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/sysctl.h> 56 #include <sys/syslog.h> 57 #include <sys/sx.h> 58 #include <sys/taskqueue.h> 59 #include <sys/vnode.h> 60 #include <sys/conf.h> 61 #include <sys/event.h> 62 #include <sys/mount.h> 63 #include <geom/geom.h> 64 65 #include <machine/atomic.h> 66 67 #include <vm/vm.h> 68 #include <vm/vm_page.h> 69 #include <vm/vm_extern.h> 70 #include <vm/pmap.h> 71 #include <vm/vm_map.h> 72 #include <vm/vm_object.h> 73 #include <vm/uma.h> 74 #include <sys/aio.h> 75 76 /* 77 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 78 * overflow. (XXX will be removed soon.) 79 */ 80 static u_long jobrefid; 81 82 /* 83 * Counter for aio_fsync. 84 */ 85 static uint64_t jobseqno; 86 87 #ifndef MAX_AIO_PER_PROC 88 #define MAX_AIO_PER_PROC 32 89 #endif 90 91 #ifndef MAX_AIO_QUEUE_PER_PROC 92 #define MAX_AIO_QUEUE_PER_PROC 256 93 #endif 94 95 #ifndef MAX_AIO_QUEUE 96 #define MAX_AIO_QUEUE 1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */ 97 #endif 98 99 #ifndef MAX_BUF_AIO 100 #define MAX_BUF_AIO 16 101 #endif 102 103 FEATURE(aio, "Asynchronous I/O"); 104 SYSCTL_DECL(_p1003_1b); 105 106 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list"); 107 static MALLOC_DEFINE(M_AIO, "aio", "structures for asynchronous I/O"); 108 109 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 110 "Async IO management"); 111 112 static int enable_aio_unsafe = 0; 113 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0, 114 "Permit asynchronous IO on all file types, not just known-safe types"); 115 116 static unsigned int unsafe_warningcnt = 1; 117 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW, 118 &unsafe_warningcnt, 0, 119 "Warnings that will be triggered upon failed IO requests on unsafe files"); 120 121 static int max_aio_procs = MAX_AIO_PROCS; 122 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0, 123 "Maximum number of kernel processes to use for handling async IO "); 124 125 static int num_aio_procs = 0; 126 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0, 127 "Number of presently active kernel processes for async IO"); 128 129 /* 130 * The code will adjust the actual number of AIO processes towards this 131 * number when it gets a chance. 132 */ 133 static int target_aio_procs = TARGET_AIO_PROCS; 134 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 135 0, 136 "Preferred number of ready kernel processes for async IO"); 137 138 static int max_queue_count = MAX_AIO_QUEUE; 139 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 140 "Maximum number of aio requests to queue, globally"); 141 142 static int num_queue_count = 0; 143 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 144 "Number of queued aio requests"); 145 146 static int num_buf_aio = 0; 147 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 148 "Number of aio requests presently handled by the buf subsystem"); 149 150 static int num_unmapped_aio = 0; 151 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio, 152 0, 153 "Number of aio requests presently handled by unmapped I/O buffers"); 154 155 /* Number of async I/O processes in the process of being started */ 156 /* XXX This should be local to aio_aqueue() */ 157 static int num_aio_resv_start = 0; 158 159 static int aiod_lifetime; 160 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 161 "Maximum lifetime for idle aiod"); 162 163 static int max_aio_per_proc = MAX_AIO_PER_PROC; 164 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 165 0, 166 "Maximum active aio requests per process"); 167 168 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 169 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 170 &max_aio_queue_per_proc, 0, 171 "Maximum queued aio requests per process"); 172 173 static int max_buf_aio = MAX_BUF_AIO; 174 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 175 "Maximum buf aio requests per process"); 176 177 /* 178 * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires 179 * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with 180 * vfs.aio.aio_listio_max. 181 */ 182 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max, 183 CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc, 184 0, "Maximum aio requests for a single lio_listio call"); 185 186 #ifdef COMPAT_FREEBSD6 187 typedef struct oaiocb { 188 int aio_fildes; /* File descriptor */ 189 off_t aio_offset; /* File offset for I/O */ 190 volatile void *aio_buf; /* I/O buffer in process space */ 191 size_t aio_nbytes; /* Number of bytes for I/O */ 192 struct osigevent aio_sigevent; /* Signal to deliver */ 193 int aio_lio_opcode; /* LIO opcode */ 194 int aio_reqprio; /* Request priority -- ignored */ 195 struct __aiocb_private _aiocb_private; 196 } oaiocb_t; 197 #endif 198 199 /* 200 * Below is a key of locks used to protect each member of struct kaiocb 201 * aioliojob and kaioinfo and any backends. 202 * 203 * * - need not protected 204 * a - locked by kaioinfo lock 205 * b - locked by backend lock, the backend lock can be null in some cases, 206 * for example, BIO belongs to this type, in this case, proc lock is 207 * reused. 208 * c - locked by aio_job_mtx, the lock for the generic file I/O backend. 209 */ 210 211 /* 212 * If the routine that services an AIO request blocks while running in an 213 * AIO kernel process it can starve other I/O requests. BIO requests 214 * queued via aio_qbio() complete asynchronously and do not use AIO kernel 215 * processes at all. Socket I/O requests use a separate pool of 216 * kprocs and also force non-blocking I/O. Other file I/O requests 217 * use the generic fo_read/fo_write operations which can block. The 218 * fsync and mlock operations can also block while executing. Ideally 219 * none of these requests would block while executing. 220 * 221 * Note that the service routines cannot toggle O_NONBLOCK in the file 222 * structure directly while handling a request due to races with 223 * userland threads. 224 */ 225 226 /* jobflags */ 227 #define KAIOCB_QUEUEING 0x01 228 #define KAIOCB_CANCELLED 0x02 229 #define KAIOCB_CANCELLING 0x04 230 #define KAIOCB_CHECKSYNC 0x08 231 #define KAIOCB_CLEARED 0x10 232 #define KAIOCB_FINISHED 0x20 233 234 /* 235 * AIO process info 236 */ 237 #define AIOP_FREE 0x1 /* proc on free queue */ 238 239 struct aioproc { 240 int aioprocflags; /* (c) AIO proc flags */ 241 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */ 242 struct proc *aioproc; /* (*) the AIO proc */ 243 }; 244 245 /* 246 * data-structure for lio signal management 247 */ 248 struct aioliojob { 249 int lioj_flags; /* (a) listio flags */ 250 int lioj_count; /* (a) count of jobs */ 251 int lioj_finished_count; /* (a) count of finished jobs */ 252 struct sigevent lioj_signal; /* (a) signal on all I/O done */ 253 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */ 254 struct knlist klist; /* (a) list of knotes */ 255 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */ 256 }; 257 258 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 259 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 260 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */ 261 262 /* 263 * per process aio data structure 264 */ 265 struct kaioinfo { 266 struct mtx kaio_mtx; /* the lock to protect this struct */ 267 int kaio_flags; /* (a) per process kaio flags */ 268 int kaio_active_count; /* (c) number of currently used AIOs */ 269 int kaio_count; /* (a) size of AIO queue */ 270 int kaio_buffer_count; /* (a) number of bio buffers */ 271 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */ 272 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */ 273 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */ 274 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */ 275 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */ 276 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */ 277 struct task kaio_task; /* (*) task to kick aio processes */ 278 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */ 279 }; 280 281 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx) 282 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx) 283 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f)) 284 #define AIO_MTX(ki) (&(ki)->kaio_mtx) 285 286 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 287 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */ 288 289 /* 290 * Operations used to interact with userland aio control blocks. 291 * Different ABIs provide their own operations. 292 */ 293 struct aiocb_ops { 294 int (*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty); 295 long (*fetch_status)(struct aiocb *ujob); 296 long (*fetch_error)(struct aiocb *ujob); 297 int (*store_status)(struct aiocb *ujob, long status); 298 int (*store_error)(struct aiocb *ujob, long error); 299 int (*store_kernelinfo)(struct aiocb *ujob, long jobref); 300 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob); 301 }; 302 303 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */ 304 static struct sema aio_newproc_sem; 305 static struct mtx aio_job_mtx; 306 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */ 307 static struct unrhdr *aiod_unr; 308 309 static void aio_biocleanup(struct bio *bp); 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 * aiocb async io jobs 343 * aiolio list io jobs 344 */ 345 static uma_zone_t kaio_zone, aiocb_zone, aiolio_zone; 346 347 /* kqueue filters for aio */ 348 static struct filterops aio_filtops = { 349 .f_isfd = 0, 350 .f_attach = filt_aioattach, 351 .f_detach = filt_aiodetach, 352 .f_event = filt_aio, 353 }; 354 static struct filterops lio_filtops = { 355 .f_isfd = 0, 356 .f_attach = filt_lioattach, 357 .f_detach = filt_liodetach, 358 .f_event = filt_lio 359 }; 360 361 static eventhandler_tag exit_tag, exec_tag; 362 363 TASKQUEUE_DEFINE_THREAD(aiod_kick); 364 365 /* 366 * Main operations function for use as a kernel module. 