1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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/param.h> 24 #include <sys/systm.h> 25 #include <sys/malloc.h> 26 #include <sys/bio.h> 27 #include <sys/buf.h> 28 #include <sys/capsicum.h> 29 #include <sys/eventhandler.h> 30 #include <sys/sysproto.h> 31 #include <sys/filedesc.h> 32 #include <sys/kernel.h> 33 #include <sys/module.h> 34 #include <sys/kthread.h> 35 #include <sys/fcntl.h> 36 #include <sys/file.h> 37 #include <sys/limits.h> 38 #include <sys/lock.h> 39 #include <sys/mutex.h> 40 #include <sys/unistd.h> 41 #include <sys/posix4.h> 42 #include <sys/proc.h> 43 #include <sys/resourcevar.h> 44 #include <sys/signalvar.h> 45 #include <sys/syscallsubr.h> 46 #include <sys/protosw.h> 47 #include <sys/rwlock.h> 48 #include <sys/sema.h> 49 #include <sys/socket.h> 50 #include <sys/socketvar.h> 51 #include <sys/syscall.h> 52 #include <sys/sysctl.h> 53 #include <sys/syslog.h> 54 #include <sys/sx.h> 55 #include <sys/taskqueue.h> 56 #include <sys/vnode.h> 57 #include <sys/conf.h> 58 #include <sys/event.h> 59 #include <sys/mount.h> 60 #include <geom/geom.h> 61 62 #include <machine/atomic.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_page.h> 66 #include <vm/vm_extern.h> 67 #include <vm/pmap.h> 68 #include <vm/vm_map.h> 69 #include <vm/vm_object.h> 70 #include <vm/vnode_pager.h> 71 #include <vm/uma.h> 72 #include <sys/aio.h> 73 74 /* 75 * Counter for aio_fsync. 76 */ 77 static uint64_t jobseqno; 78 79 #ifndef MAX_AIO_PER_PROC 80 #define MAX_AIO_PER_PROC 32 81 #endif 82 83 #ifndef MAX_AIO_QUEUE_PER_PROC 84 #define MAX_AIO_QUEUE_PER_PROC 256 85 #endif 86 87 #ifndef MAX_AIO_QUEUE 88 #define MAX_AIO_QUEUE 1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */ 89 #endif 90 91 #ifndef MAX_BUF_AIO 92 #define MAX_BUF_AIO 16 93 #endif 94 95 FEATURE(aio, "Asynchronous I/O"); 96 SYSCTL_DECL(_p1003_1b); 97 98 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list"); 99 static MALLOC_DEFINE(M_AIO, "aio", "structures for asynchronous I/O"); 100 101 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 102 "Async IO management"); 103 104 static int enable_aio_unsafe = 0; 105 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0, 106 "Permit asynchronous IO on all file types, not just known-safe types"); 107 108 static unsigned int unsafe_warningcnt = 1; 109 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW, 110 &unsafe_warningcnt, 0, 111 "Warnings that will be triggered upon failed IO requests on unsafe files"); 112 113 static int max_aio_procs = MAX_AIO_PROCS; 114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0, 115 "Maximum number of kernel processes to use for handling async IO "); 116 117 static int num_aio_procs = 0; 118 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0, 119 "Number of presently active kernel processes for async IO"); 120 121 /* 122 * The code will adjust the actual number of AIO processes towards this 123 * number when it gets a chance. 124 */ 125 static int target_aio_procs = TARGET_AIO_PROCS; 126 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 127 0, 128 "Preferred number of ready kernel processes for async IO"); 129 130 static int max_queue_count = MAX_AIO_QUEUE; 131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 132 "Maximum number of aio requests to queue, globally"); 133 134 static int num_queue_count = 0; 135 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 136 "Number of queued aio requests"); 137 138 static int num_buf_aio = 0; 139 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 140 "Number of aio requests presently handled by the buf subsystem"); 141 142 static int num_unmapped_aio = 0; 143 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio, 144 0, 145 "Number of aio requests presently handled by unmapped I/O buffers"); 146 147 /* Number of async I/O processes in the process of being started */ 148 /* XXX This should be local to aio_aqueue() */ 149 static int num_aio_resv_start = 0; 150 151 static int aiod_lifetime; 152 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 153 "Maximum lifetime for idle aiod"); 154 155 static int max_aio_per_proc = MAX_AIO_PER_PROC; 156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 157 0, 158 "Maximum active aio requests per process"); 159 160 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 162 &max_aio_queue_per_proc, 0, 163 "Maximum queued aio requests per process"); 164 165 static int max_buf_aio = MAX_BUF_AIO; 166 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 167 "Maximum buf aio requests per process"); 168 169 /* 170 * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires 171 * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with 172 * vfs.aio.aio_listio_max. 173 */ 174 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max, 175 CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc, 176 0, "Maximum aio requests for a single lio_listio call"); 177 178 #ifdef COMPAT_FREEBSD6 179 typedef struct oaiocb { 180 int aio_fildes; /* File descriptor */ 181 off_t aio_offset; /* File offset for I/O */ 182 volatile void *aio_buf; /* I/O buffer in process space */ 183 size_t aio_nbytes; /* Number of bytes for I/O */ 184 struct osigevent aio_sigevent; /* Signal to deliver */ 185 int aio_lio_opcode; /* LIO opcode */ 186 int aio_reqprio; /* Request priority -- ignored */ 187 struct __aiocb_private _aiocb_private; 188 } oaiocb_t; 189 #endif 190 191 /* 192 * Below is a key of locks used to protect each member of struct kaiocb 193 * aioliojob and kaioinfo and any backends. 194 * 195 * * - need not protected 196 * a - locked by kaioinfo lock 197 * b - locked by backend lock, the backend lock can be null in some cases, 198 * for example, BIO belongs to this type, in this case, proc lock is 199 * reused. 200 * c - locked by aio_job_mtx, the lock for the generic file I/O backend. 201 */ 202 203 /* 204 * If the routine that services an AIO request blocks while running in an 205 * AIO kernel process it can starve other I/O requests. BIO requests 206 * queued via aio_qbio() complete asynchronously and do not use AIO kernel 207 * processes at all. Socket I/O requests use a separate pool of 208 * kprocs and also force non-blocking I/O. Other file I/O requests 209 * use the generic fo_read/fo_write operations which can block. The 210 * fsync and mlock operations can also block while executing. Ideally 211 * none of these requests would block while executing. 212 * 213 * Note that the service routines cannot toggle O_NONBLOCK in the file 214 * structure directly while handling a request due to races with 215 * userland threads. 216 */ 217 218 /* jobflags */ 219 #define KAIOCB_QUEUEING 0x01 220 #define KAIOCB_CANCELLED 0x02 221 #define KAIOCB_CANCELLING 0x04 222 #define KAIOCB_CHECKSYNC 0x08 223 #define KAIOCB_CLEARED 0x10 224 #define KAIOCB_FINISHED 0x20 225 226 /* ioflags */ 227 #define KAIOCB_IO_FOFFSET 0x01 228 229 /* 230 * AIO process info 231 */ 232 #define AIOP_FREE 0x1 /* proc on free queue */ 233 234 struct aioproc { 235 int aioprocflags; /* (c) AIO proc flags */ 236 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */ 237 struct proc *aioproc; /* (*) the AIO proc */ 238 }; 239 240 /* 241 * data-structure for lio signal management 242 */ 243 struct aioliojob { 244 int lioj_flags; /* (a) listio flags */ 245 int lioj_count; /* (a) count of jobs */ 246 int lioj_finished_count; /* (a) count of finished jobs */ 247 struct sigevent lioj_signal; /* (a) signal on all I/O done */ 248 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */ 249 struct knlist klist; /* (a) list of knotes */ 250 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */ 251 }; 252 253 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 254 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 255 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */ 256 257 /* 258 * per process aio data structure 259 */ 260 struct kaioinfo { 261 struct mtx kaio_mtx; /* the lock to protect this struct */ 262 int kaio_flags; /* (a) per process kaio flags */ 263 int kaio_active_count; /* (c) number of currently used AIOs */ 264 int kaio_count; /* (a) size of AIO queue */ 265 int kaio_buffer_count; /* (a) number of bio buffers */ 266 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */ 267 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */ 268 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */ 269 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */ 270 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */ 271 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */ 272 struct task kaio_task; /* (*) task to kick aio processes */ 273 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */ 274 }; 275 276 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx) 277 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx) 278 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f)) 279 #define AIO_MTX(ki) (&(ki)->kaio_mtx) 280 281 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 282 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */ 283 284 /* 285 * Operations used to interact with userland aio control blocks. 286 * Different ABIs provide their own operations. 287 */ 288 struct aiocb_ops { 289 int (*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty); 290 long (*fetch_status)(struct aiocb *ujob); 291 long (*fetch_error)(struct aiocb *ujob); 292 int (*store_status)(struct aiocb *ujob, long status); 293 int (*store_error)(struct aiocb *ujob, long error); 294 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob); 295 }; 296 297 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */ 298 static struct sema aio_newproc_sem; 299 static struct mtx aio_job_mtx; 300 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */ 301 static struct unrhdr *aiod_unr; 302 303 static void aio_biocleanup(struct bio *bp); 304 void aio_init_aioinfo(struct proc *p); 305 static int aio_onceonly(void); 306 static int aio_free_entry(struct kaiocb *job); 307 static void aio_process_rw(struct kaiocb *job); 308 static void aio_process_sync(struct kaiocb *job); 309 static void aio_process_mlock(struct kaiocb *job); 310 static void aio_schedule_fsync(void *context, int pending); 311 static int aio_newproc(int *); 312 int aio_aqueue(struct thread *td, struct aiocb *ujob, 313 struct aioliojob *lio, int type, struct aiocb_ops *ops); 314 static int aio_queue_file(struct file *fp, struct kaiocb *job); 315 static void aio_biowakeup(struct bio *bp); 316 static void aio_proc_rundown(void *arg, struct proc *p); 317 static void aio_proc_rundown_exec(void *arg, struct proc *p, 318 struct image_params *imgp); 319 static int aio_qbio(struct proc *p, struct kaiocb *job); 320 static void aio_daemon(void *param); 321 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job); 322 static bool aio_clear_cancel_function_locked(struct kaiocb *job); 323 static int aio_kick(struct proc *userp); 324 static void aio_kick_nowait(struct proc *userp); 325 static void aio_kick_helper(void *context, int pending); 326 static int filt_aioattach(struct knote *kn); 327 static void filt_aiodetach(struct knote *kn); 328 static int filt_aio(struct knote *kn, long hint); 329 static int filt_lioattach(struct knote *kn); 330 static void filt_liodetach(struct knote *kn); 331 static int filt_lio(struct knote *kn, long hint); 332 333 /* 334 * Zones for: 335 * kaio Per process async io info 336 * aiocb async io jobs 337 * aiolio list io jobs 338 */ 339 static uma_zone_t kaio_zone, aiocb_zone, aiolio_zone; 340 341 /* kqueue filters for aio */ 342 static const struct filterops aio_filtops = { 343 .f_isfd = 0, 344 .f_attach = filt_aioattach, 345 .f_detach = filt_aiodetach, 346 .f_event = filt_aio, 347 }; 348 static const struct filterops lio_filtops = { 349 .f_isfd = 0, 350 .f_attach = filt_lioattach, 351 .f_detach = filt_liodetach, 352 .f_event = filt_lio 353 }; 354 355 static eventhandler_tag exit_tag, exec_tag; 356 357 TASKQUEUE_DEFINE_THREAD(aiod_kick); 358 359 /* 360 * Main operations function for use as a kernel module. 