1 /*- 2 * Copyright (c) 1997 John S. Dyson. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. John S. Dyson's name may not be used to endorse or promote products 10 * derived from this software without specific prior written permission. 11 * 12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything 13 * bad that happens because of using this software isn't the responsibility 14 * of the author. This software is distributed AS-IS. 15 */ 16 17 /* 18 * This file contains support for the POSIX 1003.1B AIO/LIO facility. 19 */ 20 21 #include <sys/cdefs.h> 22 __FBSDID("$FreeBSD$"); 23 24 #include <sys/param.h> 25 #include <sys/systm.h> 26 #include <sys/malloc.h> 27 #include <sys/bio.h> 28 #include <sys/buf.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/proc.h> 42 #include <sys/resourcevar.h> 43 #include <sys/signalvar.h> 44 #include <sys/protosw.h> 45 #include <sys/sema.h> 46 #include <sys/socket.h> 47 #include <sys/socketvar.h> 48 #include <sys/syscall.h> 49 #include <sys/sysent.h> 50 #include <sys/sysctl.h> 51 #include <sys/sx.h> 52 #include <sys/taskqueue.h> 53 #include <sys/vnode.h> 54 #include <sys/conf.h> 55 #include <sys/event.h> 56 #include <sys/mount.h> 57 58 #include <machine/atomic.h> 59 60 #include <posix4/posix4.h> 61 #include <vm/vm.h> 62 #include <vm/vm_extern.h> 63 #include <vm/pmap.h> 64 #include <vm/vm_map.h> 65 #include <vm/vm_object.h> 66 #include <vm/uma.h> 67 #include <sys/aio.h> 68 69 #include "opt_vfs_aio.h" 70 71 /* 72 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 73 * overflow. (XXX will be removed soon.) 74 */ 75 static u_long jobrefid; 76 77 /* 78 * Counter for aio_fsync. 79 */ 80 static uint64_t jobseqno; 81 82 #define JOBST_NULL 0 83 #define JOBST_JOBQSOCK 1 84 #define JOBST_JOBQGLOBAL 2 85 #define JOBST_JOBRUNNING 3 86 #define JOBST_JOBFINISHED 4 87 #define JOBST_JOBQBUF 5 88 #define JOBST_JOBQSYNC 6 89 90 #ifndef MAX_AIO_PER_PROC 91 #define MAX_AIO_PER_PROC 32 92 #endif 93 94 #ifndef MAX_AIO_QUEUE_PER_PROC 95 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */ 96 #endif 97 98 #ifndef MAX_AIO_PROCS 99 #define MAX_AIO_PROCS 32 100 #endif 101 102 #ifndef MAX_AIO_QUEUE 103 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */ 104 #endif 105 106 #ifndef TARGET_AIO_PROCS 107 #define TARGET_AIO_PROCS 4 108 #endif 109 110 #ifndef MAX_BUF_AIO 111 #define MAX_BUF_AIO 16 112 #endif 113 114 #ifndef AIOD_TIMEOUT_DEFAULT 115 #define AIOD_TIMEOUT_DEFAULT (10 * hz) 116 #endif 117 118 #ifndef AIOD_LIFETIME_DEFAULT 119 #define AIOD_LIFETIME_DEFAULT (30 * hz) 120 #endif 121 122 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management"); 123 124 static int max_aio_procs = MAX_AIO_PROCS; 125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, 126 CTLFLAG_RW, &max_aio_procs, 0, 127 "Maximum number of kernel threads to use for handling async IO "); 128 129 static int num_aio_procs = 0; 130 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, 131 CTLFLAG_RD, &num_aio_procs, 0, 132 "Number of presently active kernel threads for async IO"); 133 134 /* 135 * The code will adjust the actual number of AIO processes towards this 136 * number when it gets a chance. 137 */ 138 static int target_aio_procs = TARGET_AIO_PROCS; 139 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 140 0, "Preferred number of ready kernel threads for async IO"); 141 142 static int max_queue_count = MAX_AIO_QUEUE; 143 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 144 "Maximum number of aio requests to queue, globally"); 145 146 static int num_queue_count = 0; 147 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 148 "Number of queued aio requests"); 149 150 static int num_buf_aio = 0; 151 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 152 "Number of aio requests presently handled by the buf subsystem"); 153 154 /* Number of async I/O thread in the process of being started */ 155 /* XXX This should be local to aio_aqueue() */ 156 static int num_aio_resv_start = 0; 157 158 static int aiod_timeout; 159 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0, 160 "Timeout value for synchronous aio operations"); 161 162 static int aiod_lifetime; 163 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 164 "Maximum lifetime for idle aiod"); 165 166 static int unloadable = 0; 167 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0, 168 "Allow unload of aio (not recommended)"); 169 170 171 static int max_aio_per_proc = MAX_AIO_PER_PROC; 172 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 173 0, "Maximum active aio requests per process (stored in the process)"); 174 175 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 176 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 177 &max_aio_queue_per_proc, 0, 178 "Maximum queued aio requests per process (stored in the process)"); 179 180 static int max_buf_aio = MAX_BUF_AIO; 181 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 182 "Maximum buf aio requests per process (stored in the process)"); 183 184 typedef struct oaiocb { 185 int aio_fildes; /* File descriptor */ 186 off_t aio_offset; /* File offset for I/O */ 187 volatile void *aio_buf; /* I/O buffer in process space */ 188 size_t aio_nbytes; /* Number of bytes for I/O */ 189 struct osigevent aio_sigevent; /* Signal to deliver */ 190 int aio_lio_opcode; /* LIO opcode */ 191 int aio_reqprio; /* Request priority -- ignored */ 192 struct __aiocb_private _aiocb_private; 193 } oaiocb_t; 194 195 /* 196 * Below is a key of locks used to protect each member of struct aiocblist 197 * aioliojob and kaioinfo and any backends. 198 * 199 * * - need not protected 200 * a - locked by kaioinfo lock 201 * b - locked by backend lock, the backend lock can be null in some cases, 202 * for example, BIO belongs to this type, in this case, proc lock is 203 * reused. 204 * c - locked by aio_job_mtx, the lock for the generic file I/O backend. 205 */ 206 207 /* 208 * Current, there is only two backends: BIO and generic file I/O. 209 * socket I/O is served by generic file I/O, this is not a good idea, since 210 * disk file I/O and any other types without O_NONBLOCK flag can block daemon 211 * threads, if there is no thread to serve socket I/O, the socket I/O will be 212 * delayed too long or starved, we should create some threads dedicated to 213 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O 214 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file 215 * structure is not safe because there is race between userland and aio 216 * daemons. 217 */ 218 219 struct aiocblist { 220 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */ 221 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */ 222 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */ 223 int jobflags; /* (a) job flags */ 224 int jobstate; /* (b) job state */ 225 int inputcharge; /* (*) input blockes */ 226 int outputcharge; /* (*) output blockes */ 227 struct buf *bp; /* (*) private to BIO backend, 228 * buffer pointer 229 */ 230 struct proc *userproc; /* (*) user process */ 231 struct ucred *cred; /* (*) active credential when created */ 232 struct file *fd_file; /* (*) pointer to file structure */ 233 struct aioliojob *lio; /* (*) optional lio job */ 234 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */ 235 struct knlist klist; /* (a) list of knotes */ 236 struct aiocb uaiocb; /* (*) kernel I/O control block */ 237 ksiginfo_t ksi; /* (a) realtime signal info */ 238 struct task biotask; /* (*) private to BIO backend */ 239 uint64_t seqno; /* (*) job number */ 240 int pending; /* (a) number of pending I/O, aio_fsync only */ 241 }; 242 243 /* jobflags */ 244 #define AIOCBLIST_DONE 0x01 245 #define AIOCBLIST_BUFDONE 0x02 246 #define AIOCBLIST_RUNDOWN 0x04 247 #define AIOCBLIST_CHECKSYNC 0x08 248 249 /* 250 * AIO process info 251 */ 252 #define AIOP_FREE 0x1 /* proc on free queue */ 253 254 struct aiothreadlist { 255 int aiothreadflags; /* (c) AIO proc flags */ 256 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */ 257 struct thread *aiothread; /* (*) the AIO thread */ 258 }; 259 260 /* 261 * data-structure for lio signal management 262 */ 263 struct aioliojob { 264 int lioj_flags; /* (a) listio flags */ 265 int lioj_count; /* (a) listio flags */ 266 int lioj_finished_count; /* (a) listio flags */ 267 struct sigevent lioj_signal; /* (a) signal on all I/O done */ 268 