367 */ 368 static int 369 aio_modload(struct module *module, int cmd, void *arg) 370 { 371 int error = 0; 372 373 switch (cmd) { 374 case MOD_LOAD: 375 aio_onceonly(); 376 break; 377 case MOD_SHUTDOWN: 378 break; 379 default: 380 error = EOPNOTSUPP; 381 break; 382 } 383 return (error); 384 } 385 386 static moduledata_t aio_mod = { 387 "aio", 388 &aio_modload, 389 NULL 390 }; 391 392 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY); 393 MODULE_VERSION(aio, 1); 394 395 /* 396 * Startup initialization 397 */ 398 static int 399 aio_onceonly(void) 400 { 401 402 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 403 EVENTHANDLER_PRI_ANY); 404 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, 405 NULL, EVENTHANDLER_PRI_ANY); 406 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 407 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 408 TAILQ_INIT(&aio_freeproc); 409 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 410 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 411 TAILQ_INIT(&aio_jobs); 412 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 413 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 414 NULL, NULL, UMA_ALIGN_PTR, 0); 415 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL, 416 NULL, NULL, UMA_ALIGN_PTR, 0); 417 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 418 NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 419 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 420 jobrefid = 1; 421 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO); 422 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 423 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 424 425 return (0); 426 } 427 428 /* 429 * Init the per-process aioinfo structure. The aioinfo limits are set 430 * per-process for user limit (resource) management. 431 */ 432 void 433 aio_init_aioinfo(struct proc *p) 434 { 435 struct kaioinfo *ki; 436 437 ki = uma_zalloc(kaio_zone, M_WAITOK); 438 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW); 439 ki->kaio_flags = 0; 440 ki->kaio_active_count = 0; 441 ki->kaio_count = 0; 442 ki->kaio_buffer_count = 0; 443 TAILQ_INIT(&ki->kaio_all); 444 TAILQ_INIT(&ki->kaio_done); 445 TAILQ_INIT(&ki->kaio_jobqueue); 446 TAILQ_INIT(&ki->kaio_liojoblist); 447 TAILQ_INIT(&ki->kaio_syncqueue); 448 TAILQ_INIT(&ki->kaio_syncready); 449 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 450 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki); 451 PROC_LOCK(p); 452 if (p->p_aioinfo == NULL) { 453 p->p_aioinfo = ki; 454 PROC_UNLOCK(p); 455 } else { 456 PROC_UNLOCK(p); 457 mtx_destroy(&ki->kaio_mtx); 458 uma_zfree(kaio_zone, ki); 459 } 460 461 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs)) 462 aio_newproc(NULL); 463 } 464 465 static int 466 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext) 467 { 468 struct thread *td; 469 int error; 470 471 error = sigev_findtd(p, sigev, &td); 472 if (error) 473 return (error); 474 if (!KSI_ONQ(ksi)) { 475 ksiginfo_set_sigev(ksi, sigev); 476 ksi->ksi_code = SI_ASYNCIO; 477 ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0; 478 tdsendsignal(p, td, ksi->ksi_signo, ksi); 479 } 480 PROC_UNLOCK(p); 481 return (error); 482 } 483 484 /* 485 * Free a job entry. Wait for completion if it is currently active, but don't 486 * delay forever. If we delay, we return a flag that says that we have to 487 * restart the queue scan. 488 */ 489 static int 490 aio_free_entry(struct kaiocb *job) 491 { 492 struct kaioinfo *ki; 493 struct aioliojob *lj; 494 struct proc *p; 495 496 p = job->userproc; 497 MPASS(curproc == p); 498 ki = p->p_aioinfo; 499 MPASS(ki != NULL); 500 501 AIO_LOCK_ASSERT(ki, MA_OWNED); 502 MPASS(job->jobflags & KAIOCB_FINISHED); 503 504 atomic_subtract_int(&num_queue_count, 1); 505 506 ki->kaio_count--; 507 MPASS(ki->kaio_count >= 0); 508 509 TAILQ_REMOVE(&ki->kaio_done, job, plist); 510 TAILQ_REMOVE(&ki->kaio_all, job, allist); 511 512 lj = job->lio; 513 if (lj) { 514 lj->lioj_count--; 515 lj->lioj_finished_count--; 516 517 if (lj->lioj_count == 0) { 518 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 519 /* lio is going away, we need to destroy any knotes */ 520 knlist_delete(&lj->klist, curthread, 1); 521 PROC_LOCK(p); 522 sigqueue_take(&lj->lioj_ksi); 523 PROC_UNLOCK(p); 524 uma_zfree(aiolio_zone, lj); 525 } 526 } 527 528 /* job is going away, we need to destroy any knotes */ 529 knlist_delete(&job->klist, curthread, 1); 530 PROC_LOCK(p); 531 sigqueue_take(&job->ksi); 532 PROC_UNLOCK(p); 533 534 AIO_UNLOCK(ki); 535 536 /* 537 * The thread argument here is used to find the owning process 538 * and is also passed to fo_close() which may pass it to various 539 * places such as devsw close() routines. Because of that, we 540 * need a thread pointer from the process owning the job that is 541 * persistent and won't disappear out from under us or move to 542 * another process. 543 * 544 * Currently, all the callers of this function call it to remove 545 * a kaiocb from the current process' job list either via a 546 * syscall or due to the current process calling exit() or 547 * execve(). Thus, we know that p == curproc. We also know that 548 * curthread can't exit since we are curthread. 549 * 550 * Therefore, we use curthread as the thread to pass to 551 * knlist_delete(). This does mean that it is possible for the 552 * thread pointer at close time to differ from the thread pointer 553 * at open time, but this is already true of file descriptors in 554 * a multithreaded process. 555 */ 556 if (job->fd_file) 557 fdrop(job->fd_file, curthread); 558 crfree(job->cred); 559 if (job->uiop != &job->uio) 560 free(job->uiop, M_IOV); 561 uma_zfree(aiocb_zone, job); 562 AIO_LOCK(ki); 563 564 return (0); 565 } 566 567 static void 568 aio_proc_rundown_exec(void *arg, struct proc *p, 569 struct image_params *imgp __unused) 570 { 571 aio_proc_rundown(arg, p); 572 } 573 574 static int 575 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job) 576 { 577 aio_cancel_fn_t *func; 578 int cancelled; 579 580 AIO_LOCK_ASSERT(ki, MA_OWNED); 581 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED)) 582 return (0); 583 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0); 584 job->jobflags |= KAIOCB_CANCELLED; 585 586 func = job->cancel_fn; 587 588 /* 589 * If there is no cancel routine, just leave the job marked as 590 * cancelled. The job should be in active use by a caller who 591 * should complete it normally or when it fails to install a 592 * cancel routine. 593 */ 594 if (func == NULL) 595 return (0); 596 597 /* 598 * Set the CANCELLING flag so that aio_complete() will defer 599 * completions of this job. This prevents the job from being 600 * freed out from under the cancel callback. After the 601 * callback any deferred completion (whether from the callback 602 * or any other source) will be completed. 603 */ 604 job->jobflags |= KAIOCB_CANCELLING; 605 AIO_UNLOCK(ki); 606 func(job); 607 AIO_LOCK(ki); 608 job->jobflags &= ~KAIOCB_CANCELLING; 609 if (job->jobflags & KAIOCB_FINISHED) { 610 cancelled = job->uaiocb._aiocb_private.error == ECANCELED; 611 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 612 aio_bio_done_notify(p, job); 613 } else { 614 /* 615 * The cancel callback might have scheduled an 616 * operation to cancel this request, but it is 617 * only counted as cancelled if the request is 618 * cancelled when the callback returns. 619 */ 620 cancelled = 0; 621 } 622 return (cancelled); 623 } 624 625 /* 626 * Rundown the jobs for a given process. 627 */ 628 static void 629 aio_proc_rundown(void *arg, struct proc *p) 630 { 631 struct kaioinfo *ki; 632 struct aioliojob *lj; 633 struct kaiocb *job, *jobn; 634 635 KASSERT(curthread->td_proc == p, 636 ("%s: called on non-curproc", __func__)); 637 ki = p->p_aioinfo; 638 if (ki == NULL) 639 return; 640 641 AIO_LOCK(ki); 642 ki->kaio_flags |= KAIO_RUNDOWN; 643 644 restart: 645 646 /* 647 * Try to cancel all pending requests. This code simulates 648 * aio_cancel on all pending I/O requests. 649 */ 650 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 651 aio_cancel_job(p, ki, job); 652 } 653 654 /* Wait for all running I/O to be finished */ 655 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) { 656 ki->kaio_flags |= KAIO_WAKEUP; 657 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 658 goto restart; 659 } 660 661 /* Free all completed I/O requests. */ 662 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL) 663 aio_free_entry(job); 664 665 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 666 if (lj->lioj_count == 0) { 667 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 668 knlist_delete(&lj->klist, curthread, 1); 669 PROC_LOCK(p); 670 sigqueue_take(&lj->lioj_ksi); 671 PROC_UNLOCK(p); 672 uma_zfree(aiolio_zone, lj); 673 } else { 674 panic("LIO job not cleaned up: C:%d, FC:%d\n", 675 lj->lioj_count, lj->lioj_finished_count); 676 } 677 } 678 AIO_UNLOCK(ki); 679 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task); 680 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task); 681 mtx_destroy(&ki->kaio_mtx); 682 uma_zfree(kaio_zone, ki); 683 p->p_aioinfo = NULL; 684 } 685 686 /* 687 * Select a job to run (called by an AIO daemon). 688 */ 689 static struct kaiocb * 690 aio_selectjob(struct aioproc *aiop) 691 { 692 struct kaiocb *job; 693 struct kaioinfo *ki; 694 struct proc *userp; 695 696 mtx_assert(&aio_job_mtx, MA_OWNED); 697 restart: 698 TAILQ_FOREACH(job, &aio_jobs, list) { 699 userp = job->userproc; 700 ki = userp->p_aioinfo; 701 702 if (ki->kaio_active_count < max_aio_per_proc) { 703 TAILQ_REMOVE(&aio_jobs, job, list); 704 if (!aio_clear_cancel_function(job)) 705 goto restart; 706 707 /* Account for currently active jobs. */ 708 ki->kaio_active_count++; 709 break; 710 } 711 } 712 return (job); 713 } 714 715 /* 716 * Move all data to a permanent storage device. This code 717 * simulates the fsync and fdatasync syscalls. 718 */ 719 static int 720 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op) 721 { 722 struct mount *mp; 723 vm_object_t obj; 724 int error; 725 726 for (;;) { 727 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 728 if (error != 0) 729 break; 730 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 731 obj = vp->v_object; 732 if (obj != NULL) { 733 VM_OBJECT_WLOCK(obj); 734 vm_object_page_clean(obj, 0, 0, 0); 735 VM_OBJECT_WUNLOCK(obj); 736 } 737 if (op == LIO_DSYNC) 738 error = VOP_FDATASYNC(vp, td); 739 else 740 error = VOP_FSYNC(vp, MNT_WAIT, td); 741 742 VOP_UNLOCK(vp); 743 vn_finished_write(mp); 744 if (error != ERELOOKUP) 745 break; 746 } 747 return (error); 748 } 749 750 /* 751 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that 752 * does the I/O request for the non-bio version of the operations. The normal 753 * vn operations are used, and this code should work in all instances for every 754 * type of file, including pipes, sockets, fifos, and regular files. 