361 */ 362 static int 363 aio_modload(struct module *module, int cmd, void *arg) 364 { 365 int error = 0; 366 367 switch (cmd) { 368 case MOD_LOAD: 369 aio_onceonly(); 370 break; 371 case MOD_SHUTDOWN: 372 break; 373 default: 374 error = EOPNOTSUPP; 375 break; 376 } 377 return (error); 378 } 379 380 static moduledata_t aio_mod = { 381 "aio", 382 &aio_modload, 383 NULL 384 }; 385 386 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY); 387 MODULE_VERSION(aio, 1); 388 389 /* 390 * Startup initialization 391 */ 392 static int 393 aio_onceonly(void) 394 { 395 396 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 397 EVENTHANDLER_PRI_ANY); 398 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, 399 NULL, EVENTHANDLER_PRI_ANY); 400 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 401 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 402 TAILQ_INIT(&aio_freeproc); 403 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 404 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 405 TAILQ_INIT(&aio_jobs); 406 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 407 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 408 NULL, NULL, UMA_ALIGN_PTR, 0); 409 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL, 410 NULL, NULL, UMA_ALIGN_PTR, 0); 411 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 412 NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 413 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 414 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO); 415 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 416 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 417 418 return (0); 419 } 420 421 /* 422 * Init the per-process aioinfo structure. The aioinfo limits are set 423 * per-process for user limit (resource) management. 424 */ 425 void 426 aio_init_aioinfo(struct proc *p) 427 { 428 struct kaioinfo *ki; 429 430 ki = uma_zalloc(kaio_zone, M_WAITOK); 431 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW); 432 ki->kaio_flags = 0; 433 ki->kaio_active_count = 0; 434 ki->kaio_count = 0; 435 ki->kaio_buffer_count = 0; 436 TAILQ_INIT(&ki->kaio_all); 437 TAILQ_INIT(&ki->kaio_done); 438 TAILQ_INIT(&ki->kaio_jobqueue); 439 TAILQ_INIT(&ki->kaio_liojoblist); 440 TAILQ_INIT(&ki->kaio_syncqueue); 441 TAILQ_INIT(&ki->kaio_syncready); 442 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 443 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki); 444 PROC_LOCK(p); 445 if (p->p_aioinfo == NULL) { 446 p->p_aioinfo = ki; 447 PROC_UNLOCK(p); 448 } else { 449 PROC_UNLOCK(p); 450 mtx_destroy(&ki->kaio_mtx); 451 uma_zfree(kaio_zone, ki); 452 } 453 454 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs)) 455 aio_newproc(NULL); 456 } 457 458 static int 459 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext) 460 { 461 struct thread *td; 462 int error; 463 464 error = sigev_findtd(p, sigev, &td); 465 if (error) 466 return (error); 467 if (!KSI_ONQ(ksi)) { 468 ksiginfo_set_sigev(ksi, sigev); 469 ksi->ksi_code = SI_ASYNCIO; 470 ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0; 471 tdsendsignal(p, td, ksi->ksi_signo, ksi); 472 } 473 PROC_UNLOCK(p); 474 return (error); 475 } 476 477 /* 478 * Free a job entry. Wait for completion if it is currently active, but don't 479 * delay forever. If we delay, we return a flag that says that we have to 480 * restart the queue scan. 481 */ 482 static int 483 aio_free_entry(struct kaiocb *job) 484 { 485 struct kaioinfo *ki; 486 struct aioliojob *lj; 487 struct proc *p; 488 489 p = job->userproc; 490 MPASS(curproc == p); 491 ki = p->p_aioinfo; 492 MPASS(ki != NULL); 493 494 AIO_LOCK_ASSERT(ki, MA_OWNED); 495 MPASS(job->jobflags & KAIOCB_FINISHED); 496 497 atomic_subtract_int(&num_queue_count, 1); 498 499 ki->kaio_count--; 500 MPASS(ki->kaio_count >= 0); 501 502 TAILQ_REMOVE(&ki->kaio_done, job, plist); 503 TAILQ_REMOVE(&ki->kaio_all, job, allist); 504 505 lj = job->lio; 506 if (lj) { 507 lj->lioj_count--; 508 lj->lioj_finished_count--; 509 510 if (lj->lioj_count == 0) { 511 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 512 /* lio is going away, we need to destroy any knotes */ 513 knlist_delete(&lj->klist, curthread, 1); 514 PROC_LOCK(p); 515 sigqueue_take(&lj->lioj_ksi); 516 PROC_UNLOCK(p); 517 uma_zfree(aiolio_zone, lj); 518 } 519 } 520 521 /* job is going away, we need to destroy any knotes */ 522 knlist_delete(&job->klist, curthread, 1); 523 PROC_LOCK(p); 524 sigqueue_take(&job->ksi); 525 PROC_UNLOCK(p); 526 527 AIO_UNLOCK(ki); 528 529 /* 530 * The thread argument here is used to find the owning process 531 * and is also passed to fo_close() which may pass it to various 532 * places such as devsw close() routines. Because of that, we 533 * need a thread pointer from the process owning the job that is 534 * persistent and won't disappear out from under us or move to 535 * another process. 536 * 537 * Currently, all the callers of this function call it to remove 538 * a kaiocb from the current process' job list either via a 539 * syscall or due to the current process calling exit() or 540 * execve(). Thus, we know that p == curproc. We also know that 541 * curthread can't exit since we are curthread. 542 * 543 * Therefore, we use curthread as the thread to pass to 544 * knlist_delete(). This does mean that it is possible for the 545 * thread pointer at close time to differ from the thread pointer 546 * at open time, but this is already true of file descriptors in 547 * a multithreaded process. 548 */ 549 if (job->fd_file) 550 fdrop(job->fd_file, curthread); 551 crfree(job->cred); 552 if (job->uiop != &job->uio) 553 freeuio(job->uiop); 554 uma_zfree(aiocb_zone, job); 555 AIO_LOCK(ki); 556 557 return (0); 558 } 559 560 static void 561 aio_proc_rundown_exec(void *arg, struct proc *p, 562 struct image_params *imgp __unused) 563 { 564 aio_proc_rundown(arg, p); 565 } 566 567 static int 568 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job) 569 { 570 aio_cancel_fn_t *func; 571 int cancelled; 572 573 AIO_LOCK_ASSERT(ki, MA_OWNED); 574 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED)) 575 return (0); 576 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0); 577 job->jobflags |= KAIOCB_CANCELLED; 578 579 func = job->cancel_fn; 580 581 /* 582 * If there is no cancel routine, just leave the job marked as 583 * cancelled. The job should be in active use by a caller who 584 * should complete it normally or when it fails to install a 585 * cancel routine. 586 */ 587 if (func == NULL) 588 return (0); 589 590 /* 591 * Set the CANCELLING flag so that aio_complete() will defer 592 * completions of this job. This prevents the job from being 593 * freed out from under the cancel callback. After the 594 * callback any deferred completion (whether from the callback 595 * or any other source) will be completed. 596 */ 597 job->jobflags |= KAIOCB_CANCELLING; 598 AIO_UNLOCK(ki); 599 func(job); 600 AIO_LOCK(ki); 601 job->jobflags &= ~KAIOCB_CANCELLING; 602 if (job->jobflags & KAIOCB_FINISHED) { 603 cancelled = job->uaiocb._aiocb_private.error == ECANCELED; 604 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 605 aio_bio_done_notify(p, job); 606 } else { 607 /* 608 * The cancel callback might have scheduled an 609 * operation to cancel this request, but it is 610 * only counted as cancelled if the request is 611 * cancelled when the callback returns. 612 */ 613 cancelled = 0; 614 } 615 return (cancelled); 616 } 617 618 /* 619 * Rundown the jobs for a given process. 620 */ 621 static void 622 aio_proc_rundown(void *arg, struct proc *p) 623 { 624 struct kaioinfo *ki; 625 struct aioliojob *lj; 626 struct kaiocb *job, *jobn; 627 628 KASSERT(curthread->td_proc == p, 629 ("%s: called on non-curproc", __func__)); 630 ki = p->p_aioinfo; 631 if (ki == NULL) 632 return; 633 634 AIO_LOCK(ki); 635 ki->kaio_flags |= KAIO_RUNDOWN; 636 637 restart: 638 639 /* 640 * Try to cancel all pending requests. This code simulates 641 * aio_cancel on all pending I/O requests. 642 */ 643 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 644 aio_cancel_job(p, ki, job); 645 } 646 647 /* Wait for all running I/O to be finished */ 648 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) { 649 ki->kaio_flags |= KAIO_WAKEUP; 650 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 651 goto restart; 652 } 653 654 /* Free all completed I/O requests. */ 655 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL) 656 aio_free_entry(job); 657 658 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 659 if (lj->lioj_count == 0) { 660 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 661 knlist_delete(&lj->klist, curthread, 1); 662 PROC_LOCK(p); 663 sigqueue_take(&lj->lioj_ksi); 664 PROC_UNLOCK(p); 665 uma_zfree(aiolio_zone, lj); 666 } else { 667 panic("LIO job not cleaned up: C:%d, FC:%d\n", 668 lj->lioj_count, lj->lioj_finished_count); 669 } 670 } 671 AIO_UNLOCK(ki); 672 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task); 673 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task); 674 mtx_destroy(&ki->kaio_mtx); 675 uma_zfree(kaio_zone, ki); 676 p->p_aioinfo = NULL; 677 } 678 679 /* 680 * Select a job to run (called by an AIO daemon). 