 aio_schedule(job, aio_process_rw); 1819 error = 0; 1820 } else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) { 1821 AIO_LOCK(ki); 1822 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) { 1823 if (job2->fd_file == job->fd_file && 1824 ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) && 1825 job2->seqno < job->seqno) { 1826 job2->jobflags |= KAIOCB_CHECKSYNC; 1827 job->pending++; 1828 } 1829 } 1830 if (job->pending != 0) { 1831 if (!aio_set_cancel_function_locked(job, 1832 aio_cancel_sync)) { 1833 AIO_UNLOCK(ki); 1834 aio_cancel(job); 1835 return (0); 1836 } 1837 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list); 1838 AIO_UNLOCK(ki); 1839 return (0); 1840 } 1841 AIO_UNLOCK(ki); 1842 aio_schedule(job, aio_process_sync); 1843 error = 0; 1844 } else { 1845 error = EINVAL; 1846 } 1847 return (error); 1848 } 1849 1850 static void 1851 aio_kick_nowait(struct proc *userp) 1852 { 1853 struct kaioinfo *ki = userp->p_aioinfo; 1854 struct aioproc *aiop; 1855 1856 mtx_assert(&aio_job_mtx, MA_OWNED); 1857 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1858 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1859 aiop->aioprocflags &= ~AIOP_FREE; 1860 wakeup(aiop->aioproc); 1861 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1862 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) { 1863 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task); 1864 } 1865 } 1866 1867 static int 1868 aio_kick(struct proc *userp) 1869 { 1870 struct kaioinfo *ki = userp->p_aioinfo; 1871 struct aioproc *aiop; 1872 int error, ret = 0; 1873 1874 mtx_assert(&aio_job_mtx, MA_OWNED); 1875 retryproc: 1876 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1877 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1878 aiop->aioprocflags &= ~AIOP_FREE; 1879 wakeup(aiop->aioproc); 1880 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1881 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) { 1882 num_aio_resv_start++; 1883 mtx_unlock(&aio_job_mtx); 1884 error = aio_newproc(&num_aio_resv_start); 1885 mtx_lock(&aio_job_mtx); 1886 if (error) { 1887 num_aio_resv_start--; 1888 goto retryproc; 1889 } 1890 } else { 1891 ret = -1; 1892 } 1893 return (ret); 1894 } 1895 1896 static void 1897 aio_kick_helper(void *context, int pending) 1898 { 1899 struct proc *userp = context; 1900 1901 mtx_lock(&aio_job_mtx); 1902 while (--pending >= 0) { 1903 if (aio_kick(userp)) 1904 break; 1905 } 1906 mtx_unlock(&aio_job_mtx); 1907 } 1908 1909 /* 1910 * Support the aio_return system call, as a side-effect, kernel resources are 1911 * released. 1912 */ 1913 static int 1914 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 1915 { 1916 struct proc *p = td->td_proc; 1917 struct kaiocb *job; 1918 struct kaioinfo *ki; 1919 long status, error; 1920 1921 ki = p->p_aioinfo; 1922 if (ki == NULL) 1923 return (EINVAL); 1924 AIO_LOCK(ki); 1925 TAILQ_FOREACH(job, &ki->kaio_done, plist) { 1926 if (job->ujob == ujob) 1927 break; 1928 } 1929 if (job != NULL) { 1930 MPASS(job->jobflags & KAIOCB_FINISHED); 1931 status = job->uaiocb._aiocb_private.status; 1932 error = job->uaiocb._aiocb_private.error; 1933 td->td_retval[0] = status; 1934 td->td_ru.ru_oublock += job->outblock; 1935 td->td_ru.ru_inblock += job->inblock; 1936 td->td_ru.ru_msgsnd += job->msgsnd; 1937 td->td_ru.ru_msgrcv += job->msgrcv; 1938 aio_free_entry(job); 1939 AIO_UNLOCK(ki); 1940 ops->store_error(ujob, error); 1941 ops->store_status(ujob, status); 1942 } else { 1943 error = EINVAL; 1944 AIO_UNLOCK(ki); 1945 } 1946 return (error); 1947 } 1948 1949 int 1950 sys_aio_return(struct thread *td, struct aio_return_args *uap) 1951 { 1952 1953 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops)); 1954 } 1955 1956 /* 1957 * Allow a process to wakeup when any of the I/O requests are completed. 