755 * 756 * XXX I don't think it works well for socket, pipe, and fifo. 757 */ 758 static void 759 aio_process_rw(struct kaiocb *job) 760 { 761 struct ucred *td_savedcred; 762 struct thread *td; 763 struct aiocb *cb; 764 struct file *fp; 765 ssize_t cnt; 766 long msgsnd_st, msgsnd_end; 767 long msgrcv_st, msgrcv_end; 768 long oublock_st, oublock_end; 769 long inblock_st, inblock_end; 770 int error, opcode; 771 772 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ || 773 job->uaiocb.aio_lio_opcode == LIO_READV || 774 job->uaiocb.aio_lio_opcode == LIO_WRITE || 775 job->uaiocb.aio_lio_opcode == LIO_WRITEV, 776 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 777 778 aio_switch_vmspace(job); 779 td = curthread; 780 td_savedcred = td->td_ucred; 781 td->td_ucred = job->cred; 782 job->uiop->uio_td = td; 783 cb = &job->uaiocb; 784 fp = job->fd_file; 785 786 opcode = job->uaiocb.aio_lio_opcode; 787 cnt = job->uiop->uio_resid; 788 789 msgrcv_st = td->td_ru.ru_msgrcv; 790 msgsnd_st = td->td_ru.ru_msgsnd; 791 inblock_st = td->td_ru.ru_inblock; 792 oublock_st = td->td_ru.ru_oublock; 793 794 /* 795 * aio_aqueue() acquires a reference to the file that is 796 * released in aio_free_entry(). 797 */ 798 if (opcode == LIO_READ || opcode == LIO_READV) { 799 if (job->uiop->uio_resid == 0) 800 error = 0; 801 else 802 error = fo_read(fp, job->uiop, fp->f_cred, FOF_OFFSET, 803 td); 804 } else { 805 if (fp->f_type == DTYPE_VNODE) 806 bwillwrite(); 807 error = fo_write(fp, job->uiop, 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 != 0 && job->uiop->uio_resid != cnt) { 820 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 821 error = 0; 822 if (error == EPIPE && (opcode & LIO_WRITE)) { 823 PROC_LOCK(job->userproc); 824 kern_psignal(job->userproc, SIGPIPE); 825 PROC_UNLOCK(job->userproc); 826 } 827 } 828 829 cnt -= job->uiop->uio_resid; 830 td->td_ucred = td_savedcred; 831 if (error) 832 aio_complete(job, -1, error); 833 else 834 aio_complete(job, cnt, 0); 835 } 836 837 static void 838 aio_process_sync(struct kaiocb *job) 839 { 840 struct thread *td = curthread; 841 struct ucred *td_savedcred = td->td_ucred; 842 struct file *fp = job->fd_file; 843 int error = 0; 844 845 KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC, 846 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 847 848 td->td_ucred = job->cred; 849 if (fp->f_vnode != NULL) { 850 error = aio_fsync_vnode(td, fp->f_vnode, 851 job->uaiocb.aio_lio_opcode); 852 } 853 td->td_ucred = td_savedcred; 854 if (error) 855 aio_complete(job, -1, error); 856 else 857 aio_complete(job, 0, 0); 858 } 859 860 static void 861 aio_process_mlock(struct kaiocb *job) 862 { 863 struct aiocb *cb = &job->uaiocb; 864 int error; 865 866 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK, 867 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 868 869 aio_switch_vmspace(job); 870 error = kern_mlock(job->userproc, job->cred, 871 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes); 872 aio_complete(job, error != 0 ? -1 : 0, error); 873 } 874 875 static void 876 aio_bio_done_notify(struct proc *userp, struct kaiocb *job) 877 { 878 struct aioliojob *lj; 879 struct kaioinfo *ki; 880 struct kaiocb *sjob, *sjobn; 881 int lj_done; 882 bool schedule_fsync; 883 884 ki = userp->p_aioinfo; 885 AIO_LOCK_ASSERT(ki, MA_OWNED); 886 lj = job->lio; 887 lj_done = 0; 888 if (lj) { 889 lj->lioj_finished_count++; 890 if (lj->lioj_count == lj->lioj_finished_count) 891 lj_done = 1; 892 } 893 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist); 894 MPASS(job->jobflags & KAIOCB_FINISHED); 895 896 if (ki->kaio_flags & KAIO_RUNDOWN) 897 goto notification_done; 898 899 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 900 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 901 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true); 902 903 KNOTE_LOCKED(&job->klist, 1); 904 905 if (lj_done) { 906 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 907 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 908 KNOTE_LOCKED(&lj->klist, 1); 909 } 910 if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED)) 911 == LIOJ_SIGNAL && 912 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 913 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 914 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi, 915 true); 916 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 917 } 918 } 919 920 notification_done: 921 if (job->jobflags & KAIOCB_CHECKSYNC) { 922 schedule_fsync = false; 923 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) { 924 if (job->fd_file != sjob->fd_file || 925 job->seqno >= sjob->seqno) 926 continue; 927 if (--sjob->pending > 0) 928 continue; 929 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list); 930 if (!aio_clear_cancel_function_locked(sjob)) 931 continue; 932 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list); 933 schedule_fsync = true; 934 } 935 if (schedule_fsync) 936 taskqueue_enqueue(taskqueue_aiod_kick, 937 &ki->kaio_sync_task); 938 } 939 if (ki->kaio_flags & KAIO_WAKEUP) { 940 ki->kaio_flags &= ~KAIO_WAKEUP; 941 wakeup(&userp->p_aioinfo); 942 } 943 } 944 945 static void 946 aio_schedule_fsync(void *context, int pending) 947 { 948 struct kaioinfo *ki; 949 struct kaiocb *job; 950 951 ki = context; 952 AIO_LOCK(ki); 953 while (!TAILQ_EMPTY(&ki->kaio_syncready)) { 954 job = TAILQ_FIRST(&ki->kaio_syncready); 955 TAILQ_REMOVE(&ki->kaio_syncready, job, list); 956 AIO_UNLOCK(ki); 957 aio_schedule(job, aio_process_sync); 958 AIO_LOCK(ki); 959 } 960 AIO_UNLOCK(ki); 961 } 962 963 bool 964 aio_cancel_cleared(struct kaiocb *job) 965 { 966 967 /* 968 * The caller should hold the same queue lock held when 969 * aio_clear_cancel_function() was called and set this flag 970 * ensuring this check sees an up-to-date value. However, 971 * there is no way to assert that. 972 */ 973 return ((job->jobflags & KAIOCB_CLEARED) != 0); 974 } 975 976 static bool 977 aio_clear_cancel_function_locked(struct kaiocb *job) 978 { 979 980 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 981 MPASS(job->cancel_fn != NULL); 982 if (job->jobflags & KAIOCB_CANCELLING) { 983 job->jobflags |= KAIOCB_CLEARED; 984 return (false); 985 } 986 job->cancel_fn = NULL; 987 return (true); 988 } 989 990 bool 991 aio_clear_cancel_function(struct kaiocb *job) 992 { 993 struct kaioinfo *ki; 994 bool ret; 995 996 ki = job->userproc->p_aioinfo; 997 AIO_LOCK(ki); 998 ret = aio_clear_cancel_function_locked(job); 999 AIO_UNLOCK(ki); 1000 return (ret); 1001 } 1002 1003 static bool 1004 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func) 1005 { 1006 1007 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 1008 if (job->jobflags & KAIOCB_CANCELLED) 1009 return (false); 1010 job->cancel_fn = func; 1011 return (true); 1012 } 1013 1014 bool 1015 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func) 1016 { 1017 struct kaioinfo *ki; 1018 bool ret; 1019 1020 ki = job->userproc->p_aioinfo; 1021 AIO_LOCK(ki); 1022 ret = aio_set_cancel_function_locked(job, func); 1023 AIO_UNLOCK(ki); 1024 return (ret); 1025 } 1026 1027 void 1028 aio_complete(struct kaiocb *job, long status, int error) 1029 { 1030 struct kaioinfo *ki; 1031 struct proc *userp; 1032 1033 job->uaiocb._aiocb_private.error = error; 1034 job->uaiocb._aiocb_private.status = status; 1035 1036 userp = job->userproc; 1037 ki = userp->p_aioinfo; 1038 1039 AIO_LOCK(ki); 1040 KASSERT(!(job->jobflags & KAIOCB_FINISHED), 1041 ("duplicate aio_complete")); 1042 job->jobflags |= KAIOCB_FINISHED; 1043 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) { 1044 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 1045 aio_bio_done_notify(userp, job); 1046 } 1047 AIO_UNLOCK(ki); 1048 } 1049 1050 void 1051 aio_cancel(struct kaiocb *job) 1052 { 1053 1054 aio_complete(job, -1, ECANCELED); 1055 } 1056 1057 void 1058 aio_switch_vmspace(struct kaiocb *job) 1059 { 1060 1061 vmspace_switch_aio(job->userproc->p_vmspace); 1062 } 1063 1064 /* 1065 * The AIO daemon, most of the actual work is done in aio_process_*, 1066 * but the setup (and address space mgmt) is done in this routine. 1067 */ 1068 static void 1069 aio_daemon(void *_id) 1070 { 1071 struct kaiocb *job; 1072 struct aioproc *aiop; 1073 struct kaioinfo *ki; 1074 struct proc *p; 1075 struct vmspace *myvm; 1076 struct thread *td = curthread; 1077 int id = (intptr_t)_id; 1078 1079 /* 1080 * Grab an extra reference on the daemon's vmspace so that it 1081 * doesn't get freed by jobs that switch to a different 1082 * vmspace. 1083 */ 1084 p = td->td_proc; 1085 myvm = vmspace_acquire_ref(p); 1086 1087 KASSERT(p->p_textvp == NULL, ("kthread has a textvp")); 1088 1089 /* 1090 * Allocate and ready the aio control info. There is one aiop structure 1091 * per daemon. 1092 */ 1093 aiop = malloc(sizeof(*aiop), M_AIO, M_WAITOK); 1094 aiop->aioproc = p; 1095 aiop->aioprocflags = 0; 1096 1097 /* 1098 * Wakeup parent process. (Parent sleeps to keep from blasting away 1099 * and creating too many daemons.) 1100 */ 1101 sema_post(&aio_newproc_sem); 1102 1103 mtx_lock(&aio_job_mtx); 1104 for (;;) { 1105 /* 1106 * Take daemon off of free queue 1107 */ 1108 if (aiop->aioprocflags & AIOP_FREE) { 1109 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1110 aiop->aioprocflags &= ~AIOP_FREE; 1111 } 1112 1113 /* 1114 * Check for jobs. 1115 */ 1116 while ((job = aio_selectjob(aiop)) != NULL) { 1117 mtx_unlock(&aio_job_mtx); 1118 1119 ki = job->userproc->p_aioinfo; 1120 job->handle_fn(job); 1121 1122 mtx_lock(&aio_job_mtx); 1123 /* Decrement the active job count. */ 1124 ki->kaio_active_count--; 1125 } 1126 1127 /* 1128 * Disconnect from user address space. 1129 */ 1130 if (p->p_vmspace != myvm) { 1131 mtx_unlock(&aio_job_mtx); 1132 vmspace_switch_aio(myvm); 1133 mtx_lock(&aio_job_mtx); 1134 /* 1135 * We have to restart to avoid race, we only sleep if 1136 * no job can be selected. 1137 */ 1138 continue; 1139 } 1140 1141 mtx_assert(&aio_job_mtx, MA_OWNED); 1142 1143 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1144 aiop->aioprocflags |= AIOP_FREE; 1145 1146 /* 1147 * If daemon is inactive for a long time, allow it to exit, 1148 * thereby freeing resources. 1149 */ 1150 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy", 1151 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) && 1152 (aiop->aioprocflags & AIOP_FREE) && 1153 num_aio_procs > target_aio_procs) 1154 break; 1155 } 1156 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1157 num_aio_procs--; 1158 mtx_unlock(&aio_job_mtx); 1159 free(aiop, M_AIO); 1160 free_unr(aiod_unr, id); 1161 vmspace_free(myvm); 1162 1163 KASSERT(p->p_vmspace == myvm, 1164 ("AIOD: bad vmspace for exiting daemon")); 1165 KASSERT(refcount_load(&myvm->vm_refcnt) > 1, 1166 ("AIOD: bad vm refcnt for exiting daemon: %d", 1167 refcount_load(&myvm->vm_refcnt))); 1168 kproc_exit(0); 1169 } 1170 1171 /* 1172 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1173 * AIO daemon modifies its environment itself. 