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */ 269 struct knlist klist; /* (a) list of knotes */ 270 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */ 271 }; 272 273 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 274 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 275 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */ 276 277 /* 278 * per process aio data structure 279 */ 280 struct kaioinfo { 281 struct mtx kaio_mtx; /* the lock to protect this struct */ 282 int kaio_flags; /* (a) per process kaio flags */ 283 int kaio_maxactive_count; /* (*) maximum number of AIOs */ 284 int kaio_active_count; /* (c) number of currently used AIOs */ 285 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */ 286 int kaio_count; /* (a) size of AIO queue */ 287 int kaio_ballowed_count; /* (*) maximum number of buffers */ 288 int kaio_buffer_count; /* (a) number of physio buffers */ 289 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */ 290 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */ 291 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */ 292 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */ 293 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */ 294 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets, 295 * NOT USED YET. 296 */ 297 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */ 298 struct task kaio_task; /* (*) task to kick aio threads */ 299 }; 300 301 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx) 302 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx) 303 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f)) 304 #define AIO_MTX(ki) (&(ki)->kaio_mtx) 305 306 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 307 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */ 308 309 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */ 310 static struct sema aio_newproc_sem; 311 static struct mtx aio_job_mtx; 312 static struct mtx aio_sock_mtx; 313 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */ 314 static struct unrhdr *aiod_unr; 315 316 static void aio_init_aioinfo(struct proc *p); 317 static void aio_onceonly(void); 318 static int aio_free_entry(struct aiocblist *aiocbe); 319 static void aio_process(struct aiocblist *aiocbe); 320 static int aio_newproc(int *); 321 static int aio_aqueue(struct thread *td, struct aiocb *job, 322 struct aioliojob *lio, int type, int osigev); 323 static void aio_physwakeup(struct buf *bp); 324 static void aio_proc_rundown(void *arg, struct proc *p); 325 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp); 326 static int aio_qphysio(struct proc *p, struct aiocblist *iocb); 327 static void biohelper(void *, int); 328 static void aio_daemon(void *param); 329 static void aio_swake_cb(struct socket *, struct sockbuf *); 330 static int aio_unload(void); 331 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type); 332 #define DONE_BUF 1 333 #define DONE_QUEUE 2 334 static int do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev); 335 static int aio_kick(struct proc *userp); 336 static void aio_kick_nowait(struct proc *userp); 337 static void aio_kick_helper(void *context, int pending); 338 static int filt_aioattach(struct knote *kn); 339 static void filt_aiodetach(struct knote *kn); 340 static int filt_aio(struct knote *kn, long hint); 341 static int filt_lioattach(struct knote *kn); 342 static void filt_liodetach(struct knote *kn); 343 static int filt_lio(struct knote *kn, long hint); 344 345 /* 346 * Zones for: 347 * kaio Per process async io info 348 * aiop async io thread data 349 * aiocb async io jobs 350 * aiol list io job pointer - internal to aio_suspend XXX 351 * aiolio list io jobs 352 */ 353 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone; 354 355 /* kqueue filters for aio */ 356 static struct filterops aio_filtops = 357 { 0, filt_aioattach, filt_aiodetach, filt_aio }; 358 static struct filterops lio_filtops = 359 { 0, filt_lioattach, filt_liodetach, filt_lio }; 360 361 static eventhandler_tag exit_tag, exec_tag; 362 363 TASKQUEUE_DEFINE_THREAD(aiod_bio); 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_UNLOAD: 378 error = aio_unload(); 379 break; 380 case MOD_SHUTDOWN: 381 break; 382 default: 383 error = EINVAL; 384 break; 385 } 386 return (error); 387 } 388 389 static moduledata_t aio_mod = { 390 "aio", 391 &aio_modload, 392 NULL 393 }; 394 395 SYSCALL_MODULE_HELPER(aio_cancel); 396 SYSCALL_MODULE_HELPER(aio_error); 397 SYSCALL_MODULE_HELPER(aio_fsync); 398 SYSCALL_MODULE_HELPER(aio_read); 399 SYSCALL_MODULE_HELPER(aio_return); 400 SYSCALL_MODULE_HELPER(aio_suspend); 401 SYSCALL_MODULE_HELPER(aio_waitcomplete); 402 SYSCALL_MODULE_HELPER(aio_write); 403 SYSCALL_MODULE_HELPER(lio_listio); 404 SYSCALL_MODULE_HELPER(oaio_read); 405 SYSCALL_MODULE_HELPER(oaio_write); 406 SYSCALL_MODULE_HELPER(olio_listio); 407 408 DECLARE_MODULE(aio, aio_mod, 409 SI_SUB_VFS, SI_ORDER_ANY); 410 MODULE_VERSION(aio, 1); 411 412 /* 413 * Startup initialization 414 */ 415 static void 416 aio_onceonly(void) 417 { 418 419 /* XXX: should probably just use so->callback */ 420 aio_swake = &aio_swake_cb; 421 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 422 EVENTHANDLER_PRI_ANY); 423 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL, 424 EVENTHANDLER_PRI_ANY); 425 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 426 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 427 TAILQ_INIT(&aio_freeproc); 428 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 429 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 430 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF); 431 TAILQ_INIT(&aio_jobs); 432 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 433 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 434 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 435 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL, 436 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 437 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL, 438 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 439 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL, 440 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 441 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 442 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 443 aiod_timeout = AIOD_TIMEOUT_DEFAULT; 444 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 445 jobrefid = 1; 446 async_io_version = _POSIX_VERSION; 447 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX); 448 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 449 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 450 } 451 452 /* 453 * Callback for unload of AIO when used as a module. 454 */ 455 static int 456 aio_unload(void) 457 { 458 int error; 459 460 /* 461 * XXX: no unloads by default, it's too dangerous. 462 * perhaps we could do it if locked out callers and then 463 * did an aio_proc_rundown() on each process. 464 * 465 * jhb: aio_proc_rundown() needs to run on curproc though, 466 * so I don't think that would fly. 467 */ 468 if (!unloadable) 469 return (EOPNOTSUPP); 470 471 error = kqueue_del_filteropts(EVFILT_AIO); 472 if (error) 473 return error; 474 error = kqueue_del_filteropts(EVFILT_LIO); 475 if (error) 476 return error; 477 async_io_version = 0; 478 aio_swake = NULL; 479 taskqueue_free(taskqueue_aiod_bio); 480 delete_unrhdr(aiod_unr); 481 uma_zdestroy(kaio_zone); 482 uma_zdestroy(aiop_zone); 483 uma_zdestroy(aiocb_zone); 484 uma_zdestroy(aiol_zone); 485 uma_zdestroy(aiolio_zone); 486 EVENTHANDLER_DEREGISTER(process_exit, exit_tag); 487 EVENTHANDLER_DEREGISTER(process_exec, exec_tag); 488 mtx_destroy(&aio_job_mtx); 489 mtx_destroy(&aio_sock_mtx); 490 sema_destroy(&aio_newproc_sem); 491 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1); 492 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1); 493 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1); 494 return (0); 495 } 496 497 /* 498 * Init the per-process aioinfo structure. The aioinfo limits are set 499 * per-process for user limit (resource) management. 