681 */ 682 static struct kaiocb * 683 aio_selectjob(struct aioproc *aiop) 684 { 685 struct kaiocb *job; 686 struct kaioinfo *ki; 687 struct proc *userp; 688 689 mtx_assert(&aio_job_mtx, MA_OWNED); 690 restart: 691 TAILQ_FOREACH(job, &aio_jobs, list) { 692 userp = job->userproc; 693 ki = userp->p_aioinfo; 694 695 if (ki->kaio_active_count < max_aio_per_proc) { 696 TAILQ_REMOVE(&aio_jobs, job, list); 697 if (!aio_clear_cancel_function(job)) 698 goto restart; 699 700 /* Account for currently active jobs. */ 701 ki->kaio_active_count++; 702 break; 703 } 704 } 705 return (job); 706 } 707 708 /* 709 * Move all data to a permanent storage device. This code 710 * simulates the fsync and fdatasync syscalls. 711 */ 712 static int 713 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op) 714 { 715 struct mount *mp; 716 int error; 717 718 for (;;) { 719 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 720 if (error != 0) 721 break; 722 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 723 vnode_pager_clean_async(vp); 724 if (op == LIO_DSYNC) 725 error = VOP_FDATASYNC(vp, td); 726 else 727 error = VOP_FSYNC(vp, MNT_WAIT, td); 728 729 VOP_UNLOCK(vp); 730 vn_finished_write(mp); 731 if (error != ERELOOKUP) 732 break; 733 } 734 return (error); 735 } 736 737 /* 738 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that 739 * does the I/O request for the non-bio version of the operations. The normal 740 * vn operations are used, and this code should work in all instances for every 741 * type of file, including pipes, sockets, fifos, and regular files. 742 * 743 * XXX I don't think it works well for socket, pipe, and fifo. 744 */ 745 static void 746 aio_process_rw(struct kaiocb *job) 747 { 748 struct ucred *td_savedcred; 749 struct thread *td; 750 struct file *fp; 751 ssize_t cnt; 752 long msgsnd_st, msgsnd_end; 753 long msgrcv_st, msgrcv_end; 754 long oublock_st, oublock_end; 755 long inblock_st, inblock_end; 756 int error, opcode; 757 758 opcode = job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET; 759 KASSERT(opcode == LIO_READ || opcode == LIO_READV || 760 opcode == LIO_WRITE || opcode == LIO_WRITEV, 761 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 762 763 aio_switch_vmspace(job); 764 td = curthread; 765 td_savedcred = td->td_ucred; 766 td->td_ucred = job->cred; 767 job->uiop->uio_td = td; 768 fp = job->fd_file; 769 770 cnt = job->uiop->uio_resid; 771 772 msgrcv_st = td->td_ru.ru_msgrcv; 773 msgsnd_st = td->td_ru.ru_msgsnd; 774 inblock_st = td->td_ru.ru_inblock; 775 oublock_st = td->td_ru.ru_oublock; 776 777 /* 778 * aio_aqueue() acquires a reference to the file that is 779 * released in aio_free_entry(). 780 */ 781 if (opcode == LIO_READ || opcode == LIO_READV) { 782 if (job->uiop->uio_resid == 0) 783 error = 0; 784 else 785 error = fo_read(fp, job->uiop, fp->f_cred, 786 (job->ioflags & KAIOCB_IO_FOFFSET) != 0 ? 0 : 787 FOF_OFFSET, td); 788 } else { 789 if (fp->f_type == DTYPE_VNODE) 790 bwillwrite(); 791 error = fo_write(fp, job->uiop, fp->f_cred, (job->ioflags & 792 KAIOCB_IO_FOFFSET) != 0 ? 0 : FOF_OFFSET, td); 793 } 794 msgrcv_end = td->td_ru.ru_msgrcv; 795 msgsnd_end = td->td_ru.ru_msgsnd; 796 inblock_end = td->td_ru.ru_inblock; 797 oublock_end = td->td_ru.ru_oublock; 798 799 job->msgrcv = msgrcv_end - msgrcv_st; 800 job->msgsnd = msgsnd_end - msgsnd_st; 801 job->inblock = inblock_end - inblock_st; 802 job->outblock = oublock_end - oublock_st; 803 804 if (error != 0 && job->uiop->uio_resid != cnt) { 805 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 806 error = 0; 807 if (error == EPIPE && (opcode & LIO_WRITE)) { 808 PROC_LOCK(job->userproc); 809 kern_psignal(job->userproc, SIGPIPE); 810 PROC_UNLOCK(job->userproc); 811 } 812 } 813 814 cnt -= job->uiop->uio_resid; 815 td->td_ucred = td_savedcred; 816 if (error) 817 aio_complete(job, -1, error); 818 else 819 aio_complete(job, cnt, 0); 820 } 821 822 static void 823 aio_process_sync(struct kaiocb *job) 824 { 825 struct thread *td = curthread; 826 struct ucred *td_savedcred = td->td_ucred; 827 struct file *fp = job->fd_file; 828 int error = 0; 829 830 KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC, 831 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 832 833 td->td_ucred = job->cred; 834 if (fp->f_vnode != NULL) { 835 error = aio_fsync_vnode(td, fp->f_vnode, 836 job->uaiocb.aio_lio_opcode); 837 } 838 td->td_ucred = td_savedcred; 839 if (error) 840 aio_complete(job, -1, error); 841 else 842 aio_complete(job, 0, 0); 843 } 844 845 static void 846 aio_process_mlock(struct kaiocb *job) 847 { 848 struct aiocb *cb = &job->uaiocb; 849 int error; 850 851 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK, 852 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 853 854 aio_switch_vmspace(job); 855 error = kern_mlock(job->userproc, job->cred, 856 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes); 857 aio_complete(job, error != 0 ? -1 : 0, error); 858 } 859 860 static void 861 aio_bio_done_notify(struct proc *userp, struct kaiocb *job) 862 { 863 struct aioliojob *lj; 864 struct kaioinfo *ki; 865 struct kaiocb *sjob, *sjobn; 866 int lj_done; 867 bool schedule_fsync; 868 869 ki = userp->p_aioinfo; 870 AIO_LOCK_ASSERT(ki, MA_OWNED); 871 lj = job->lio; 872 lj_done = 0; 873 if (lj) { 874 lj->lioj_finished_count++; 875 if (lj->lioj_count == lj->lioj_finished_count) 876 lj_done = 1; 877 } 878 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist); 879 MPASS(job->jobflags & KAIOCB_FINISHED); 880 881 if (ki->kaio_flags & KAIO_RUNDOWN) 882 goto notification_done; 883 884 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 885 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 886 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true); 887 888 KNOTE_LOCKED(&job->klist, 1); 889 890 if (lj_done) { 891 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 892 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 893 KNOTE_LOCKED(&lj->klist, 1); 894 } 895 if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED)) 896 == LIOJ_SIGNAL && 897 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 898 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 899 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi, 900 true); 901 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 902 } 903 } 904 905 notification_done: 906 if (job->jobflags & KAIOCB_CHECKSYNC) { 907 schedule_fsync = false; 908 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) { 909 if (job->fd_file != sjob->fd_file || 910 job->seqno >= sjob->seqno) 911 continue; 912 if (--sjob->pending > 0) 913 continue; 914 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list); 915 if (!aio_clear_cancel_function_locked(sjob)) 916 continue; 917 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list); 918 schedule_fsync = true; 919 } 920 if (schedule_fsync) 921 taskqueue_enqueue(taskqueue_aiod_kick, 922 &ki->kaio_sync_task); 923 } 924 if (ki->kaio_flags & KAIO_WAKEUP) { 925 ki->kaio_flags &= ~KAIO_WAKEUP; 926 wakeup(&userp->p_aioinfo); 927 } 928 } 929 930 static void 931 aio_schedule_fsync(void *context, int pending) 932 { 933 struct kaioinfo *ki; 934 struct kaiocb *job; 935 936 ki = context; 937 AIO_LOCK(ki); 938 while (!TAILQ_EMPTY(&ki->kaio_syncready)) { 939 job = TAILQ_FIRST(&ki->kaio_syncready); 940 TAILQ_REMOVE(&ki->kaio_syncready, job, list); 941 AIO_UNLOCK(ki); 942 aio_schedule(job, aio_process_sync); 943 AIO_LOCK(ki); 944 } 945 AIO_UNLOCK(ki); 946 } 947 948 bool 949 aio_cancel_cleared(struct kaiocb *job) 950 { 951 952 /* 953 * The caller should hold the same queue lock held when 954 * aio_clear_cancel_function() was called and set this flag 955 * ensuring this check sees an up-to-date value. However, 956 * there is no way to assert that. 957 */ 958 return ((job->jobflags & KAIOCB_CLEARED) != 0); 959 } 960 961 static bool 962 aio_clear_cancel_function_locked(struct kaiocb *job) 963 { 964 965 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 966 MPASS(job->cancel_fn != NULL); 967 if (job->jobflags & KAIOCB_CANCELLING) { 968 job->jobflags |= KAIOCB_CLEARED; 969 return (false); 970 } 971 job->cancel_fn = NULL; 972 return (true); 973 } 974 975 bool 976 aio_clear_cancel_function(struct kaiocb *job) 977 { 978 struct kaioinfo *ki; 979 bool ret; 980 981 ki = job->userproc->p_aioinfo; 982 AIO_LOCK(ki); 983 ret = aio_clear_cancel_function_locked(job); 984 AIO_UNLOCK(ki); 985 return (ret); 986 } 987 988 static bool 989 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func) 990 { 991 992 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 993 if (job->jobflags & KAIOCB_CANCELLED) 994 return (false); 995 job->cancel_fn = func; 996 return (true); 997 } 998 999 bool 1000 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func) 1001 { 1002 struct kaioinfo *ki; 1003 bool ret; 1004 1005 ki = job->userproc->p_aioinfo; 1006 AIO_LOCK(ki); 1007 ret = aio_set_cancel_function_locked(job, func); 1008 AIO_UNLOCK(ki); 1009 return (ret); 1010 } 1011 1012 void 1013 aio_complete(struct kaiocb *job, long status, int error) 1014 { 1015 struct kaioinfo *ki; 1016 struct proc *userp; 1017 1018 job->uaiocb._aiocb_private.error = error; 1019 job->uaiocb._aiocb_private.status = status; 1020 1021 userp = job->userproc; 1022 ki = userp->p_aioinfo; 1023 1024 AIO_LOCK(ki); 1025 KASSERT(!(job->jobflags & KAIOCB_FINISHED), 1026 ("duplicate aio_complete")); 1027 job->jobflags |= KAIOCB_FINISHED; 1028 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) { 1029 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 1030 aio_bio_done_notify(userp, job); 1031 } 1032 AIO_UNLOCK(ki); 1033 } 1034 1035 void 1036 aio_cancel(struct kaiocb *job) 1037 { 1038 1039 aio_complete(job, -1, ECANCELED); 1040 } 1041 1042 void 1043 aio_switch_vmspace(struct kaiocb *job) 1044 { 1045 1046 vmspace_switch_aio(job->userproc->p_vmspace); 1047 } 1048 1049 /* 1050 * The AIO daemon, most of the actual work is done in aio_process_*, 1051 * but the setup (and address space mgmt) is done in this routine. 1052 */ 1053 static void 1054 aio_daemon(void *_id) 1055 { 1056 struct kaiocb *job; 1057 struct aioproc *aiop; 1058 struct kaioinfo *ki; 1059 struct proc *p; 1060 struct vmspace *myvm; 1061 struct thread *td = curthread; 1062 int id = (intptr_t)_id; 1063 1064 /* 1065 * Grab an extra reference on the daemon's vmspace so that it 1066 * doesn't get freed by jobs that switch to a different 1067 * vmspace. 1068 */ 1069 p = td->td_proc; 1070 myvm = vmspace_acquire_ref(p); 1071 1072 KASSERT(p->p_textvp == NULL, ("kthread has a textvp")); 1073 1074 /* 1075 * Allocate and ready the aio control info. There is one aiop structure 1076 * per daemon. 1077 */ 1078 aiop = malloc(sizeof(*aiop), M_AIO, M_WAITOK); 1079 aiop->aioproc = p; 1080 aiop->aioprocflags = 0; 1081 1082 /* 1083 * Wakeup parent process. (Parent sleeps to keep from blasting away 1084 * and creating too many daemons.) 