1958 */ 1959 static int 1960 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist, 1961 struct timespec *ts) 1962 { 1963 struct proc *p = td->td_proc; 1964 struct timeval atv; 1965 struct kaioinfo *ki; 1966 struct kaiocb *firstjob, *job; 1967 int error, i, timo; 1968 1969 timo = 0; 1970 if (ts) { 1971 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000) 1972 return (EINVAL); 1973 1974 TIMESPEC_TO_TIMEVAL(&atv, ts); 1975 if (itimerfix(&atv)) 1976 return (EINVAL); 1977 timo = tvtohz(&atv); 1978 } 1979 1980 ki = p->p_aioinfo; 1981 if (ki == NULL) 1982 return (EAGAIN); 1983 1984 if (njoblist == 0) 1985 return (0); 1986 1987 AIO_LOCK(ki); 1988 for (;;) { 1989 firstjob = NULL; 1990 error = 0; 1991 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 1992 for (i = 0; i < njoblist; i++) { 1993 if (job->ujob == ujoblist[i]) { 1994 if (firstjob == NULL) 1995 firstjob = job; 1996 if (job->jobflags & KAIOCB_FINISHED) 1997 goto RETURN; 1998 } 1999 } 2000 } 2001 /* All tasks were finished. */ 2002 if (firstjob == NULL) 2003 break; 2004 2005 ki->kaio_flags |= KAIO_WAKEUP; 2006 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2007 "aiospn", timo); 2008 if (error == ERESTART) 2009 error = EINTR; 2010 if (error) 2011 break; 2012 } 2013 RETURN: 2014 AIO_UNLOCK(ki); 2015 return (error); 2016 } 2017 2018 int 2019 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap) 2020 { 2021 struct timespec ts, *tsp; 2022 struct aiocb **ujoblist; 2023 int error; 2024 2025 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 2026 return (EINVAL); 2027 2028 if (uap->timeout) { 2029 /* Get timespec struct. */ 2030 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 2031 return (error); 2032 tsp = &ts; 2033 } else 2034 tsp = NULL; 2035 2036 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK); 2037 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0])); 2038 if (error == 0) 2039 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2040 free(ujoblist, M_AIOS); 2041 return (error); 2042 } 2043 2044 /* 2045 * aio_cancel cancels any non-bio aio operations not currently in progress. 2046 */ 2047 int 2048 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap) 2049 { 2050 struct proc *p = td->td_proc; 2051 struct kaioinfo *ki; 2052 struct kaiocb *job, *jobn; 2053 struct file *fp; 2054 int error; 2055 int cancelled = 0; 2056 int notcancelled = 0; 2057 struct vnode *vp; 2058 2059 /* Lookup file object. */ 2060 error = fget(td, uap->fd, &cap_no_rights, &fp); 2061 if (error) 2062 return (error); 2063 2064 ki = p->p_aioinfo; 2065 if (ki == NULL) 2066 goto done; 2067 2068 if (fp->f_type == DTYPE_VNODE) { 2069 vp = fp->f_vnode; 2070 if (vn_isdisk(vp)) { 2071 fdrop(fp, td); 2072 td->td_retval[0] = AIO_NOTCANCELED; 2073 return (0); 2074 } 2075 } 2076 2077 AIO_LOCK(ki); 2078 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 2079 if ((uap->fd == job->uaiocb.aio_fildes) && 2080 ((uap->aiocbp == NULL) || 2081 (uap->aiocbp == job->ujob))) { 2082 if (aio_cancel_job(p, ki, job)) { 2083 cancelled++; 2084 } else { 2085 notcancelled++; 2086 } 2087 if (uap->aiocbp != NULL) 2088 break; 2089 } 2090 } 2091 AIO_UNLOCK(ki); 2092 2093 done: 2094 fdrop(fp, td); 2095 2096 if (uap->aiocbp != NULL) { 2097 if (cancelled) { 2098 td->td_retval[0] = AIO_CANCELED; 2099 return (0); 2100 } 2101 } 2102 2103 if (notcancelled) { 2104 td->td_retval[0] = AIO_NOTCANCELED; 2105 return (0); 2106 } 2107 2108 if (cancelled) { 2109 td->td_retval[0] = AIO_CANCELED; 2110 return (0); 2111 } 2112 2113 td->td_retval[0] = AIO_ALLDONE; 2114 2115 return (0); 2116 } 2117 2118 /* 2119 * aio_error is implemented in the kernel level for compatibility purposes 2120 * only. For a user mode async implementation, it would be best to do it in 2121 * a userland subroutine. 2122 */ 2123 static int 2124 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 2125 { 2126 struct proc *p = td->td_proc; 2127 struct kaiocb *job; 2128 struct kaioinfo *ki; 2129 int status; 2130 2131 ki = p->p_aioinfo; 2132 if (ki == NULL) { 2133 td->td_retval[0] = EINVAL; 2134 return (0); 2135 } 2136 2137 AIO_LOCK(ki); 2138 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 2139 if (job->ujob == ujob) { 2140 if (job->jobflags & KAIOCB_FINISHED) 2141 td->td_retval[0] = 2142 job->uaiocb._aiocb_private.error; 2143 else 2144 td->td_retval[0] = EINPROGRESS; 2145 AIO_UNLOCK(ki); 2146 return (0); 2147 } 2148 } 2149 AIO_UNLOCK(ki); 2150 2151 /* 2152 * Hack for failure of aio_aqueue. 