500 */ 501 static void 502 aio_init_aioinfo(struct proc *p) 503 { 504 struct kaioinfo *ki; 505 506 ki = uma_zalloc(kaio_zone, M_WAITOK); 507 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF); 508 ki->kaio_flags = 0; 509 ki->kaio_maxactive_count = max_aio_per_proc; 510 ki->kaio_active_count = 0; 511 ki->kaio_qallowed_count = max_aio_queue_per_proc; 512 ki->kaio_count = 0; 513 ki->kaio_ballowed_count = max_buf_aio; 514 ki->kaio_buffer_count = 0; 515 TAILQ_INIT(&ki->kaio_all); 516 TAILQ_INIT(&ki->kaio_done); 517 TAILQ_INIT(&ki->kaio_jobqueue); 518 TAILQ_INIT(&ki->kaio_bufqueue); 519 TAILQ_INIT(&ki->kaio_liojoblist); 520 TAILQ_INIT(&ki->kaio_sockqueue); 521 TAILQ_INIT(&ki->kaio_syncqueue); 522 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 523 PROC_LOCK(p); 524 if (p->p_aioinfo == NULL) { 525 p->p_aioinfo = ki; 526 PROC_UNLOCK(p); 527 } else { 528 PROC_UNLOCK(p); 529 mtx_destroy(&ki->kaio_mtx); 530 uma_zfree(kaio_zone, ki); 531 } 532 533 while (num_aio_procs < target_aio_procs) 534 aio_newproc(NULL); 535 } 536 537 static int 538 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi) 539 { 540 int ret = 0; 541 542 PROC_LOCK(p); 543 if (!KSI_ONQ(ksi)) { 544 ksi->ksi_code = SI_ASYNCIO; 545 ksi->ksi_flags |= KSI_EXT | KSI_INS; 546 ret = psignal_event(p, sigev, ksi); 547 } 548 PROC_UNLOCK(p); 549 return (ret); 550 } 551 552 /* 553 * Free a job entry. Wait for completion if it is currently active, but don't 554 * delay forever. If we delay, we return a flag that says that we have to 555 * restart the queue scan. 556 */ 557 static int 558 aio_free_entry(struct aiocblist *aiocbe) 559 { 560 struct kaioinfo *ki; 561 struct aioliojob *lj; 562 struct proc *p; 563 564 p = aiocbe->userproc; 565 MPASS(curproc == p); 566 ki = p->p_aioinfo; 567 MPASS(ki != NULL); 568 569 AIO_LOCK_ASSERT(ki, MA_OWNED); 570 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED); 571 572 atomic_subtract_int(&num_queue_count, 1); 573 574 ki->kaio_count--; 575 MPASS(ki->kaio_count >= 0); 576 577 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist); 578 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist); 579 580 lj = aiocbe->lio; 581 if (lj) { 582 lj->lioj_count--; 583 lj->lioj_finished_count--; 584 585 if (lj->lioj_count == 0) { 586 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 587 /* lio is going away, we need to destroy any knotes */ 588 knlist_delete(&lj->klist, curthread, 1); 589 PROC_LOCK(p); 590 sigqueue_take(&lj->lioj_ksi); 591 PROC_UNLOCK(p); 592 uma_zfree(aiolio_zone, lj); 593 } 594 } 595 596 /* aiocbe is going away, we need to destroy any knotes */ 597 knlist_delete(&aiocbe->klist, curthread, 1); 598 PROC_LOCK(p); 599 sigqueue_take(&aiocbe->ksi); 600 PROC_UNLOCK(p); 601 602 MPASS(aiocbe->bp == NULL); 603 aiocbe->jobstate = JOBST_NULL; 604 AIO_UNLOCK(ki); 605 606 /* 607 * The thread argument here is used to find the owning process 608 * and is also passed to fo_close() which may pass it to various 609 * places such as devsw close() routines. Because of that, we 610 * need a thread pointer from the process owning the job that is 611 * persistent and won't disappear out from under us or move to 612 * another process. 613 * 614 * Currently, all the callers of this function call it to remove 615 * an aiocblist from the current process' job list either via a 616 * syscall or due to the current process calling exit() or 617 * execve(). Thus, we know that p == curproc. We also know that 618 * curthread can't exit since we are curthread. 619 * 620 * Therefore, we use curthread as the thread to pass to 621 * knlist_delete(). This does mean that it is possible for the 622 * thread pointer at close time to differ from the thread pointer 623 * at open time, but this is already true of file descriptors in 624 * a multithreaded process. 625 */ 626 fdrop(aiocbe->fd_file, curthread); 627 crfree(aiocbe->cred); 628 uma_zfree(aiocb_zone, aiocbe); 629 AIO_LOCK(ki); 630 631 return (0); 632 } 633 634 static void 635 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused) 636 { 637 aio_proc_rundown(arg, p); 638 } 639 640 /* 641 * Rundown the jobs for a given process. 642 */ 643 static void 644 aio_proc_rundown(void *arg, struct proc *p) 645 { 646 struct kaioinfo *ki; 647 struct aioliojob *lj; 648 struct aiocblist *cbe, *cbn; 649 struct file *fp; 650 struct socket *so; 651 int remove; 652 653 KASSERT(curthread->td_proc == p, 654 ("%s: called on non-curproc", __func__)); 655 ki = p->p_aioinfo; 656 if (ki == NULL) 657 return; 658 659 AIO_LOCK(ki); 660 ki->kaio_flags |= KAIO_RUNDOWN; 661 662 restart: 663 664 /* 665 * Try to cancel all pending requests. This code simulates 666 * aio_cancel on all pending I/O requests. 667 */ 668 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) { 669 remove = 0; 670 mtx_lock(&aio_job_mtx); 671 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 672 TAILQ_REMOVE(&aio_jobs, cbe, list); 673 remove = 1; 674 } else if (cbe->jobstate == JOBST_JOBQSOCK) { 675 fp = cbe->fd_file; 676 MPASS(fp->f_type == DTYPE_SOCKET); 677 so = fp->f_data; 678 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 679 remove = 1; 680 } else if (cbe->jobstate == JOBST_JOBQSYNC) { 681 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list); 682 remove = 1; 683 } 684 mtx_unlock(&aio_job_mtx); 685 686 if (remove) { 687 cbe->jobstate = JOBST_JOBFINISHED; 688 cbe->uaiocb._aiocb_private.status = -1; 689 cbe->uaiocb._aiocb_private.error = ECANCELED; 690 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 691 aio_bio_done_notify(p, cbe, DONE_QUEUE); 692 } 693 } 694 695 /* Wait for all running I/O to be finished */ 696 if (TAILQ_FIRST(&ki->kaio_bufqueue) || 697 TAILQ_FIRST(&ki->kaio_jobqueue)) { 698 ki->kaio_flags |= KAIO_WAKEUP; 699 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 700 goto restart; 701 } 702 703 /* Free all completed I/O requests. */ 704 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL) 705 aio_free_entry(cbe); 706 707 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 708 if (lj->lioj_count == 0) { 709 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 710 knlist_delete(&lj->klist, curthread, 1); 711 PROC_LOCK(p); 712 sigqueue_take(&lj->lioj_ksi); 713 PROC_UNLOCK(p); 714 uma_zfree(aiolio_zone, lj); 715 } else { 716 panic("LIO job not cleaned up: C:%d, FC:%d\n", 717 lj->lioj_count, lj->lioj_finished_count); 718 } 719 } 720 AIO_UNLOCK(ki); 721 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task); 722 uma_zfree(kaio_zone, ki); 723 p->p_aioinfo = NULL; 724 } 725 726 /* 727 * Select a job to run (called by an AIO daemon). 728 */ 729 static struct aiocblist * 730 aio_selectjob(struct aiothreadlist *aiop) 731 { 732 struct aiocblist *aiocbe; 733 struct kaioinfo *ki; 734 struct proc *userp; 735 736 mtx_assert(&aio_job_mtx, MA_OWNED); 737 TAILQ_FOREACH(aiocbe, &aio_jobs, list) { 738 userp = aiocbe->userproc; 739 ki = userp->p_aioinfo; 740 741 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 742 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 743 /* Account for currently active jobs. */ 744 ki->kaio_active_count++; 745 aiocbe->jobstate = JOBST_JOBRUNNING; 746 break; 747 } 748 } 749 return (aiocbe); 750 } 751 752 /* 753 * Move all data to a permanent storage device, this code 754 * simulates fsync syscall. 755 */ 756 static int 757 aio_fsync_vnode(struct thread *td, struct vnode *vp) 758 { 759 struct mount *mp; 760 int vfslocked; 761 int error; 762 763 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 764 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 765 goto drop; 766 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 767 if (vp->v_object != NULL) { 768 VM_OBJECT_LOCK(vp->v_object); 769 vm_object_page_clean(vp->v_object, 0, 0, 0); 770 VM_OBJECT_UNLOCK(vp->v_object); 771 } 772 error = VOP_FSYNC(vp, MNT_WAIT, td); 773 774 VOP_UNLOCK(vp, 0, td); 775 vn_finished_write(mp); 776 drop: 777 VFS_UNLOCK_GIANT(vfslocked); 778 return (error); 779 } 780 781 /* 782 * The AIO processing activity. This is the code that does the I/O request for 783 * the non-physio version of the operations. The normal vn operations are used, 784 * and this code should work in all instances for every type of file, including 785 * pipes, sockets, fifos, and regular files. 