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 * $FreeBSD: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $ 17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.32 2006/12/23 00:35:04 swildner Exp $ 18 */ 19 20 /* 21 * This file contains support for the POSIX 1003.1B AIO/LIO facility. 22 */ 23 24 #include <sys/param.h> 25 #include <sys/systm.h> 26 #include <sys/buf.h> 27 #include <sys/sysproto.h> 28 #include <sys/filedesc.h> 29 #include <sys/kernel.h> 30 #include <sys/fcntl.h> 31 #include <sys/file.h> 32 #include <sys/lock.h> 33 #include <sys/unistd.h> 34 #include <sys/proc.h> 35 #include <sys/resourcevar.h> 36 #include <sys/signalvar.h> 37 #include <sys/protosw.h> 38 #include <sys/socketvar.h> 39 #include <sys/sysctl.h> 40 #include <sys/vnode.h> 41 #include <sys/conf.h> 42 #include <sys/event.h> 43 44 #include <vm/vm.h> 45 #include <vm/vm_extern.h> 46 #include <vm/pmap.h> 47 #include <vm/vm_map.h> 48 #include <vm/vm_zone.h> 49 #include <sys/aio.h> 50 #include <sys/file2.h> 51 #include <sys/buf2.h> 52 #include <sys/thread2.h> 53 54 #include <machine/limits.h> 55 #include "opt_vfs_aio.h" 56 57 #ifdef VFS_AIO 58 59 /* 60 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 61 * overflow. 62 */ 63 static long jobrefid; 64 65 #define JOBST_NULL 0x0 66 #define JOBST_JOBQGLOBAL 0x2 67 #define JOBST_JOBRUNNING 0x3 68 #define JOBST_JOBFINISHED 0x4 69 #define JOBST_JOBQBUF 0x5 70 #define JOBST_JOBBFINISHED 0x6 71 72 #ifndef MAX_AIO_PER_PROC 73 #define MAX_AIO_PER_PROC 32 74 #endif 75 76 #ifndef MAX_AIO_QUEUE_PER_PROC 77 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */ 78 #endif 79 80 #ifndef MAX_AIO_PROCS 81 #define MAX_AIO_PROCS 32 82 #endif 83 84 #ifndef MAX_AIO_QUEUE 85 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */ 86 #endif 87 88 #ifndef TARGET_AIO_PROCS 89 #define TARGET_AIO_PROCS 4 90 #endif 91 92 #ifndef MAX_BUF_AIO 93 #define MAX_BUF_AIO 16 94 #endif 95 96 #ifndef AIOD_TIMEOUT_DEFAULT 97 #define AIOD_TIMEOUT_DEFAULT (10 * hz) 98 #endif 99 100 #ifndef AIOD_LIFETIME_DEFAULT 101 #define AIOD_LIFETIME_DEFAULT (30 * hz) 102 #endif 103 104 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management"); 105 106 static int max_aio_procs = MAX_AIO_PROCS; 107 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, 108 CTLFLAG_RW, &max_aio_procs, 0, 109 "Maximum number of kernel threads to use for handling async IO"); 110 111 static int num_aio_procs = 0; 112 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, 113 CTLFLAG_RD, &num_aio_procs, 0, 114 "Number of presently active kernel threads for async IO"); 115 116 /* 117 * The code will adjust the actual number of AIO processes towards this 118 * number when it gets a chance. 119 */ 120 static int target_aio_procs = TARGET_AIO_PROCS; 121 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 122 0, "Preferred number of ready kernel threads for async IO"); 123 124 static int max_queue_count = MAX_AIO_QUEUE; 125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 126 "Maximum number of aio requests to queue, globally"); 127 128 static int num_queue_count = 0; 129 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 130 "Number of queued aio requests"); 131 132 static int num_buf_aio = 0; 133 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 134 "Number of aio requests presently handled by the buf subsystem"); 135 136 /* Number of async I/O thread in the process of being started */ 137 /* XXX This should be local to _aio_aqueue() */ 138 static int num_aio_resv_start = 0; 139 140 static int aiod_timeout; 141 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0, 142 "Timeout value for synchronous aio operations"); 143 144 static int aiod_lifetime; 145 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 146 "Maximum lifetime for idle aiod"); 147 148 static int max_aio_per_proc = MAX_AIO_PER_PROC; 149 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 150 0, "Maximum active aio requests per process (stored in the process)"); 151 152 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 153 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 154 &max_aio_queue_per_proc, 0, 155 "Maximum queued aio requests per process (stored in the process)"); 156 157 static int max_buf_aio = MAX_BUF_AIO; 158 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 159 "Maximum buf aio requests per process (stored in the process)"); 160 161 /* 162 * AIO process info 163 */ 164 #define AIOP_FREE 0x1 /* proc on free queue */ 165 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */ 166 167 struct aioproclist { 168 int aioprocflags; /* AIO proc flags */ 169 TAILQ_ENTRY(aioproclist) list; /* List of processes */ 170 struct proc *aioproc; /* The AIO thread */ 171 }; 172 173 /* 174 * data-structure for lio signal management 175 */ 176 struct aio_liojob { 177 int lioj_flags; 178 int lioj_buffer_count; 179 int lioj_buffer_finished_count; 180 int lioj_queue_count; 181 int lioj_queue_finished_count; 182 struct sigevent lioj_signal; /* signal on all I/O done */ 183 TAILQ_ENTRY(aio_liojob) lioj_list; 184 struct kaioinfo *lioj_ki; 185 }; 186 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 187 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 188 189 /* 190 * per process aio data structure 191 */ 192 struct kaioinfo { 193 int kaio_flags; /* per process kaio flags */ 194 int kaio_maxactive_count; /* maximum number of AIOs */ 195 int kaio_active_count; /* number of currently used AIOs */ 196 int kaio_qallowed_count; /* maxiumu size of AIO queue */ 197 int kaio_queue_count; /* size of AIO queue */ 198 int kaio_ballowed_count; /* maximum number of buffers */ 199 int kaio_queue_finished_count; /* number of daemon jobs finished */ 200 int kaio_buffer_count; /* number of physio buffers */ 201 int kaio_buffer_finished_count; /* count of I/O done */ 202 struct proc *kaio_p; /* process that uses this kaio block */ 203 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */ 204 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */ 205 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */ 206 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */ 207 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */ 208 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */ 209 }; 210 211 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 212 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */ 213 214 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc; 215 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */ 216 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */ 217 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */ 218 219 static void aio_init_aioinfo(struct proc *p); 220 static void aio_onceonly(void *); 221 static int aio_free_entry(struct aiocblist *aiocbe); 222 static void aio_process(struct aiocblist *aiocbe); 223 static int aio_newproc(void); 224 static int aio_aqueue(struct aiocb *job, int type); 225 static void aio_physwakeup(struct bio *bio); 226 static int aio_fphysio(struct aiocblist *aiocbe); 227 static int aio_qphysio(struct proc *p, struct aiocblist *iocb); 228 static void aio_daemon(void *uproc); 229 static void process_signal(void *aioj); 230 231 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL); 232 233 /* 234 * Zones for: 235 * kaio Per process async io info 236 * aiop async io thread data 237 * aiocb async io jobs 238 * aiol list io job pointer - internal to aio_suspend XXX 239 * aiolio list io jobs 240 */ 241 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone; 242 243 /* 244 * Startup initialization 245 */ 246 static void 247 aio_onceonly(void *na) 248 { 249 TAILQ_INIT(&aio_freeproc); 250 TAILQ_INIT(&aio_activeproc); 251 TAILQ_INIT(&aio_jobs); 252 TAILQ_INIT(&aio_bufjobs); 253 TAILQ_INIT(&aio_freejobs); 254 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1); 255 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1); 256 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1); 257 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1); 258 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1); 259 aiod_timeout = AIOD_TIMEOUT_DEFAULT; 260 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 261 jobrefid = 1; 262 } 263 264 /* 265 * Init the per-process aioinfo structure. The aioinfo limits are set 266 * per-process for user limit (resource) management. 