1174 */ 1175 static int 1176 aio_newproc(int *start) 1177 { 1178 int error; 1179 struct proc *p; 1180 int id; 1181 1182 id = alloc_unr(aiod_unr); 1183 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p, 1184 RFNOWAIT, 0, "aiod%d", id); 1185 if (error == 0) { 1186 /* 1187 * Wait until daemon is started. 1188 */ 1189 sema_wait(&aio_newproc_sem); 1190 mtx_lock(&aio_job_mtx); 1191 num_aio_procs++; 1192 if (start != NULL) 1193 (*start)--; 1194 mtx_unlock(&aio_job_mtx); 1195 } else { 1196 free_unr(aiod_unr, id); 1197 } 1198 return (error); 1199 } 1200 1201 /* 1202 * Try the high-performance, low-overhead bio method for eligible 1203 * VCHR devices. This method doesn't use an aio helper thread, and 1204 * thus has very low overhead. 1205 * 1206 * Assumes that the caller, aio_aqueue(), has incremented the file 1207 * structure's reference count, preventing its deallocation for the 1208 * duration of this call. 1209 */ 1210 static int 1211 aio_qbio(struct proc *p, struct kaiocb *job) 1212 { 1213 struct aiocb *cb; 1214 struct file *fp; 1215 struct buf *pbuf; 1216 struct vnode *vp; 1217 struct cdevsw *csw; 1218 struct cdev *dev; 1219 struct kaioinfo *ki; 1220 struct bio **bios = NULL; 1221 off_t offset; 1222 int bio_cmd, error, i, iovcnt, opcode, poff, ref; 1223 vm_prot_t prot; 1224 bool use_unmapped; 1225 1226 cb = &job->uaiocb; 1227 fp = job->fd_file; 1228 opcode = cb->aio_lio_opcode; 1229 1230 if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV || 1231 opcode == LIO_READ || opcode == LIO_READV)) 1232 return (-1); 1233 if (fp == NULL || fp->f_type != DTYPE_VNODE) 1234 return (-1); 1235 1236 vp = fp->f_vnode; 1237 if (vp->v_type != VCHR) 1238 return (-1); 1239 if (vp->v_bufobj.bo_bsize == 0) 1240 return (-1); 1241 1242 bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ; 1243 iovcnt = job->uiop->uio_iovcnt; 1244 if (iovcnt > max_buf_aio) 1245 return (-1); 1246 for (i = 0; i < iovcnt; i++) { 1247 if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0) 1248 return (-1); 1249 if (job->uiop->uio_iov[i].iov_len > maxphys) { 1250 error = -1; 1251 return (-1); 1252 } 1253 } 1254 offset = cb->aio_offset; 1255 1256 ref = 0; 1257 csw = devvn_refthread(vp, &dev, &ref); 1258 if (csw == NULL) 1259 return (ENXIO); 1260 1261 if ((csw->d_flags & D_DISK) == 0) { 1262 error = -1; 1263 goto unref; 1264 } 1265 if (job->uiop->uio_resid > dev->si_iosize_max) { 1266 error = -1; 1267 goto unref; 1268 } 1269 1270 ki = p->p_aioinfo; 1271 job->error = 0; 1272 1273 use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed; 1274 if (!use_unmapped) { 1275 AIO_LOCK(ki); 1276 if (ki->kaio_buffer_count + iovcnt > max_buf_aio) { 1277 AIO_UNLOCK(ki); 1278 error = EAGAIN; 1279 goto unref; 1280 } 1281 ki->kaio_buffer_count += iovcnt; 1282 AIO_UNLOCK(ki); 1283 } 1284 1285 bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK); 1286 atomic_store_int(&job->nbio, iovcnt); 1287 for (i = 0; i < iovcnt; i++) { 1288 struct vm_page** pages; 1289 struct bio *bp; 1290 void *buf; 1291 size_t nbytes; 1292 int npages; 1293 1294 buf = job->uiop->uio_iov[i].iov_base; 1295 nbytes = job->uiop->uio_iov[i].iov_len; 1296 1297 bios[i] = g_alloc_bio(); 1298 bp = bios[i]; 1299 1300 poff = (vm_offset_t)buf & PAGE_MASK; 1301 if (use_unmapped) { 1302 pbuf = NULL; 1303 pages = malloc(sizeof(vm_page_t) * (atop(round_page( 1304 nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO); 1305 } else { 1306 pbuf = uma_zalloc(pbuf_zone, M_WAITOK); 1307 BUF_KERNPROC(pbuf); 1308 pages = pbuf->b_pages; 1309 } 1310 1311 bp->bio_length = nbytes; 1312 bp->bio_bcount = nbytes; 1313 bp->bio_done = aio_biowakeup; 1314 bp->bio_offset = offset; 1315 bp->bio_cmd = bio_cmd; 1316 bp->bio_dev = dev; 1317 bp->bio_caller1 = job; 1318 bp->bio_caller2 = pbuf; 1319 1320 prot = VM_PROT_READ; 1321 if (opcode == LIO_READ || opcode == LIO_READV) 1322 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 1323 npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 1324 (vm_offset_t)buf, bp->bio_length, prot, pages, 1325 atop(maxphys) + 1); 1326 if (npages < 0) { 1327 if (pbuf != NULL) 1328 uma_zfree(pbuf_zone, pbuf); 1329 else 1330 free(pages, M_TEMP); 1331 error = EFAULT; 1332 g_destroy_bio(bp); 1333 i--; 1334 goto destroy_bios; 1335 } 1336 if (pbuf != NULL) { 1337 pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages); 1338 bp->bio_data = pbuf->b_data + poff; 1339 pbuf->b_npages = npages; 1340 atomic_add_int(&num_buf_aio, 1); 1341 } else { 1342 bp->bio_ma = pages; 1343 bp->bio_ma_n = npages; 1344 bp->bio_ma_offset = poff; 1345 bp->bio_data = unmapped_buf; 1346 bp->bio_flags |= BIO_UNMAPPED; 1347 atomic_add_int(&num_unmapped_aio, 1); 1348 } 1349 1350 offset += nbytes; 1351 } 1352 1353 /* Perform transfer. */ 1354 for (i = 0; i < iovcnt; i++) 1355 csw->d_strategy(bios[i]); 1356 free(bios, M_TEMP); 1357 1358 dev_relthread(dev, ref); 1359 return (0); 1360 1361 destroy_bios: 1362 for (; i >= 0; i--) 1363 aio_biocleanup(bios[i]); 1364 free(bios, M_TEMP); 1365 unref: 1366 dev_relthread(dev, ref); 1367 return (error); 1368 } 1369 1370 #ifdef COMPAT_FREEBSD6 1371 static int 1372 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig) 1373 { 1374 1375 /* 1376 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 1377 * supported by AIO with the old sigevent structure. 1378 */ 1379 nsig->sigev_notify = osig->sigev_notify; 1380 switch (nsig->sigev_notify) { 1381 case SIGEV_NONE: 1382 break; 1383 case SIGEV_SIGNAL: 1384 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 1385 break; 1386 case SIGEV_KEVENT: 1387 nsig->sigev_notify_kqueue = 1388 osig->__sigev_u.__sigev_notify_kqueue; 1389 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr; 1390 break; 1391 default: 1392 return (EINVAL); 1393 } 1394 return (0); 1395 } 1396 1397 static int 1398 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob, 1399 int type __unused) 1400 { 1401 struct oaiocb *ojob; 1402 struct aiocb *kcb = &kjob->uaiocb; 1403 int error; 1404 1405 bzero(kcb, sizeof(struct aiocb)); 1406 error = copyin(ujob, kcb, sizeof(struct oaiocb)); 1407 if (error) 1408 return (error); 1409 /* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */ 1410 ojob = (struct oaiocb *)kcb; 1411 return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent)); 1412 } 1413 #endif 1414 1415 static int 1416 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type) 1417 { 1418 struct aiocb *kcb = &kjob->uaiocb; 1419 int error; 1420 1421 error = copyin(ujob, kcb, sizeof(struct aiocb)); 1422 if (error) 1423 return (error); 1424 if (type & LIO_VECTORED) { 1425 /* malloc a uio and copy in the iovec */ 1426 error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov), 1427 kcb->aio_iovcnt, &kjob->uiop); 1428 } 1429 1430 return (error); 1431 } 1432 1433 static long 1434 aiocb_fetch_status(struct aiocb *ujob) 1435 { 1436 1437 return (fuword(&ujob->_aiocb_private.status)); 1438 } 1439 1440 static long 1441 aiocb_fetch_error(struct aiocb *ujob) 1442 { 1443 1444 return (fuword(&ujob->_aiocb_private.error)); 1445 } 1446 1447 static int 1448 aiocb_store_status(struct aiocb *ujob, long status) 1449 { 1450 1451 return (suword(&ujob->_aiocb_private.status, status)); 1452 } 1453 1454 static int 1455 aiocb_store_error(struct aiocb *ujob, long error) 1456 { 1457 1458 return (suword(&ujob->_aiocb_private.error, error)); 1459 } 1460 1461 static int 1462 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref) 1463 { 1464 1465 return (suword(&ujob->_aiocb_private.kernelinfo, jobref)); 1466 } 1467 1468 static int 1469 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 1470 { 1471 1472 return (suword(ujobp, (long)ujob)); 1473 } 1474 1475 static struct aiocb_ops aiocb_ops = { 1476 .aio_copyin = aiocb_copyin, 1477 .fetch_status = aiocb_fetch_status, 1478 .fetch_error = aiocb_fetch_error, 1479 .store_status = aiocb_store_status, 1480 .store_error = aiocb_store_error, 1481 .store_kernelinfo = aiocb_store_kernelinfo, 1482 .store_aiocb = aiocb_store_aiocb, 1483 }; 1484 1485 #ifdef COMPAT_FREEBSD6 1486 static struct aiocb_ops aiocb_ops_osigevent = { 1487 .aio_copyin = aiocb_copyin_old_sigevent, 1488 .fetch_status = aiocb_fetch_status, 1489 .fetch_error = aiocb_fetch_error, 1490 .store_status = aiocb_store_status, 1491 .store_error = aiocb_store_error, 1492 .store_kernelinfo = aiocb_store_kernelinfo, 1493 .store_aiocb = aiocb_store_aiocb, 1494 }; 1495 #endif 1496 1497 /* 1498 * Queue a new AIO request. Choosing either the threaded or direct bio VCHR 1499 * technique is done in this code. 1500 */ 1501 int 1502 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj, 1503 int type, struct aiocb_ops *ops) 1504 { 1505 struct proc *p = td->td_proc; 1506 struct file *fp = NULL; 1507 struct kaiocb *job; 1508 struct kaioinfo *ki; 1509 struct kevent kev; 1510 int opcode; 1511 int error; 1512 int fd, kqfd; 1513 int jid; 1514 u_short evflags; 1515 1516 if (p->p_aioinfo == NULL) 1517 aio_init_aioinfo(p); 1518 1519 ki = p->p_aioinfo; 1520 1521 ops->store_status(ujob, -1); 1522 ops->store_error(ujob, 0); 1523 ops->store_kernelinfo(ujob, -1); 1524 1525 if (num_queue_count >= max_queue_count || 1526 ki->kaio_count >= max_aio_queue_per_proc) { 1527 error = EAGAIN; 1528 goto err1; 1529 } 1530 1531 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1532 knlist_init_mtx(&job->klist, AIO_MTX(ki)); 1533 1534 error = ops->aio_copyin(ujob, job, type); 1535 if (error) 1536 goto err2; 1537 1538 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) { 1539 error = EINVAL; 1540 goto err2; 1541 } 1542 1543 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1544 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1545 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1546 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1547 error = EINVAL; 1548 goto err2; 1549 } 1550 1551 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1552 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1553 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) { 1554 error = EINVAL; 1555 goto err2; 1556 } 1557 1558 /* Get the opcode. */ 1559 if (type == LIO_NOP) { 1560 switch (job->uaiocb.aio_lio_opcode) { 1561 case LIO_WRITE: 1562 case LIO_NOP: 1563 case LIO_READ: 1564 opcode = job->uaiocb.aio_lio_opcode; 1565 break; 1566 default: 1567 error = EINVAL; 1568 goto err2; 1569 } 1570 } else 1571 opcode = job->uaiocb.aio_lio_opcode = type; 1572 1573 ksiginfo_init(&job->ksi); 1574 1575 /* Save userspace address of the job info. */ 1576 job->ujob = ujob; 1577 1578 /* 1579 * Validate the opcode and fetch the file object for the specified 1580 * file descriptor. 