786 * 787 * XXX I don't think it works well for socket, pipe, and fifo. 788 */ 789 static void 790 aio_process(struct aiocblist *aiocbe) 791 { 792 struct ucred *td_savedcred; 793 struct thread *td; 794 struct proc *mycp; 795 struct aiocb *cb; 796 struct file *fp; 797 struct socket *so; 798 struct uio auio; 799 struct iovec aiov; 800 int cnt; 801 int error; 802 int oublock_st, oublock_end; 803 int inblock_st, inblock_end; 804 805 td = curthread; 806 td_savedcred = td->td_ucred; 807 td->td_ucred = aiocbe->cred; 808 mycp = td->td_proc; 809 cb = &aiocbe->uaiocb; 810 fp = aiocbe->fd_file; 811 812 if (cb->aio_lio_opcode == LIO_SYNC) { 813 error = 0; 814 cnt = 0; 815 if (fp->f_vnode != NULL) 816 error = aio_fsync_vnode(td, fp->f_vnode); 817 cb->_aiocb_private.error = error; 818 cb->_aiocb_private.status = 0; 819 td->td_ucred = td_savedcred; 820 return; 821 } 822 823 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 824 aiov.iov_len = cb->aio_nbytes; 825 826 auio.uio_iov = &aiov; 827 auio.uio_iovcnt = 1; 828 auio.uio_offset = cb->aio_offset; 829 auio.uio_resid = cb->aio_nbytes; 830 cnt = cb->aio_nbytes; 831 auio.uio_segflg = UIO_USERSPACE; 832 auio.uio_td = td; 833 834 inblock_st = mycp->p_stats->p_ru.ru_inblock; 835 oublock_st = mycp->p_stats->p_ru.ru_oublock; 836 /* 837 * aio_aqueue() acquires a reference to the file that is 838 * released in aio_free_entry(). 839 */ 840 if (cb->aio_lio_opcode == LIO_READ) { 841 auio.uio_rw = UIO_READ; 842 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td); 843 } else { 844 if (fp->f_type == DTYPE_VNODE) 845 bwillwrite(); 846 auio.uio_rw = UIO_WRITE; 847 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td); 848 } 849 inblock_end = mycp->p_stats->p_ru.ru_inblock; 850 oublock_end = mycp->p_stats->p_ru.ru_oublock; 851 852 aiocbe->inputcharge = inblock_end - inblock_st; 853 aiocbe->outputcharge = oublock_end - oublock_st; 854 855 if ((error) && (auio.uio_resid != cnt)) { 856 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 857 error = 0; 858 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) { 859 int sigpipe = 1; 860 if (fp->f_type == DTYPE_SOCKET) { 861 so = fp->f_data; 862 if (so->so_options & SO_NOSIGPIPE) 863 sigpipe = 0; 864 } 865 if (sigpipe) { 866 PROC_LOCK(aiocbe->userproc); 867 psignal(aiocbe->userproc, SIGPIPE); 868 PROC_UNLOCK(aiocbe->userproc); 869 } 870 } 871 } 872 873 cnt -= auio.uio_resid; 874 cb->_aiocb_private.error = error; 875 cb->_aiocb_private.status = cnt; 876 td->td_ucred = td_savedcred; 877 } 878 879 static void 880 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type) 881 { 882 struct aioliojob *lj; 883 struct kaioinfo *ki; 884 struct aiocblist *scb, *scbn; 885 int lj_done; 886 887 ki = userp->p_aioinfo; 888 AIO_LOCK_ASSERT(ki, MA_OWNED); 889 lj = aiocbe->lio; 890 lj_done = 0; 891 if (lj) { 892 lj->lioj_finished_count++; 893 if (lj->lioj_count == lj->lioj_finished_count) 894 lj_done = 1; 895 } 896 if (type == DONE_QUEUE) { 897 aiocbe->jobflags |= AIOCBLIST_DONE; 898 } else { 899 aiocbe->jobflags |= AIOCBLIST_BUFDONE; 900 } 901 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist); 902 aiocbe->jobstate = JOBST_JOBFINISHED; 903 904 if (ki->kaio_flags & KAIO_RUNDOWN) 905 goto notification_done; 906 907 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 908 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 909 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi); 910 911 KNOTE_LOCKED(&aiocbe->klist, 1); 912 913 if (lj_done) { 914 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 915 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 916 KNOTE_LOCKED(&lj->klist, 1); 917 } 918 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 919 == LIOJ_SIGNAL 920 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 921 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 922 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi); 923 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 924 } 925 } 926 927 notification_done: 928 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) { 929 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) { 930 if (aiocbe->fd_file == scb->fd_file && 931 aiocbe->seqno < scb->seqno) { 932 if (--scb->pending == 0) { 933 mtx_lock(&aio_job_mtx); 934 scb->jobstate = JOBST_JOBQGLOBAL; 935 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list); 936 TAILQ_INSERT_TAIL(&aio_jobs, scb, list); 937 aio_kick_nowait(userp); 938 mtx_unlock(&aio_job_mtx); 939 } 940 } 941 } 942 } 943 if (ki->kaio_flags & KAIO_WAKEUP) { 944 ki->kaio_flags &= ~KAIO_WAKEUP; 945 wakeup(&userp->p_aioinfo); 946 } 947 } 948 949 /* 950 * The AIO daemon, most of the actual work is done in aio_process, 951 * but the setup (and address space mgmt) is done in this routine. 952 */ 953 static void 954 aio_daemon(void *_id) 955 { 956 struct aiocblist *aiocbe; 957 struct aiothreadlist *aiop; 958 struct kaioinfo *ki; 959 struct proc *curcp, *mycp, *userp; 960 struct vmspace *myvm, *tmpvm; 961 struct thread *td = curthread; 962 int id = (intptr_t)_id; 963 964 /* 965 * Local copies of curproc (cp) and vmspace (myvm) 966 */ 967 mycp = td->td_proc; 968 myvm = mycp->p_vmspace; 969 970 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp")); 971 972 /* 973 * Allocate and ready the aio control info. There is one aiop structure 974 * per daemon. 975 */ 976 aiop = uma_zalloc(aiop_zone, M_WAITOK); 977 aiop->aiothread = td; 978 aiop->aiothreadflags = 0; 979 980 /* The daemon resides in its own pgrp. */ 981 setsid(td, NULL); 982 983 /* 984 * Wakeup parent process. (Parent sleeps to keep from blasting away 985 * and creating too many daemons.) 986 */ 987 sema_post(&aio_newproc_sem); 988 989 mtx_lock(&aio_job_mtx); 990 for (;;) { 991 /* 992 * curcp is the current daemon process context. 993 * userp is the current user process context. 994 */ 995 curcp = mycp; 996 997 /* 998 * Take daemon off of free queue 999 */ 1000 if (aiop->aiothreadflags & AIOP_FREE) { 1001 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1002 aiop->aiothreadflags &= ~AIOP_FREE; 1003 } 1004 1005 /* 1006 * Check for jobs. 1007 */ 1008 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 1009 mtx_unlock(&aio_job_mtx); 1010 userp = aiocbe->userproc; 1011 1012 /* 1013 * Connect to process address space for user program. 1014 */ 1015 if (userp != curcp) { 1016 /* 1017 * Save the current address space that we are 1018 * connected to. 1019 */ 1020 tmpvm = mycp->p_vmspace; 1021 1022 /* 1023 * Point to the new user address space, and 1024 * refer to it. 1025 */ 1026 mycp->p_vmspace = userp->p_vmspace; 1027 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1); 1028 1029 /* Activate the new mapping. */ 1030 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 1031 1032 /* 1033 * If the old address space wasn't the daemons 1034 * own address space, then we need to remove the 1035 * daemon's reference from the other process 1036 * that it was acting on behalf of. 1037 */ 1038 if (tmpvm != myvm) { 1039 vmspace_free(tmpvm); 1040 } 1041 curcp = userp; 1042 } 1043 1044 ki = userp->p_aioinfo; 1045 1046 /* Do the I/O function. */ 1047 aio_process(aiocbe); 1048 1049 mtx_lock(&aio_job_mtx); 1050 /* Decrement the active job count. */ 1051 ki->kaio_active_count--; 1052 mtx_unlock(&aio_job_mtx); 1053 1054 AIO_LOCK(ki); 1055 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 1056 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE); 1057 AIO_UNLOCK(ki); 1058 1059 mtx_lock(&aio_job_mtx); 1060 } 1061 1062 /* 1063 * Disconnect from user address space. 1064 */ 1065 if (curcp != mycp) { 1066 1067 mtx_unlock(&aio_job_mtx); 1068 1069 /* Get the user address space to disconnect from. */ 1070 tmpvm = mycp->p_vmspace; 1071 1072 /* Get original address space for daemon. */ 1073 mycp->p_vmspace = myvm; 1074 1075 /* Activate the daemon's address space. */ 1076 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 1077 #ifdef DIAGNOSTIC 1078 if (tmpvm == myvm) { 1079 printf("AIOD: vmspace problem -- %d\n", 1080 mycp->p_pid); 1081 } 1082 #endif 1083 /* Remove our vmspace reference. */ 1084 vmspace_free(tmpvm); 1085 1086 curcp = mycp; 1087 1088 mtx_lock(&aio_job_mtx); 1089 /* 1090 * We have to restart to avoid race, we only sleep if 1091 * no job can be selected, that should be 1092 * curcp == mycp. 1093 */ 1094 continue; 1095 } 1096 1097 mtx_assert(&aio_job_mtx, MA_OWNED); 1098 1099 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1100 aiop->aiothreadflags |= AIOP_FREE; 1101 1102 /* 1103 * If daemon is inactive for a long time, allow it to exit, 1104 * thereby freeing resources. 