1085 */ 1086 sema_post(&aio_newproc_sem); 1087 1088 mtx_lock(&aio_job_mtx); 1089 for (;;) { 1090 /* 1091 * Take daemon off of free queue 1092 */ 1093 if (aiop->aioprocflags & AIOP_FREE) { 1094 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1095 aiop->aioprocflags &= ~AIOP_FREE; 1096 } 1097 1098 /* 1099 * Check for jobs. 1100 */ 1101 while ((job = aio_selectjob(aiop)) != NULL) { 1102 mtx_unlock(&aio_job_mtx); 1103 1104 ki = job->userproc->p_aioinfo; 1105 job->handle_fn(job); 1106 1107 mtx_lock(&aio_job_mtx); 1108 /* Decrement the active job count. */ 1109 ki->kaio_active_count--; 1110 } 1111 1112 /* 1113 * Disconnect from user address space. 1114 */ 1115 if (p->p_vmspace != myvm) { 1116 mtx_unlock(&aio_job_mtx); 1117 vmspace_switch_aio(myvm); 1118 mtx_lock(&aio_job_mtx); 1119 /* 1120 * We have to restart to avoid race, we only sleep if 1121 * no job can be selected. 1122 */ 1123 continue; 1124 } 1125 1126 mtx_assert(&aio_job_mtx, MA_OWNED); 1127 1128 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1129 aiop->aioprocflags |= AIOP_FREE; 1130 1131 /* 1132 * If daemon is inactive for a long time, allow it to exit, 1133 * thereby freeing resources. 1134 */ 1135 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy", 1136 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) && 1137 (aiop->aioprocflags & AIOP_FREE) && 1138 num_aio_procs > target_aio_procs) 1139 break; 1140 } 1141 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1142 num_aio_procs--; 1143 mtx_unlock(&aio_job_mtx); 1144 free(aiop, M_AIO); 1145 free_unr(aiod_unr, id); 1146 vmspace_free(myvm); 1147 1148 KASSERT(p->p_vmspace == myvm, 1149 ("AIOD: bad vmspace for exiting daemon")); 1150 KASSERT(refcount_load(&myvm->vm_refcnt) > 1, 1151 ("AIOD: bad vm refcnt for exiting daemon: %d", 1152 refcount_load(&myvm->vm_refcnt))); 1153 kproc_exit(0); 1154 } 1155 1156 /* 1157 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1158 * AIO daemon modifies its environment itself. 1159 */ 1160 static int 1161 aio_newproc(int *start) 1162 { 1163 int error; 1164 struct proc *p; 1165 int id; 1166 1167 id = alloc_unr(aiod_unr); 1168 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p, 1169 RFNOWAIT, 0, "aiod%d", id); 1170 if (error == 0) { 1171 /* 1172 * Wait until daemon is started. 1173 */ 1174 sema_wait(&aio_newproc_sem); 1175 mtx_lock(&aio_job_mtx); 1176 num_aio_procs++; 1177 if (start != NULL) 1178 (*start)--; 1179 mtx_unlock(&aio_job_mtx); 1180 } else { 1181 free_unr(aiod_unr, id); 1182 } 1183 return (error); 1184 } 1185 1186 /* 1187 * Try the high-performance, low-overhead bio method for eligible 1188 * VCHR devices. This method doesn't use an aio helper thread, and 1189 * thus has very low overhead. 1190 * 1191 * Assumes that the caller, aio_aqueue(), has incremented the file 1192 * structure's reference count, preventing its deallocation for the 1193 * duration of this call. 1194 */ 1195 static int 1196 aio_qbio(struct proc *p, struct kaiocb *job) 1197 { 1198 struct aiocb *cb; 1199 struct file *fp; 1200 struct buf *pbuf; 1201 struct vnode *vp; 1202 struct cdevsw *csw; 1203 struct cdev *dev; 1204 struct kaioinfo *ki; 1205 struct bio **bios = NULL; 1206 off_t offset; 1207 int bio_cmd, error, i, iovcnt, opcode, poff, ref; 1208 vm_prot_t prot; 1209 bool use_unmapped; 1210 1211 cb = &job->uaiocb; 1212 fp = job->fd_file; 1213 opcode = cb->aio_lio_opcode; 1214 1215 if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV || 1216 opcode == LIO_READ || opcode == LIO_READV)) 1217 return (-1); 1218 if (fp == NULL || fp->f_type != DTYPE_VNODE) 1219 return (-1); 1220 1221 vp = fp->f_vnode; 1222 if (vp->v_type != VCHR) 1223 return (-1); 1224 if (vp->v_bufobj.bo_bsize == 0) 1225 return (-1); 1226 1227 bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ; 1228 iovcnt = job->uiop->uio_iovcnt; 1229 if (iovcnt > max_buf_aio) 1230 return (-1); 1231 for (i = 0; i < iovcnt; i++) { 1232 if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0) 1233 return (-1); 1234 if (job->uiop->uio_iov[i].iov_len > maxphys) { 1235 error = -1; 1236 return (-1); 1237 } 1238 } 1239 offset = cb->aio_offset; 1240 1241 ref = 0; 1242 csw = devvn_refthread(vp, &dev, &ref); 1243 if (csw == NULL) 1244 return (ENXIO); 1245 1246 if ((csw->d_flags & D_DISK) == 0) { 1247 error = -1; 1248 goto unref; 1249 } 1250 if (job->uiop->uio_resid > dev->si_iosize_max) { 1251 error = -1; 1252 goto unref; 1253 } 1254 1255 ki = p->p_aioinfo; 1256 job->error = 0; 1257 1258 use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed; 1259 if (!use_unmapped) { 1260 AIO_LOCK(ki); 1261 if (ki->kaio_buffer_count + iovcnt > max_buf_aio) { 1262 AIO_UNLOCK(ki); 1263 error = EAGAIN; 1264 goto unref; 1265 } 1266 ki->kaio_buffer_count += iovcnt; 1267 AIO_UNLOCK(ki); 1268 } 1269 1270 bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK); 1271 refcount_init(&job->nbio, iovcnt); 1272 for (i = 0; i < iovcnt; i++) { 1273 struct vm_page** pages; 1274 struct bio *bp; 1275 void *buf; 1276 size_t nbytes; 1277 int npages; 1278 1279 buf = job->uiop->uio_iov[i].iov_base; 1280 nbytes = job->uiop->uio_iov[i].iov_len; 1281 1282 bios[i] = g_alloc_bio(); 1283 bp = bios[i]; 1284 1285 poff = (vm_offset_t)buf & PAGE_MASK; 1286 if (use_unmapped) { 1287 pbuf = NULL; 1288 pages = malloc(sizeof(vm_page_t) * (atop(round_page( 1289 nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO); 1290 } else { 1291 pbuf = uma_zalloc(pbuf_zone, M_WAITOK); 1292 BUF_KERNPROC(pbuf); 1293 pages = pbuf->b_pages; 1294 } 1295 1296 bp->bio_length = nbytes; 1297 bp->bio_bcount = nbytes; 1298 bp->bio_done = aio_biowakeup; 1299 bp->bio_offset = offset; 1300 bp->bio_cmd = bio_cmd; 1301 bp->bio_dev = dev; 1302 bp->bio_caller1 = job; 1303 bp->bio_caller2 = pbuf; 1304 1305 prot = VM_PROT_READ; 1306 if (opcode == LIO_READ || opcode == LIO_READV) 1307 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 1308 npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 1309 (vm_offset_t)buf, bp->bio_length, prot, pages, 1310 atop(maxphys) + 1); 1311 if (npages < 0) { 1312 if (pbuf != NULL) 1313 uma_zfree(pbuf_zone, pbuf); 1314 else 1315 free(pages, M_TEMP); 1316 error = EFAULT; 1317 g_destroy_bio(bp); 1318 i--; 1319 goto destroy_bios; 1320 } 1321 if (pbuf != NULL) { 1322 pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages); 1323 bp->bio_data = pbuf->b_data + poff; 1324 pbuf->b_npages = npages; 1325 atomic_add_int(&num_buf_aio, 1); 1326 } else { 1327 bp->bio_ma = pages; 1328 bp->bio_ma_n = npages; 1329 bp->bio_ma_offset = poff; 1330 bp->bio_data = unmapped_buf; 1331 bp->bio_flags |= BIO_UNMAPPED; 1332 atomic_add_int(&num_unmapped_aio, 1); 1333 } 1334 1335 offset += nbytes; 1336 } 1337 1338 /* Perform transfer. */ 1339 for (i = 0; i < iovcnt; i++) 1340 csw->d_strategy(bios[i]); 1341 free(bios, M_TEMP); 1342 1343 dev_relthread(dev, ref); 1344 return (0); 1345 1346 destroy_bios: 1347 for (; i >= 0; i--) 1348 aio_biocleanup(bios[i]); 1349 free(bios, M_TEMP); 1350 unref: 1351 dev_relthread(dev, ref); 1352 return (error); 1353 } 1354 1355 #ifdef COMPAT_FREEBSD6 1356 static int 1357 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig) 1358 { 1359 1360 /* 1361 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 1362 * supported by AIO with the old sigevent structure. 1363 */ 1364 nsig->sigev_notify = osig->sigev_notify; 1365 switch (nsig->sigev_notify) { 1366 case SIGEV_NONE: 1367 break; 1368 case SIGEV_SIGNAL: 1369 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 1370 break; 1371 case SIGEV_KEVENT: 1372 nsig->sigev_notify_kqueue = 1373 osig->__sigev_u.__sigev_notify_kqueue; 1374 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr; 1375 break; 1376 default: 1377 return (EINVAL); 1378 } 1379 return (0); 1380 } 1381 1382 static int 1383 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob, 1384 int type __unused) 1385 { 1386 struct oaiocb *ojob; 1387 struct aiocb *kcb = &kjob->uaiocb; 1388 int error; 1389 1390 bzero(kcb, sizeof(struct aiocb)); 1391 error = copyin(ujob, kcb, sizeof(struct oaiocb)); 1392 if (error) 1393 return (error); 1394 /* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */ 1395 ojob = (struct oaiocb *)kcb; 1396 return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent)); 1397 } 1398 #endif 1399 1400 static int 1401 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type) 1402 { 1403 struct aiocb *kcb = &kjob->uaiocb; 1404 int error; 1405 1406 error = copyin(ujob, kcb, sizeof(struct aiocb)); 1407 if (error) 1408 return (error); 1409 if (type == LIO_NOP) 1410 type = kcb->aio_lio_opcode; 1411 if (type & LIO_VECTORED) { 1412 /* malloc a uio and copy in the iovec */ 1413 error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov), 1414 kcb->aio_iovcnt, &kjob->uiop); 1415 } 1416 1417 return (error); 1418 } 1419 1420 static long 1421 aiocb_fetch_status(struct aiocb *ujob) 1422 { 1423 1424 return (fuword(&ujob->_aiocb_private.status)); 1425 } 1426 1427 static long 1428 aiocb_fetch_error(struct aiocb *ujob) 1429 { 1430 1431 return (fuword(&ujob->_aiocb_private.error)); 1432 } 1433 1434 static int 1435 aiocb_store_status(struct aiocb *ujob, long status) 1436 { 1437 1438 return (suword(&ujob->_aiocb_private.status, status)); 1439 } 1440 1441 static int 1442 aiocb_store_error(struct aiocb *ujob, long error) 1443 { 1444 1445 return (suword(&ujob->_aiocb_private.error, error)); 1446 } 1447 1448 static int 1449 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 1450 { 1451 1452 return (suword(ujobp, (long)ujob)); 1453 } 1454 1455 static struct aiocb_ops aiocb_ops = { 1456 .aio_copyin = aiocb_copyin, 1457 .fetch_status = aiocb_fetch_status, 1458 .fetch_error = aiocb_fetch_error, 1459 .store_status = aiocb_store_status, 1460 .store_error = aiocb_store_error, 1461 .store_aiocb = aiocb_store_aiocb, 1462 }; 1463 1464 #ifdef COMPAT_FREEBSD6 1465 static struct aiocb_ops aiocb_ops_osigevent = { 1466 .aio_copyin = aiocb_copyin_old_sigevent, 1467 .fetch_status = aiocb_fetch_status, 1468 .fetch_error = aiocb_fetch_error, 1469 .store_status = aiocb_store_status, 1470 .store_error = aiocb_store_error, 1471 .store_aiocb = aiocb_store_aiocb, 1472 }; 1473 #endif 1474 1475 /* 1476 * Queue a new AIO request. Choosing either the threaded or direct bio VCHR 1477 * technique is done in this code. 1478 */ 1479 int 1480 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj, 1481 int type, struct aiocb_ops *ops) 1482 { 1483 struct proc *p = td->td_proc; 1484 struct file *fp = NULL; 1485 struct kaiocb *job; 1486 struct kaioinfo *ki; 1487 struct kevent kev; 1488 int opcode; 1489 int error; 1490 int fd, kqfd; 1491 u_short evflags; 1492 1493 if (p->p_aioinfo == NULL) 1494 aio_init_aioinfo(p); 1495 1496 ki = p->p_aioinfo; 1497 1498 ops->store_status(ujob, -1); 1499 ops->store_error(ujob, 0); 