2153 */ 2154 status = ops->fetch_status(ujob); 2155 if (status == -1) { 2156 td->td_retval[0] = ops->fetch_error(ujob); 2157 return (0); 2158 } 2159 2160 td->td_retval[0] = EINVAL; 2161 return (0); 2162 } 2163 2164 int 2165 sys_aio_error(struct thread *td, struct aio_error_args *uap) 2166 { 2167 2168 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops)); 2169 } 2170 2171 /* syscall - asynchronous read from a file (REALTIME) */ 2172 #ifdef COMPAT_FREEBSD6 2173 int 2174 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap) 2175 { 2176 2177 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2178 &aiocb_ops_osigevent)); 2179 } 2180 #endif 2181 2182 int 2183 sys_aio_read(struct thread *td, struct aio_read_args *uap) 2184 { 2185 2186 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops)); 2187 } 2188 2189 int 2190 sys_aio_readv(struct thread *td, struct aio_readv_args *uap) 2191 { 2192 2193 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops)); 2194 } 2195 2196 /* syscall - asynchronous write to a file (REALTIME) */ 2197 #ifdef COMPAT_FREEBSD6 2198 int 2199 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap) 2200 { 2201 2202 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2203 &aiocb_ops_osigevent)); 2204 } 2205 #endif 2206 2207 int 2208 sys_aio_write(struct thread *td, struct aio_write_args *uap) 2209 { 2210 2211 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops)); 2212 } 2213 2214 int 2215 sys_aio_writev(struct thread *td, struct aio_writev_args *uap) 2216 { 2217 2218 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops)); 2219 } 2220 2221 int 2222 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap) 2223 { 2224 2225 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops)); 2226 } 2227 2228 static int 2229 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list, 2230 struct aiocb **acb_list, int nent, struct sigevent *sig, 2231 struct aiocb_ops *ops) 2232 { 2233 struct proc *p = td->td_proc; 2234 struct aiocb *job; 2235 struct kaioinfo *ki; 2236 struct aioliojob *lj; 2237 struct kevent kev; 2238 int error; 2239 int nagain, nerror; 2240 int i; 2241 2242 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT)) 2243 return (EINVAL); 2244 2245 if (nent < 0 || nent > max_aio_queue_per_proc) 2246 return (EINVAL); 2247 2248 if (p->p_aioinfo == NULL) 2249 aio_init_aioinfo(p); 2250 2251 ki = p->p_aioinfo; 2252 2253 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2254 lj->lioj_flags = 0; 2255 lj->lioj_count = 0; 2256 lj->lioj_finished_count = 0; 2257 lj->lioj_signal.sigev_notify = SIGEV_NONE; 2258 knlist_init_mtx(&lj->klist, AIO_MTX(ki)); 2259 ksiginfo_init(&lj->lioj_ksi); 2260 2261 /* 2262 * Setup signal. 2263 */ 2264 if (sig && (mode == LIO_NOWAIT)) { 2265 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal)); 2266 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2267 /* Assume only new style KEVENT */ 2268 memset(&kev, 0, sizeof(kev)); 2269 kev.filter = EVFILT_LIO; 2270 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2271 kev.ident = (uintptr_t)uacb_list; /* something unique */ 2272 kev.data = (intptr_t)lj; 2273 /* pass user defined sigval data */ 2274 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2275 error = kqfd_register( 2276 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 2277 M_WAITOK); 2278 if (error) { 2279 uma_zfree(aiolio_zone, lj); 2280 return (error); 2281 } 2282 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2283 ; 2284 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2285 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2286 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2287 uma_zfree(aiolio_zone, lj); 2288 return EINVAL; 2289 } 2290 lj->lioj_flags |= LIOJ_SIGNAL; 2291 } else { 2292 uma_zfree(aiolio_zone, lj); 2293 return EINVAL; 2294 } 2295 } 2296 2297 AIO_LOCK(ki); 2298 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2299 /* 2300 * Add extra aiocb count to avoid the lio to be freed 2301 * by other threads doing aio_waitcomplete or aio_return, 2302 * and prevent event from being sent until we have queued 2303 * all tasks. 2304 */ 2305 lj->lioj_count = 1; 2306 AIO_UNLOCK(ki); 2307 2308 /* 2309 * Get pointers to the list of I/O requests. 2310 */ 2311 nagain = 0; 2312 nerror = 0; 2313 for (i = 0; i < nent; i++) { 2314 job = acb_list[i]; 2315 if (job != NULL) { 2316 error = aio_aqueue(td, job, lj, LIO_NOP, ops); 2317 if (error == EAGAIN) 2318 nagain++; 2319 else if (error != 0) 2320 nerror++; 2321 } 2322 } 2323 2324 error = 0; 2325 AIO_LOCK(ki); 2326 if (mode == LIO_WAIT) { 2327 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2328 ki->kaio_flags |= KAIO_WAKEUP; 2329 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2330 PRIBIO | PCATCH, "aiospn", 0); 2331 if (error == ERESTART) 2332 error = EINTR; 2333 if (error) 2334 break; 2335 } 2336 } else { 2337 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2338 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2339 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2340 KNOTE_LOCKED(&lj->klist, 1); 