1105 */ 1106 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy", 1107 aiod_lifetime)) { 1108 if (TAILQ_EMPTY(&aio_jobs)) { 1109 if ((aiop->aiothreadflags & AIOP_FREE) && 1110 (num_aio_procs > target_aio_procs)) { 1111 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1112 num_aio_procs--; 1113 mtx_unlock(&aio_job_mtx); 1114 uma_zfree(aiop_zone, aiop); 1115 free_unr(aiod_unr, id); 1116 #ifdef DIAGNOSTIC 1117 if (mycp->p_vmspace->vm_refcnt <= 1) { 1118 printf("AIOD: bad vm refcnt for" 1119 " exiting daemon: %d\n", 1120 mycp->p_vmspace->vm_refcnt); 1121 } 1122 #endif 1123 kthread_exit(0); 1124 } 1125 } 1126 } 1127 } 1128 mtx_unlock(&aio_job_mtx); 1129 panic("shouldn't be here\n"); 1130 } 1131 1132 /* 1133 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1134 * AIO daemon modifies its environment itself. 1135 */ 1136 static int 1137 aio_newproc(int *start) 1138 { 1139 int error; 1140 struct proc *p; 1141 int id; 1142 1143 id = alloc_unr(aiod_unr); 1144 error = kthread_create(aio_daemon, (void *)(intptr_t)id, &p, 1145 RFNOWAIT, 0, "aiod%d", id); 1146 if (error == 0) { 1147 /* 1148 * Wait until daemon is started. 1149 */ 1150 sema_wait(&aio_newproc_sem); 1151 mtx_lock(&aio_job_mtx); 1152 num_aio_procs++; 1153 if (start != NULL) 1154 (*start)--; 1155 mtx_unlock(&aio_job_mtx); 1156 } else { 1157 free_unr(aiod_unr, id); 1158 } 1159 return (error); 1160 } 1161 1162 /* 1163 * Try the high-performance, low-overhead physio method for eligible 1164 * VCHR devices. This method doesn't use an aio helper thread, and 1165 * thus has very low overhead. 1166 * 1167 * Assumes that the caller, aio_aqueue(), has incremented the file 1168 * structure's reference count, preventing its deallocation for the 1169 * duration of this call. 1170 */ 1171 static int 1172 aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 1173 { 1174 struct aiocb *cb; 1175 struct file *fp; 1176 struct buf *bp; 1177 struct vnode *vp; 1178 struct kaioinfo *ki; 1179 struct aioliojob *lj; 1180 int error; 1181 1182 cb = &aiocbe->uaiocb; 1183 fp = aiocbe->fd_file; 1184 1185 if (fp->f_type != DTYPE_VNODE) 1186 return (-1); 1187 1188 vp = fp->f_vnode; 1189 1190 /* 1191 * If its not a disk, we don't want to return a positive error. 1192 * It causes the aio code to not fall through to try the thread 1193 * way when you're talking to a regular file. 1194 */ 1195 if (!vn_isdisk(vp, &error)) { 1196 if (error == ENOTBLK) 1197 return (-1); 1198 else 1199 return (error); 1200 } 1201 1202 if (vp->v_bufobj.bo_bsize == 0) 1203 return (-1); 1204 1205 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize) 1206 return (-1); 1207 1208 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max) 1209 return (-1); 1210 1211 if (cb->aio_nbytes > 1212 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 1213 return (-1); 1214 1215 ki = p->p_aioinfo; 1216 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 1217 return (-1); 1218 1219 /* Create and build a buffer header for a transfer. */ 1220 bp = (struct buf *)getpbuf(NULL); 1221 BUF_KERNPROC(bp); 1222 1223 AIO_LOCK(ki); 1224 ki->kaio_count++; 1225 ki->kaio_buffer_count++; 1226 lj = aiocbe->lio; 1227 if (lj) 1228 lj->lioj_count++; 1229 AIO_UNLOCK(ki); 1230 1231 /* 1232 * Get a copy of the kva from the physical buffer. 1233 */ 1234 error = 0; 1235 1236 bp->b_bcount = cb->aio_nbytes; 1237 bp->b_bufsize = cb->aio_nbytes; 1238 bp->b_iodone = aio_physwakeup; 1239 bp->b_saveaddr = bp->b_data; 1240 bp->b_data = (void *)(uintptr_t)cb->aio_buf; 1241 bp->b_offset = cb->aio_offset; 1242 bp->b_iooffset = cb->aio_offset; 1243 bp->b_blkno = btodb(cb->aio_offset); 1244 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ; 1245 1246 /* 1247 * Bring buffer into kernel space. 1248 */ 1249 if (vmapbuf(bp) < 0) { 1250 error = EFAULT; 1251 goto doerror; 1252 } 1253 1254 AIO_LOCK(ki); 1255 aiocbe->bp = bp; 1256 bp->b_caller1 = (void *)aiocbe; 1257 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 1258 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1259 aiocbe->jobstate = JOBST_JOBQBUF; 1260 cb->_aiocb_private.status = cb->aio_nbytes; 1261 AIO_UNLOCK(ki); 1262 1263 atomic_add_int(&num_queue_count, 1); 1264 atomic_add_int(&num_buf_aio, 1); 1265 1266 bp->b_error = 0; 1267 1268 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe); 1269 1270 /* Perform transfer. */ 1271 dev_strategy(vp->v_rdev, bp); 1272 return (0); 1273 1274 doerror: 1275 AIO_LOCK(ki); 1276 ki->kaio_count--; 1277 ki->kaio_buffer_count--; 1278 if (lj) 1279 lj->lioj_count--; 1280 aiocbe->bp = NULL; 1281 AIO_UNLOCK(ki); 1282 relpbuf(bp, NULL); 1283 return (error); 1284 } 1285 1286 /* 1287 * Wake up aio requests that may be serviceable now. 1288 */ 1289 static void 1290 aio_swake_cb(struct socket *so, struct sockbuf *sb) 1291 { 1292 struct aiocblist *cb, *cbn; 1293 int opcode; 1294 1295 if (sb == &so->so_snd) 1296 opcode = LIO_WRITE; 1297 else 1298 opcode = LIO_READ; 1299 1300 SOCKBUF_LOCK(sb); 1301 sb->sb_flags &= ~SB_AIO; 1302 mtx_lock(&aio_job_mtx); 1303 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) { 1304 if (opcode == cb->uaiocb.aio_lio_opcode) { 1305 if (cb->jobstate != JOBST_JOBQSOCK) 1306 panic("invalid queue value"); 1307 /* XXX 1308 * We don't have actual sockets backend yet, 1309 * so we simply move the requests to the generic 1310 * file I/O backend. 1311 */ 1312 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1313 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1314 aio_kick_nowait(cb->userproc); 1315 } 1316 } 1317 mtx_unlock(&aio_job_mtx); 1318 SOCKBUF_UNLOCK(sb); 1319 } 1320 1321 /* 1322 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1323 * technique is done in this code. 1324 */ 1325 static int 1326 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj, 1327 int type, int oldsigev) 1328 { 1329 struct proc *p = td->td_proc; 1330 struct file *fp; 1331 struct socket *so; 1332 struct aiocblist *aiocbe, *cb; 1333 struct kaioinfo *ki; 1334 struct kevent kev; 1335 struct kqueue *kq; 1336 struct file *kq_fp; 1337 struct sockbuf *sb; 1338 int opcode; 1339 int error; 1340 int fd; 1341 int jid; 1342 1343 if (p->p_aioinfo == NULL) 1344 aio_init_aioinfo(p); 1345 1346 ki = p->p_aioinfo; 1347 1348 suword(&job->_aiocb_private.status, -1); 1349 suword(&job->_aiocb_private.error, 0); 1350 suword(&job->_aiocb_private.kernelinfo, -1); 1351 1352 if (num_queue_count >= max_queue_count || 1353 ki->kaio_count >= ki->kaio_qallowed_count) { 1354 suword(&job->_aiocb_private.error, EAGAIN); 1355 return (EAGAIN); 1356 } 1357 1358 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1359 aiocbe->inputcharge = 0; 1360 aiocbe->outputcharge = 0; 1361 knlist_init(&aiocbe->klist, AIO_MTX(ki), NULL, NULL, NULL); 1362 1363 if (oldsigev) { 1364 bzero(&aiocbe->uaiocb, sizeof(struct aiocb)); 1365 error = copyin(job, &aiocbe->uaiocb, sizeof(struct oaiocb)); 1366 bcopy(&aiocbe->uaiocb.__spare__, &aiocbe->uaiocb.aio_sigevent, 1367 sizeof(struct osigevent)); 1368 } else { 1369 error = copyin(job, &aiocbe->uaiocb, sizeof(struct aiocb)); 1370 } 1371 if (error) { 1372 suword(&job->_aiocb_private.error, error); 1373 uma_zfree(aiocb_zone, aiocbe); 1374 return (error); 1375 } 1376 1377 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1378 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1379 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1380 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1381 suword(&job->_aiocb_private.error, EINVAL); 1382 uma_zfree(aiocb_zone, aiocbe); 1383 return (EINVAL); 1384 } 1385 1386 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1387 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1388 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1389 uma_zfree(aiocb_zone, aiocbe); 1390 return (EINVAL); 1391 } 1392 1393 ksiginfo_init(&aiocbe->ksi); 1394 1395 /* Save userspace address of the job info. */ 1396 aiocbe->uuaiocb = job; 1397 1398 /* Get the opcode. */ 1399 if (type != LIO_NOP) 1400 aiocbe->uaiocb.aio_lio_opcode = type; 1401 opcode = aiocbe->uaiocb.aio_lio_opcode; 1402 1403 /* Fetch the file object for the specified file descriptor. */ 1404 fd = aiocbe->uaiocb.aio_fildes; 1405 switch (opcode) { 1406 case LIO_WRITE: 1407 error = fget_write(td, fd, &fp); 1408 break; 1409 case LIO_READ: 1410 error = fget_read(td, fd, &fp); 1411 break; 1412 default: 1413 error = fget(td, fd, &fp); 1414 } 1415 if (error) { 1416 uma_zfree(aiocb_zone, aiocbe); 1417 suword(&job->_aiocb_private.error, error); 1418 return (error); 1419 } 1420 1421 if (opcode == LIO_SYNC && fp->f_vnode == NULL) { 1422 error = EINVAL; 1423 goto aqueue_fail; 1424 } 1425 1426 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) { 1427 error = EINVAL; 1428 goto aqueue_fail; 1429 } 1430 1431 aiocbe->fd_file = fp; 1432 1433 mtx_lock(&aio_job_mtx); 1434 jid = jobrefid++; 1435 aiocbe->seqno = jobseqno++; 1436 mtx_unlock(&aio_job_mtx); 1437 error = suword(&job->_aiocb_private.kernelinfo, jid); 1438 if (error) { 1439 error = EINVAL; 1440 goto aqueue_fail; 1441 } 1442 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid; 1443 1444 if (opcode == LIO_NOP) { 1445 fdrop(fp, td); 1446 uma_zfree(aiocb_zone, aiocbe); 1447 return (0); 1448 } 1449 if ((opcode != LIO_READ) && (opcode != LIO_WRITE) && 1450 (opcode != LIO_SYNC)) { 1451 error = EINVAL; 1452 goto aqueue_fail; 1453 } 1454 1455 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) 1456 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1457 else 1458 goto no_kqueue; 1459 error = fget(td, (u_int)kev.ident, &kq_fp); 1460 if (error) 1461 goto aqueue_fail; 1462 if (kq_fp->f_type != DTYPE_KQUEUE) { 1463 fdrop(kq_fp, td); 1464 error = EBADF; 1465 goto aqueue_fail; 1466 } 1467 kq = kq_fp->f_data; 1468 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1469 kev.filter = EVFILT_AIO; 1470 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 1471 kev.data = (intptr_t)aiocbe; 1472 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1473 error = kqueue_register(kq, &kev, td, 1); 1474 fdrop(kq_fp, td); 1475 aqueue_fail: 1476 if (error) { 1477 fdrop(fp, td); 1478 uma_zfree(aiocb_zone, aiocbe); 1479 suword(&job->_aiocb_private.error, error); 1480 goto done; 1481 } 1482 no_kqueue: 1483 1484 suword(&job->_aiocb_private.error, EINPROGRESS); 1485 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1486 aiocbe->userproc = p; 1487 aiocbe->cred = crhold(td->td_ucred); 1488 aiocbe->jobflags = 0; 1489 aiocbe->lio = lj; 1490 1491 if (opcode == LIO_SYNC) 1492 goto queueit; 1493 1494 if (fp->f_type == DTYPE_SOCKET) { 1495 /* 1496 * Alternate queueing for socket ops: Reach down into the 1497 * descriptor to get the socket data. Then check to see if the 1498 * socket is ready to be read or written (based on the requested 1499 * operation). 1500 * 1501 * If it is not ready for io, then queue the aiocbe on the 1502 * socket, and set the flags so we get a call when sbnotify() 1503 * happens. 1504 * 1505 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock 1506 * and unlock the snd sockbuf for no reason. 1507 */ 1508 so = fp->f_data; 1509 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd; 1510 SOCKBUF_LOCK(sb); 1511 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1512 LIO_WRITE) && (!sowriteable(so)))) { 1513 sb->sb_flags |= SB_AIO; 1514 1515 mtx_lock(&aio_job_mtx); 1516 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1517 mtx_unlock(&aio_job_mtx); 1518 1519 AIO_LOCK(ki); 1520 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1521 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1522 aiocbe->jobstate = JOBST_JOBQSOCK; 1523 ki->kaio_count++; 1524 if (lj) 1525 lj->lioj_count++; 1526 AIO_UNLOCK(ki); 1527 SOCKBUF_UNLOCK(sb); 1528 atomic_add_int(&num_queue_count, 1); 1529 error = 0; 1530 goto done; 1531 } 1532 SOCKBUF_UNLOCK(sb); 1533 } 1534 1535 if ((error = aio_qphysio(p, aiocbe)) == 0) 1536 goto done; 1537 #if 0 1538 if (error > 0) { 1539 aiocbe->uaiocb._aiocb_private.error = error; 1540 suword(&job->_aiocb_private.error, error); 1541 goto done; 1542 } 1543 #endif 1544 queueit: 1545 /* No buffer for daemon I/O. */ 1546 aiocbe->bp = NULL; 1547 atomic_add_int(&num_queue_count, 1); 1548 1549 AIO_LOCK(ki); 1550 ki->kaio_count++; 1551 if (lj) 1552 lj->lioj_count++; 1553 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1554 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1555 if (opcode == LIO_SYNC) { 1556 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) { 1557 if (cb->fd_file == aiocbe->fd_file && 1558 cb->uaiocb.aio_lio_opcode != LIO_SYNC && 1559 cb->seqno < aiocbe->seqno) { 1560 cb->jobflags |= AIOCBLIST_CHECKSYNC; 1561 aiocbe->pending++; 1562 } 1563 } 1564 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) { 1565 if (cb->fd_file == aiocbe->fd_file && 1566 cb->uaiocb.aio_lio_opcode != LIO_SYNC && 1567 cb->seqno < aiocbe->seqno) { 1568 cb->jobflags |= AIOCBLIST_CHECKSYNC; 1569 aiocbe->pending++; 1570 } 1571 } 1572 if (aiocbe->pending != 0) { 1573 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list); 1574 aiocbe->jobstate = JOBST_JOBQSYNC; 1575 AIO_UNLOCK(ki); 1576 goto done; 1577 } 1578 } 1579 mtx_lock(&aio_job_mtx); 1580 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1581 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1582 aio_kick_nowait(p); 1583 mtx_unlock(&aio_job_mtx); 1584 AIO_UNLOCK(ki); 1585 error = 0; 1586 done: 1587 return (error); 1588 } 1589 1590 static void 1591 aio_kick_nowait(struct proc *userp) 1592 { 1593 struct kaioinfo *ki = userp->p_aioinfo; 1594 struct aiothreadlist *aiop; 1595 1596 mtx_assert(&aio_job_mtx, MA_OWNED); 1597 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1598 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1599 aiop->aiothreadflags &= ~AIOP_FREE; 1600 wakeup(aiop->aiothread); 1601 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1602 ((ki->kaio_active_count + num_aio_resv_start) < 1603 ki->kaio_maxactive_count)) { 1604 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task); 1605 } 1606 } 1607 1608 static int 1609 aio_kick(struct proc *userp) 1610 { 1611 struct kaioinfo *ki = userp->p_aioinfo; 1612 struct aiothreadlist *aiop; 1613 int error, ret = 0; 1614 1615 mtx_assert(&aio_job_mtx, MA_OWNED); 1616 retryproc: 1617 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1618 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1619 aiop->aiothreadflags &= ~AIOP_FREE; 1620 wakeup(aiop->aiothread); 1621 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1622 ((ki->kaio_active_count + num_aio_resv_start) < 1623 ki->kaio_maxactive_count)) { 1624 num_aio_resv_start++; 1625 mtx_unlock(&aio_job_mtx); 1626 error = aio_newproc(&num_aio_resv_start); 1627 mtx_lock(&aio_job_mtx); 1628 if (error) { 1629 num_aio_resv_start--; 1630 goto retryproc; 1631 } 1632 } else { 1633 ret = -1; 1634 } 1635 return (ret); 1636 } 1637 1638 static void 1639 aio_kick_helper(void *context, int pending) 1640 { 1641 struct proc *userp = context; 1642 1643 mtx_lock(&aio_job_mtx); 1644 while (--pending >= 0) { 1645 if (aio_kick(userp)) 1646 break; 1647 } 1648 mtx_unlock(&aio_job_mtx); 1649 } 1650 1651 /* 1652 * Support the aio_return system call, as a side-effect, kernel resources are 1653 * released. 1654 */ 1655 int 1656 aio_return(struct thread *td, struct aio_return_args *uap) 1657 { 1658 struct proc *p = td->td_proc; 1659 struct aiocblist *cb; 1660 struct aiocb *uaiocb; 1661 struct kaioinfo *ki; 1662 int status, error; 1663 1664 ki = p->p_aioinfo; 1665 if (ki == NULL) 1666 return (EINVAL); 1667 uaiocb = uap->aiocbp; 1668 AIO_LOCK(ki); 1669 TAILQ_FOREACH(cb, &ki->kaio_done, plist) { 1670 if (cb->uuaiocb == uaiocb) 1671 break; 1672 } 1673 if (cb != NULL) { 1674 MPASS(cb->jobstate == JOBST_JOBFINISHED); 1675 status = cb->uaiocb._aiocb_private.status; 1676 error = cb->uaiocb._aiocb_private.error; 1677 td->td_retval[0] = status; 1678 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1679 p->p_stats->p_ru.ru_oublock += 1680 cb->outputcharge; 1681 cb->outputcharge = 0; 1682 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1683 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 1684 cb->inputcharge = 0; 1685 } 1686 aio_free_entry(cb); 1687 AIO_UNLOCK(ki); 1688 suword(&uaiocb->_aiocb_private.error, error); 1689 suword(&uaiocb->_aiocb_private.status, status); 1690 } else { 1691 error = EINVAL; 1692 AIO_UNLOCK(ki); 1693 } 1694 return (error); 1695 } 1696 1697 /* 1698 * Allow a process to wakeup when any of the I/O requests are completed. 