1581 * 1582 * XXXRW: Moved the opcode validation up here so that we don't 1583 * retrieve a file descriptor without knowing what the capabiltity 1584 * should be. 1585 */ 1586 fd = job->uaiocb.aio_fildes; 1587 switch (opcode) { 1588 case LIO_WRITE: 1589 case LIO_WRITEV: 1590 error = fget_write(td, fd, &cap_pwrite_rights, &fp); 1591 break; 1592 case LIO_READ: 1593 case LIO_READV: 1594 error = fget_read(td, fd, &cap_pread_rights, &fp); 1595 break; 1596 case LIO_SYNC: 1597 case LIO_DSYNC: 1598 error = fget(td, fd, &cap_fsync_rights, &fp); 1599 break; 1600 case LIO_MLOCK: 1601 break; 1602 case LIO_NOP: 1603 error = fget(td, fd, &cap_no_rights, &fp); 1604 break; 1605 default: 1606 error = EINVAL; 1607 } 1608 if (error) 1609 goto err3; 1610 1611 if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) { 1612 error = EINVAL; 1613 goto err3; 1614 } 1615 1616 if ((opcode == LIO_READ || opcode == LIO_READV || 1617 opcode == LIO_WRITE || opcode == LIO_WRITEV) && 1618 job->uaiocb.aio_offset < 0 && 1619 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) { 1620 error = EINVAL; 1621 goto err3; 1622 } 1623 1624 if (fp != NULL && fp->f_ops == &path_fileops) { 1625 error = EBADF; 1626 goto err3; 1627 } 1628 1629 job->fd_file = fp; 1630 1631 mtx_lock(&aio_job_mtx); 1632 jid = jobrefid++; 1633 job->seqno = jobseqno++; 1634 mtx_unlock(&aio_job_mtx); 1635 error = ops->store_kernelinfo(ujob, jid); 1636 if (error) { 1637 error = EINVAL; 1638 goto err3; 1639 } 1640 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid; 1641 1642 if (opcode == LIO_NOP) { 1643 fdrop(fp, td); 1644 MPASS(job->uiop == &job->uio || job->uiop == NULL); 1645 uma_zfree(aiocb_zone, job); 1646 return (0); 1647 } 1648 1649 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT) 1650 goto no_kqueue; 1651 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags; 1652 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) { 1653 error = EINVAL; 1654 goto err3; 1655 } 1656 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue; 1657 memset(&kev, 0, sizeof(kev)); 1658 kev.ident = (uintptr_t)job->ujob; 1659 kev.filter = EVFILT_AIO; 1660 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags; 1661 kev.data = (intptr_t)job; 1662 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1663 error = kqfd_register(kqfd, &kev, td, M_WAITOK); 1664 if (error) 1665 goto err3; 1666 1667 no_kqueue: 1668 1669 ops->store_error(ujob, EINPROGRESS); 1670 job->uaiocb._aiocb_private.error = EINPROGRESS; 1671 job->userproc = p; 1672 job->cred = crhold(td->td_ucred); 1673 job->jobflags = KAIOCB_QUEUEING; 1674 job->lio = lj; 1675 1676 if (opcode & LIO_VECTORED) { 1677 /* Use the uio copied in by aio_copyin */ 1678 MPASS(job->uiop != &job->uio && job->uiop != NULL); 1679 } else { 1680 /* Setup the inline uio */ 1681 job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf; 1682 job->iov[0].iov_len = job->uaiocb.aio_nbytes; 1683 job->uio.uio_iov = job->iov; 1684 job->uio.uio_iovcnt = 1; 1685 job->uio.uio_resid = job->uaiocb.aio_nbytes; 1686 job->uio.uio_segflg = UIO_USERSPACE; 1687 job->uiop = &job->uio; 1688 } 1689 switch (opcode & (LIO_READ | LIO_WRITE)) { 1690 case LIO_READ: 1691 job->uiop->uio_rw = UIO_READ; 1692 break; 1693 case LIO_WRITE: 1694 job->uiop->uio_rw = UIO_WRITE; 1695 break; 1696 } 1697 job->uiop->uio_offset = job->uaiocb.aio_offset; 1698 job->uiop->uio_td = td; 1699 1700 if (opcode == LIO_MLOCK) { 1701 aio_schedule(job, aio_process_mlock); 1702 error = 0; 1703 } else if (fp->f_ops->fo_aio_queue == NULL) 1704 error = aio_queue_file(fp, job); 1705 else 1706 error = fo_aio_queue(fp, job); 1707 if (error) 1708 goto err4; 1709 1710 AIO_LOCK(ki); 1711 job->jobflags &= ~KAIOCB_QUEUEING; 1712 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist); 1713 ki->kaio_count++; 1714 if (lj) 1715 lj->lioj_count++; 1716 atomic_add_int(&num_queue_count, 1); 1717 if (job->jobflags & KAIOCB_FINISHED) { 1718 /* 1719 * The queue callback completed the request synchronously. 1720 * The bulk of the completion is deferred in that case 1721 * until this point. 1722 */ 1723 aio_bio_done_notify(p, job); 1724 } else 1725 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist); 1726 AIO_UNLOCK(ki); 1727 return (0); 1728 1729 err4: 1730 crfree(job->cred); 1731 err3: 1732 if (fp) 1733 fdrop(fp, td); 1734 knlist_delete(&job->klist, curthread, 0); 1735 err2: 1736 if (job->uiop != &job->uio) 1737 free(job->uiop, M_IOV); 1738 uma_zfree(aiocb_zone, job); 1739 err1: 1740 ops->store_error(ujob, error); 1741 return (error); 1742 } 1743 1744 static void 1745 aio_cancel_daemon_job(struct kaiocb *job) 1746 { 1747 1748 mtx_lock(&aio_job_mtx); 1749 if (!aio_cancel_cleared(job)) 1750 TAILQ_REMOVE(&aio_jobs, job, list); 1751 mtx_unlock(&aio_job_mtx); 1752 aio_cancel(job); 1753 } 1754 1755 void 1756 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func) 1757 { 1758 1759 mtx_lock(&aio_job_mtx); 1760 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) { 1761 mtx_unlock(&aio_job_mtx); 1762 aio_cancel(job); 1763 return; 1764 } 1765 job->handle_fn = func; 1766 TAILQ_INSERT_TAIL(&aio_jobs, job, list); 1767 aio_kick_nowait(job->userproc); 1768 mtx_unlock(&aio_job_mtx); 1769 } 1770 1771 static void 1772 aio_cancel_sync(struct kaiocb *job) 1773 { 1774 struct kaioinfo *ki; 1775 1776 ki = job->userproc->p_aioinfo; 1777 AIO_LOCK(ki); 1778 if (!aio_cancel_cleared(job)) 1779 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list); 1780 AIO_UNLOCK(ki); 1781 aio_cancel(job); 1782 } 1783 1784 int 1785 aio_queue_file(struct file *fp, struct kaiocb *job) 1786 { 1787 struct kaioinfo *ki; 1788 struct kaiocb *job2; 1789 struct vnode *vp; 1790 struct mount *mp; 1791 int error; 1792 bool safe; 1793 1794 ki = job->userproc->p_aioinfo; 1795 error = aio_qbio(job->userproc, job); 1796 if (error >= 0) 1797 return (error); 1798 safe = false; 1799 if (fp->f_type == DTYPE_VNODE) { 1800 vp = fp->f_vnode; 1801 if (vp->v_type == VREG || vp->v_type == VDIR) { 1802 mp = fp->f_vnode->v_mount; 1803 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0) 1804 safe = true; 1805 } 1806 } 1807 if (!(safe || enable_aio_unsafe)) { 1808 counted_warning(&unsafe_warningcnt, 1809 "is attempting to use unsafe AIO requests"); 1810 return (EOPNOTSUPP); 1811 } 1812 1813 if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) { 1814 aio_schedule(job, aio_process_rw); 1815 error = 0; 1816 } else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) { 1817 AIO_LOCK(ki); 1818 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) { 1819 if (job2->fd_file == job->fd_file && 1820 ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) && 1821 job2->seqno < job->seqno) { 1822 job2->jobflags |= KAIOCB_CHECKSYNC; 1823 job->pending++; 1824 } 1825 } 1826 if (job->pending != 0) { 1827 if (!aio_set_cancel_function_locked(job, 1828 aio_cancel_sync)) { 1829 AIO_UNLOCK(ki); 1830 aio_cancel(job); 1831 return (0); 1832 } 1833 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list); 1834 AIO_UNLOCK(ki); 1835 return (0); 1836 } 1837 AIO_UNLOCK(ki); 1838 aio_schedule(job, aio_process_sync); 1839 error = 0; 1840 } else { 1841 error = EINVAL; 1842 } 1843 return (error); 1844 } 1845 1846 static void 1847 aio_kick_nowait(struct proc *userp) 1848 { 1849 struct kaioinfo *ki = userp->p_aioinfo; 1850 struct aioproc *aiop; 1851 1852 mtx_assert(&aio_job_mtx, MA_OWNED); 1853 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1854 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1855 aiop->aioprocflags &= ~AIOP_FREE; 1856 wakeup(aiop->aioproc); 1857 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1858 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) { 1859 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task); 1860 } 1861 } 1862 1863 static int 1864 aio_kick(struct proc *userp) 1865 { 1866 struct kaioinfo *ki = userp->p_aioinfo; 1867 struct aioproc *aiop; 1868 int error, ret = 0; 1869 1870 mtx_assert(&aio_job_mtx, MA_OWNED); 1871 retryproc: 1872 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1873 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1874 aiop->aioprocflags &= ~AIOP_FREE; 1875 wakeup(aiop->aioproc); 1876 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1877 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) { 1878 num_aio_resv_start++; 1879 mtx_unlock(&aio_job_mtx); 1880 error = aio_newproc(&num_aio_resv_start); 1881 mtx_lock(&aio_job_mtx); 1882 if (error) { 1883 num_aio_resv_start--; 1884 goto retryproc; 1885 } 1886 } else { 1887 ret = -1; 1888 } 1889 return (ret); 1890 } 1891 1892 static void 1893 aio_kick_helper(void *context, int pending) 1894 { 1895 struct proc *userp = context; 1896 1897 mtx_lock(&aio_job_mtx); 1898 while (--pending >= 0) { 1899 if (aio_kick(userp)) 1900 break; 1901 } 1902 mtx_unlock(&aio_job_mtx); 1903 } 1904 1905 /* 1906 * Support the aio_return system call, as a side-effect, kernel resources are 1907 * released. 1908 */ 1909 static int 1910 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 1911 { 1912 struct proc *p = td->td_proc; 1913 struct kaiocb *job; 1914 struct kaioinfo *ki; 1915 long status, error; 1916 1917 ki = p->p_aioinfo; 1918 if (ki == NULL) 1919 return (EINVAL); 1920 AIO_LOCK(ki); 1921 TAILQ_FOREACH(job, &ki->kaio_done, plist) { 1922 if (job->ujob == ujob) 1923 break; 1924 } 1925 if (job != NULL) { 1926 MPASS(job->jobflags & KAIOCB_FINISHED); 1927 status = job->uaiocb._aiocb_private.status; 1928 error = job->uaiocb._aiocb_private.error; 1929 td->td_retval[0] = status; 1930 td->td_ru.ru_oublock += job->outblock; 1931 td->td_ru.ru_inblock += job->inblock; 1932 td->td_ru.ru_msgsnd += job->msgsnd; 1933 td->td_ru.ru_msgrcv += job->msgrcv; 1934 aio_free_entry(job); 1935 AIO_UNLOCK(ki); 1936 ops->store_error(ujob, error); 1937 ops->store_status(ujob, status); 1938 } else { 1939 error = EINVAL; 1940 AIO_UNLOCK(ki); 1941 } 1942 return (error); 1943 } 1944 1945 int 1946 sys_aio_return(struct thread *td, struct aio_return_args *uap) 1947 { 1948 1949 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops)); 1950 } 1951 1952 /* 1953 * Allow a process to wakeup when any of the I/O requests are completed. 