267 */ 268 static void 269 aio_init_aioinfo(struct proc *p) 270 { 271 struct kaioinfo *ki; 272 if (p->p_aioinfo == NULL) { 273 ki = zalloc(kaio_zone); 274 p->p_aioinfo = ki; 275 ki->kaio_flags = 0; 276 ki->kaio_maxactive_count = max_aio_per_proc; 277 ki->kaio_active_count = 0; 278 ki->kaio_qallowed_count = max_aio_queue_per_proc; 279 ki->kaio_queue_count = 0; 280 ki->kaio_ballowed_count = max_buf_aio; 281 ki->kaio_buffer_count = 0; 282 ki->kaio_buffer_finished_count = 0; 283 ki->kaio_p = p; 284 TAILQ_INIT(&ki->kaio_jobdone); 285 TAILQ_INIT(&ki->kaio_jobqueue); 286 TAILQ_INIT(&ki->kaio_bufdone); 287 TAILQ_INIT(&ki->kaio_bufqueue); 288 TAILQ_INIT(&ki->kaio_liojoblist); 289 TAILQ_INIT(&ki->kaio_sockqueue); 290 } 291 292 while (num_aio_procs < target_aio_procs) 293 aio_newproc(); 294 } 295 296 /* 297 * Free a job entry. Wait for completion if it is currently active, but don't 298 * delay forever. If we delay, we return a flag that says that we have to 299 * restart the queue scan. 300 */ 301 static int 302 aio_free_entry(struct aiocblist *aiocbe) 303 { 304 struct kaioinfo *ki; 305 struct aio_liojob *lj; 306 struct proc *p; 307 int error; 308 309 if (aiocbe->jobstate == JOBST_NULL) 310 panic("aio_free_entry: freeing already free job"); 311 312 p = aiocbe->userproc; 313 ki = p->p_aioinfo; 314 lj = aiocbe->lio; 315 if (ki == NULL) 316 panic("aio_free_entry: missing p->p_aioinfo"); 317 318 while (aiocbe->jobstate == JOBST_JOBRUNNING) { 319 aiocbe->jobflags |= AIOCBLIST_RUNDOWN; 320 tsleep(aiocbe, 0, "jobwai", 0); 321 } 322 if (aiocbe->bp == NULL) { 323 if (ki->kaio_queue_count <= 0) 324 panic("aio_free_entry: process queue size <= 0"); 325 if (num_queue_count <= 0) 326 panic("aio_free_entry: system wide queue size <= 0"); 327 328 if (lj) { 329 lj->lioj_queue_count--; 330 if (aiocbe->jobflags & AIOCBLIST_DONE) 331 lj->lioj_queue_finished_count--; 332 } 333 ki->kaio_queue_count--; 334 if (aiocbe->jobflags & AIOCBLIST_DONE) 335 ki->kaio_queue_finished_count--; 336 num_queue_count--; 337 } else { 338 if (lj) { 339 lj->lioj_buffer_count--; 340 if (aiocbe->jobflags & AIOCBLIST_DONE) 341 lj->lioj_buffer_finished_count--; 342 } 343 if (aiocbe->jobflags & AIOCBLIST_DONE) 344 ki->kaio_buffer_finished_count--; 345 ki->kaio_buffer_count--; 346 num_buf_aio--; 347 } 348 349 /* aiocbe is going away, we need to destroy any knotes */ 350 knote_remove(p->p_thread, &aiocbe->klist); 351 352 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN) 353 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) { 354 ki->kaio_flags &= ~KAIO_WAKEUP; 355 wakeup(p); 356 } 357 358 if (aiocbe->jobstate == JOBST_JOBQBUF) { 359 if ((error = aio_fphysio(aiocbe)) != 0) 360 return error; 361 if (aiocbe->jobstate != JOBST_JOBBFINISHED) 362 panic("aio_free_entry: invalid physio finish-up state"); 363 crit_enter(); 364 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist); 365 crit_exit(); 366 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) { 367 crit_enter(); 368 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 369 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 370 crit_exit(); 371 } else if (aiocbe->jobstate == JOBST_JOBFINISHED) 372 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist); 373 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) { 374 crit_enter(); 375 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist); 376 crit_exit(); 377 if (aiocbe->bp) { 378 vunmapbuf(aiocbe->bp); 379 relpbuf(aiocbe->bp, NULL); 380 aiocbe->bp = NULL; 381 } 382 } 383 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) { 384 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 385 zfree(aiolio_zone, lj); 386 } 387 aiocbe->jobstate = JOBST_NULL; 388 callout_stop(&aiocbe->timeout); 389 fdrop(aiocbe->fd_file); 390 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 391 return 0; 392 } 393 #endif /* VFS_AIO */ 394 395 /* 396 * Rundown the jobs for a given process. 397 */ 398 void 399 aio_proc_rundown(struct proc *p) 400 { 401 #ifndef VFS_AIO 402 return; 403 #else 404 struct kaioinfo *ki; 405 struct aio_liojob *lj, *ljn; 406 struct aiocblist *aiocbe, *aiocbn; 407 struct file *fp; 408 struct socket *so; 409 410 ki = p->p_aioinfo; 411 if (ki == NULL) 412 return; 413 414 ki->kaio_flags |= LIOJ_SIGNAL_POSTED; 415 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count > 416 ki->kaio_buffer_finished_count)) { 417 ki->kaio_flags |= KAIO_RUNDOWN; 418 if (tsleep(p, 0, "kaiowt", aiod_timeout)) 419 break; 420 } 421 422 /* 423 * Move any aio ops that are waiting on socket I/O to the normal job 424 * queues so they are cleaned up with any others. 425 */ 426 crit_enter(); 427 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe = 428 aiocbn) { 429 aiocbn = TAILQ_NEXT(aiocbe, plist); 430 fp = aiocbe->fd_file; 431 if (fp != NULL) { 432 so = (struct socket *)fp->f_data; 433 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list); 434 if (TAILQ_EMPTY(&so->so_aiojobq)) { 435 so->so_snd.sb_flags &= ~SB_AIO; 436 so->so_rcv.sb_flags &= ~SB_AIO; 437 } 438 } 439 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist); 440 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list); 441 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist); 442 } 443 crit_exit(); 444 445 restart1: 446 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) { 447 aiocbn = TAILQ_NEXT(aiocbe, plist); 448 if (aio_free_entry(aiocbe)) 449 goto restart1; 450 } 451 452 restart2: 453 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe = 454 aiocbn) { 455 aiocbn = TAILQ_NEXT(aiocbe, plist); 456 if (aio_free_entry(aiocbe)) 457 goto restart2; 458 } 459 460 restart3: 461 crit_enter(); 462 while (TAILQ_FIRST(&ki->kaio_bufqueue)) { 463 ki->kaio_flags |= KAIO_WAKEUP; 464 tsleep(p, 0, "aioprn", 0); 465 crit_exit(); 466 goto restart3; 467 } 468 crit_exit(); 469 470 restart4: 471 crit_enter(); 472 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) { 473 aiocbn = TAILQ_NEXT(aiocbe, plist); 474 if (aio_free_entry(aiocbe)) { 475 crit_exit(); 476 goto restart4; 477 } 478 } 479 crit_exit(); 480 481 /* 482 * If we've slept, jobs might have moved from one queue to another. 483 * Retry rundown if we didn't manage to empty the queues. 