1500 1501 if (num_queue_count >= max_queue_count || 1502 ki->kaio_count >= max_aio_queue_per_proc) { 1503 error = EAGAIN; 1504 goto err1; 1505 } 1506 1507 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1508 knlist_init_mtx(&job->klist, AIO_MTX(ki)); 1509 1510 error = ops->aio_copyin(ujob, job, type); 1511 if (error) 1512 goto err2; 1513 1514 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) { 1515 error = EINVAL; 1516 goto err2; 1517 } 1518 1519 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1520 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1521 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1522 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1523 error = EINVAL; 1524 goto err2; 1525 } 1526 1527 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1528 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1529 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) { 1530 error = EINVAL; 1531 goto err2; 1532 } 1533 1534 /* Get the opcode. */ 1535 if (type == LIO_NOP) { 1536 switch (job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET) { 1537 case LIO_WRITE: 1538 case LIO_WRITEV: 1539 case LIO_NOP: 1540 case LIO_READ: 1541 case LIO_READV: 1542 opcode = job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET; 1543 if ((job->uaiocb.aio_lio_opcode & LIO_FOFFSET) != 0) 1544 job->ioflags |= KAIOCB_IO_FOFFSET; 1545 break; 1546 default: 1547 error = EINVAL; 1548 goto err2; 1549 } 1550 } else 1551 opcode = job->uaiocb.aio_lio_opcode = type; 1552 1553 ksiginfo_init(&job->ksi); 1554 1555 /* Save userspace address of the job info. */ 1556 job->ujob = ujob; 1557 1558 /* 1559 * Validate the opcode and fetch the file object for the specified 1560 * file descriptor. 1561 * 1562 * XXXRW: Moved the opcode validation up here so that we don't 1563 * retrieve a file descriptor without knowing what the capabiltity 1564 * should be. 1565 */ 1566 fd = job->uaiocb.aio_fildes; 1567 switch (opcode) { 1568 case LIO_WRITE: 1569 case LIO_WRITEV: 1570 error = fget_write(td, fd, &cap_pwrite_rights, &fp); 1571 break; 1572 case LIO_READ: 1573 case LIO_READV: 1574 error = fget_read(td, fd, &cap_pread_rights, &fp); 1575 break; 1576 case LIO_SYNC: 1577 case LIO_DSYNC: 1578 error = fget(td, fd, &cap_fsync_rights, &fp); 1579 break; 1580 case LIO_MLOCK: 1581 break; 1582 case LIO_NOP: 1583 error = fget(td, fd, &cap_no_rights, &fp); 1584 break; 1585 default: 1586 error = EINVAL; 1587 } 1588 if (error) 1589 goto err3; 1590 1591 if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) { 1592 error = EINVAL; 1593 goto err3; 1594 } 1595 1596 if ((opcode == LIO_READ || opcode == LIO_READV || 1597 opcode == LIO_WRITE || opcode == LIO_WRITEV) && 1598 job->uaiocb.aio_offset < 0 && 1599 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) { 1600 error = EINVAL; 1601 goto err3; 1602 } 1603 1604 if (fp != NULL && fp->f_ops == &path_fileops) { 1605 error = EBADF; 1606 goto err3; 1607 } 1608 1609 job->fd_file = fp; 1610 1611 mtx_lock(&aio_job_mtx); 1612 job->seqno = jobseqno++; 1613 mtx_unlock(&aio_job_mtx); 1614 if (opcode == LIO_NOP) { 1615 fdrop(fp, td); 1616 MPASS(job->uiop == &job->uio || job->uiop == NULL); 1617 uma_zfree(aiocb_zone, job); 1618 return (0); 1619 } 1620 1621 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT) 1622 goto no_kqueue; 1623 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags; 1624 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) { 1625 error = EINVAL; 1626 goto err3; 1627 } 1628 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue; 1629 memset(&kev, 0, sizeof(kev)); 1630 kev.ident = (uintptr_t)job->ujob; 1631 kev.filter = EVFILT_AIO; 1632 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags; 1633 kev.data = (intptr_t)job; 1634 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1635 error = kqfd_register(kqfd, &kev, td, M_WAITOK); 1636 if (error) 1637 goto err3; 1638 1639 no_kqueue: 1640 1641 ops->store_error(ujob, EINPROGRESS); 1642 job->uaiocb._aiocb_private.error = EINPROGRESS; 1643 job->userproc = p; 1644 job->cred = crhold(td->td_ucred); 1645 job->jobflags = KAIOCB_QUEUEING; 1646 job->lio = lj; 1647 1648 if (opcode & LIO_VECTORED) { 1649 /* Use the uio copied in by aio_copyin */ 1650 MPASS(job->uiop != &job->uio && job->uiop != NULL); 1651 } else { 1652 /* Setup the inline uio */ 1653 job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf; 1654 job->iov[0].iov_len = job->uaiocb.aio_nbytes; 1655 job->uio.uio_iov = job->iov; 1656 job->uio.uio_iovcnt = 1; 1657 job->uio.uio_resid = job->uaiocb.aio_nbytes; 1658 job->uio.uio_segflg = UIO_USERSPACE; 1659 job->uiop = &job->uio; 1660 } 1661 switch (opcode & (LIO_READ | LIO_WRITE)) { 1662 case LIO_READ: 1663 job->uiop->uio_rw = UIO_READ; 1664 break; 1665 case LIO_WRITE: 1666 job->uiop->uio_rw = UIO_WRITE; 1667 break; 1668 } 1669 job->uiop->uio_offset = job->uaiocb.aio_offset; 1670 job->uiop->uio_td = td; 1671 1672 if (opcode == LIO_MLOCK) { 1673 aio_schedule(job, aio_process_mlock); 1674 error = 0; 1675 } else if (fp->f_ops->fo_aio_queue == NULL) 1676 error = aio_queue_file(fp, job); 1677 else 1678 error = fo_aio_queue(fp, job); 1679 if (error) 1680 goto err4; 1681 1682 AIO_LOCK(ki); 1683 job->jobflags &= ~KAIOCB_QUEUEING; 1684 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist); 1685 ki->kaio_count++; 1686 if (lj) 1687 lj->lioj_count++; 1688 atomic_add_int(&num_queue_count, 1); 1689 if (job->jobflags & KAIOCB_FINISHED) { 1690 /* 1691 * The queue callback completed the request synchronously. 1692 * The bulk of the completion is deferred in that case 1693 * until this point. 1694 */ 1695 aio_bio_done_notify(p, job); 1696 } else 1697 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist); 1698 AIO_UNLOCK(ki); 1699 return (0); 1700 1701 err4: 1702 crfree(job->cred); 1703 err3: 1704 if (fp) 1705 fdrop(fp, td); 1706 knlist_delete(&job->klist, curthread, 0); 1707 err2: 1708 if (job->uiop != &job->uio) 1709 freeuio(job->uiop); 1710 uma_zfree(aiocb_zone, job); 1711 err1: 1712 ops->store_error(ujob, error); 1713 return (error); 1714 } 1715 1716 static void 1717 aio_cancel_daemon_job(struct kaiocb *job) 1718 { 1719 1720 mtx_lock(&aio_job_mtx); 1721 if (!aio_cancel_cleared(job)) 1722 TAILQ_REMOVE(&aio_jobs, job, list); 1723 mtx_unlock(&aio_job_mtx); 1724 aio_cancel(job); 1725 } 1726 1727 void 1728 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func) 1729 { 1730 1731 mtx_lock(&aio_job_mtx); 1732 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) { 1733 mtx_unlock(&aio_job_mtx); 1734 aio_cancel(job); 1735 return; 1736 } 1737 job->handle_fn = func; 1738 TAILQ_INSERT_TAIL(&aio_jobs, job, list); 1739 aio_kick_nowait(job->userproc); 1740 mtx_unlock(&aio_job_mtx); 1741 } 1742 1743 static void 1744 aio_cancel_sync(struct kaiocb *job) 1745 { 1746 struct kaioinfo *ki; 1747 1748 ki = job->userproc->p_aioinfo; 1749 AIO_LOCK(ki); 1750 if (!aio_cancel_cleared(job)) 1751 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list); 1752 AIO_UNLOCK(ki); 1753 aio_cancel(job); 1754 } 1755 1756 int 1757 aio_queue_file(struct file *fp, struct kaiocb *job) 1758 { 1759 struct kaioinfo *ki; 1760 struct kaiocb *job2; 1761 struct vnode *vp; 1762 struct mount *mp; 1763 int error; 1764 bool safe; 1765 1766 ki = job->userproc->p_aioinfo; 1767 error = aio_qbio(job->userproc, job); 1768 if (error >= 0) 1769 return (error); 1770 safe = false; 1771 if (fp->f_type == DTYPE_VNODE) { 1772 vp = fp->f_vnode; 1773 if (vp->v_type == VREG || vp->v_type == VDIR) { 1774 mp = fp->f_vnode->v_mount; 1775 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0) 1776 safe = true; 1777 } 1778 } 1779 if (!(safe || enable_aio_unsafe)) { 1780 counted_warning(&unsafe_warningcnt, 1781 "is attempting to use unsafe AIO requests"); 1782 return (EOPNOTSUPP); 1783 } 1784 1785 if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) { 1786 aio_schedule(job, aio_process_rw); 1787 error = 0; 1788 } else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) { 1789 AIO_LOCK(ki); 1790 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) { 1791 if (job2->fd_file == job->fd_file && 1792 ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) && 1793 job2->seqno < job->seqno) { 1794 job2->jobflags |= KAIOCB_CHECKSYNC; 1795 job->pending++; 1796 } 1797 } 1798 if (job->pending != 0) { 1799 if (!aio_set_cancel_function_locked(job, 1800 aio_cancel_sync)) { 1801 AIO_UNLOCK(ki); 1802 aio_cancel(job); 1803 return (0); 1804 } 1805 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list); 1806 AIO_UNLOCK(ki); 1807 return (0); 1808 } 1809 AIO_UNLOCK(ki); 1810 aio_schedule(job, aio_process_sync); 1811 error = 0; 1812 } else { 1813 error = EINVAL; 1814 } 1815 return (error); 1816 } 1817 1818 static void 1819 aio_kick_nowait(struct proc *userp) 1820 { 1821 struct kaioinfo *ki = userp->p_aioinfo; 1822 struct aioproc *aiop; 1823 1824 mtx_assert(&aio_job_mtx, MA_OWNED); 1825 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1826 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1827 aiop->aioprocflags &= ~AIOP_FREE; 1828 wakeup(aiop->aioproc); 1829 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1830 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) { 1831 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task); 1832 } 1833 } 1834 1835 static int 1836 aio_kick(struct proc *userp) 1837 { 1838 struct kaioinfo *ki = userp->p_aioinfo; 1839 struct aioproc *aiop; 1840 int error, ret = 0; 1841 1842 mtx_assert(&aio_job_mtx, MA_OWNED); 1843 retryproc: 1844 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1845 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1846 aiop->aioprocflags &= ~AIOP_FREE; 1847 wakeup(aiop->aioproc); 1848 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1849 ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) { 1850 num_aio_resv_start++; 1851 mtx_unlock(&aio_job_mtx); 1852 error = aio_newproc(&num_aio_resv_start); 1853 mtx_lock(&aio_job_mtx); 1854 if (error) { 1855 num_aio_resv_start--; 1856 goto retryproc; 1857 } 1858 } else { 1859 ret = -1; 1860 } 1861 return (ret); 1862 } 1863 1864 static void 1865 aio_kick_helper(void *context, int pending) 1866 { 1867 struct proc *userp = context; 1868 1869 mtx_lock(&aio_job_mtx); 1870 while (--pending >= 0) { 1871 if (aio_kick(userp)) 1872 break; 1873 } 1874 mtx_unlock(&aio_job_mtx); 1875 } 1876 1877 /* 1878 * Support the aio_return system call, as a side-effect, kernel resources are 1879 * released. 