1954 */ 1955 static int 1956 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist, 1957 struct timespec *ts) 1958 { 1959 struct proc *p = td->td_proc; 1960 struct timeval atv; 1961 struct kaioinfo *ki; 1962 struct kaiocb *firstjob, *job; 1963 int error, i, timo; 1964 1965 timo = 0; 1966 if (ts) { 1967 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000) 1968 return (EINVAL); 1969 1970 TIMESPEC_TO_TIMEVAL(&atv, ts); 1971 if (itimerfix(&atv)) 1972 return (EINVAL); 1973 timo = tvtohz(&atv); 1974 } 1975 1976 ki = p->p_aioinfo; 1977 if (ki == NULL) 1978 return (EAGAIN); 1979 1980 if (njoblist == 0) 1981 return (0); 1982 1983 AIO_LOCK(ki); 1984 for (;;) { 1985 firstjob = NULL; 1986 error = 0; 1987 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 1988 for (i = 0; i < njoblist; i++) { 1989 if (job->ujob == ujoblist[i]) { 1990 if (firstjob == NULL) 1991 firstjob = job; 1992 if (job->jobflags & KAIOCB_FINISHED) 1993 goto RETURN; 1994 } 1995 } 1996 } 1997 /* All tasks were finished. */ 1998 if (firstjob == NULL) 1999 break; 2000 2001 ki->kaio_flags |= KAIO_WAKEUP; 2002 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2003 "aiospn", timo); 2004 if (error == ERESTART) 2005 error = EINTR; 2006 if (error) 2007 break; 2008 } 2009 RETURN: 2010 AIO_UNLOCK(ki); 2011 return (error); 2012 } 2013 2014 int 2015 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap) 2016 { 2017 struct timespec ts, *tsp; 2018 struct aiocb **ujoblist; 2019 int error; 2020 2021 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 2022 return (EINVAL); 2023 2024 if (uap->timeout) { 2025 /* Get timespec struct. */ 2026 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 2027 return (error); 2028 tsp = &ts; 2029 } else 2030 tsp = NULL; 2031 2032 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK); 2033 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0])); 2034 if (error == 0) 2035 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2036 free(ujoblist, M_AIO); 2037 return (error); 2038 } 2039 2040 /* 2041 * aio_cancel cancels any non-bio aio operations not currently in progress. 2042 */ 2043 int 2044 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap) 2045 { 2046 struct proc *p = td->td_proc; 2047 struct kaioinfo *ki; 2048 struct kaiocb *job, *jobn; 2049 struct file *fp; 2050 int error; 2051 int cancelled = 0; 2052 int notcancelled = 0; 2053 struct vnode *vp; 2054 2055 /* Lookup file object. */ 2056 error = fget(td, uap->fd, &cap_no_rights, &fp); 2057 if (error) 2058 return (error); 2059 2060 ki = p->p_aioinfo; 2061 if (ki == NULL) 2062 goto done; 2063 2064 if (fp->f_type == DTYPE_VNODE) { 2065 vp = fp->f_vnode; 2066 if (vn_isdisk(vp)) { 2067 fdrop(fp, td); 2068 td->td_retval[0] = AIO_NOTCANCELED; 2069 return (0); 2070 } 2071 } 2072 2073 AIO_LOCK(ki); 2074 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 2075 if ((uap->fd == job->uaiocb.aio_fildes) && 2076 ((uap->aiocbp == NULL) || 2077 (uap->aiocbp == job->ujob))) { 2078 if (aio_cancel_job(p, ki, job)) { 2079 cancelled++; 2080 } else { 2081 notcancelled++; 2082 } 2083 if (uap->aiocbp != NULL) 2084 break; 2085 } 2086 } 2087 AIO_UNLOCK(ki); 2088 2089 done: 2090 fdrop(fp, td); 2091 2092 if (uap->aiocbp != NULL) { 2093 if (cancelled) { 2094 td->td_retval[0] = AIO_CANCELED; 2095 return (0); 2096 } 2097 } 2098 2099 if (notcancelled) { 2100 td->td_retval[0] = AIO_NOTCANCELED; 2101 return (0); 2102 } 2103 2104 if (cancelled) { 2105 td->td_retval[0] = AIO_CANCELED; 2106 return (0); 2107 } 2108 2109 td->td_retval[0] = AIO_ALLDONE; 2110 2111 return (0); 2112 } 2113 2114 /* 2115 * aio_error is implemented in the kernel level for compatibility purposes 2116 * only. For a user mode async implementation, it would be best to do it in 2117 * a userland subroutine. 2118 */ 2119 static int 2120 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 2121 { 2122 struct proc *p = td->td_proc; 2123 struct kaiocb *job; 2124 struct kaioinfo *ki; 2125 int status; 2126 2127 ki = p->p_aioinfo; 2128 if (ki == NULL) { 2129 td->td_retval[0] = EINVAL; 2130 return (0); 2131 } 2132 2133 AIO_LOCK(ki); 2134 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 2135 if (job->ujob == ujob) { 2136 if (job->jobflags & KAIOCB_FINISHED) 2137 td->td_retval[0] = 2138 job->uaiocb._aiocb_private.error; 2139 else 2140 td->td_retval[0] = EINPROGRESS; 2141 AIO_UNLOCK(ki); 2142 return (0); 2143 } 2144 } 2145 AIO_UNLOCK(ki); 2146 2147 /* 2148 * Hack for failure of aio_aqueue. 2149 */ 2150 status = ops->fetch_status(ujob); 2151 if (status == -1) { 2152 td->td_retval[0] = ops->fetch_error(ujob); 2153 return (0); 2154 } 2155 2156 td->td_retval[0] = EINVAL; 2157 return (0); 2158 } 2159 2160 int 2161 sys_aio_error(struct thread *td, struct aio_error_args *uap) 2162 { 2163 2164 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops)); 2165 } 2166 2167 /* syscall - asynchronous read from a file (REALTIME) */ 2168 #ifdef COMPAT_FREEBSD6 2169 int 2170 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap) 2171 { 2172 2173 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2174 &aiocb_ops_osigevent)); 2175 } 2176 #endif 2177 2178 int 2179 sys_aio_read(struct thread *td, struct aio_read_args *uap) 2180 { 2181 2182 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops)); 2183 } 2184 2185 int 2186 sys_aio_readv(struct thread *td, struct aio_readv_args *uap) 2187 { 2188 2189 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops)); 2190 } 2191 2192 /* syscall - asynchronous write to a file (REALTIME) */ 2193 #ifdef COMPAT_FREEBSD6 2194 int 2195 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap) 2196 { 2197 2198 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2199 &aiocb_ops_osigevent)); 2200 } 2201 #endif 2202 2203 int 2204 sys_aio_write(struct thread *td, struct aio_write_args *uap) 2205 { 2206 2207 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops)); 2208 } 2209 2210 int 2211 sys_aio_writev(struct thread *td, struct aio_writev_args *uap) 2212 { 2213 2214 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops)); 2215 } 2216 2217 int 2218 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap) 2219 { 2220 2221 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops)); 2222 } 2223 2224 static int 2225 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list, 2226 struct aiocb **acb_list, int nent, struct sigevent *sig, 2227 struct aiocb_ops *ops) 2228 { 2229 struct proc *p = td->td_proc; 2230 struct aiocb *job; 2231 struct kaioinfo *ki; 2232 struct aioliojob *lj; 2233 struct kevent kev; 2234 int error; 2235 int nagain, nerror; 2236 int i; 2237 2238 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT)) 2239 return (EINVAL); 2240 2241 if (nent < 0 || nent > max_aio_queue_per_proc) 2242 return (EINVAL); 2243 2244 if (p->p_aioinfo == NULL) 2245 aio_init_aioinfo(p); 2246 2247 ki = p->p_aioinfo; 2248 2249 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2250 lj->lioj_flags = 0; 2251 lj->lioj_count = 0; 2252 lj->lioj_finished_count = 0; 2253 lj->lioj_signal.sigev_notify = SIGEV_NONE; 2254 knlist_init_mtx(&lj->klist, AIO_MTX(ki)); 2255 ksiginfo_init(&lj->lioj_ksi); 2256 2257 /* 2258 * Setup signal. 2259 */ 2260 if (sig && (mode == LIO_NOWAIT)) { 2261 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal)); 2262 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2263 /* Assume only new style KEVENT */ 2264 memset(&kev, 0, sizeof(kev)); 2265 kev.filter = EVFILT_LIO; 2266 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2267 kev.ident = (uintptr_t)uacb_list; /* something unique */ 2268 kev.data = (intptr_t)lj; 2269 /* pass user defined sigval data */ 2270 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2271 error = kqfd_register( 2272 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 2273 M_WAITOK); 2274 if (error) { 2275 uma_zfree(aiolio_zone, lj); 2276 return (error); 2277 } 2278 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2279 ; 2280 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2281 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2282 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2283 uma_zfree(aiolio_zone, lj); 2284 return EINVAL; 2285 } 2286 lj->lioj_flags |= LIOJ_SIGNAL; 2287 } else { 2288 uma_zfree(aiolio_zone, lj); 2289 return EINVAL; 2290 } 2291 } 2292 2293 AIO_LOCK(ki); 2294 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2295 /* 2296 * Add extra aiocb count to avoid the lio to be freed 2297 * by other threads doing aio_waitcomplete or aio_return, 2298 * and prevent event from being sent until we have queued 2299 * all tasks. 2300 */ 2301 lj->lioj_count = 1; 2302 AIO_UNLOCK(ki); 2303 2304 /* 2305 * Get pointers to the list of I/O requests. 2306 */ 2307 nagain = 0; 2308 nerror = 0; 2309 for (i = 0; i < nent; i++) { 2310 job = acb_list[i]; 2311 if (job != NULL) { 2312 error = aio_aqueue(td, job, lj, LIO_NOP, ops); 2313 if (error == EAGAIN) 2314 nagain++; 2315 else if (error != 0) 2316 nerror++; 2317 } 2318 } 2319 2320 error = 0; 2321 AIO_LOCK(ki); 2322 if (mode == LIO_WAIT) { 2323 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2324 ki->kaio_flags |= KAIO_WAKEUP; 2325 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2326 PRIBIO | PCATCH, "aiospn", 0); 2327 if (error == ERESTART) 2328 error = EINTR; 2329 if (error) 2330 break; 2331 } 2332 } else { 2333 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2334 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2335 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2336 KNOTE_LOCKED(&lj->klist, 1); 2337 } 2338 if ((lj->lioj_flags & (LIOJ_SIGNAL | 2339 LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL && 2340 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2341 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2342 aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi, 2343 lj->lioj_count != 1); 2344 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2345 } 2346 } 2347 } 2348 lj->lioj_count--; 2349 if (lj->lioj_count == 0) { 2350 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2351 knlist_delete(&lj->klist, curthread, 1); 2352 PROC_LOCK(p); 2353 sigqueue_take(&lj->lioj_ksi); 2354 PROC_UNLOCK(p); 2355 AIO_UNLOCK(ki); 2356 uma_zfree(aiolio_zone, lj); 2357 } else 2358 AIO_UNLOCK(ki); 2359 2360 if (nerror) 2361 return (EIO); 2362 else if (nagain) 2363 return (EAGAIN); 2364 else 2365 return (error); 2366 } 2367 2368 /* syscall - list directed I/O (REALTIME) */ 2369 #ifdef COMPAT_FREEBSD6 2370 int 2371 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap) 2372 { 2373 struct aiocb **acb_list; 2374 struct sigevent *sigp, sig; 2375 struct osigevent osig; 2376 int error, nent; 2377 2378 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2379 return (EINVAL); 2380 2381 nent = uap->nent; 2382 if (nent < 0 || nent > max_aio_queue_per_proc) 2383 return (EINVAL); 2384 2385 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2386 error = copyin(uap->sig, &osig, sizeof(osig)); 2387 if (error) 2388 return (error); 2389 error = convert_old_sigevent(&osig, &sig); 2390 if (error) 2391 return (error); 2392 sigp = &sig; 2393 } else 2394 sigp = NULL; 2395 2396 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2397 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2398 if (error == 0) 2399 error = kern_lio_listio(td, uap->mode, 2400 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2401 &aiocb_ops_osigevent); 2402 free(acb_list, M_LIO); 2403 return (error); 2404 } 2405 #endif 2406 2407 /* syscall - list directed I/O (REALTIME) */ 2408 int 2409 sys_lio_listio(struct thread *td, struct lio_listio_args *uap) 2410 { 2411 struct aiocb **acb_list; 2412 struct sigevent *sigp, sig; 2413 int error, nent; 2414 2415 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2416 return (EINVAL); 2417 2418 nent = uap->nent; 2419 if (nent < 0 || nent > max_aio_queue_per_proc) 2420 return (EINVAL); 2421 2422 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2423 error = copyin(uap->sig, &sig, sizeof(sig)); 2424 if (error) 2425 return (error); 2426 sigp = &sig; 2427 } else 2428 sigp = NULL; 2429 2430 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2431 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2432 if (error == 0) 2433 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list, 2434 nent, sigp, &aiocb_ops); 2435 free(acb_list, M_LIO); 2436 return (error); 2437 } 2438 2439 static void 2440 aio_biocleanup(struct bio *bp) 2441 { 2442 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2443 struct kaioinfo *ki; 2444 struct buf *pbuf = (struct buf *)bp->bio_caller2; 2445 2446 /* Release mapping into kernel space. */ 2447 if (pbuf != NULL) { 2448 MPASS(pbuf->b_npages <= atop(maxphys) + 1); 2449 pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages); 2450 vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages); 2451 uma_zfree(pbuf_zone, pbuf); 2452 atomic_subtract_int(&num_buf_aio, 1); 2453 ki = job->userproc->p_aioinfo; 2454 AIO_LOCK(ki); 2455 ki->kaio_buffer_count--; 2456 AIO_UNLOCK(ki); 2457 } else { 2458 MPASS(bp->bio_ma_n <= atop(maxphys) + 1); 2459 vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n); 2460 free(bp->bio_ma, M_TEMP); 2461 atomic_subtract_int(&num_unmapped_aio, 1); 2462 } 2463 g_destroy_bio(bp); 2464 } 2465 2466 static void 2467 aio_biowakeup(struct bio *bp) 2468 { 2469 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2470 size_t nbytes; 2471 long bcount = bp->bio_bcount; 2472 long resid = bp->bio_resid; 2473 int error, opcode, nblks; 2474 int bio_error = bp->bio_error; 2475 uint16_t flags = bp->bio_flags; 2476 2477 opcode = job->uaiocb.aio_lio_opcode; 2478 2479 aio_biocleanup(bp); 2480 2481 nbytes =bcount - resid; 2482 atomic_add_acq_long(&job->nbytes, nbytes); 2483 nblks = btodb(nbytes); 2484 error = 0; 2485 /* 2486 * If multiple bios experienced an error, the job will reflect the 2487 * error of whichever failed bio completed last. 2488 */ 2489 if (flags & BIO_ERROR) 2490 atomic_set_int(&job->error, bio_error); 2491 if (opcode & LIO_WRITE) 2492 atomic_add_int(&job->outblock, nblks); 2493 else 2494 atomic_add_int(&job->inblock, nblks); 2495 atomic_subtract_int(&job->nbio, 1); 2496 2497 2498 if (atomic_load_int(&job->nbio) == 0) { 2499 if (atomic_load_int(&job->error)) 2500 aio_complete(job, -1, job->error); 2501 else 2502 aio_complete(job, atomic_load_long(&job->nbytes), 0); 2503 } 2504 } 2505 2506 /* syscall - wait for the next completion of an aio request */ 2507 static int 2508 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp, 2509 struct timespec *ts, struct aiocb_ops *ops) 2510 { 2511 struct proc *p = td->td_proc; 2512 struct timeval atv; 2513 struct kaioinfo *ki; 2514 struct kaiocb *job; 2515 struct aiocb *ujob; 2516 long error, status; 2517 int timo; 2518 2519 ops->store_aiocb(ujobp, NULL); 2520 2521 if (ts == NULL) { 2522 timo = 0; 2523 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) { 2524 timo = -1; 2525 } else { 2526 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000)) 2527 return (EINVAL); 2528 2529 TIMESPEC_TO_TIMEVAL(&atv, ts); 2530 if (itimerfix(&atv)) 2531 return (EINVAL); 2532 timo = tvtohz(&atv); 2533 } 2534 2535 if (p->p_aioinfo == NULL) 2536 aio_init_aioinfo(p); 2537 ki = p->p_aioinfo; 2538 2539 error = 0; 2540 job = NULL; 2541 AIO_LOCK(ki); 2542 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2543 if (timo == -1) { 2544 error = EWOULDBLOCK; 2545 break; 2546 } 2547 ki->kaio_flags |= KAIO_WAKEUP; 2548 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2549 "aiowc", timo); 2550 if (timo && error == ERESTART) 2551 error = EINTR; 2552 if (error) 2553 break; 2554 } 2555 2556 if (job != NULL) { 2557 MPASS(job->jobflags & KAIOCB_FINISHED); 2558 ujob = job->ujob; 2559 status = job->uaiocb._aiocb_private.status; 2560 error = job->uaiocb._aiocb_private.error; 2561 td->td_retval[0] = status; 2562 td->td_ru.ru_oublock += job->outblock; 2563 td->td_ru.ru_inblock += job->inblock; 2564 td->td_ru.ru_msgsnd += job->msgsnd; 2565 td->td_ru.ru_msgrcv += job->msgrcv; 2566 aio_free_entry(job); 2567 AIO_UNLOCK(ki); 2568 ops->store_aiocb(ujobp, ujob); 2569 ops->store_error(ujob, error); 2570 ops->store_status(ujob, status); 2571 } else 2572 AIO_UNLOCK(ki); 2573 2574 return (error); 2575 } 2576 2577 int 2578 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2579 { 2580 struct timespec ts, *tsp; 2581 int error; 2582 2583 if (uap->timeout) { 2584 /* Get timespec struct. */ 2585 error = copyin(uap->timeout, &ts, sizeof(ts)); 2586 if (error) 2587 return (error); 2588 tsp = &ts; 2589 } else 2590 tsp = NULL; 2591 2592 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops)); 2593 } 2594 2595 static int 2596 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob, 2597 struct aiocb_ops *ops) 2598 { 2599 int listop; 2600 2601 switch (op) { 2602 case O_SYNC: 2603 listop = LIO_SYNC; 2604 break; 2605 case O_DSYNC: 2606 listop = LIO_DSYNC; 2607 break; 2608 default: 2609 return (EINVAL); 2610 } 2611 2612 return (aio_aqueue(td, ujob, NULL, listop, ops)); 2613 } 2614 2615 int 2616 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2617 { 2618 2619 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops)); 2620 } 2621 2622 /* kqueue attach function */ 2623 static int 2624 filt_aioattach(struct knote *kn) 2625 { 2626 struct kaiocb *job; 2627 2628 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata; 2629 2630 /* 2631 * The job pointer must be validated before using it, so 2632 * registration is restricted to the kernel; the user cannot 2633 * set EV_FLAG1. 2634 */ 2635 if ((kn->kn_flags & EV_FLAG1) == 0) 2636 return (EPERM); 2637 kn->kn_ptr.p_aio = job; 2638 kn->kn_flags &= ~EV_FLAG1; 2639 2640 knlist_add(&job->klist, kn, 0); 2641 2642 return (0); 2643 } 2644 2645 /* kqueue detach function */ 2646 static void 2647 filt_aiodetach(struct knote *kn) 2648 { 2649 struct knlist *knl; 2650 2651 knl = &kn->kn_ptr.p_aio->klist; 2652 knl->kl_lock(knl->kl_lockarg); 2653 if (!knlist_empty(knl)) 2654 knlist_remove(knl, kn, 1); 2655 knl->kl_unlock(knl->kl_lockarg); 2656 } 2657 2658 /* kqueue filter function */ 2659 /*ARGSUSED*/ 2660 static int 2661 filt_aio(struct knote *kn, long hint) 2662 { 2663 struct kaiocb *job = kn->kn_ptr.p_aio; 2664 2665 kn->kn_data = job->uaiocb._aiocb_private.error; 2666 if (!(job->jobflags & KAIOCB_FINISHED)) 2667 return (0); 2668 kn->kn_flags |= EV_EOF; 2669 return (1); 2670 } 2671 2672 /* kqueue attach function */ 2673 static int 2674 filt_lioattach(struct knote *kn) 2675 { 2676 struct aioliojob *lj; 2677 2678 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata; 2679 2680 /* 2681 * The aioliojob pointer must be validated before using it, so 2682 * registration is restricted to the kernel; the user cannot 2683 * set EV_FLAG1. 2684 */ 2685 if ((kn->kn_flags & EV_FLAG1) == 0) 2686 return (EPERM); 2687 kn->kn_ptr.p_lio = lj; 2688 kn->kn_flags &= ~EV_FLAG1; 2689 2690 knlist_add(&lj->klist, kn, 0); 2691 2692 return (0); 2693 } 2694 2695 /* kqueue detach function */ 2696 static void 2697 filt_liodetach(struct knote *kn) 2698 { 2699 struct knlist *knl; 2700 2701 knl = &kn->kn_ptr.p_lio->klist; 2702 knl->kl_lock(knl->kl_lockarg); 2703 if (!knlist_empty(knl)) 2704 knlist_remove(knl, kn, 1); 2705 knl->kl_unlock(knl->kl_lockarg); 2706 } 2707 2708 /* kqueue filter function */ 2709 /*ARGSUSED*/ 2710 static int 2711 filt_lio(struct knote *kn, long hint) 2712 { 2713 struct aioliojob * lj = kn->kn_ptr.p_lio; 2714 2715 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2716 } 2717 2718 #ifdef COMPAT_FREEBSD32 2719 #include <sys/mount.h> 2720 #include <sys/socket.h> 2721 #include <sys/sysent.h> 2722 #include <compat/freebsd32/freebsd32.h> 2723 #include <compat/freebsd32/freebsd32_proto.h> 2724 #include <compat/freebsd32/freebsd32_signal.h> 2725 #include <compat/freebsd32/freebsd32_syscall.h> 2726 #include <compat/freebsd32/freebsd32_util.h> 2727 2728 struct __aiocb_private32 { 2729 int32_t status; 2730 int32_t error; 2731 uint32_t kernelinfo; 2732 }; 2733 2734 #ifdef COMPAT_FREEBSD6 2735 typedef struct oaiocb32 { 2736 int aio_fildes; /* File descriptor */ 2737 uint64_t aio_offset __packed; /* File offset for I/O */ 2738 uint32_t aio_buf; /* I/O buffer in process space */ 2739 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2740 struct osigevent32 aio_sigevent; /* Signal to deliver */ 2741 int aio_lio_opcode; /* LIO opcode */ 2742 int aio_reqprio; /* Request priority -- ignored */ 2743 struct __aiocb_private32 _aiocb_private; 2744 } oaiocb32_t; 2745 #endif 2746 2747 typedef struct aiocb32 { 2748 int32_t aio_fildes; /* File descriptor */ 2749 uint64_t aio_offset __packed; /* File offset for I/O */ 2750 uint32_t aio_buf; /* I/O buffer in process space */ 2751 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2752 int __spare__[2]; 2753 uint32_t __spare2__; 2754 int aio_lio_opcode; /* LIO opcode */ 2755 int aio_reqprio; /* Request priority -- ignored */ 2756 struct __aiocb_private32 _aiocb_private; 2757 struct sigevent32 aio_sigevent; /* Signal to deliver */ 2758 } aiocb32_t; 2759 2760 #ifdef COMPAT_FREEBSD6 2761 static int 2762 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig) 2763 { 2764 2765 /* 2766 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 2767 * supported by AIO with the old sigevent structure. 