484 */ 485 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL || 486 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL || 487 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL || 488 TAILQ_FIRST(&ki->kaio_bufdone) != NULL) 489 goto restart1; 490 491 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) { 492 ljn = TAILQ_NEXT(lj, lioj_list); 493 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 494 0)) { 495 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 496 zfree(aiolio_zone, lj); 497 } else { 498 #ifdef DIAGNOSTIC 499 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, " 500 "QF:%d\n", lj->lioj_buffer_count, 501 lj->lioj_buffer_finished_count, 502 lj->lioj_queue_count, 503 lj->lioj_queue_finished_count); 504 #endif 505 } 506 } 507 508 zfree(kaio_zone, ki); 509 p->p_aioinfo = NULL; 510 #endif /* VFS_AIO */ 511 } 512 513 #ifdef VFS_AIO 514 /* 515 * Select a job to run (called by an AIO daemon). 516 */ 517 static struct aiocblist * 518 aio_selectjob(struct aioproclist *aiop) 519 { 520 struct aiocblist *aiocbe; 521 struct kaioinfo *ki; 522 struct proc *userp; 523 524 crit_enter(); 525 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe = 526 TAILQ_NEXT(aiocbe, list)) { 527 userp = aiocbe->userproc; 528 ki = userp->p_aioinfo; 529 530 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 531 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 532 crit_exit(); 533 return aiocbe; 534 } 535 } 536 crit_exit(); 537 538 return NULL; 539 } 540 541 /* 542 * The AIO processing activity. This is the code that does the I/O request for 543 * the non-physio version of the operations. The normal vn operations are used, 544 * and this code should work in all instances for every type of file, including 545 * pipes, sockets, fifos, and regular files. 546 */ 547 static void 548 aio_process(struct aiocblist *aiocbe) 549 { 550 struct thread *mytd; 551 struct aiocb *cb; 552 struct file *fp; 553 struct uio auio; 554 struct iovec aiov; 555 int cnt; 556 int error; 557 int oublock_st, oublock_end; 558 int inblock_st, inblock_end; 559 560 mytd = curthread; 561 cb = &aiocbe->uaiocb; 562 fp = aiocbe->fd_file; 563 564 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 565 aiov.iov_len = cb->aio_nbytes; 566 567 auio.uio_iov = &aiov; 568 auio.uio_iovcnt = 1; 569 auio.uio_offset = cb->aio_offset; 570 auio.uio_resid = cb->aio_nbytes; 571 cnt = cb->aio_nbytes; 572 auio.uio_segflg = UIO_USERSPACE; 573 auio.uio_td = mytd; 574 575 inblock_st = mytd->td_proc->p_stats->p_ru.ru_inblock; 576 oublock_st = mytd->td_proc->p_stats->p_ru.ru_oublock; 577 /* 578 * _aio_aqueue() acquires a reference to the file that is 579 * released in aio_free_entry(). 580 */ 581 if (cb->aio_lio_opcode == LIO_READ) { 582 auio.uio_rw = UIO_READ; 583 error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET); 584 } else { 585 auio.uio_rw = UIO_WRITE; 586 error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET); 587 } 588 inblock_end = mytd->td_proc->p_stats->p_ru.ru_inblock; 589 oublock_end = mytd->td_proc->p_stats->p_ru.ru_oublock; 590 591 aiocbe->inputcharge = inblock_end - inblock_st; 592 aiocbe->outputcharge = oublock_end - oublock_st; 593 594 if ((error) && (auio.uio_resid != cnt)) { 595 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 596 error = 0; 597 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) 598 ksignal(aiocbe->userproc, SIGPIPE); 599 } 600 601 cnt -= auio.uio_resid; 602 cb->_aiocb_private.error = error; 603 cb->_aiocb_private.status = cnt; 604 } 605 606 /* 607 * The AIO daemon, most of the actual work is done in aio_process, 608 * but the setup (and address space mgmt) is done in this routine. 609 * 610 * The MP lock is held on entry. 611 */ 612 static void 613 aio_daemon(void *uproc) 614 { 615 struct aio_liojob *lj; 616 struct aiocb *cb; 617 struct aiocblist *aiocbe; 618 struct aioproclist *aiop; 619 struct kaioinfo *ki; 620 struct proc *curcp, *mycp, *userp; 621 struct vmspace *myvm, *tmpvm; 622 struct ucred *cr; 623 624 /* 625 * Local copies of curproc (cp) and vmspace (myvm) 626 */ 627 mycp = curproc; 628 myvm = mycp->p_vmspace; 629 630 if (mycp->p_textvp) { 631 vrele(mycp->p_textvp); 632 mycp->p_textvp = NULL; 633 } 634 635 /* 636 * Allocate and ready the aio control info. There is one aiop structure 637 * per daemon. 638 */ 639 aiop = zalloc(aiop_zone); 640 aiop->aioproc = mycp; 641 aiop->aioprocflags |= AIOP_FREE; 642 643 crit_enter(); 644 645 /* 646 * Place thread (lightweight process) onto the AIO free thread list. 647 */ 648 if (TAILQ_EMPTY(&aio_freeproc)) 649 wakeup(&aio_freeproc); 650 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 651 652 crit_exit(); 653 654 /* Make up a name for the daemon. */ 655 strcpy(mycp->p_comm, "aiod"); 656 657 /* 658 * Get rid of our current filedescriptors. AIOD's don't need any 659 * filedescriptors, except as temporarily inherited from the client. 660 * Credentials are also cloned, and made equivalent to "root". 661 */ 662 fdfree(mycp); 663 mycp->p_fd = NULL; 664 cr = cratom(&mycp->p_ucred); 665 cr->cr_uid = 0; 666 uireplace(&cr->cr_uidinfo, uifind(0)); 667 cr->cr_ngroups = 1; 668 cr->cr_groups[0] = 1; 669 670 /* The daemon resides in its own pgrp. */ 671 enterpgrp(mycp, mycp->p_pid, 1); 672 673 /* Mark special process type. */ 674 mycp->p_flag |= P_SYSTEM | P_KTHREADP; 675 676 /* 677 * Wakeup parent process. (Parent sleeps to keep from blasting away 678 * and creating too many daemons.) 679 */ 680 wakeup(mycp); 681 682 for (;;) { 683 /* 684 * curcp is the current daemon process context. 685 * userp is the current user process context. 686 */ 687 curcp = mycp; 688 689 /* 690 * Take daemon off of free queue 691 */ 692 if (aiop->aioprocflags & AIOP_FREE) { 693 crit_enter(); 694 TAILQ_REMOVE(&aio_freeproc, aiop, list); 695 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 696 aiop->aioprocflags &= ~AIOP_FREE; 697 crit_exit(); 698 } 699 aiop->aioprocflags &= ~AIOP_SCHED; 700 701 /* 702 * Check for jobs. 703 */ 704 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 705 cb = &aiocbe->uaiocb; 706 userp = aiocbe->userproc; 707 708 aiocbe->jobstate = JOBST_JOBRUNNING; 709 710 /* 711 * Connect to process address space for user program. 712 */ 713 if (userp != curcp) { 714 /* 715 * Save the current address space that we are 716 * connected to. 717 */ 718 tmpvm = mycp->p_vmspace; 719 720 /* 721 * Point to the new user address space, and 722 * refer to it. 723 */ 724 mycp->p_vmspace = userp->p_vmspace; 725 mycp->p_vmspace->vm_refcnt++; 726 727 /* Activate the new mapping. */ 728 pmap_activate(mycp); 729 730 /* 731 * If the old address space wasn't the daemons 732 * own address space, then we need to remove the 733 * daemon's reference from the other process 734 * that it was acting on behalf of. 735 */ 736 if (tmpvm != myvm) { 737 vmspace_free(tmpvm); 738 } 739 curcp = userp; 740 } 741 742 ki = userp->p_aioinfo; 743 lj = aiocbe->lio; 744 745 /* Account for currently active jobs. */ 746 ki->kaio_active_count++; 747 748 /* Do the I/O function. */ 749 aio_process(aiocbe); 750 751 /* Decrement the active job count. */ 752 ki->kaio_active_count--; 753 754 /* 755 * Increment the completion count for wakeup/signal 756 * comparisons. 757 */ 758 aiocbe->jobflags |= AIOCBLIST_DONE; 759 ki->kaio_queue_finished_count++; 760 if (lj) 761 lj->lioj_queue_finished_count++; 762 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags 763 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) { 764 ki->kaio_flags &= ~KAIO_WAKEUP; 765 wakeup(userp); 766 } 767 768 crit_enter(); 769 if (lj && (lj->lioj_flags & 770 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) { 771 if ((lj->lioj_queue_finished_count == 772 lj->lioj_queue_count) && 773 (lj->lioj_buffer_finished_count == 774 lj->lioj_buffer_count)) { 775 ksignal(userp, 776 lj->lioj_signal.sigev_signo); 777 lj->lioj_flags |= 778 LIOJ_SIGNAL_POSTED; 779 } 780 } 781 crit_exit(); 782 783 aiocbe->jobstate = JOBST_JOBFINISHED; 784 785 crit_enter(); 786 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 787 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist); 788 crit_exit(); 789 KNOTE(&aiocbe->klist, 0); 790 791 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) { 792 wakeup(aiocbe); 793 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN; 794 } 795 796 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 797 ksignal(userp, cb->aio_sigevent.sigev_signo); 798 } 799 } 800 801 /* 802 * Disconnect from user address space. 