1880 */ 1881 static int 1882 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 1883 { 1884 struct proc *p = td->td_proc; 1885 struct kaiocb *job; 1886 struct kaioinfo *ki; 1887 long status, error; 1888 1889 ki = p->p_aioinfo; 1890 if (ki == NULL) 1891 return (EINVAL); 1892 AIO_LOCK(ki); 1893 TAILQ_FOREACH(job, &ki->kaio_done, plist) { 1894 if (job->ujob == ujob) 1895 break; 1896 } 1897 if (job != NULL) { 1898 MPASS(job->jobflags & KAIOCB_FINISHED); 1899 status = job->uaiocb._aiocb_private.status; 1900 error = job->uaiocb._aiocb_private.error; 1901 td->td_retval[0] = status; 1902 td->td_ru.ru_oublock += job->outblock; 1903 td->td_ru.ru_inblock += job->inblock; 1904 td->td_ru.ru_msgsnd += job->msgsnd; 1905 td->td_ru.ru_msgrcv += job->msgrcv; 1906 aio_free_entry(job); 1907 AIO_UNLOCK(ki); 1908 ops->store_error(ujob, error); 1909 ops->store_status(ujob, status); 1910 } else { 1911 error = EINVAL; 1912 AIO_UNLOCK(ki); 1913 } 1914 return (error); 1915 } 1916 1917 int 1918 sys_aio_return(struct thread *td, struct aio_return_args *uap) 1919 { 1920 1921 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops)); 1922 } 1923 1924 /* 1925 * Allow a process to wakeup when any of the I/O requests are completed. 1926 */ 1927 static int 1928 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist, 1929 struct timespec *ts) 1930 { 1931 struct proc *p = td->td_proc; 1932 struct timeval atv; 1933 struct kaioinfo *ki; 1934 struct kaiocb *firstjob, *job; 1935 int error, i, timo; 1936 1937 timo = 0; 1938 if (ts) { 1939 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000) 1940 return (EINVAL); 1941 1942 TIMESPEC_TO_TIMEVAL(&atv, ts); 1943 if (itimerfix(&atv)) 1944 return (EINVAL); 1945 timo = tvtohz(&atv); 1946 } 1947 1948 ki = p->p_aioinfo; 1949 if (ki == NULL) 1950 return (EAGAIN); 1951 1952 if (njoblist == 0) 1953 return (0); 1954 1955 AIO_LOCK(ki); 1956 for (;;) { 1957 firstjob = NULL; 1958 error = 0; 1959 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 1960 for (i = 0; i < njoblist; i++) { 1961 if (job->ujob == ujoblist[i]) { 1962 if (firstjob == NULL) 1963 firstjob = job; 1964 if (job->jobflags & KAIOCB_FINISHED) 1965 goto RETURN; 1966 } 1967 } 1968 } 1969 /* All tasks were finished. */ 1970 if (firstjob == NULL) 1971 break; 1972 1973 ki->kaio_flags |= KAIO_WAKEUP; 1974 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 1975 "aiospn", timo); 1976 if (error == ERESTART) 1977 error = EINTR; 1978 if (error) 1979 break; 1980 } 1981 RETURN: 1982 AIO_UNLOCK(ki); 1983 return (error); 1984 } 1985 1986 int 1987 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap) 1988 { 1989 struct timespec ts, *tsp; 1990 struct aiocb **ujoblist; 1991 int error; 1992 1993 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 1994 return (EINVAL); 1995 1996 if (uap->timeout) { 1997 /* Get timespec struct. */ 1998 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1999 return (error); 2000 tsp = &ts; 2001 } else 2002 tsp = NULL; 2003 2004 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK); 2005 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0])); 2006 if (error == 0) 2007 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2008 free(ujoblist, M_AIO); 2009 return (error); 2010 } 2011 2012 /* 2013 * aio_cancel cancels any non-bio aio operations not currently in progress. 2014 */ 2015 int 2016 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap) 2017 { 2018 struct proc *p = td->td_proc; 2019 struct kaioinfo *ki; 2020 struct kaiocb *job, *jobn; 2021 struct file *fp; 2022 int error; 2023 int cancelled = 0; 2024 int notcancelled = 0; 2025 struct vnode *vp; 2026 2027 /* Lookup file object. */ 2028 error = fget(td, uap->fd, &cap_no_rights, &fp); 2029 if (error) 2030 return (error); 2031 2032 ki = p->p_aioinfo; 2033 if (ki == NULL) 2034 goto done; 2035 2036 if (fp->f_type == DTYPE_VNODE) { 2037 vp = fp->f_vnode; 2038 if (vn_isdisk(vp)) { 2039 fdrop(fp, td); 2040 td->td_retval[0] = AIO_NOTCANCELED; 2041 return (0); 2042 } 2043 } 2044 2045 AIO_LOCK(ki); 2046 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 2047 if ((uap->fd == job->uaiocb.aio_fildes) && 2048 ((uap->aiocbp == NULL) || 2049 (uap->aiocbp == job->ujob))) { 2050 if (aio_cancel_job(p, ki, job)) { 2051 cancelled++; 2052 } else { 2053 notcancelled++; 2054 } 2055 if (uap->aiocbp != NULL) 2056 break; 2057 } 2058 } 2059 AIO_UNLOCK(ki); 2060 2061 done: 2062 fdrop(fp, td); 2063 2064 if (uap->aiocbp != NULL) { 2065 if (cancelled) { 2066 td->td_retval[0] = AIO_CANCELED; 2067 return (0); 2068 } 2069 } 2070 2071 if (notcancelled) { 2072 td->td_retval[0] = AIO_NOTCANCELED; 2073 return (0); 2074 } 2075 2076 if (cancelled) { 2077 td->td_retval[0] = AIO_CANCELED; 2078 return (0); 2079 } 2080 2081 td->td_retval[0] = AIO_ALLDONE; 2082 2083 return (0); 2084 } 2085 2086 /* 2087 * aio_error is implemented in the kernel level for compatibility purposes 2088 * only. For a user mode async implementation, it would be best to do it in 2089 * a userland subroutine. 2090 */ 2091 static int 2092 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 2093 { 2094 struct proc *p = td->td_proc; 2095 struct kaiocb *job; 2096 struct kaioinfo *ki; 2097 int status; 2098 2099 ki = p->p_aioinfo; 2100 if (ki == NULL) { 2101 td->td_retval[0] = EINVAL; 2102 return (0); 2103 } 2104 2105 AIO_LOCK(ki); 2106 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 2107 if (job->ujob == ujob) { 2108 if (job->jobflags & KAIOCB_FINISHED) 2109 td->td_retval[0] = 2110 job->uaiocb._aiocb_private.error; 2111 else 2112 td->td_retval[0] = EINPROGRESS; 2113 AIO_UNLOCK(ki); 2114 return (0); 2115 } 2116 } 2117 AIO_UNLOCK(ki); 2118 2119 /* 2120 * Hack for failure of aio_aqueue. 2121 */ 2122 status = ops->fetch_status(ujob); 2123 if (status == -1) { 2124 td->td_retval[0] = ops->fetch_error(ujob); 2125 return (0); 2126 } 2127 2128 td->td_retval[0] = EINVAL; 2129 return (0); 2130 } 2131 2132 int 2133 sys_aio_error(struct thread *td, struct aio_error_args *uap) 2134 { 2135 2136 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops)); 2137 } 2138 2139 /* syscall - asynchronous read from a file (REALTIME) */ 2140 #ifdef COMPAT_FREEBSD6 2141 int 2142 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap) 2143 { 2144 2145 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2146 &aiocb_ops_osigevent)); 2147 } 2148 #endif 2149 2150 int 2151 sys_aio_read(struct thread *td, struct aio_read_args *uap) 2152 { 2153 2154 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops)); 2155 } 2156 2157 int 2158 sys_aio_readv(struct thread *td, struct aio_readv_args *uap) 2159 { 2160 2161 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops)); 2162 } 2163 2164 /* syscall - asynchronous write to a file (REALTIME) */ 2165 #ifdef COMPAT_FREEBSD6 2166 int 2167 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap) 2168 { 2169 2170 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2171 &aiocb_ops_osigevent)); 2172 } 2173 #endif 2174 2175 int 2176 sys_aio_write(struct thread *td, struct aio_write_args *uap) 2177 { 2178 2179 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops)); 2180 } 2181 2182 int 2183 sys_aio_writev(struct thread *td, struct aio_writev_args *uap) 2184 { 2185 2186 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops)); 2187 } 2188 2189 int 2190 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap) 2191 { 2192 2193 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops)); 2194 } 2195 2196 static int 2197 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list, 2198 struct aiocb **acb_list, int nent, struct sigevent *sig, 2199 struct aiocb_ops *ops) 2200 { 2201 struct proc *p = td->td_proc; 2202 struct aiocb *job; 2203 struct kaioinfo *ki; 2204 struct aioliojob *lj; 2205 struct kevent kev; 2206 int error; 2207 int nagain, nerror; 2208 int i; 2209 2210 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT)) 2211 return (EINVAL); 2212 2213 if (nent < 0 || nent > max_aio_queue_per_proc) 2214 return (EINVAL); 2215 2216 if (p->p_aioinfo == NULL) 2217 aio_init_aioinfo(p); 2218 2219 ki = p->p_aioinfo; 2220 2221 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2222 lj->lioj_flags = 0; 2223 lj->lioj_count = 0; 2224 lj->lioj_finished_count = 0; 2225 lj->lioj_signal.sigev_notify = SIGEV_NONE; 2226 knlist_init_mtx(&lj->klist, AIO_MTX(ki)); 2227 ksiginfo_init(&lj->lioj_ksi); 2228 2229 /* 2230 * Setup signal. 2231 */ 2232 if (sig && (mode == LIO_NOWAIT)) { 2233 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal)); 2234 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2235 /* Assume only new style KEVENT */ 2236 memset(&kev, 0, sizeof(kev)); 2237 kev.filter = EVFILT_LIO; 2238 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2239 kev.ident = (uintptr_t)uacb_list; /* something unique */ 2240 kev.data = (intptr_t)lj; 2241 /* pass user defined sigval data */ 2242 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2243 error = kqfd_register( 2244 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 2245 M_WAITOK); 2246 if (error) { 2247 uma_zfree(aiolio_zone, lj); 2248 return (error); 2249 } 2250 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2251 ; 2252 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2253 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2254 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2255 uma_zfree(aiolio_zone, lj); 2256 return EINVAL; 2257 } 2258 lj->lioj_flags |= LIOJ_SIGNAL; 2259 } else { 2260 uma_zfree(aiolio_zone, lj); 2261 return EINVAL; 2262 } 2263 } 2264 2265 AIO_LOCK(ki); 2266 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2267 /* 2268 * Add extra aiocb count to avoid the lio to be freed 2269 * by other threads doing aio_waitcomplete or aio_return, 2270 * and prevent event from being sent until we have queued 2271 * all tasks. 2272 */ 2273 lj->lioj_count = 1; 2274 AIO_UNLOCK(ki); 2275 2276 /* 2277 * Get pointers to the list of I/O requests. 