2341 } 2342 if ((lj->lioj_flags & (LIOJ_SIGNAL | 2343 LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL && 2344 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2345 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2346 aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi, 2347 lj->lioj_count != 1); 2348 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2349 } 2350 } 2351 } 2352 lj->lioj_count--; 2353 if (lj->lioj_count == 0) { 2354 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2355 knlist_delete(&lj->klist, curthread, 1); 2356 PROC_LOCK(p); 2357 sigqueue_take(&lj->lioj_ksi); 2358 PROC_UNLOCK(p); 2359 AIO_UNLOCK(ki); 2360 uma_zfree(aiolio_zone, lj); 2361 } else 2362 AIO_UNLOCK(ki); 2363 2364 if (nerror) 2365 return (EIO); 2366 else if (nagain) 2367 return (EAGAIN); 2368 else 2369 return (error); 2370 } 2371 2372 /* syscall - list directed I/O (REALTIME) */ 2373 #ifdef COMPAT_FREEBSD6 2374 int 2375 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap) 2376 { 2377 struct aiocb **acb_list; 2378 struct sigevent *sigp, sig; 2379 struct osigevent osig; 2380 int error, nent; 2381 2382 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2383 return (EINVAL); 2384 2385 nent = uap->nent; 2386 if (nent < 0 || nent > max_aio_queue_per_proc) 2387 return (EINVAL); 2388 2389 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2390 error = copyin(uap->sig, &osig, sizeof(osig)); 2391 if (error) 2392 return (error); 2393 error = convert_old_sigevent(&osig, &sig); 2394 if (error) 2395 return (error); 2396 sigp = &sig; 2397 } else 2398 sigp = NULL; 2399 2400 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2401 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2402 if (error == 0) 2403 error = kern_lio_listio(td, uap->mode, 2404 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2405 &aiocb_ops_osigevent); 2406 free(acb_list, M_LIO); 2407 return (error); 2408 } 2409 #endif 2410 2411 /* syscall - list directed I/O (REALTIME) */ 2412 int 2413 sys_lio_listio(struct thread *td, struct lio_listio_args *uap) 2414 { 2415 struct aiocb **acb_list; 2416 struct sigevent *sigp, sig; 2417 int error, nent; 2418 2419 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2420 return (EINVAL); 2421 2422 nent = uap->nent; 2423 if (nent < 0 || nent > max_aio_queue_per_proc) 2424 return (EINVAL); 2425 2426 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2427 error = copyin(uap->sig, &sig, sizeof(sig)); 2428 if (error) 2429 return (error); 2430 sigp = &sig; 2431 } else 2432 sigp = NULL; 2433 2434 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2435 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2436 if (error == 0) 2437 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list, 2438 nent, sigp, &aiocb_ops); 2439 free(acb_list, M_LIO); 2440 return (error); 2441 } 2442 2443 static void 2444 aio_biocleanup(struct bio *bp) 2445 { 2446 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2447 struct kaioinfo *ki; 2448 struct buf *pbuf = (struct buf *)bp->bio_caller2; 2449 2450 /* Release mapping into kernel space. */ 2451 if (pbuf != NULL) { 2452 MPASS(pbuf->b_npages <= atop(maxphys) + 1); 2453 pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages); 2454 vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages); 2455 uma_zfree(pbuf_zone, pbuf); 2456 atomic_subtract_int(&num_buf_aio, 1); 2457 ki = job->userproc->p_aioinfo; 2458 AIO_LOCK(ki); 2459 ki->kaio_buffer_count--; 2460 AIO_UNLOCK(ki); 2461 } else { 2462 MPASS(bp->bio_ma_n <= atop(maxphys) + 1); 2463 vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n); 2464 free(bp->bio_ma, M_TEMP); 2465 atomic_subtract_int(&num_unmapped_aio, 1); 2466 } 2467 g_destroy_bio(bp); 2468 } 2469 2470 static void 2471 aio_biowakeup(struct bio *bp) 2472 { 2473 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2474 size_t nbytes; 2475 long bcount = bp->bio_bcount; 2476 long resid = bp->bio_resid; 2477 int error, opcode, nblks; 2478 int bio_error = bp->bio_error; 2479 uint16_t flags = bp->bio_flags; 2480 2481 opcode = job->uaiocb.aio_lio_opcode; 2482 2483 aio_biocleanup(bp); 2484 2485 nbytes =bcount - resid; 2486 atomic_add_acq_long(&job->nbytes, nbytes); 2487 nblks = btodb(nbytes); 2488 error = 0; 2489 /* 2490 * If multiple bios experienced an error, the job will reflect the 2491 * error of whichever failed bio completed last. 2492 */ 2493 if (flags & BIO_ERROR) 2494 atomic_set_int(&job->error, bio_error); 