803 */ 804 if (curcp != mycp) { 805 /* Get the user address space to disconnect from. */ 806 tmpvm = mycp->p_vmspace; 807 808 /* Get original address space for daemon. */ 809 mycp->p_vmspace = myvm; 810 811 /* Activate the daemon's address space. */ 812 pmap_activate(mycp); 813 #ifdef DIAGNOSTIC 814 if (tmpvm == myvm) { 815 kprintf("AIOD: vmspace problem -- %d\n", 816 mycp->p_pid); 817 } 818 #endif 819 /* Remove our vmspace reference. */ 820 vmspace_free(tmpvm); 821 822 curcp = mycp; 823 } 824 825 /* 826 * If we are the first to be put onto the free queue, wakeup 827 * anyone waiting for a daemon. 828 */ 829 crit_enter(); 830 TAILQ_REMOVE(&aio_activeproc, aiop, list); 831 if (TAILQ_EMPTY(&aio_freeproc)) 832 wakeup(&aio_freeproc); 833 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 834 aiop->aioprocflags |= AIOP_FREE; 835 crit_exit(); 836 837 /* 838 * If daemon is inactive for a long time, allow it to exit, 839 * thereby freeing resources. 840 */ 841 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp, 842 0, "aiordy", aiod_lifetime)) { 843 crit_enter(); 844 if (TAILQ_EMPTY(&aio_jobs)) { 845 if ((aiop->aioprocflags & AIOP_FREE) && 846 (num_aio_procs > target_aio_procs)) { 847 TAILQ_REMOVE(&aio_freeproc, aiop, list); 848 crit_exit(); 849 zfree(aiop_zone, aiop); 850 num_aio_procs--; 851 #ifdef DIAGNOSTIC 852 if (mycp->p_vmspace->vm_refcnt <= 1) { 853 kprintf("AIOD: bad vm refcnt for" 854 " exiting daemon: %d\n", 855 mycp->p_vmspace->vm_refcnt); 856 } 857 #endif 858 exit1(0); 859 } 860 } 861 crit_exit(); 862 } 863 } 864 } 865 866 /* 867 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 868 * AIO daemon modifies its environment itself. 869 */ 870 static int 871 aio_newproc() 872 { 873 int error; 874 struct lwp *lp, *nlp; 875 struct proc *np; 876 877 lp = &proc0.p_lwp; 878 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np); 879 if (error) 880 return error; 881 nlp = LIST_FIRST(&np->p_lwps); 882 cpu_set_fork_handler(nlp, aio_daemon, curproc); 883 start_forked_proc(lp, np); 884 885 /* 886 * Wait until daemon is started, but continue on just in case to 887 * handle error conditions. 888 */ 889 error = tsleep(np, 0, "aiosta", aiod_timeout); 890 num_aio_procs++; 891 892 return error; 893 } 894 895 /* 896 * Try the high-performance, low-overhead physio method for eligible 897 * VCHR devices. This method doesn't use an aio helper thread, and 898 * thus has very low overhead. 899 * 900 * Assumes that the caller, _aio_aqueue(), has incremented the file 901 * structure's reference count, preventing its deallocation for the 902 * duration of this call. 903 */ 904 static int 905 aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 906 { 907 int error; 908 struct aiocb *cb; 909 struct file *fp; 910 struct buf *bp; 911 struct vnode *vp; 912 struct kaioinfo *ki; 913 struct aio_liojob *lj; 914 int notify; 915 916 cb = &aiocbe->uaiocb; 917 fp = aiocbe->fd_file; 918 919 if (fp->f_type != DTYPE_VNODE) 920 return (-1); 921 922 vp = (struct vnode *)fp->f_data; 923 924 /* 925 * If its not a disk, we don't want to return a positive error. 926 * It causes the aio code to not fall through to try the thread 927 * way when you're talking to a regular file. 928 */ 929 if (!vn_isdisk(vp, &error)) { 930 if (error == ENOTBLK) 931 return (-1); 932 else 933 return (error); 934 } 935 936 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys) 937 return (-1); 938 939 if (cb->aio_nbytes > 940 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 941 return (-1); 942 943 ki = p->p_aioinfo; 944 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 945 return (-1); 946 947 ki->kaio_buffer_count++; 948 949 lj = aiocbe->lio; 950 if (lj) 951 lj->lioj_buffer_count++; 952 953 /* Create and build a buffer header for a transfer. */ 954 bp = getpbuf(NULL); 955 BUF_KERNPROC(bp); 956 957 /* 958 * Get a copy of the kva from the physical buffer. 959 */ 960 bp->b_bio1.bio_caller_info1.ptr = p; 961 error = 0; 962 963 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ? 964 BUF_CMD_WRITE : BUF_CMD_READ; 965 bp->b_bio1.bio_done = aio_physwakeup; 966 bp->b_bio1.bio_offset = cb->aio_offset; 967 968 /* Bring buffer into kernel space. */ 969 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) { 970 error = EFAULT; 971 goto doerror; 972 } 973 974 crit_enter(); 975 976 aiocbe->bp = bp; 977 bp->b_bio1.bio_caller_info2.ptr = aiocbe; 978 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list); 979 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 980 aiocbe->jobstate = JOBST_JOBQBUF; 981 cb->_aiocb_private.status = cb->aio_nbytes; 982 num_buf_aio++; 983 bp->b_error = 0; 984 985 crit_exit(); 986 987 /* Perform transfer. */ 988 dev_dstrategy(vp->v_rdev, &bp->b_bio1); 989 990 notify = 0; 991 crit_enter(); 992 993 /* 994 * If we had an error invoking the request, or an error in processing 995 * the request before we have returned, we process it as an error in 996 * transfer. Note that such an I/O error is not indicated immediately, 997 * but is returned using the aio_error mechanism. In this case, 998 * aio_suspend will return immediately. 999 */ 1000 if (bp->b_error || (bp->b_flags & B_ERROR)) { 1001 struct aiocb *job = aiocbe->uuaiocb; 1002 1003 aiocbe->uaiocb._aiocb_private.status = 0; 1004 suword(&job->_aiocb_private.status, 0); 1005 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 1006 suword(&job->_aiocb_private.error, bp->b_error); 1007 1008 ki->kaio_buffer_finished_count++; 1009 1010 if (aiocbe->jobstate != JOBST_JOBBFINISHED) { 1011 aiocbe->jobstate = JOBST_JOBBFINISHED; 1012 aiocbe->jobflags |= AIOCBLIST_DONE; 1013 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 1014 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 1015 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 1016 notify = 1; 1017 } 1018 } 1019 crit_exit(); 1020 if (notify) 1021 KNOTE(&aiocbe->klist, 0); 1022 return 0; 1023 1024 doerror: 1025 ki->kaio_buffer_count--; 1026 if (lj) 1027 lj->lioj_buffer_count--; 1028 aiocbe->bp = NULL; 1029 relpbuf(bp, NULL); 1030 return error; 1031 } 1032 1033 /* 1034 * This waits/tests physio completion. 