2278 */ 2279 nagain = 0; 2280 nerror = 0; 2281 for (i = 0; i < nent; i++) { 2282 job = acb_list[i]; 2283 if (job != NULL) { 2284 error = aio_aqueue(td, job, lj, LIO_NOP, ops); 2285 if (error == EAGAIN) 2286 nagain++; 2287 else if (error != 0) 2288 nerror++; 2289 } 2290 } 2291 2292 error = 0; 2293 AIO_LOCK(ki); 2294 if (mode == LIO_WAIT) { 2295 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2296 ki->kaio_flags |= KAIO_WAKEUP; 2297 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2298 PRIBIO | PCATCH, "aiospn", 0); 2299 if (error == ERESTART) 2300 error = EINTR; 2301 if (error) 2302 break; 2303 } 2304 } else { 2305 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2306 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2307 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2308 KNOTE_LOCKED(&lj->klist, 1); 2309 } 2310 if ((lj->lioj_flags & (LIOJ_SIGNAL | 2311 LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL && 2312 (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2313 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2314 aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi, 2315 lj->lioj_count != 1); 2316 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2317 } 2318 } 2319 } 2320 lj->lioj_count--; 2321 if (lj->lioj_count == 0) { 2322 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2323 knlist_delete(&lj->klist, curthread, 1); 2324 PROC_LOCK(p); 2325 sigqueue_take(&lj->lioj_ksi); 2326 PROC_UNLOCK(p); 2327 AIO_UNLOCK(ki); 2328 uma_zfree(aiolio_zone, lj); 2329 } else 2330 AIO_UNLOCK(ki); 2331 2332 if (nerror) 2333 return (EIO); 2334 else if (nagain) 2335 return (EAGAIN); 2336 else 2337 return (error); 2338 } 2339 2340 /* syscall - list directed I/O (REALTIME) */ 2341 #ifdef COMPAT_FREEBSD6 2342 int 2343 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap) 2344 { 2345 struct aiocb **acb_list; 2346 struct sigevent *sigp, sig; 2347 struct osigevent osig; 2348 int error, nent; 2349 2350 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2351 return (EINVAL); 2352 2353 nent = uap->nent; 2354 if (nent < 0 || nent > max_aio_queue_per_proc) 2355 return (EINVAL); 2356 2357 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2358 error = copyin(uap->sig, &osig, sizeof(osig)); 2359 if (error) 2360 return (error); 2361 error = convert_old_sigevent(&osig, &sig); 2362 if (error) 2363 return (error); 2364 sigp = &sig; 2365 } else 2366 sigp = NULL; 2367 2368 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2369 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2370 if (error == 0) 2371 error = kern_lio_listio(td, uap->mode, 2372 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2373 &aiocb_ops_osigevent); 2374 free(acb_list, M_LIO); 2375 return (error); 2376 } 2377 #endif 2378 2379 /* syscall - list directed I/O (REALTIME) */ 2380 int 2381 sys_lio_listio(struct thread *td, struct lio_listio_args *uap) 2382 { 2383 struct aiocb **acb_list; 2384 struct sigevent *sigp, sig; 2385 int error, nent; 2386 2387 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2388 return (EINVAL); 2389 2390 nent = uap->nent; 2391 if (nent < 0 || nent > max_aio_queue_per_proc) 2392 return (EINVAL); 2393 2394 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2395 error = copyin(uap->sig, &sig, sizeof(sig)); 2396 if (error) 2397 return (error); 2398 sigp = &sig; 2399 } else 2400 sigp = NULL; 2401 2402 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2403 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2404 if (error == 0) 2405 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list, 2406 nent, sigp, &aiocb_ops); 2407 free(acb_list, M_LIO); 2408 return (error); 2409 } 2410 2411 static void 2412 aio_biocleanup(struct bio *bp) 2413 { 2414 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2415 struct kaioinfo *ki; 2416 struct buf *pbuf = (struct buf *)bp->bio_caller2; 2417 2418 /* Release mapping into kernel space. */ 2419 if (pbuf != NULL) { 2420 MPASS(pbuf->b_npages <= atop(maxphys) + 1); 2421 pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages); 2422 vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages); 2423 uma_zfree(pbuf_zone, pbuf); 2424 atomic_subtract_int(&num_buf_aio, 1); 2425 ki = job->userproc->p_aioinfo; 2426 AIO_LOCK(ki); 2427 ki->kaio_buffer_count--; 2428 AIO_UNLOCK(ki); 2429 } else { 2430 MPASS(bp->bio_ma_n <= atop(maxphys) + 1); 2431 vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n); 2432 free(bp->bio_ma, M_TEMP); 2433 atomic_subtract_int(&num_unmapped_aio, 1); 2434 } 2435 g_destroy_bio(bp); 2436 } 2437 2438 static void 2439 aio_biowakeup(struct bio *bp) 2440 { 2441 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2442 size_t nbytes; 2443 long bcount = bp->bio_bcount; 2444 long resid = bp->bio_resid; 2445 int opcode, nblks; 2446 int bio_error = bp->bio_error; 2447 uint16_t flags = bp->bio_flags; 2448 2449 opcode = job->uaiocb.aio_lio_opcode; 2450 2451 aio_biocleanup(bp); 2452 2453 nbytes = bcount - resid; 2454 atomic_add_acq_long(&job->nbytes, nbytes); 2455 nblks = btodb(nbytes); 2456 2457 /* 2458 * If multiple bios experienced an error, the job will reflect the 2459 * error of whichever failed bio completed last. 2460 */ 2461 if (flags & BIO_ERROR) 2462 atomic_store_int(&job->error, bio_error); 2463 if (opcode & LIO_WRITE) 2464 atomic_add_int(&job->outblock, nblks); 2465 else 2466 atomic_add_int(&job->inblock, nblks); 2467 2468 if (refcount_release(&job->nbio)) { 2469 bio_error = atomic_load_int(&job->error); 2470 if (bio_error != 0) 2471 aio_complete(job, -1, bio_error); 2472 else 2473 aio_complete(job, atomic_load_long(&job->nbytes), 0); 2474 } 2475 } 2476 2477 /* syscall - wait for the next completion of an aio request */ 2478 static int 2479 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp, 2480 struct timespec *ts, struct aiocb_ops *ops) 2481 { 2482 struct proc *p = td->td_proc; 2483 struct timeval atv; 2484 struct kaioinfo *ki; 2485 struct kaiocb *job; 2486 struct aiocb *ujob; 2487 long error, status; 2488 int timo; 2489 2490 ops->store_aiocb(ujobp, NULL); 2491 2492 if (ts == NULL) { 2493 timo = 0; 2494 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) { 2495 timo = -1; 2496 } else { 2497 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000)) 2498 return (EINVAL); 2499 2500 TIMESPEC_TO_TIMEVAL(&atv, ts); 2501 if (itimerfix(&atv)) 2502 return (EINVAL); 2503 timo = tvtohz(&atv); 2504 } 2505 2506 if (p->p_aioinfo == NULL) 2507 aio_init_aioinfo(p); 2508 ki = p->p_aioinfo; 2509 2510 error = 0; 2511 job = NULL; 2512 AIO_LOCK(ki); 2513 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2514 if (timo == -1) { 2515 error = EWOULDBLOCK; 2516 break; 2517 } 2518 ki->kaio_flags |= KAIO_WAKEUP; 2519 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2520 "aiowc", timo); 2521 if (timo && error == ERESTART) 2522 error = EINTR; 2523 if (error) 2524 break; 2525 } 2526 2527 if (job != NULL) { 2528 MPASS(job->jobflags & KAIOCB_FINISHED); 2529 ujob = job->ujob; 2530 status = job->uaiocb._aiocb_private.status; 2531 error = job->uaiocb._aiocb_private.error; 2532 td->td_retval[0] = status; 2533 td->td_ru.ru_oublock += job->outblock; 2534 td->td_ru.ru_inblock += job->inblock; 2535 td->td_ru.ru_msgsnd += job->msgsnd; 2536 td->td_ru.ru_msgrcv += job->msgrcv; 2537 aio_free_entry(job); 2538 AIO_UNLOCK(ki); 2539 ops->store_aiocb(ujobp, ujob); 2540 ops->store_error(ujob, error); 2541 ops->store_status(ujob, status); 2542 } else 2543 AIO_UNLOCK(ki); 2544 2545 return (error); 2546 } 2547 2548 int 2549 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2550 { 2551 struct timespec ts, *tsp; 2552 int error; 2553 2554 if (uap->timeout) { 2555 /* Get timespec struct. */ 2556 error = copyin(uap->timeout, &ts, sizeof(ts)); 2557 if (error) 2558 return (error); 2559 tsp = &ts; 2560 } else 2561 tsp = NULL; 2562 2563 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops)); 2564 } 2565 2566 static int 2567 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob, 2568 struct aiocb_ops *ops) 2569 { 2570 int listop; 2571 2572 switch (op) { 2573 case O_SYNC: 2574 listop = LIO_SYNC; 2575 break; 2576 case O_DSYNC: 2577 listop = LIO_DSYNC; 2578 break; 2579 default: 2580 return (EINVAL); 2581 } 2582 2583 return (aio_aqueue(td, ujob, NULL, listop, ops)); 2584 } 2585 2586 int 2587 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2588 { 2589 2590 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops)); 2591 } 2592 2593 /* kqueue attach function */ 2594 static int 2595 filt_aioattach(struct knote *kn) 2596 { 2597 struct kaiocb *job; 2598 2599 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata; 2600 2601 /* 2602 * The job pointer must be validated before using it, so 2603 * registration is restricted to the kernel; the user cannot 2604 * set EV_FLAG1. 2605 */ 2606 if ((kn->kn_flags & EV_FLAG1) == 0) 2607 return (EPERM); 2608 kn->kn_ptr.p_aio = job; 2609 kn->kn_flags &= ~EV_FLAG1; 2610 2611 knlist_add(&job->klist, kn, 0); 2612 2613 return (0); 2614 } 2615 2616 /* kqueue detach function */ 2617 static void 2618 filt_aiodetach(struct knote *kn) 2619 { 2620 struct knlist *knl; 2621 2622 knl = &kn->kn_ptr.p_aio->klist; 2623 knl->kl_lock(knl->kl_lockarg); 2624 if (!knlist_empty(knl)) 2625 knlist_remove(knl, kn, 1); 2626 knl->kl_unlock(knl->kl_lockarg); 2627 } 2628 2629 /* kqueue filter function */ 2630 /*ARGSUSED*/ 2631 static int 2632 filt_aio(struct knote *kn, long hint) 2633 { 2634 struct kaiocb *job = kn->kn_ptr.p_aio; 2635 2636 kn->kn_data = job->uaiocb._aiocb_private.error; 2637 if (!(job->jobflags & KAIOCB_FINISHED)) 2638 return (0); 2639 kn->kn_flags |= EV_EOF; 2640 return (1); 2641 } 2642 2643 /* kqueue attach function */ 2644 static int 2645 filt_lioattach(struct knote *kn) 2646 { 2647 struct aioliojob *lj; 2648 2649 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata; 2650 2651 /* 2652 * The aioliojob pointer must be validated before using it, so 2653 * registration is restricted to the kernel; the user cannot 2654 * set EV_FLAG1. 2655 */ 2656 if ((kn->kn_flags & EV_FLAG1) == 0) 2657 return (EPERM); 2658 kn->kn_ptr.p_lio = lj; 2659 kn->kn_flags &= ~EV_FLAG1; 2660 2661 knlist_add(&lj->klist, kn, 0); 2662 2663 return (0); 2664 } 2665 2666 /* kqueue detach function */ 2667 static void 2668 filt_liodetach(struct knote *kn) 2669 { 2670 struct knlist *knl; 2671 2672 knl = &kn->kn_ptr.p_lio->klist; 2673 knl->kl_lock(knl->kl_lockarg); 2674 if (!knlist_empty(knl)) 2675 knlist_remove(knl, kn, 1); 2676 knl->kl_unlock(knl->kl_lockarg); 2677 } 2678 2679 /* kqueue filter function */ 2680 /*ARGSUSED*/ 2681 static int 2682 filt_lio(struct knote *kn, long hint) 2683 { 2684 struct aioliojob * lj = kn->kn_ptr.p_lio; 2685 2686 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2687 } 2688 2689 #ifdef COMPAT_FREEBSD32 2690 #include <sys/mount.h> 2691 #include <sys/socket.h> 2692 #include <sys/sysent.h> 2693 #include <compat/freebsd32/freebsd32.h> 2694 #include <compat/freebsd32/freebsd32_proto.h> 2695 #include <compat/freebsd32/freebsd32_signal.h> 2696 #include <compat/freebsd32/freebsd32_syscall.h> 2697 #include <compat/freebsd32/freebsd32_util.h> 2698 2699 struct __aiocb_private32 { 2700 int32_t status; 2701 int32_t error; 2702 uint32_t spare; 2703 }; 2704 2705 #ifdef COMPAT_FREEBSD6 2706 typedef struct oaiocb32 { 2707 int aio_fildes; /* File descriptor */ 2708 uint64_t aio_offset __packed; /* File offset for I/O */ 2709 uint32_t aio_buf; /* I/O buffer in process space */ 2710 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2711 struct osigevent32 aio_sigevent; /* Signal to deliver */ 2712 int aio_lio_opcode; /* LIO opcode */ 2713 int aio_reqprio; /* Request priority -- ignored */ 2714 struct __aiocb_private32 _aiocb_private; 2715 } oaiocb32_t; 2716 #endif 2717 2718 typedef struct aiocb32 { 2719 int32_t aio_fildes; /* File descriptor */ 2720 uint64_t aio_offset __packed; /* File offset for I/O */ 2721 uint32_t aio_buf; /* I/O buffer in process space */ 2722 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2723 int __spare__[2]; 2724 uint32_t __spare2__; 2725 int aio_lio_opcode; /* LIO opcode */ 2726 int aio_reqprio; /* Request priority -- ignored */ 2727 struct __aiocb_private32 _aiocb_private; 2728 struct sigevent32 aio_sigevent; /* Signal to deliver */ 2729 } aiocb32_t; 2730 2731 #ifdef COMPAT_FREEBSD6 2732 static int 2733 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig) 2734 { 2735 2736 /* 2737 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 2738 * supported by AIO with the old sigevent structure. 