2495 if (opcode & LIO_WRITE) 2496 atomic_add_int(&job->outblock, nblks); 2497 else 2498 atomic_add_int(&job->inblock, nblks); 2499 atomic_subtract_int(&job->nbio, 1); 2500 2501 2502 if (atomic_load_int(&job->nbio) == 0) { 2503 if (atomic_load_int(&job->error)) 2504 aio_complete(job, -1, job->error); 2505 else 2506 aio_complete(job, atomic_load_long(&job->nbytes), 0); 2507 } 2508 } 2509 2510 /* syscall - wait for the next completion of an aio request */ 2511 static int 2512 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp, 2513 struct timespec *ts, struct aiocb_ops *ops) 2514 { 2515 struct proc *p = td->td_proc; 2516 struct timeval atv; 2517 struct kaioinfo *ki; 2518 struct kaiocb *job; 2519 struct aiocb *ujob; 2520 long error, status; 2521 int timo; 2522 2523 ops->store_aiocb(ujobp, NULL); 2524 2525 if (ts == NULL) { 2526 timo = 0; 2527 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) { 2528 timo = -1; 2529 } else { 2530 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000)) 2531 return (EINVAL); 2532 2533 TIMESPEC_TO_TIMEVAL(&atv, ts); 2534 if (itimerfix(&atv)) 2535 return (EINVAL); 2536 timo = tvtohz(&atv); 2537 } 2538 2539 if (p->p_aioinfo == NULL) 2540 aio_init_aioinfo(p); 2541 ki = p->p_aioinfo; 2542 2543 error = 0; 2544 job = NULL; 2545 AIO_LOCK(ki); 2546 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2547 if (timo == -1) { 2548 error = EWOULDBLOCK; 2549 break; 2550 } 2551 ki->kaio_flags |= KAIO_WAKEUP; 2552 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2553 "aiowc", timo); 2554 if (timo && error == ERESTART) 2555 error = EINTR; 2556 if (error) 2557 break; 2558 } 2559 2560 if (job != NULL) { 2561 MPASS(job->jobflags & KAIOCB_FINISHED); 2562 ujob = job->ujob; 2563 status = job->uaiocb._aiocb_private.status; 2564 error = job->uaiocb._aiocb_private.error; 2565 td->td_retval[0] = status; 2566 td->td_ru.ru_oublock += job->outblock; 2567 td->td_ru.ru_inblock += job->inblock; 2568 td->td_ru.ru_msgsnd += job->msgsnd; 2569 td->td_ru.ru_msgrcv += job->msgrcv; 2570 aio_free_entry(job); 2571 AIO_UNLOCK(ki); 2572 ops->store_aiocb(ujobp, ujob); 2573 ops->store_error(ujob, error); 2574 ops->store_status(ujob, status); 2575 } else 2576 AIO_UNLOCK(ki); 2577 2578 return (error); 2579 } 2580 2581 int 2582 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2583 { 2584 struct timespec ts, *tsp; 2585 int error; 2586 2587 if (uap->timeout) { 2588 /* Get timespec struct. */ 2589 error = copyin(uap->timeout, &ts, sizeof(ts)); 2590 if (error) 2591 return (error); 2592 tsp = &ts; 2593 } else 2594 tsp = NULL; 2595 2596 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops)); 2597 } 2598 2599 static int 2600 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob, 2601 struct aiocb_ops *ops) 2602 { 2603 int listop; 2604 2605 switch (op) { 2606 case O_SYNC: 2607 listop = LIO_SYNC; 2608 break; 2609 case O_DSYNC: 2610 listop = LIO_DSYNC; 2611 break; 2612 default: 2613 return (EINVAL); 2614 } 2615 2616 return (aio_aqueue(td, ujob, NULL, listop, ops)); 2617 } 2618 2619 int 2620 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2621 { 2622 2623 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops)); 2624 } 2625 2626 /* kqueue attach function */ 2627 static int 2628 filt_aioattach(struct knote *kn) 2629 { 2630 struct kaiocb *job; 2631 2632 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata; 2633 2634 /* 2635 * The job pointer must be validated before using it, so 2636 * registration is restricted to the kernel; the user cannot 2637 * set EV_FLAG1. 2638 */ 2639 if ((kn->kn_flags & EV_FLAG1) == 0) 2640 return (EPERM); 2641 kn->kn_ptr.p_aio = job; 2642 kn->kn_flags &= ~EV_FLAG1; 2643 2644 knlist_add(&job->klist, kn, 0); 2645 2646 return (0); 2647 } 2648 2649 /* kqueue detach function */ 2650 static void 2651 filt_aiodetach(struct knote *kn) 2652 { 2653 struct knlist *knl; 2654 2655 knl = &kn->kn_ptr.p_aio->klist; 2656 knl->kl_lock(knl->kl_lockarg); 2657 if (!knlist_empty(knl)) 2658 knlist_remove(knl, kn, 1); 2659 knl->kl_unlock(knl->kl_lockarg); 2660 } 2661 2662 /* kqueue filter function */ 2663 /*ARGSUSED*/ 2664 static int 2665 filt_aio(struct knote *kn, long hint) 2666 { 2667 struct kaiocb *job = kn->kn_ptr.p_aio; 2668 2669 kn->kn_data = job->uaiocb._aiocb_private.error; 2670 if (!