1035 */ 1036 static int 1037 aio_fphysio(struct aiocblist *iocb) 1038 { 1039 struct buf *bp; 1040 int error; 1041 1042 bp = iocb->bp; 1043 1044 crit_enter(); 1045 while (bp->b_cmd != BUF_CMD_DONE) { 1046 if (tsleep(bp, 0, "physstr", aiod_timeout)) { 1047 if (bp->b_cmd != BUF_CMD_DONE) { 1048 crit_exit(); 1049 return EINPROGRESS; 1050 } else { 1051 break; 1052 } 1053 } 1054 } 1055 crit_exit(); 1056 1057 /* Release mapping into kernel space. */ 1058 vunmapbuf(bp); 1059 iocb->bp = 0; 1060 1061 error = 0; 1062 1063 /* Check for an error. */ 1064 if (bp->b_flags & B_ERROR) 1065 error = bp->b_error; 1066 1067 relpbuf(bp, NULL); 1068 return (error); 1069 } 1070 #endif /* VFS_AIO */ 1071 1072 /* 1073 * Wake up aio requests that may be serviceable now. 1074 */ 1075 void 1076 aio_swake(struct socket *so, struct sockbuf *sb) 1077 { 1078 #ifndef VFS_AIO 1079 return; 1080 #else 1081 struct aiocblist *cb,*cbn; 1082 struct proc *p; 1083 struct kaioinfo *ki = NULL; 1084 int opcode, wakecount = 0; 1085 struct aioproclist *aiop; 1086 1087 if (sb == &so->so_snd) { 1088 opcode = LIO_WRITE; 1089 so->so_snd.sb_flags &= ~SB_AIO; 1090 } else { 1091 opcode = LIO_READ; 1092 so->so_rcv.sb_flags &= ~SB_AIO; 1093 } 1094 1095 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) { 1096 cbn = TAILQ_NEXT(cb, list); 1097 if (opcode == cb->uaiocb.aio_lio_opcode) { 1098 p = cb->userproc; 1099 ki = p->p_aioinfo; 1100 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1101 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist); 1102 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1103 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist); 1104 wakecount++; 1105 if (cb->jobstate != JOBST_JOBQGLOBAL) 1106 panic("invalid queue value"); 1107 } 1108 } 1109 1110 while (wakecount--) { 1111 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) { 1112 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1113 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1114 aiop->aioprocflags &= ~AIOP_FREE; 1115 wakeup(aiop->aioproc); 1116 } 1117 } 1118 #endif /* VFS_AIO */ 1119 } 1120 1121 #ifdef VFS_AIO 1122 /* 1123 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1124 * technique is done in this code. 1125 */ 1126 static int 1127 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type) 1128 { 1129 struct proc *p = curproc; 1130 struct filedesc *fdp; 1131 struct file *fp; 1132 unsigned int fd; 1133 struct socket *so; 1134 int error; 1135 int opcode, user_opcode; 1136 struct aiocblist *aiocbe; 1137 struct aioproclist *aiop; 1138 struct kaioinfo *ki; 1139 struct kevent kev; 1140 struct kqueue *kq; 1141 struct file *kq_fp; 1142 1143 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL) 1144 TAILQ_REMOVE(&aio_freejobs, aiocbe, list); 1145 else 1146 aiocbe = zalloc (aiocb_zone); 1147 1148 aiocbe->inputcharge = 0; 1149 aiocbe->outputcharge = 0; 1150 callout_init(&aiocbe->timeout); 1151 SLIST_INIT(&aiocbe->klist); 1152 1153 suword(&job->_aiocb_private.status, -1); 1154 suword(&job->_aiocb_private.error, 0); 1155 suword(&job->_aiocb_private.kernelinfo, -1); 1156 1157 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb)); 1158 if (error) { 1159 suword(&job->_aiocb_private.error, error); 1160 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1161 return error; 1162 } 1163 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL && 1164 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1165 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1166 return EINVAL; 1167 } 1168 1169 /* Save userspace address of the job info. */ 1170 aiocbe->uuaiocb = job; 1171 1172 /* Get the opcode. */ 1173 user_opcode = aiocbe->uaiocb.aio_lio_opcode; 1174 if (type != LIO_NOP) 1175 aiocbe->uaiocb.aio_lio_opcode = type; 1176 opcode = aiocbe->uaiocb.aio_lio_opcode; 1177 1178 /* Get the fd info for process. */ 1179 fdp = p->p_fd; 1180 1181 /* 1182 * Range check file descriptor. 1183 */ 1184 fd = aiocbe->uaiocb.aio_fildes; 1185 if (fd >= fdp->fd_nfiles) { 1186 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1187 if (type == 0) 1188 suword(&job->_aiocb_private.error, EBADF); 1189 return EBADF; 1190 } 1191 1192 fp = aiocbe->fd_file = fdp->fd_files[fd].fp; 1193 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 1194 0))) { 1195 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1196 if (type == 0) 1197 suword(&job->_aiocb_private.error, EBADF); 1198 return EBADF; 1199 } 1200 fhold(fp); 1201 1202 if (aiocbe->uaiocb.aio_offset == -1LL) { 1203 error = EINVAL; 1204 goto aqueue_fail; 1205 } 1206 error = suword(&job->_aiocb_private.kernelinfo, jobrefid); 1207 if (error) { 1208 error = EINVAL; 1209 goto aqueue_fail; 1210 } 1211 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid; 1212 if (jobrefid == LONG_MAX) 1213 jobrefid = 1; 1214 else 1215 jobrefid++; 1216 1217 if (opcode == LIO_NOP) { 1218 fdrop(fp); 1219 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1220 if (type == 0) { 1221 suword(&job->_aiocb_private.error, 0); 1222 suword(&job->_aiocb_private.status, 0); 1223 suword(&job->_aiocb_private.kernelinfo, 0); 1224 } 1225 return 0; 1226 } 1227 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) { 1228 if (type == 0) 1229 suword(&job->_aiocb_private.status, 0); 1230 error = EINVAL; 1231 goto aqueue_fail; 1232 } 1233 1234 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) { 1235 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1236 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr; 1237 } 1238 else { 1239 /* 1240 * This method for requesting kevent-based notification won't 1241 * work on the alpha, since we're passing in a pointer 1242 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT- 1243 * based method instead. 1244 */ 1245 if (user_opcode == LIO_NOP || user_opcode == LIO_READ || 1246 user_opcode == LIO_WRITE) 1247 goto no_kqueue; 1248 1249 error = copyin((struct kevent *)(uintptr_t)user_opcode, 1250 &kev, sizeof(kev)); 1251 if (error) 1252 goto aqueue_fail; 1253 } 1254 if ((u_int)kev.ident >= fdp->fd_nfiles || 1255 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL || 1256 (kq_fp->f_type != DTYPE_KQUEUE)) { 1257 error = EBADF; 1258 goto aqueue_fail; 1259 } 1260 kq = (struct kqueue *)kq_fp->f_data; 1261 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1262 kev.filter = EVFILT_AIO; 1263 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 1264 kev.data = (intptr_t)aiocbe; 1265 error = kqueue_register(kq, &kev, p->p_thread); 1266 aqueue_fail: 1267 if (error) { 1268 fdrop(fp); 1269 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1270 if (type == 0) 1271 suword(&job->_aiocb_private.error, error); 1272 goto done; 1273 } 1274 no_kqueue: 1275 1276 suword(&job->_aiocb_private.error, EINPROGRESS); 1277 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1278 aiocbe->userproc = p; 1279 aiocbe->jobflags = 0; 1280 aiocbe->lio = lj; 1281 ki = p->p_aioinfo; 1282 1283 if (fp->f_type == DTYPE_SOCKET) { 1284 /* 1285 * Alternate queueing for socket ops: Reach down into the 1286 * descriptor to get the socket data. Then check to see if the 1287 * socket is ready to be read or written (based on the requested 1288 * operation). 1289 * 1290 * If it is not ready for io, then queue the aiocbe on the 1291 * socket, and set the flags so we get a call when sbnotify() 1292 * happens. 