1699 */ 1700 int 1701 aio_suspend(struct thread *td, struct aio_suspend_args *uap) 1702 { 1703 struct proc *p = td->td_proc; 1704 struct timeval atv; 1705 struct timespec ts; 1706 struct aiocb *const *cbptr, *cbp; 1707 struct kaioinfo *ki; 1708 struct aiocblist *cb, *cbfirst; 1709 struct aiocb **ujoblist; 1710 int njoblist; 1711 int error; 1712 int timo; 1713 int i; 1714 1715 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX) 1716 return (EINVAL); 1717 1718 timo = 0; 1719 if (uap->timeout) { 1720 /* Get timespec struct. */ 1721 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1722 return (error); 1723 1724 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) 1725 return (EINVAL); 1726 1727 TIMESPEC_TO_TIMEVAL(&atv, &ts); 1728 if (itimerfix(&atv)) 1729 return (EINVAL); 1730 timo = tvtohz(&atv); 1731 } 1732 1733 ki = p->p_aioinfo; 1734 if (ki == NULL) 1735 return (EAGAIN); 1736 1737 njoblist = 0; 1738 ujoblist = uma_zalloc(aiol_zone, M_WAITOK); 1739 cbptr = uap->aiocbp; 1740 1741 for (i = 0; i < uap->nent; i++) { 1742 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1743 if (cbp == 0) 1744 continue; 1745 ujoblist[njoblist] = cbp; 1746 njoblist++; 1747 } 1748 1749 if (njoblist == 0) { 1750 uma_zfree(aiol_zone, ujoblist); 1751 return (0); 1752 } 1753 1754 AIO_LOCK(ki); 1755 for (;;) { 1756 cbfirst = NULL; 1757 error = 0; 1758 TAILQ_FOREACH(cb, &ki->kaio_all, allist) { 1759 for (i = 0; i < njoblist; i++) { 1760 if (cb->uuaiocb == ujoblist[i]) { 1761 if (cbfirst == NULL) 1762 cbfirst = cb; 1763 if (cb->jobstate == JOBST_JOBFINISHED) 1764 goto RETURN; 1765 } 1766 } 1767 } 1768 /* All tasks were finished. */ 1769 if (cbfirst == NULL) 1770 break; 1771 1772 ki->kaio_flags |= KAIO_WAKEUP; 1773 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 1774 "aiospn", timo); 1775 if (error == ERESTART) 1776 error = EINTR; 1777 if (error) 1778 break; 1779 } 1780 RETURN: 1781 AIO_UNLOCK(ki); 1782 uma_zfree(aiol_zone, ujoblist); 1783 return (error); 1784 } 1785 1786 /* 1787 * aio_cancel cancels any non-physio aio operations not currently in 1788 * progress. 1789 */ 1790 int 1791 aio_cancel(struct thread *td, struct aio_cancel_args *uap) 1792 { 1793 struct proc *p = td->td_proc; 1794 struct kaioinfo *ki; 1795 struct aiocblist *cbe, *cbn; 1796 struct file *fp; 1797 struct socket *so; 1798 int error; 1799 int remove; 1800 int cancelled = 0; 1801 int notcancelled = 0; 1802 struct vnode *vp; 1803 1804 /* Lookup file object. */ 1805 error = fget(td, uap->fd, &fp); 1806 if (error) 1807 return (error); 1808 1809 ki = p->p_aioinfo; 1810 if (ki == NULL) 1811 goto done; 1812 1813 if (fp->f_type == DTYPE_VNODE) { 1814 vp = fp->f_vnode; 1815 if (vn_isdisk(vp, &error)) { 1816 fdrop(fp, td); 1817 td->td_retval[0] = AIO_NOTCANCELED; 1818 return (0); 1819 } 1820 } 1821 1822 AIO_LOCK(ki); 1823 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) { 1824 if ((uap->fd == cbe->uaiocb.aio_fildes) && 1825 ((uap->aiocbp == NULL) || 1826 (uap->aiocbp == cbe->uuaiocb))) { 1827 remove = 0; 1828 1829 mtx_lock(&aio_job_mtx); 1830 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 1831 TAILQ_REMOVE(&aio_jobs, cbe, list); 1832 remove = 1; 1833 } else if (cbe->jobstate == JOBST_JOBQSOCK) { 1834 MPASS(fp->f_type == DTYPE_SOCKET); 1835 so = fp->f_data; 1836 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 1837 remove = 1; 1838 } else if (cbe->jobstate == JOBST_JOBQSYNC) { 1839 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list); 1840 remove = 1; 1841 } 1842 mtx_unlock(&aio_job_mtx); 1843 1844 if (remove) { 1845 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 1846 cbe->uaiocb._aiocb_private.status = -1; 1847 cbe->uaiocb._aiocb_private.error = ECANCELED; 1848 aio_bio_done_notify(p, cbe, DONE_QUEUE); 1849 cancelled++; 1850 } else { 1851 notcancelled++; 1852 } 1853 if (uap->aiocbp != NULL) 1854 break; 1855 } 1856 } 1857 AIO_UNLOCK(ki); 1858 1859 done: 1860 fdrop(fp, td); 1861 1862 if (uap->aiocbp != NULL) { 1863 if (cancelled) { 1864 td->td_retval[0] = AIO_CANCELED; 1865 return (0); 1866 } 1867 } 1868 1869 if (notcancelled) { 1870 td->td_retval[0] = AIO_NOTCANCELED; 1871 return (0); 1872 } 1873 1874 if (cancelled) { 1875 td->td_retval[0] = AIO_CANCELED; 1876 return (0); 1877 } 1878 1879 td->td_retval[0] = AIO_ALLDONE; 1880 1881 return (0); 1882 } 1883 1884 /* 1885 * aio_error is implemented in the kernel level for compatibility purposes only. 1886 * For a user mode async implementation, it would be best to do it in a userland 1887 * subroutine. 1888 */ 1889 int 1890 aio_error(struct thread *td, struct aio_error_args *uap) 1891 { 1892 struct proc *p = td->td_proc; 1893 struct aiocblist *cb; 1894 struct kaioinfo *ki; 1895 int status; 1896 1897 ki = p->p_aioinfo; 1898 if (ki == NULL) { 1899 td->td_retval[0] = EINVAL; 1900 return (0); 1901 } 1902 1903 AIO_LOCK(ki); 1904 TAILQ_FOREACH(cb, &ki->kaio_all, allist) { 1905 if (cb->uuaiocb == uap->aiocbp) { 1906 if (cb->jobstate == JOBST_JOBFINISHED) 1907 td->td_retval[0] = 1908 cb->uaiocb._aiocb_private.error; 1909 else 1910 td->td_retval[0] = EINPROGRESS; 1911 AIO_UNLOCK(ki); 1912 return (0); 1913 } 1914 } 1915 AIO_UNLOCK(ki); 1916 1917 /* 1918 * Hack for failure of aio_aqueue. 1919 */ 1920 status = fuword(&uap->aiocbp->_aiocb_private.status); 1921 if (status == -1) { 1922 td->td_retval[0] = fuword(&uap->aiocbp->_aiocb_private.error); 1923 return (0); 1924 } 1925 1926 td->td_retval[0] = EINVAL; 1927 return (0); 1928 } 1929 1930 /* syscall - asynchronous read from a file (REALTIME) */ 1931 int 1932 oaio_read(struct thread *td, struct oaio_read_args *uap) 1933 { 1934 1935 return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 1); 1936 } 1937 1938 int 1939 aio_read(struct thread *td, struct aio_read_args *uap) 1940 { 1941 1942 return aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, 0); 1943 } 1944 1945 /* syscall - asynchronous write to a file (REALTIME) */ 1946 int 1947 oaio_write(struct thread *td, struct oaio_write_args *uap) 1948 { 1949 1950 return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 1); 1951 } 1952 1953 int 1954 aio_write(struct thread *td, struct aio_write_args *uap) 1955 { 1956 1957 return aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, 0); 1958 } 1959 1960 /* syscall - list directed I/O (REALTIME) */ 1961 int 1962 olio_listio(struct thread *td, struct olio_listio_args *uap) 1963 { 1964 return do_lio_listio(td, (struct lio_listio_args *)uap, 1); 1965 } 1966 1967 /* syscall - list directed I/O (REALTIME) */ 1968 int 1969 lio_listio(struct thread *td, struct lio_listio_args *uap) 1970 { 1971 return do_lio_listio(td, uap, 0); 1972 } 1973 1974 static int 1975 do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev) 1976 { 1977 struct proc *p = td->td_proc; 1978 struct aiocb *iocb, * const *cbptr; 1979 struct kaioinfo *ki; 1980 struct aioliojob *lj; 1981 struct kevent kev; 1982 struct kqueue * kq; 1983 struct file *kq_fp; 1984 int nent; 1985 int error; 1986 int nerror; 1987 int i; 1988 1989 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 1990 return (EINVAL); 1991 1992 nent = uap->nent; 1993 if (nent < 0 || nent > AIO_LISTIO_MAX) 1994 return (EINVAL); 1995 1996 if (p->p_aioinfo == NULL) 1997 aio_init_aioinfo(p); 1998 1999 ki = p->p_aioinfo; 2000 2001 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2002 lj->lioj_flags = 0; 2003 lj->lioj_count = 0; 2004 lj->lioj_finished_count = 0; 2005 knlist_init(&lj->klist, AIO_MTX(ki), NULL, NULL, NULL); 2006 ksiginfo_init(&lj->lioj_ksi); 2007 2008 /* 2009 * Setup signal. 2010 */ 2011 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2012 bzero(&lj->lioj_signal, sizeof(&lj->lioj_signal)); 2013 error = copyin(uap->sig, &lj->lioj_signal, 2014 oldsigev ? sizeof(struct osigevent) : 2015 sizeof(struct sigevent)); 2016 if (error) { 2017 uma_zfree(aiolio_zone, lj); 2018 return (error); 2019 } 2020 2021 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2022 /* Assume only new style KEVENT */ 2023 error = fget(td, lj->lioj_signal.sigev_notify_kqueue, 2024 &kq_fp); 2025 if (error) { 2026 uma_zfree(aiolio_zone, lj); 2027 return (error); 2028 } 2029 if (kq_fp->f_type != DTYPE_KQUEUE) { 2030 fdrop(kq_fp, td); 2031 uma_zfree(aiolio_zone, lj); 2032 return (EBADF); 2033 } 2034 kq = (struct kqueue *)kq_fp->f_data; 2035 kev.filter = EVFILT_LIO; 2036 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2037 kev.ident = (uintptr_t)uap->acb_list; /* something unique */ 2038 kev.data = (intptr_t)lj; 2039 /* pass user defined sigval data */ 2040 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2041 error = kqueue_register(kq, &kev, td, 1); 2042 fdrop(kq_fp, td); 2043 if (error) { 2044 uma_zfree(aiolio_zone, lj); 2045 return (error); 2046 } 2047 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2048 ; 2049 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2050 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2051 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2052 uma_zfree(aiolio_zone, lj); 2053 return EINVAL; 2054 } 2055 lj->lioj_flags |= LIOJ_SIGNAL; 2056 } else { 2057 uma_zfree(aiolio_zone, lj); 2058 return EINVAL; 2059 } 2060 } 2061 2062 AIO_LOCK(ki); 2063 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2064 /* 2065 * Add extra aiocb count to avoid the lio to be freed 2066 * by other threads doing aio_waitcomplete or aio_return, 2067 * and prevent event from being sent until we have queued 2068 * all tasks. 