(job->jobflags & KAIOCB_FINISHED)) 2671 return (0); 2672 kn->kn_flags |= EV_EOF; 2673 return (1); 2674 } 2675 2676 /* kqueue attach function */ 2677 static int 2678 filt_lioattach(struct knote *kn) 2679 { 2680 struct aioliojob *lj; 2681 2682 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata; 2683 2684 /* 2685 * The aioliojob pointer must be validated before using it, so 2686 * registration is restricted to the kernel; the user cannot 2687 * set EV_FLAG1. 2688 */ 2689 if ((kn->kn_flags & EV_FLAG1) == 0) 2690 return (EPERM); 2691 kn->kn_ptr.p_lio = lj; 2692 kn->kn_flags &= ~EV_FLAG1; 2693 2694 knlist_add(&lj->klist, kn, 0); 2695 2696 return (0); 2697 } 2698 2699 /* kqueue detach function */ 2700 static void 2701 filt_liodetach(struct knote *kn) 2702 { 2703 struct knlist *knl; 2704 2705 knl = &kn->kn_ptr.p_lio->klist; 2706 knl->kl_lock(knl->kl_lockarg); 2707 if (!knlist_empty(knl)) 2708 knlist_remove(knl, kn, 1); 2709 knl->kl_unlock(knl->kl_lockarg); 2710 } 2711 2712 /* kqueue filter function */ 2713 /*ARGSUSED*/ 2714 static int 2715 filt_lio(struct knote *kn, long hint) 2716 { 2717 struct aioliojob * lj = kn->kn_ptr.p_lio; 2718 2719 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2720 } 2721 2722 #ifdef COMPAT_FREEBSD32 2723 #include <sys/mount.h> 2724 #include <sys/socket.h> 2725 #include <compat/freebsd32/freebsd32.h> 2726 #include <compat/freebsd32/freebsd32_proto.h> 2727 #include <compat/freebsd32/freebsd32_signal.h> 2728 #include <compat/freebsd32/freebsd32_syscall.h> 2729 #include <compat/freebsd32/freebsd32_util.h> 2730 2731 struct __aiocb_private32 { 2732 int32_t status; 2733 int32_t error; 2734 uint32_t kernelinfo; 2735 }; 2736 2737 #ifdef COMPAT_FREEBSD6 2738 typedef struct oaiocb32 { 2739 int aio_fildes; /* File descriptor */ 2740 uint64_t aio_offset __packed; /* File offset for I/O */ 2741 uint32_t aio_buf; /* I/O buffer in process space */ 2742 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2743 struct osigevent32 aio_sigevent; /* Signal to deliver */ 2744 int aio_lio_opcode; /* LIO opcode */ 2745 int aio_reqprio; /* Request priority -- ignored */ 2746 struct __aiocb_private32 _aiocb_private; 2747 } oaiocb32_t; 2748 #endif 2749 2750 typedef struct aiocb32 { 2751 int32_t aio_fildes; /* File descriptor */ 2752 uint64_t aio_offset __packed; /* File offset for I/O */ 2753 uint32_t aio_buf; /* I/O buffer in process space */ 2754 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2755 int __spare__[2]; 2756 uint32_t __spare2__; 2757 int aio_lio_opcode; /* LIO opcode */ 2758 int aio_reqprio; /* Request priority -- ignored */ 2759 struct __aiocb_private32 _aiocb_private; 2760 struct sigevent32 aio_sigevent; /* Signal to deliver */ 2761 } aiocb32_t; 2762 2763 #ifdef COMPAT_FREEBSD6 2764 static int 2765 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig) 2766 { 2767 2768 /* 2769 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 2770 * supported by AIO with the old sigevent structure. 2771 */ 2772 CP(*osig, *nsig, sigev_notify); 2773 switch (nsig->sigev_notify) { 2774 case SIGEV_NONE: 2775 break; 2776 case SIGEV_SIGNAL: 2777 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 2778 break; 2779 case SIGEV_KEVENT: 2780 nsig->sigev_notify_kqueue = 2781 osig->__sigev_u.__sigev_notify_kqueue; 2782 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr); 2783 break; 2784 default: 2785 return (EINVAL); 2786 } 2787 return (0); 2788 } 2789 2790 static int 2791 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob, 2792 int type __unused) 2793 { 2794 struct oaiocb32 job32; 2795 struct aiocb *kcb = &kjob->uaiocb; 2796 int error; 2797 2798 bzero(kcb, sizeof(struct aiocb)); 2799 error = copyin(ujob, &job32, sizeof(job32)); 2800 if (error) 2801 return (error); 2802 2803 /* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */ 2804 2805 CP(job32, *kcb, aio_fildes); 2806 CP(job32, *kcb, aio_offset); 2807 PTRIN_CP(job32, *kcb, aio_buf); 2808 CP(job32, *kcb, aio_nbytes); 2809 CP(job32, *kcb, aio_lio_opcode); 2810 CP(job32, *kcb, aio_reqprio); 2811 CP(job32, *kcb, _aiocb_private.status); 2812 CP(job32, *kcb, _aiocb_private.error); 2813 PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo); 2814 return (convert_old_sigevent32(&job32.aio_sigevent, 2815 &kcb->aio_sigevent)); 2816 } 2817 #endif 2818 2819 static int 2820 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type) 2821 { 2822 struct aiocb32 job32; 2823 struct aiocb *kcb = &kjob->uaiocb; 2824 struct iovec32 *iov32; 2825 int error; 2826 2827 error = copyin(ujob, &job32, sizeof(job32)); 2828 if (error) 2829 return (error); 2830 