1293 */ 1294 so = (struct socket *)fp->f_data; 1295 crit_enter(); 1296 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1297 LIO_WRITE) && (!sowriteable(so)))) { 1298 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1299 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist); 1300 if (opcode == LIO_READ) 1301 so->so_rcv.sb_flags |= SB_AIO; 1302 else 1303 so->so_snd.sb_flags |= SB_AIO; 1304 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */ 1305 ki->kaio_queue_count++; 1306 num_queue_count++; 1307 crit_exit(); 1308 error = 0; 1309 goto done; 1310 } 1311 crit_exit(); 1312 } 1313 1314 if ((error = aio_qphysio(p, aiocbe)) == 0) 1315 goto done; 1316 if (error > 0) { 1317 suword(&job->_aiocb_private.status, 0); 1318 aiocbe->uaiocb._aiocb_private.error = error; 1319 suword(&job->_aiocb_private.error, error); 1320 goto done; 1321 } 1322 1323 /* No buffer for daemon I/O. */ 1324 aiocbe->bp = NULL; 1325 1326 ki->kaio_queue_count++; 1327 if (lj) 1328 lj->lioj_queue_count++; 1329 crit_enter(); 1330 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1331 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1332 crit_exit(); 1333 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1334 1335 num_queue_count++; 1336 error = 0; 1337 1338 /* 1339 * If we don't have a free AIO process, and we are below our quota, then 1340 * start one. Otherwise, depend on the subsequent I/O completions to 1341 * pick-up this job. If we don't successfully create the new process 1342 * (thread) due to resource issues, we return an error for now (EAGAIN), 1343 * which is likely not the correct thing to do. 1344 */ 1345 crit_enter(); 1346 retryproc: 1347 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1348 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1349 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1350 aiop->aioprocflags &= ~AIOP_FREE; 1351 wakeup(aiop->aioproc); 1352 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1353 ((ki->kaio_active_count + num_aio_resv_start) < 1354 ki->kaio_maxactive_count)) { 1355 num_aio_resv_start++; 1356 if ((error = aio_newproc()) == 0) { 1357 num_aio_resv_start--; 1358 goto retryproc; 1359 } 1360 num_aio_resv_start--; 1361 } 1362 crit_exit(); 1363 done: 1364 return error; 1365 } 1366 1367 /* 1368 * This routine queues an AIO request, checking for quotas. 1369 */ 1370 static int 1371 aio_aqueue(struct aiocb *job, int type) 1372 { 1373 struct proc *p = curproc; 1374 struct kaioinfo *ki; 1375 1376 if (p->p_aioinfo == NULL) 1377 aio_init_aioinfo(p); 1378 1379 if (num_queue_count >= max_queue_count) 1380 return EAGAIN; 1381 1382 ki = p->p_aioinfo; 1383 if (ki->kaio_queue_count >= ki->kaio_qallowed_count) 1384 return EAGAIN; 1385 1386 return _aio_aqueue(job, NULL, type); 1387 } 1388 #endif /* VFS_AIO */ 1389 1390 /* 1391 * Support the aio_return system call, as a side-effect, kernel resources are 1392 * released. 1393 */ 1394 int 1395 sys_aio_return(struct aio_return_args *uap) 1396 { 1397 #ifndef VFS_AIO 1398 return ENOSYS; 1399 #else 1400 struct proc *p = curproc; 1401 long jobref; 1402 struct aiocblist *cb, *ncb; 1403 struct aiocb *ujob; 1404 struct kaioinfo *ki; 1405 1406 ki = p->p_aioinfo; 1407 if (ki == NULL) 1408 return EINVAL; 1409 1410 ujob = uap->aiocbp; 1411 1412 jobref = fuword(&ujob->_aiocb_private.kernelinfo); 1413 if (jobref == -1 || jobref == 0) 1414 return EINVAL; 1415 1416 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1417 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == 1418 jobref) { 1419 if (ujob == cb->uuaiocb) { 1420 uap->sysmsg_result = 1421 cb->uaiocb._aiocb_private.status; 1422 } else 1423 uap->sysmsg_result = EFAULT; 1424 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1425 p->p_stats->p_ru.ru_oublock += 1426 cb->outputcharge; 1427 cb->outputcharge = 0; 1428 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1429 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 1430 cb->inputcharge = 0; 1431 } 1432 aio_free_entry(cb); 1433 return 0; 1434 } 1435 } 1436 crit_enter(); 1437 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) { 1438 ncb = TAILQ_NEXT(cb, plist); 1439 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) 1440 == jobref) { 1441 crit_exit(); 1442 if (ujob == cb->uuaiocb) { 1443 uap->sysmsg_result = 1444 cb->uaiocb._aiocb_private.status; 1445 } else 1446 uap->sysmsg_result = EFAULT; 1447 aio_free_entry(cb); 1448 return 0; 1449 } 1450 } 1451 crit_exit(); 1452 1453 return (EINVAL); 1454 #endif /* VFS_AIO */ 1455 } 1456 1457 /* 1458 * Allow a process to wakeup when any of the I/O requests are completed. 1459 */ 1460 int 1461 sys_aio_suspend(struct aio_suspend_args *uap) 1462 { 1463 #ifndef VFS_AIO 1464 return ENOSYS; 1465 #else 1466 struct proc *p = curproc; 1467 struct timeval atv; 1468 struct timespec ts; 1469 struct aiocb *const *cbptr, *cbp; 1470 struct kaioinfo *ki; 1471 struct aiocblist *cb; 1472 int i; 1473 int njoblist; 1474 int error, timo; 1475 long *ijoblist; 1476 struct aiocb **ujoblist; 1477 1478 if (uap->nent > AIO_LISTIO_MAX) 1479 return EINVAL; 1480 1481 timo = 0; 1482 if (uap->timeout) { 1483 /* Get timespec struct. */ 1484 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1485 return error; 1486 1487 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) 1488 return (EINVAL); 1489 1490 TIMESPEC_TO_TIMEVAL(&atv, &ts); 1491 if (itimerfix(&atv)) 1492 return (EINVAL); 1493 timo = tvtohz_high(&atv); 1494 } 1495 1496 ki = p->p_aioinfo; 1497 if (ki == NULL) 1498 return EAGAIN; 1499 1500 njoblist = 0; 1501 ijoblist = zalloc(aiol_zone); 1502 ujoblist = zalloc(aiol_zone); 1503 cbptr = uap->aiocbp; 1504 1505 for (i = 0; i < uap->nent; i++) { 1506 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1507 if (cbp == 0) 1508 continue; 1509 ujoblist[njoblist] = cbp; 1510 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo); 1511 njoblist++; 1512 } 1513 1514 if (njoblist == 0) { 1515 zfree(aiol_zone, ijoblist); 1516 zfree(aiol_zone, ujoblist); 1517 return 0; 1518 } 1519 1520 error = 0; 1521 for (;;) { 1522 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1523 for (i = 0; i < njoblist; i++) { 1524 if (((intptr_t) 1525 cb->uaiocb._aiocb_private.kernelinfo) == 1526 ijoblist[i]) { 1527 if (ujoblist[i] != cb->uuaiocb) 1528 error = EINVAL; 1529 zfree(aiol_zone, ijoblist); 1530 zfree(aiol_zone, ujoblist); 1531 return error; 1532 } 1533 } 1534 } 1535 1536 crit_enter(); 1537 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = 1538 TAILQ_NEXT(cb, plist)) { 1539 for (i = 0; i < njoblist; i++) { 1540 if (((intptr_t) 1541 cb->uaiocb._aiocb_private.kernelinfo) == 1542 ijoblist[i]) { 1543 crit_exit(); 1544 if (ujoblist[i] != cb->uuaiocb) 1545 error = EINVAL; 1546 zfree(aiol_zone, ijoblist); 1547 zfree(aiol_zone, ujoblist); 1548 return error; 1549 } 1550 } 1551 } 1552 1553 ki->kaio_flags |= KAIO_WAKEUP; 1554 error = tsleep(p, PCATCH, "aiospn", timo); 1555 crit_exit(); 1556 1557 if (error == ERESTART || error == EINTR) { 1558 zfree(aiol_zone, ijoblist); 1559 zfree(aiol_zone, ujoblist); 1560 return EINTR; 1561 } else if (error == EWOULDBLOCK) { 1562 zfree(aiol_zone, ijoblist); 1563 zfree(aiol_zone, ujoblist); 1564 return EAGAIN; 1565 } 1566 } 1567 1568 /* NOTREACHED */ 1569 return EINVAL; 1570 #endif /* VFS_AIO */ 1571 } 1572 1573 /* 1574 * aio_cancel cancels any non-physio aio operations not currently in 1575 * progress. 1576 */ 1577 int 1578 sys_aio_cancel(struct aio_cancel_args *uap) 1579 { 1580 #ifndef VFS_AIO 1581 return ENOSYS; 1582 #else 1583 struct proc *p = curproc; 1584 struct kaioinfo *ki; 1585 struct aiocblist *cbe, *cbn; 1586 struct file *fp; 1587 struct filedesc *fdp; 1588 struct socket *so; 1589 struct proc *po; 1590 int error; 1591 int cancelled=0; 1592 int notcancelled=0; 1593 struct vnode *vp; 1594 1595 fdp = p->p_fd; 1596 if ((u_int)uap->fd >= fdp->fd_nfiles || 1597 (fp = fdp->fd_files[uap->fd].fp) == NULL) 1598 return (EBADF); 1599 1600 if (fp->f_type == DTYPE_VNODE) { 1601 vp = (struct vnode *)fp->f_data; 1602 1603 if (vn_isdisk(vp,&error)) { 1604 uap->sysmsg_result = AIO_NOTCANCELED; 1605 return 0; 1606 } 1607 } else if (fp->f_type == DTYPE_SOCKET) { 1608 so = (struct socket *)fp->f_data; 1609 1610 crit_enter(); 1611 1612 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) { 1613 cbn = TAILQ_NEXT(cbe, list); 1614 if ((uap->aiocbp == NULL) || 1615 (uap->aiocbp == cbe->uuaiocb) ) { 1616 po = cbe->userproc; 1617 ki = po->p_aioinfo; 1618 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 1619 