2069 */ 2070 lj->lioj_count = 1; 2071 AIO_UNLOCK(ki); 2072 2073 /* 2074 * Get pointers to the list of I/O requests. 2075 */ 2076 nerror = 0; 2077 cbptr = uap->acb_list; 2078 for (i = 0; i < uap->nent; i++) { 2079 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 2080 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) { 2081 error = aio_aqueue(td, iocb, lj, LIO_NOP, oldsigev); 2082 if (error != 0) 2083 nerror++; 2084 } 2085 } 2086 2087 error = 0; 2088 AIO_LOCK(ki); 2089 if (uap->mode == LIO_WAIT) { 2090 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2091 ki->kaio_flags |= KAIO_WAKEUP; 2092 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2093 PRIBIO | PCATCH, "aiospn", 0); 2094 if (error == ERESTART) 2095 error = EINTR; 2096 if (error) 2097 break; 2098 } 2099 } else { 2100 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2101 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2102 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2103 KNOTE_LOCKED(&lj->klist, 1); 2104 } 2105 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 2106 == LIOJ_SIGNAL 2107 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2108 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2109 aio_sendsig(p, &lj->lioj_signal, 2110 &lj->lioj_ksi); 2111 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2112 } 2113 } 2114 } 2115 lj->lioj_count--; 2116 if (lj->lioj_count == 0) { 2117 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2118 knlist_delete(&lj->klist, curthread, 1); 2119 PROC_LOCK(p); 2120 sigqueue_take(&lj->lioj_ksi); 2121 PROC_UNLOCK(p); 2122 AIO_UNLOCK(ki); 2123 uma_zfree(aiolio_zone, lj); 2124 } else 2125 AIO_UNLOCK(ki); 2126 2127 if (nerror) 2128 return (EIO); 2129 return (error); 2130 } 2131 2132 /* 2133 * Called from interrupt thread for physio, we should return as fast 2134 * as possible, so we schedule a biohelper task. 2135 */ 2136 static void 2137 aio_physwakeup(struct buf *bp) 2138 { 2139 struct aiocblist *aiocbe; 2140 2141 aiocbe = (struct aiocblist *)bp->b_caller1; 2142 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask); 2143 } 2144 2145 /* 2146 * Task routine to perform heavy tasks, process wakeup, and signals. 2147 */ 2148 static void 2149 biohelper(void *context, int pending) 2150 { 2151 struct aiocblist *aiocbe = context; 2152 struct buf *bp; 2153 struct proc *userp; 2154 struct kaioinfo *ki; 2155 int nblks; 2156 2157 bp = aiocbe->bp; 2158 userp = aiocbe->userproc; 2159 ki = userp->p_aioinfo; 2160 AIO_LOCK(ki); 2161 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 2162 aiocbe->uaiocb._aiocb_private.error = 0; 2163 if (bp->b_ioflags & BIO_ERROR) 2164 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 2165 nblks = btodb(aiocbe->uaiocb.aio_nbytes); 2166 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE) 2167 aiocbe->outputcharge += nblks; 2168 else 2169 aiocbe->inputcharge += nblks; 2170 aiocbe->bp = NULL; 2171 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist); 2172 ki->kaio_buffer_count--; 2173 aio_bio_done_notify(userp, aiocbe, DONE_BUF); 2174 AIO_UNLOCK(ki); 2175 2176 /* Release mapping into kernel space. */ 2177 vunmapbuf(bp); 2178 relpbuf(bp, NULL); 2179 atomic_subtract_int(&num_buf_aio, 1); 2180 } 2181 2182 /* syscall - wait for the next completion of an aio request */ 2183 int 2184 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2185 { 2186 struct proc *p = td->td_proc; 2187 struct timeval atv; 2188 struct timespec ts; 2189 struct kaioinfo *ki; 2190 struct aiocblist *cb; 2191 struct aiocb *uuaiocb; 2192 int error, status, timo; 2193 2194 suword(uap->aiocbp, (long)NULL); 2195 2196 timo = 0; 2197 if (uap->timeout) { 2198 /* Get timespec struct. */ 2199 error = copyin(uap->timeout, &ts, sizeof(ts)); 2200 if (error) 2201 return (error); 2202 2203 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000)) 2204 return (EINVAL); 2205 2206 TIMESPEC_TO_TIMEVAL(&atv, &ts); 2207 if (itimerfix(&atv)) 2208 return (EINVAL); 2209 timo = tvtohz(&atv); 2210 } 2211 2212 if (p->p_aioinfo == NULL) 2213 aio_init_aioinfo(p); 2214 ki = p->p_aioinfo; 2215 2216 error = 0; 2217 cb = NULL; 2218 AIO_LOCK(ki); 2219 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2220 ki->kaio_flags |= KAIO_WAKEUP; 2221 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2222 "aiowc", timo); 2223 if (timo && error == ERESTART) 2224 error = EINTR; 2225 if (error) 2226 break; 2227 } 2228 2229 if (cb != NULL) { 2230 MPASS(cb->jobstate == JOBST_JOBFINISHED); 2231 uuaiocb = cb->uuaiocb; 2232 status = cb->uaiocb._aiocb_private.status; 2233 error = cb->uaiocb._aiocb_private.error; 2234 td->td_retval[0] = status; 2235 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2236 p->p_stats->p_ru.ru_oublock += cb->outputcharge; 2237 cb->outputcharge = 0; 2238 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2239 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 2240 cb->inputcharge = 0; 2241 } 2242 aio_free_entry(cb); 2243 AIO_UNLOCK(ki); 2244 suword(uap->aiocbp, (long)uuaiocb); 2245 suword(&uuaiocb->_aiocb_private.error, error); 2246 suword(&uuaiocb->_aiocb_private.status, status); 2247 } else 2248 AIO_UNLOCK(ki); 2249 2250 return (error); 2251 } 2252 2253 int 2254 aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2255 { 2256 struct proc *p = td->td_proc; 2257 struct kaioinfo *ki; 2258 2259 if (uap->op != O_SYNC) /* XXX lack of O_DSYNC */ 2260 return (EINVAL); 2261 ki = p->p_aioinfo; 2262 if (ki == NULL) 2263 aio_init_aioinfo(p); 2264 return aio_aqueue(td, uap->aiocbp, NULL, LIO_SYNC, 0); 2265 } 2266 2267 /* kqueue attach function */ 2268 static int 2269 filt_aioattach(struct knote *kn) 2270 { 2271 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2272 2273 /* 2274 * The aiocbe pointer must be validated before using it, so 2275 * registration is restricted to the kernel; the user cannot 2276 * set EV_FLAG1. 2277 */ 2278 if ((kn->kn_flags & EV_FLAG1) == 0) 2279 return (EPERM); 2280 kn->kn_flags &= ~EV_FLAG1; 2281 2282 knlist_add(&aiocbe->klist, kn, 0); 2283 2284 return (0); 2285 } 2286 2287 /* kqueue detach function */ 2288 static void 2289 filt_aiodetach(struct knote *kn) 2290 { 2291 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2292 2293 if (!knlist_empty(&aiocbe->klist)) 2294 knlist_remove(&aiocbe->klist, kn, 0); 2295 } 2296 2297 /* kqueue filter function */ 2298 /*ARGSUSED*/ 2299 static int 2300 filt_aio(struct knote *kn, long hint) 2301 { 2302 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2303 2304 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2305 if (aiocbe->jobstate != JOBST_JOBFINISHED) 2306 return (0); 2307 kn->kn_flags |= EV_EOF; 2308 return (1); 2309 } 2310 2311 /* kqueue attach function */ 2312 static int 2313 filt_lioattach(struct knote *kn) 2314 { 2315 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata; 2316 2317 /* 2318 * The aioliojob pointer must be validated before using it, so 2319 * registration is restricted to the kernel; the user cannot 2320 * set EV_FLAG1. 2321 */ 2322 if ((kn->kn_flags & EV_FLAG1) == 0) 2323 return (EPERM); 2324 kn->kn_flags &= ~EV_FLAG1; 2325 2326 knlist_add(&lj->klist, kn, 0); 2327 2328 return (0); 2329 } 2330 2331 /* kqueue detach function */ 2332 static void 2333 filt_liodetach(struct knote *kn) 2334 { 2335 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata; 2336 2337 if (!knlist_empty(&lj->klist)) 2338 knlist_remove(&lj->klist, kn, 0); 2339 } 2340 2341 /* kqueue filter function */ 2342 /*ARGSUSED*/ 2343 static int 2344 filt_lio(struct knote *kn, long hint) 2345 { 2346 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata; 2347 2348 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2349 } 2350