CP(job32, *kcb, aio_fildes); 2831 CP(job32, *kcb, aio_offset); 2832 CP(job32, *kcb, aio_lio_opcode); 2833 if (type & LIO_VECTORED) { 2834 iov32 = PTRIN(job32.aio_iov); 2835 CP(job32, *kcb, aio_iovcnt); 2836 /* malloc a uio and copy in the iovec */ 2837 error = freebsd32_copyinuio(iov32, 2838 kcb->aio_iovcnt, &kjob->uiop); 2839 if (error) 2840 return (error); 2841 } else { 2842 PTRIN_CP(job32, *kcb, aio_buf); 2843 CP(job32, *kcb, aio_nbytes); 2844 } 2845 CP(job32, *kcb, aio_reqprio); 2846 CP(job32, *kcb, _aiocb_private.status); 2847 CP(job32, *kcb, _aiocb_private.error); 2848 PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo); 2849 error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent); 2850 2851 return (error); 2852 } 2853 2854 static long 2855 aiocb32_fetch_status(struct aiocb *ujob) 2856 { 2857 struct aiocb32 *ujob32; 2858 2859 ujob32 = (struct aiocb32 *)ujob; 2860 return (fuword32(&ujob32->_aiocb_private.status)); 2861 } 2862 2863 static long 2864 aiocb32_fetch_error(struct aiocb *ujob) 2865 { 2866 struct aiocb32 *ujob32; 2867 2868 ujob32 = (struct aiocb32 *)ujob; 2869 return (fuword32(&ujob32->_aiocb_private.error)); 2870 } 2871 2872 static int 2873 aiocb32_store_status(struct aiocb *ujob, long status) 2874 { 2875 struct aiocb32 *ujob32; 2876 2877 ujob32 = (struct aiocb32 *)ujob; 2878 return (suword32(&ujob32->_aiocb_private.status, status)); 2879 } 2880 2881 static int 2882 aiocb32_store_error(struct aiocb *ujob, long error) 2883 { 2884 struct aiocb32 *ujob32; 2885 2886 ujob32 = (struct aiocb32 *)ujob; 2887 return (suword32(&ujob32->_aiocb_private.error, error)); 2888 } 2889 2890 static int 2891 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref) 2892 { 2893 struct aiocb32 *ujob32; 2894 2895 ujob32 = (struct aiocb32 *)ujob; 2896 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref)); 2897 } 2898 2899 static int 2900 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 2901 { 2902 2903 return (suword32(ujobp, (long)ujob)); 2904 } 2905 2906 static struct aiocb_ops aiocb32_ops = { 2907 .aio_copyin = aiocb32_copyin, 2908 .fetch_status = aiocb32_fetch_status, 2909 .fetch_error = aiocb32_fetch_error, 2910 .store_status = aiocb32_store_status, 2911 .store_error = aiocb32_store_error, 2912 .store_kernelinfo = aiocb32_store_kernelinfo, 2913 .store_aiocb = aiocb32_store_aiocb, 2914 }; 2915 2916 #ifdef COMPAT_FREEBSD6 2917 static struct aiocb_ops aiocb32_ops_osigevent = { 2918 .aio_copyin = aiocb32_copyin_old_sigevent, 2919 .fetch_status = aiocb32_fetch_status, 2920 .fetch_error = aiocb32_fetch_error, 2921 .store_status = aiocb32_store_status, 2922 .store_error = aiocb32_store_error, 2923 .store_kernelinfo = aiocb32_store_kernelinfo, 2924 .store_aiocb = aiocb32_store_aiocb, 2925 }; 2926 #endif 2927 2928 int 2929 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap) 2930 { 2931 2932 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2933 } 2934 2935 int 2936 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap) 2937 { 2938 struct timespec32 ts32; 2939 struct timespec ts, *tsp; 2940 struct aiocb **ujoblist; 2941 uint32_t *ujoblist32; 2942 int error, i; 2943 2944 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 2945 return (EINVAL); 2946 2947 if (uap->timeout) { 2948 /* Get timespec struct. */ 2949 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0) 2950 return (error); 2951 CP(ts32, ts, tv_sec); 2952 CP(ts32, ts, tv_nsec); 2953 tsp = &ts; 2954 } else 2955 tsp = NULL; 2956 2957 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK); 2958 ujoblist32 = (uint32_t *)ujoblist; 2959 error = copyin(uap->aiocbp, ujoblist32, uap->nent * 2960 sizeof(ujoblist32[0])); 2961 if (error == 0) { 2962 for (i = uap->nent - 1; i >= 0; i--) 2963 ujoblist[i] = PTRIN(ujoblist32[i]); 2964 2965 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2966 } 2967 free(ujoblist, M_AIOS); 2968 return (error); 2969 } 2970 2971 int 2972 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap) 2973 { 2974 2975 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2976 } 2977 2978 #ifdef COMPAT_FREEBSD6 2979 int 2980 freebsd6_freebsd32_aio_read(struct thread *td, 2981 struct freebsd6_freebsd32_aio_read_args *uap) 2982 { 2983 2984 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2985 &aiocb32_ops_osigevent)); 2986 } 2987 #endif 2988 2989 int 2990 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap) 2991 { 2992 2993 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2994 &aiocb32_ops)); 2995 } 2996 2997 int 2998 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap) 2999 { 3000 3001 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV, 3002 &aiocb32_ops)); 3003 } 3004 3005 #ifdef COMPAT_FREEBSD6 3006 int 3007 freebsd6_freebsd32_aio_write(struct thread *td, 3008 struct freebsd6_freebsd32_aio_write_args *uap) 3009 { 3010 3011 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 3012 &aiocb32_ops_osigevent)); 3013 } 3014 #endif 3015 3016 int 3017 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap) 3018 { 3019 3020 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 3021 &aiocb32_ops)); 3022 } 3023 3024 int 3025 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap) 3026 { 3027 3028 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV, 3029 &aiocb32_ops)); 3030 } 3031 3032 int 3033 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap) 3034 { 3035 3036 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK, 3037 &aiocb32_ops)); 3038 } 3039 3040 int 3041 freebsd32_aio_waitcomplete(struct thread *td, 3042 struct freebsd32_aio_waitcomplete_args *uap) 3043 { 3044 struct timespec32 ts32; 3045 struct timespec ts, *tsp; 3046 int error; 3047 3048 if (uap->timeout) { 3049 /* Get timespec struct. */ 3050 error = copyin(uap->timeout, &ts32, sizeof(ts32)); 3051 if (error) 3052 return (error); 3053 CP(ts32, ts, tv_sec); 3054 CP(ts32, ts, tv_nsec); 3055 tsp = &ts; 3056 } else 3057 tsp = NULL; 3058 3059 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp, 3060 &aiocb32_ops)); 3061 } 3062 3063 int 3064 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap) 3065 { 3066 3067 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp, 3068 &aiocb32_ops)); 3069 } 3070 3071 #ifdef COMPAT_FREEBSD6 3072 int 3073 freebsd6_freebsd32_lio_listio(struct thread *td, 3074 struct freebsd6_freebsd32_lio_listio_args *uap) 3075 { 3076 struct aiocb **acb_list; 3077 struct sigevent *sigp, sig; 3078 struct osigevent32 osig; 3079 uint32_t *acb_list32; 3080 int error, i, nent; 3081 3082 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3083 return (EINVAL); 3084 3085 nent = uap->nent; 3086 if (nent < 0 || nent > max_aio_queue_per_proc) 3087 return (EINVAL); 3088 3089 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3090 error = copyin(uap->sig, &osig, sizeof(osig)); 3091 if (error) 3092 return (error); 3093 error = convert_old_sigevent32(&osig, &sig); 3094 if (error) 3095 return (error); 3096 sigp = &sig; 3097 } else 3098 sigp = NULL; 3099 3100 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3101 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3102 if (error) { 3103 free(acb_list32, M_LIO); 3104 return (error); 3105 } 3106 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3107 for (i = 0; i < nent; i++) 3108 acb_list[i] = PTRIN(acb_list32[i]); 3109 free(acb_list32, M_LIO); 3110 3111 error = kern_lio_listio(td, uap->mode, 3112 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3113 &aiocb32_ops_osigevent); 3114 free(acb_list, M_LIO); 3115 return (error); 3116 } 3117 #endif 3118 3119 int 3120 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap) 3121 { 3122 struct aiocb **acb_list; 3123 struct sigevent *sigp, sig; 3124 struct sigevent32 sig32; 3125 uint32_t *acb_list32; 3126 int error, i, nent; 3127 3128 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3129 return (EINVAL); 3130 3131 nent = uap->nent; 3132 if (nent < 0 || nent > max_aio_queue_per_proc) 3133 return (EINVAL); 3134 3135 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3136 error = copyin(uap->sig, &sig32, sizeof(sig32)); 3137 if (error) 3138 return (error); 3139 error = convert_sigevent32(&sig32, &sig); 3140 if (error) 3141 return (error); 3142 sigp = &sig; 3143 } else 3144 sigp = NULL; 3145 3146 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3147 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3148 if (error) { 3149 free(acb_list32, M_LIO); 3150 return (error); 3151 } 3152 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3153 for (i = 0; i < nent; i++) 3154 acb_list[i] = PTRIN(acb_list32[i]); 3155 free(acb_list32, M_LIO); 3156 3157 error = kern_lio_listio(td, uap->mode, 3158 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3159 &aiocb32_ops); 3160 free(acb_list, M_LIO); 3161 return (error); 3162 } 3163 3164 #endif
Cache object: 2d1a3187994fff03d0e2b66c4f306a6f
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]