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist); 1620 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist); 1621 if (ki->kaio_flags & KAIO_WAKEUP) { 1622 wakeup(po); 1623 } 1624 cbe->jobstate = JOBST_JOBFINISHED; 1625 cbe->uaiocb._aiocb_private.status=-1; 1626 cbe->uaiocb._aiocb_private.error=ECANCELED; 1627 cancelled++; 1628 /* XXX cancelled, knote? */ 1629 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1630 SIGEV_SIGNAL) 1631 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1632 if (uap->aiocbp) 1633 break; 1634 } 1635 } 1636 crit_exit(); 1637 1638 if ((cancelled) && (uap->aiocbp)) { 1639 uap->sysmsg_result = AIO_CANCELED; 1640 return 0; 1641 } 1642 } 1643 ki=p->p_aioinfo; 1644 if (ki == NULL) 1645 goto done; 1646 crit_enter(); 1647 1648 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) { 1649 cbn = TAILQ_NEXT(cbe, plist); 1650 1651 if ((uap->fd == cbe->uaiocb.aio_fildes) && 1652 ((uap->aiocbp == NULL ) || 1653 (uap->aiocbp == cbe->uuaiocb))) { 1654 1655 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 1656 TAILQ_REMOVE(&aio_jobs, cbe, list); 1657 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 1658 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, 1659 plist); 1660 cancelled++; 1661 ki->kaio_queue_finished_count++; 1662 cbe->jobstate = JOBST_JOBFINISHED; 1663 cbe->uaiocb._aiocb_private.status = -1; 1664 cbe->uaiocb._aiocb_private.error = ECANCELED; 1665 /* XXX cancelled, knote? */ 1666 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1667 SIGEV_SIGNAL) 1668 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1669 } else { 1670 notcancelled++; 1671 } 1672 } 1673 } 1674 crit_exit(); 1675 done: 1676 if (notcancelled) { 1677 uap->sysmsg_result = AIO_NOTCANCELED; 1678 return 0; 1679 } 1680 if (cancelled) { 1681 uap->sysmsg_result = AIO_CANCELED; 1682 return 0; 1683 } 1684 uap->sysmsg_result = AIO_ALLDONE; 1685 1686 return 0; 1687 #endif /* VFS_AIO */ 1688 } 1689 1690 /* 1691 * aio_error is implemented in the kernel level for compatibility purposes only. 1692 * For a user mode async implementation, it would be best to do it in a userland 1693 * subroutine. 1694 */ 1695 int 1696 sys_aio_error(struct aio_error_args *uap) 1697 { 1698 #ifndef VFS_AIO 1699 return ENOSYS; 1700 #else 1701 struct proc *p = curproc; 1702 struct aiocblist *cb; 1703 struct kaioinfo *ki; 1704 long jobref; 1705 1706 ki = p->p_aioinfo; 1707 if (ki == NULL) 1708 return EINVAL; 1709 1710 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo); 1711 if ((jobref == -1) || (jobref == 0)) 1712 return EINVAL; 1713 1714 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1715 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1716 jobref) { 1717 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1718 return 0; 1719 } 1720 } 1721 1722 crit_enter(); 1723 1724 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb, 1725 plist)) { 1726 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1727 jobref) { 1728 uap->sysmsg_result = EINPROGRESS; 1729 crit_exit(); 1730 return 0; 1731 } 1732 } 1733 1734 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb, 1735 plist)) { 1736 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1737 jobref) { 1738 uap->sysmsg_result = EINPROGRESS; 1739 crit_exit(); 1740 return 0; 1741 } 1742 } 1743 crit_exit(); 1744 1745 crit_enter(); 1746 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb, 1747 plist)) { 1748 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1749 jobref) { 1750 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1751 crit_exit(); 1752 return 0; 1753 } 1754 } 1755 1756 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb, 1757 plist)) { 1758 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1759 jobref) { 1760 uap->sysmsg_result = EINPROGRESS; 1761 crit_exit(); 1762 return 0; 1763 } 1764 } 1765 crit_exit(); 1766 1767 #if (0) 1768 /* 1769 * Hack for lio. 1770 */ 1771 status = fuword(&uap->aiocbp->_aiocb_private.status); 1772 if (status == -1) 1773 return fuword(&uap->aiocbp->_aiocb_private.error); 1774 #endif 1775 return EINVAL; 1776 #endif /* VFS_AIO */ 1777 } 1778 1779 /* syscall - asynchronous read from a file (REALTIME) */ 1780 int 1781 sys_aio_read(struct aio_read_args *uap) 1782 { 1783 #ifndef VFS_AIO 1784 return ENOSYS; 1785 #else 1786 return aio_aqueue(uap->aiocbp, LIO_READ); 1787 #endif /* VFS_AIO */ 1788 } 1789 1790 /* syscall - asynchronous write to a file (REALTIME) */ 1791 int 1792 sys_aio_write(struct aio_write_args *uap) 1793 { 1794 #ifndef VFS_AIO 1795 return ENOSYS; 1796 #else 1797 return aio_aqueue(uap->aiocbp, LIO_WRITE); 1798 #endif /* VFS_AIO */ 1799 } 1800 1801 /* syscall - XXX undocumented */ 1802 int 1803 sys_lio_listio(struct lio_listio_args *uap) 1804 { 1805 #ifndef VFS_AIO 1806 return ENOSYS; 1807 #else 1808 struct proc *p = curproc; 1809 int nent, nentqueued; 1810 struct aiocb *iocb, * const *cbptr; 1811 struct aiocblist *cb; 1812 struct kaioinfo *ki; 1813 struct aio_liojob *lj; 1814 int error, runningcode; 1815 int nerror; 1816 int i; 1817 1818 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 1819 return EINVAL; 1820 1821 nent = uap->nent; 1822 if (nent > AIO_LISTIO_MAX) 1823 return EINVAL; 1824 1825 if (p->p_aioinfo == NULL) 1826 aio_init_aioinfo(p); 1827 1828 if ((nent + num_queue_count) > max_queue_count) 1829 return EAGAIN; 1830 1831 ki = p->p_aioinfo; 1832 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) 1833 return EAGAIN; 1834 1835 lj = zalloc(aiolio_zone); 1836 if (!lj) 1837 return EAGAIN; 1838 1839 lj->lioj_flags = 0; 1840 lj->lioj_buffer_count = 0; 1841 lj->lioj_buffer_finished_count = 0; 1842 lj->lioj_queue_count = 0; 1843 lj->lioj_queue_finished_count = 0; 1844 lj->lioj_ki = ki; 1845 1846 /* 1847 * Setup signal. 1848 */ 1849 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 1850 error = copyin(uap->sig, &lj->lioj_signal, 1851 sizeof(lj->lioj_signal)); 1852 if (error) { 1853 zfree(aiolio_zone, lj); 1854 return error; 1855 } 1856 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 1857 zfree(aiolio_zone, lj); 1858 return EINVAL; 1859 } 1860 lj->lioj_flags |= LIOJ_SIGNAL; 1861 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED; 1862 } else 1863 lj->lioj_flags &= ~LIOJ_SIGNAL; 1864 1865 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 1866 /* 1867 * Get pointers to the list of I/O requests. 1868 */ 1869 nerror = 0; 1870 nentqueued = 0; 1871 cbptr = uap->acb_list; 1872 for (i = 0; i < uap->nent; i++) { 1873 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1874 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) { 1875 error = _aio_aqueue(iocb, lj, 0); 1876 if (error == 0) 1877 nentqueued++; 1878 else 1879 nerror++; 1880 } 1881 } 1882 1883 /* 1884 * If we haven't queued any, then just return error. 1885 */ 1886 if (nentqueued == 0) 1887 return 0; 1888 1889 /* 1890 * Calculate the appropriate error return. 1891 */ 1892 runningcode = 0; 1893 if (nerror) 1894 runningcode = EIO; 1895 1896 if (uap->mode == LIO_WAIT) { 1897 int command, found, jobref; 1898 1899 for (;;) { 1900 found = 0; 1901 for (i = 0; i < uap->nent; i++) { 1902 /* 1903 * Fetch address of the control buf pointer in 1904 * user space. 1905 */ 1906 iocb = (struct aiocb *) 1907 (intptr_t)fuword(&cbptr[i]); 1908 if (((intptr_t)iocb == -1) || ((intptr_t)iocb 1909 == 0)) 1910 continue; 1911 1912 /* 1913 * Fetch the associated command from user space. 