2739 */ 2740 CP(*osig, *nsig, sigev_notify); 2741 switch (nsig->sigev_notify) { 2742 case SIGEV_NONE: 2743 break; 2744 case SIGEV_SIGNAL: 2745 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 2746 break; 2747 case SIGEV_KEVENT: 2748 nsig->sigev_notify_kqueue = 2749 osig->__sigev_u.__sigev_notify_kqueue; 2750 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr); 2751 break; 2752 default: 2753 return (EINVAL); 2754 } 2755 return (0); 2756 } 2757 2758 static int 2759 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob, 2760 int type __unused) 2761 { 2762 struct oaiocb32 job32; 2763 struct aiocb *kcb = &kjob->uaiocb; 2764 int error; 2765 2766 bzero(kcb, sizeof(struct aiocb)); 2767 error = copyin(ujob, &job32, sizeof(job32)); 2768 if (error) 2769 return (error); 2770 2771 /* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */ 2772 2773 CP(job32, *kcb, aio_fildes); 2774 CP(job32, *kcb, aio_offset); 2775 PTRIN_CP(job32, *kcb, aio_buf); 2776 CP(job32, *kcb, aio_nbytes); 2777 CP(job32, *kcb, aio_lio_opcode); 2778 CP(job32, *kcb, aio_reqprio); 2779 CP(job32, *kcb, _aiocb_private.status); 2780 CP(job32, *kcb, _aiocb_private.error); 2781 return (convert_old_sigevent32(&job32.aio_sigevent, 2782 &kcb->aio_sigevent)); 2783 } 2784 #endif 2785 2786 static int 2787 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type) 2788 { 2789 struct aiocb32 job32; 2790 struct aiocb *kcb = &kjob->uaiocb; 2791 struct iovec32 *iov32; 2792 int error; 2793 2794 error = copyin(ujob, &job32, sizeof(job32)); 2795 if (error) 2796 return (error); 2797 CP(job32, *kcb, aio_fildes); 2798 CP(job32, *kcb, aio_offset); 2799 CP(job32, *kcb, aio_lio_opcode); 2800 if (type == LIO_NOP) 2801 type = kcb->aio_lio_opcode; 2802 if (type & LIO_VECTORED) { 2803 iov32 = PTRIN(job32.aio_iov); 2804 CP(job32, *kcb, aio_iovcnt); 2805 /* malloc a uio and copy in the iovec */ 2806 error = freebsd32_copyinuio(iov32, 2807 kcb->aio_iovcnt, &kjob->uiop); 2808 if (error) 2809 return (error); 2810 } else { 2811 PTRIN_CP(job32, *kcb, aio_buf); 2812 CP(job32, *kcb, aio_nbytes); 2813 } 2814 CP(job32, *kcb, aio_reqprio); 2815 CP(job32, *kcb, _aiocb_private.status); 2816 CP(job32, *kcb, _aiocb_private.error); 2817 error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent); 2818 2819 return (error); 2820 } 2821 2822 static long 2823 aiocb32_fetch_status(struct aiocb *ujob) 2824 { 2825 struct aiocb32 *ujob32; 2826 2827 ujob32 = (struct aiocb32 *)ujob; 2828 return (fuword32(&ujob32->_aiocb_private.status)); 2829 } 2830 2831 static long 2832 aiocb32_fetch_error(struct aiocb *ujob) 2833 { 2834 struct aiocb32 *ujob32; 2835 2836 ujob32 = (struct aiocb32 *)ujob; 2837 return (fuword32(&ujob32->_aiocb_private.error)); 2838 } 2839 2840 static int 2841 aiocb32_store_status(struct aiocb *ujob, long status) 2842 { 2843 struct aiocb32 *ujob32; 2844 2845 ujob32 = (struct aiocb32 *)ujob; 2846 return (suword32(&ujob32->_aiocb_private.status, status)); 2847 } 2848 2849 static int 2850 aiocb32_store_error(struct aiocb *ujob, long error) 2851 { 2852 struct aiocb32 *ujob32; 2853 2854 ujob32 = (struct aiocb32 *)ujob; 2855 return (suword32(&ujob32->_aiocb_private.error, error)); 2856 } 2857 2858 static int 2859 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 2860 { 2861 2862 return (suword32(ujobp, (long)ujob)); 2863 } 2864 2865 static struct aiocb_ops aiocb32_ops = { 2866 .aio_copyin = aiocb32_copyin, 2867 .fetch_status = aiocb32_fetch_status, 2868 .fetch_error = aiocb32_fetch_error, 2869 .store_status = aiocb32_store_status, 2870 .store_error = aiocb32_store_error, 2871 .store_aiocb = aiocb32_store_aiocb, 2872 }; 2873 2874 #ifdef COMPAT_FREEBSD6 2875 static struct aiocb_ops aiocb32_ops_osigevent = { 2876 .aio_copyin = aiocb32_copyin_old_sigevent, 2877 .fetch_status = aiocb32_fetch_status, 2878 .fetch_error = aiocb32_fetch_error, 2879 .store_status = aiocb32_store_status, 2880 .store_error = aiocb32_store_error, 2881 .store_aiocb = aiocb32_store_aiocb, 2882 }; 2883 #endif 2884 2885 int 2886 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap) 2887 { 2888 2889 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2890 } 2891 2892 int 2893 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap) 2894 { 2895 struct timespec32 ts32; 2896 struct timespec ts, *tsp; 2897 struct aiocb **ujoblist; 2898 uint32_t *ujoblist32; 2899 int error, i; 2900 2901 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 2902 return (EINVAL); 2903 2904 if (uap->timeout) { 2905 /* Get timespec struct. */ 2906 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0) 2907 return (error); 2908 CP(ts32, ts, tv_sec); 2909 CP(ts32, ts, tv_nsec); 2910 tsp = &ts; 2911 } else 2912 tsp = NULL; 2913 2914 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK); 2915 ujoblist32 = (uint32_t *)ujoblist; 2916 error = copyin(uap->aiocbp, ujoblist32, uap->nent * 2917 sizeof(ujoblist32[0])); 2918 if (error == 0) { 2919 for (i = uap->nent - 1; i >= 0; i--) 2920 ujoblist[i] = PTRIN(ujoblist32[i]); 2921 2922 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2923 } 2924 free(ujoblist, M_AIO); 2925 return (error); 2926 } 2927 2928 int 2929 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap) 2930 { 2931 2932 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2933 } 2934 2935 #ifdef COMPAT_FREEBSD6 2936 int 2937 freebsd6_freebsd32_aio_read(struct thread *td, 2938 struct freebsd6_freebsd32_aio_read_args *uap) 2939 { 2940 2941 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2942 &aiocb32_ops_osigevent)); 2943 } 2944 #endif 2945 2946 int 2947 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap) 2948 { 2949 2950 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2951 &aiocb32_ops)); 2952 } 2953 2954 int 2955 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap) 2956 { 2957 2958 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV, 2959 &aiocb32_ops)); 2960 } 2961 2962 #ifdef COMPAT_FREEBSD6 2963 int 2964 freebsd6_freebsd32_aio_write(struct thread *td, 2965 struct freebsd6_freebsd32_aio_write_args *uap) 2966 { 2967 2968 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2969 &aiocb32_ops_osigevent)); 2970 } 2971 #endif 2972 2973 int 2974 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap) 2975 { 2976 2977 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2978 &aiocb32_ops)); 2979 } 2980 2981 int 2982 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap) 2983 { 2984 2985 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV, 2986 &aiocb32_ops)); 2987 } 2988 2989 int 2990 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap) 2991 { 2992 2993 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK, 2994 &aiocb32_ops)); 2995 } 2996 2997 int 2998 freebsd32_aio_waitcomplete(struct thread *td, 2999 struct freebsd32_aio_waitcomplete_args *uap) 3000 { 3001 struct timespec32 ts32; 3002 struct timespec ts, *tsp; 3003 int error; 3004 3005 if (uap->timeout) { 3006 /* Get timespec struct. */ 3007 error = copyin(uap->timeout, &ts32, sizeof(ts32)); 3008 if (error) 3009 return (error); 3010 CP(ts32, ts, tv_sec); 3011 CP(ts32, ts, tv_nsec); 3012 tsp = &ts; 3013 } else 3014 tsp = NULL; 3015 3016 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp, 3017 &aiocb32_ops)); 3018 } 3019 3020 int 3021 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap) 3022 { 3023 3024 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp, 3025 &aiocb32_ops)); 3026 } 3027 3028 #ifdef COMPAT_FREEBSD6 3029 int 3030 freebsd6_freebsd32_lio_listio(struct thread *td, 3031 struct freebsd6_freebsd32_lio_listio_args *uap) 3032 { 3033 struct aiocb **acb_list; 3034 struct sigevent *sigp, sig; 3035 struct osigevent32 osig; 3036 uint32_t *acb_list32; 3037 int error, i, nent; 3038 3039 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3040 return (EINVAL); 3041 3042 nent = uap->nent; 3043 if (nent < 0 || nent > max_aio_queue_per_proc) 3044 return (EINVAL); 3045 3046 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3047 error = copyin(uap->sig, &osig, sizeof(osig)); 3048 if (error) 3049 return (error); 3050 error = convert_old_sigevent32(&osig, &sig); 3051 if (error) 3052 return (error); 3053 sigp = &sig; 3054 } else 3055 sigp = NULL; 3056 3057 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3058 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3059 if (error) { 3060 free(acb_list32, M_LIO); 3061 return (error); 3062 } 3063 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3064 for (i = 0; i < nent; i++) 3065 acb_list[i] = PTRIN(acb_list32[i]); 3066 free(acb_list32, M_LIO); 3067 3068 error = kern_lio_listio(td, uap->mode, 3069 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3070 &aiocb32_ops_osigevent); 3071 free(acb_list, M_LIO); 3072 return (error); 3073 } 3074 #endif 3075 3076 int 3077 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap) 3078 { 3079 struct aiocb **acb_list; 3080 struct sigevent *sigp, sig; 3081 struct sigevent32 sig32; 3082 uint32_t *acb_list32; 3083 int error, i, nent; 3084 3085 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3086 return (EINVAL); 3087 3088 nent = uap->nent; 3089 if (nent < 0 || nent > max_aio_queue_per_proc) 3090 return (EINVAL); 3091 3092 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3093 error = copyin(uap->sig, &sig32, sizeof(sig32)); 3094 if (error) 3095 return (error); 3096 error = convert_sigevent32(&sig32, &sig); 3097 if (error) 3098 return (error); 3099 sigp = &sig; 3100 } else 3101 sigp = NULL; 3102 3103 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3104 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3105 if (error) { 3106 free(acb_list32, M_LIO); 3107 return (error); 3108 } 3109 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3110 for (i = 0; i < nent; i++) 3111 acb_list[i] = PTRIN(acb_list32[i]); 3112 free(acb_list32, M_LIO); 3113 3114 error = kern_lio_listio(td, uap->mode, 3115 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3116 &aiocb32_ops); 3117 free(acb_list, M_LIO); 3118 return (error); 3119 } 3120 3121 #endif 3122