2768 */ 2769 CP(*osig, *nsig, sigev_notify); 2770 switch (nsig->sigev_notify) { 2771 case SIGEV_NONE: 2772 break; 2773 case SIGEV_SIGNAL: 2774 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 2775 break; 2776 case SIGEV_KEVENT: 2777 nsig->sigev_notify_kqueue = 2778 osig->__sigev_u.__sigev_notify_kqueue; 2779 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr); 2780 break; 2781 default: 2782 return (EINVAL); 2783 } 2784 return (0); 2785 } 2786 2787 static int 2788 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob, 2789 int type __unused) 2790 { 2791 struct oaiocb32 job32; 2792 struct aiocb *kcb = &kjob->uaiocb; 2793 int error; 2794 2795 bzero(kcb, sizeof(struct aiocb)); 2796 error = copyin(ujob, &job32, sizeof(job32)); 2797 if (error) 2798 return (error); 2799 2800 /* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */ 2801 2802 CP(job32, *kcb, aio_fildes); 2803 CP(job32, *kcb, aio_offset); 2804 PTRIN_CP(job32, *kcb, aio_buf); 2805 CP(job32, *kcb, aio_nbytes); 2806 CP(job32, *kcb, aio_lio_opcode); 2807 CP(job32, *kcb, aio_reqprio); 2808 CP(job32, *kcb, _aiocb_private.status); 2809 CP(job32, *kcb, _aiocb_private.error); 2810 PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo); 2811 return (convert_old_sigevent32(&job32.aio_sigevent, 2812 &kcb->aio_sigevent)); 2813 } 2814 #endif 2815 2816 static int 2817 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type) 2818 { 2819 struct aiocb32 job32; 2820 struct aiocb *kcb = &kjob->uaiocb; 2821 struct iovec32 *iov32; 2822 int error; 2823 2824 error = copyin(ujob, &job32, sizeof(job32)); 2825 if (error) 2826 return (error); 2827 CP(job32, *kcb, aio_fildes); 2828 CP(job32, *kcb, aio_offset); 2829 CP(job32, *kcb, aio_lio_opcode); 2830 if (type & LIO_VECTORED) { 2831 iov32 = PTRIN(job32.aio_iov); 2832 CP(job32, *kcb, aio_iovcnt); 2833 /* malloc a uio and copy in the iovec */ 2834 error = freebsd32_copyinuio(iov32, 2835 kcb->aio_iovcnt, &kjob->uiop); 2836 if (error) 2837 return (error); 2838 } else { 2839 PTRIN_CP(job32, *kcb, aio_buf); 2840 CP(job32, *kcb, aio_nbytes); 2841 } 2842 CP(job32, *kcb, aio_reqprio); 2843 CP(job32, *kcb, _aiocb_private.status); 2844 CP(job32, *kcb, _aiocb_private.error); 2845 PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo); 2846 error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent); 2847 2848 return (error); 2849 } 2850 2851 static long 2852 aiocb32_fetch_status(struct aiocb *ujob) 2853 { 2854 struct aiocb32 *ujob32; 2855 2856 ujob32 = (struct aiocb32 *)ujob; 2857 return (fuword32(&ujob32->_aiocb_private.status)); 2858 } 2859 2860 static long 2861 aiocb32_fetch_error(struct aiocb *ujob) 2862 { 2863 struct aiocb32 *ujob32; 2864 2865 ujob32 = (struct aiocb32 *)ujob; 2866 return (fuword32(&ujob32->_aiocb_private.error)); 2867 } 2868 2869 static int 2870 aiocb32_store_status(struct aiocb *ujob, long status) 2871 { 2872 struct aiocb32 *ujob32; 2873 2874 ujob32 = (struct aiocb32 *)ujob; 2875 return (suword32(&ujob32->_aiocb_private.status, status)); 2876 } 2877 2878 static int 2879 aiocb32_store_error(struct aiocb *ujob, long error) 2880 { 2881 struct aiocb32 *ujob32; 2882 2883 ujob32 = (struct aiocb32 *)ujob; 2884 return (suword32(&ujob32->_aiocb_private.error, error)); 2885 } 2886 2887 static int 2888 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref) 2889 { 2890 struct aiocb32 *ujob32; 2891 2892 ujob32 = (struct aiocb32 *)ujob; 2893 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref)); 2894 } 2895 2896 static int 2897 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 2898 { 2899 2900 return (suword32(ujobp, (long)ujob)); 2901 } 2902 2903 static struct aiocb_ops aiocb32_ops = { 2904 .aio_copyin = aiocb32_copyin, 2905 .fetch_status = aiocb32_fetch_status, 2906 .fetch_error = aiocb32_fetch_error, 2907 .store_status = aiocb32_store_status, 2908 .store_error = aiocb32_store_error, 2909 .store_kernelinfo = aiocb32_store_kernelinfo, 2910 .store_aiocb = aiocb32_store_aiocb, 2911 }; 2912 2913 #ifdef COMPAT_FREEBSD6 2914 static struct aiocb_ops aiocb32_ops_osigevent = { 2915 .aio_copyin = aiocb32_copyin_old_sigevent, 2916 .fetch_status = aiocb32_fetch_status, 2917 .fetch_error = aiocb32_fetch_error, 2918 .store_status = aiocb32_store_status, 2919 .store_error = aiocb32_store_error, 2920 .store_kernelinfo = aiocb32_store_kernelinfo, 2921 .store_aiocb = aiocb32_store_aiocb, 2922 }; 2923 #endif 2924 2925 int 2926 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap) 2927 { 2928 2929 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2930 } 2931 2932 int 2933 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap) 2934 { 2935 struct timespec32 ts32; 2936 struct timespec ts, *tsp; 2937 struct aiocb **ujoblist; 2938 uint32_t *ujoblist32; 2939 int error, i; 2940 2941 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 2942 return (EINVAL); 2943 2944 if (uap->timeout) { 2945 /* Get timespec struct. */ 2946 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0) 2947 return (error); 2948 CP(ts32, ts, tv_sec); 2949 CP(ts32, ts, tv_nsec); 2950 tsp = &ts; 2951 } else 2952 tsp = NULL; 2953 2954 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK); 2955 ujoblist32 = (uint32_t *)ujoblist; 2956 error = copyin(uap->aiocbp, ujoblist32, uap->nent * 2957 sizeof(ujoblist32[0])); 2958 if (error == 0) { 2959 for (i = uap->nent - 1; i >= 0; i--) 2960 ujoblist[i] = PTRIN(ujoblist32[i]); 2961 2962 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2963 } 2964 free(ujoblist, M_AIO); 2965 return (error); 2966 } 2967 2968 int 2969 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap) 2970 { 2971 2972 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2973 } 2974 2975 #ifdef COMPAT_FREEBSD6 2976 int 2977 freebsd6_freebsd32_aio_read(struct thread *td, 2978 struct freebsd6_freebsd32_aio_read_args *uap) 2979 { 2980 2981 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2982 &aiocb32_ops_osigevent)); 2983 } 2984 #endif 2985 2986 int 2987 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap) 2988 { 2989 2990 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2991 &aiocb32_ops)); 2992 } 2993 2994 int 2995 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap) 2996 { 2997 2998 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV, 2999 &aiocb32_ops)); 3000 } 3001 3002 #ifdef COMPAT_FREEBSD6 3003 int 3004 freebsd6_freebsd32_aio_write(struct thread *td, 3005 struct freebsd6_freebsd32_aio_write_args *uap) 3006 { 3007 3008 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 3009 &aiocb32_ops_osigevent)); 3010 } 3011 #endif 3012 3013 int 3014 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap) 3015 { 3016 3017 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 3018 &aiocb32_ops)); 3019 } 3020 3021 int 3022 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap) 3023 { 3024 3025 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV, 3026 &aiocb32_ops)); 3027 } 3028 3029 int 3030 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap) 3031 { 3032 3033 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK, 3034 &aiocb32_ops)); 3035 } 3036 3037 int 3038 freebsd32_aio_waitcomplete(struct thread *td, 3039 struct freebsd32_aio_waitcomplete_args *uap) 3040 { 3041 struct timespec32 ts32; 3042 struct timespec ts, *tsp; 3043 int error; 3044 3045 if (uap->timeout) { 3046 /* Get timespec struct. */ 3047 error = copyin(uap->timeout, &ts32, sizeof(ts32)); 3048 if (error) 3049 return (error); 3050 CP(ts32, ts, tv_sec); 3051 CP(ts32, ts, tv_nsec); 3052 tsp = &ts; 3053 } else 3054 tsp = NULL; 3055 3056 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp, 3057 &aiocb32_ops)); 3058 } 3059 3060 int 3061 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap) 3062 { 3063 3064 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp, 3065 &aiocb32_ops)); 3066 } 3067 3068 #ifdef COMPAT_FREEBSD6 3069 int 3070 freebsd6_freebsd32_lio_listio(struct thread *td, 3071 struct freebsd6_freebsd32_lio_listio_args *uap) 3072 { 3073 struct aiocb **acb_list; 3074 struct sigevent *sigp, sig; 3075 struct osigevent32 osig; 3076 uint32_t *acb_list32; 3077 int error, i, nent; 3078 3079 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3080 return (EINVAL); 3081 3082 nent = uap->nent; 3083 if (nent < 0 || nent > max_aio_queue_per_proc) 3084 return (EINVAL); 3085 3086 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3087 error = copyin(uap->sig, &osig, sizeof(osig)); 3088 if (error) 3089 return (error); 3090 error = convert_old_sigevent32(&osig, &sig); 3091 if (error) 3092 return (error); 3093 sigp = &sig; 3094 } else 3095 sigp = NULL; 3096 3097 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3098 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3099 if (error) { 3100 free(acb_list32, M_LIO); 3101 return (error); 3102 } 3103 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3104 for (i = 0; i < nent; i++) 3105 acb_list[i] = PTRIN(acb_list32[i]); 3106 free(acb_list32, M_LIO); 3107 3108 error = kern_lio_listio(td, uap->mode, 3109 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3110 &aiocb32_ops_osigevent); 3111 free(acb_list, M_LIO); 3112 return (error); 3113 } 3114 #endif 3115 3116 int 3117 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap) 3118 { 3119 struct aiocb **acb_list; 3120 struct sigevent *sigp, sig; 3121 struct sigevent32 sig32; 3122 uint32_t *acb_list32; 3123 int error, i, nent; 3124 3125 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3126 return (EINVAL); 3127 3128 nent = uap->nent; 3129 if (nent < 0 || nent > max_aio_queue_per_proc) 3130 return (EINVAL); 3131 3132 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3133 error = copyin(uap->sig, &sig32, sizeof(sig32)); 3134 if (error) 3135 return (error); 3136 error = convert_sigevent32(&sig32, &sig); 3137 if (error) 3138 return (error); 3139 sigp = &sig; 3140 } else 3141 sigp = NULL; 3142 3143 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3144 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3145 if (error) { 3146 free(acb_list32, M_LIO); 3147 return (error); 3148 } 3149 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3150 for (i = 0; i < nent; i++) 3151 acb_list[i] = PTRIN(acb_list32[i]); 3152 free(acb_list32, M_LIO); 3153 3154 error = kern_lio_listio(td, uap->mode, 3155 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3156 &aiocb32_ops); 3157 free(acb_list, M_LIO); 3158 return (error); 3159 } 3160 3161 #endif
Cache object: 104ff0000c1614746046abd87664ec96
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]