1914 */ 1915 command = fuword(&iocb->aio_lio_opcode); 1916 if (command == LIO_NOP) { 1917 found++; 1918 continue; 1919 } 1920 1921 jobref = fuword(&iocb->_aiocb_private.kernelinfo); 1922 1923 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1924 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1925 == jobref) { 1926 if (cb->uaiocb.aio_lio_opcode 1927 == LIO_WRITE) { 1928 p->p_stats->p_ru.ru_oublock 1929 += 1930 cb->outputcharge; 1931 cb->outputcharge = 0; 1932 } else if (cb->uaiocb.aio_lio_opcode 1933 == LIO_READ) { 1934 p->p_stats->p_ru.ru_inblock 1935 += cb->inputcharge; 1936 cb->inputcharge = 0; 1937 } 1938 found++; 1939 break; 1940 } 1941 } 1942 1943 crit_enter(); 1944 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) { 1945 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1946 == jobref) { 1947 found++; 1948 break; 1949 } 1950 } 1951 crit_exit(); 1952 } 1953 1954 /* 1955 * If all I/Os have been disposed of, then we can 1956 * return. 1957 */ 1958 if (found == nentqueued) 1959 return runningcode; 1960 1961 ki->kaio_flags |= KAIO_WAKEUP; 1962 error = tsleep(p, PCATCH, "aiospn", 0); 1963 1964 if (error == EINTR) 1965 return EINTR; 1966 else if (error == EWOULDBLOCK) 1967 return EAGAIN; 1968 } 1969 } 1970 1971 return runningcode; 1972 #endif /* VFS_AIO */ 1973 } 1974 1975 #ifdef VFS_AIO 1976 /* 1977 * This is a weird hack so that we can post a signal. It is safe to do so from 1978 * a timeout routine, but *not* from an interrupt routine. 1979 */ 1980 static void 1981 process_signal(void *aioj) 1982 { 1983 struct aiocblist *aiocbe = aioj; 1984 struct aio_liojob *lj = aiocbe->lio; 1985 struct aiocb *cb = &aiocbe->uaiocb; 1986 1987 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) && 1988 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) { 1989 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo); 1990 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 1991 } 1992 1993 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) 1994 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo); 1995 } 1996 1997 /* 1998 * Interrupt handler for physio, performs the necessary process wakeups, and 1999 * signals. 2000 */ 2001 static void 2002 aio_physwakeup(struct bio *bio) 2003 { 2004 struct buf *bp = bio->bio_buf; 2005 struct aiocblist *aiocbe; 2006 struct proc *p; 2007 struct kaioinfo *ki; 2008 struct aio_liojob *lj; 2009 2010 aiocbe = bio->bio_caller_info2.ptr; 2011 2012 if (aiocbe) { 2013 p = bio->bio_caller_info1.ptr; 2014 2015 aiocbe->jobstate = JOBST_JOBBFINISHED; 2016 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 2017 aiocbe->uaiocb._aiocb_private.error = 0; 2018 aiocbe->jobflags |= AIOCBLIST_DONE; 2019 2020 if (bp->b_flags & B_ERROR) 2021 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 2022 2023 lj = aiocbe->lio; 2024 if (lj) { 2025 lj->lioj_buffer_finished_count++; 2026 2027 /* 2028 * wakeup/signal if all of the interrupt jobs are done. 2029 */ 2030 if (lj->lioj_buffer_finished_count == 2031 lj->lioj_buffer_count) { 2032 /* 2033 * Post a signal if it is called for. 2034 */ 2035 if ((lj->lioj_flags & 2036 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == 2037 LIOJ_SIGNAL) { 2038 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2039 callout_reset(&aiocbe->timeout, 0, 2040 process_signal, aiocbe); 2041 } 2042 } 2043 } 2044 2045 ki = p->p_aioinfo; 2046 if (ki) { 2047 ki->kaio_buffer_finished_count++; 2048 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 2049 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 2050 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 2051 2052 KNOTE(&aiocbe->klist, 0); 2053 /* Do the wakeup. */ 2054 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) { 2055 ki->kaio_flags &= ~KAIO_WAKEUP; 2056 wakeup(p); 2057 } 2058 } 2059 2060 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 2061 callout_reset(&aiocbe->timeout, 0, 2062 process_signal, aiocbe); 2063 } 2064 } 2065 bp->b_cmd = BUF_CMD_DONE; 2066 wakeup(bp); 2067 } 2068 #endif /* VFS_AIO */ 2069 2070 /* syscall - wait for the next completion of an aio request */ 2071 int 2072 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap) 2073 { 2074 #ifndef VFS_AIO 2075 return ENOSYS; 2076 #else 2077 struct proc *p = curproc; 2078 struct timeval atv; 2079 struct timespec ts; 2080 struct kaioinfo *ki; 2081 struct aiocblist *cb = NULL; 2082 int error, timo; 2083 2084 suword(uap->aiocbp, (int)NULL); 2085 2086 timo = 0; 2087 if (uap->timeout) { 2088 /* Get timespec struct. */ 2089 error = copyin(uap->timeout, &ts, sizeof(ts)); 2090 if (error) 2091 return error; 2092 2093 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000)) 2094 return (EINVAL); 2095 2096 TIMESPEC_TO_TIMEVAL(&atv, &ts); 2097 if (itimerfix(&atv)) 2098 return (EINVAL); 2099 timo = tvtohz_high(&atv); 2100 } 2101 2102 ki = p->p_aioinfo; 2103 if (ki == NULL) 2104 return EAGAIN; 2105 2106 for (;;) { 2107 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) { 2108 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2109 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2110 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2111 p->p_stats->p_ru.ru_oublock += 2112 cb->outputcharge; 2113 cb->outputcharge = 0; 2114 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2115 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 2116 cb->inputcharge = 0; 2117 } 2118 aio_free_entry(cb); 2119 return cb->uaiocb._aiocb_private.error; 2120 } 2121 2122 crit_enter(); 2123 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) { 2124 crit_exit(); 2125 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2126 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2127 aio_free_entry(cb); 2128 return cb->uaiocb._aiocb_private.error; 2129 } 2130 2131 ki->kaio_flags |= KAIO_WAKEUP; 2132 error = tsleep(p, PCATCH, "aiowc", timo); 2133 crit_exit(); 2134 2135 if (error == ERESTART) 2136 return EINTR; 2137 else if (error < 0) 2138 return error; 2139 else if (error == EINTR) 2140 return EINTR; 2141 else if (error == EWOULDBLOCK) 2142 return EAGAIN; 2143 } 2144 #endif /* VFS_AIO */ 2145 } 2146 2147 #ifndef VFS_AIO 2148 static int 2149 filt_aioattach(struct knote *kn) 2150 { 2151 2152 return (ENXIO); 2153 } 2154 2155 struct filterops aio_filtops = 2156 { 0, filt_aioattach, NULL, NULL }; 2157 2158 #else 2159 /* kqueue attach function */ 2160 static int 2161 filt_aioattach(struct knote *kn) 2162 { 2163 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2164 2165 /* 2166 * The aiocbe pointer must be validated before using it, so 2167 * registration is restricted to the kernel; the user cannot 2168 * set EV_FLAG1. 2169 */ 2170 if ((kn->kn_flags & EV_FLAG1) == 0) 2171 return (EPERM); 2172 kn->kn_flags &= ~EV_FLAG1; 2173 2174 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext); 2175 2176 return (0); 2177 } 2178 2179 /* kqueue detach function */ 2180 static void 2181 filt_aiodetach(struct knote *kn) 2182 { 2183 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2184 2185 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext); 2186 } 2187 2188 /* kqueue filter function */ 2189 /*ARGSUSED*/ 2190 static int 2191 filt_aio(struct knote *kn, long hint) 2192 { 2193 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2194 2195 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2196 if (aiocbe->jobstate != JOBST_JOBFINISHED && 2197 aiocbe->jobstate != JOBST_JOBBFINISHED) 2198 return (0); 2199 kn->kn_flags |= EV_EOF; 2200 return (1); 2201 } 2202 2203 struct filterops aio_filtops = 2204 { 0, filt_aioattach, filt_aiodetach, filt_aio }; 2205 #endif /* VFS_AIO */ 2206