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 1.19 2005/10/11 09:59:56 corecode 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 buf *bp); 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, curthread); 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 printf("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, FOF_OFFSET, mytd); 584 } else { 585 auio.uio_rw = UIO_WRITE; 586 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, mytd); 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 psignal(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 psignal(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 psignal(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 printf("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 printf("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; 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 cpu_set_fork_handler(np, aio_daemon, curproc); 882 start_forked_proc(lp, np); 883 884 /* 885 * Wait until daemon is started, but continue on just in case to 886 * handle error conditions. 887 */ 888 error = tsleep(np, 0, "aiosta", aiod_timeout); 889 num_aio_procs++; 890 891 return error; 892 } 893 894 /* 895 * Try the high-performance, low-overhead physio method for eligible 896 * VCHR devices. This method doesn't use an aio helper thread, and 897 * thus has very low overhead. 898 * 899 * Assumes that the caller, _aio_aqueue(), has incremented the file 900 * structure's reference count, preventing its deallocation for the 901 * duration of this call. 902 */ 903 static int 904 aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 905 { 906 int error; 907 struct aiocb *cb; 908 struct file *fp; 909 struct buf *bp; 910 struct vnode *vp; 911 struct kaioinfo *ki; 912 struct aio_liojob *lj; 913 int notify; 914 915 cb = &aiocbe->uaiocb; 916 fp = aiocbe->fd_file; 917 918 if (fp->f_type != DTYPE_VNODE) 919 return (-1); 920 921 vp = (struct vnode *)fp->f_data; 922 923 /* 924 * If its not a disk, we don't want to return a positive error. 925 * It causes the aio code to not fall through to try the thread 926 * way when you're talking to a regular file. 927 */ 928 if (!vn_isdisk(vp, &error)) { 929 if (error == ENOTBLK) 930 return (-1); 931 else 932 return (error); 933 } 934 935 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys) 936 return (-1); 937 938 if (cb->aio_nbytes > 939 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 940 return (-1); 941 942 ki = p->p_aioinfo; 943 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 944 return (-1); 945 946 ki->kaio_buffer_count++; 947 948 lj = aiocbe->lio; 949 if (lj) 950 lj->lioj_buffer_count++; 951 952 /* Create and build a buffer header for a transfer. */ 953 bp = (struct buf *)getpbuf(NULL); 954 BUF_KERNPROC(bp); 955 956 /* 957 * Get a copy of the kva from the physical buffer. 958 */ 959 bp->b_caller1 = p; 960 bp->b_dev = vp->v_rdev; 961 error = 0; 962 963 bp->b_bcount = cb->aio_nbytes; 964 bp->b_bufsize = cb->aio_nbytes; 965 bp->b_flags = B_PHYS | (cb->aio_lio_opcode == LIO_WRITE ? 966 B_WRITE : B_READ); 967 bp->b_iodone = aio_physwakeup; 968 bp->b_saveaddr = bp->b_data; 969 bp->b_data = (void *)(uintptr_t)cb->aio_buf; 970 bp->b_blkno = btodb(cb->aio_offset); 971 972 /* Bring buffer into kernel space. */ 973 if (vmapbuf(bp) < 0) { 974 error = EFAULT; 975 goto doerror; 976 } 977 978 crit_enter(); 979 980 aiocbe->bp = bp; 981 bp->b_spc = (void *)aiocbe; 982 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list); 983 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 984 aiocbe->jobstate = JOBST_JOBQBUF; 985 cb->_aiocb_private.status = cb->aio_nbytes; 986 num_buf_aio++; 987 bp->b_error = 0; 988 989 crit_exit(); 990 991 /* Perform transfer. */ 992 BUF_STRATEGY(bp, 0); 993 994 notify = 0; 995 crit_enter(); 996 997 /* 998 * If we had an error invoking the request, or an error in processing 999 * the request before we have returned, we process it as an error in 1000 * transfer. Note that such an I/O error is not indicated immediately, 1001 * but is returned using the aio_error mechanism. In this case, 1002 * aio_suspend will return immediately. 1003 */ 1004 if (bp->b_error || (bp->b_flags & B_ERROR)) { 1005 struct aiocb *job = aiocbe->uuaiocb; 1006 1007 aiocbe->uaiocb._aiocb_private.status = 0; 1008 suword(&job->_aiocb_private.status, 0); 1009 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 1010 suword(&job->_aiocb_private.error, bp->b_error); 1011 1012 ki->kaio_buffer_finished_count++; 1013 1014 if (aiocbe->jobstate != JOBST_JOBBFINISHED) { 1015 aiocbe->jobstate = JOBST_JOBBFINISHED; 1016 aiocbe->jobflags |= AIOCBLIST_DONE; 1017 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 1018 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 1019 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 1020 notify = 1; 1021 } 1022 } 1023 crit_exit(); 1024 if (notify) 1025 KNOTE(&aiocbe->klist, 0); 1026 return 0; 1027 1028 doerror: 1029 ki->kaio_buffer_count--; 1030 if (lj) 1031 lj->lioj_buffer_count--; 1032 aiocbe->bp = NULL; 1033 relpbuf(bp, NULL); 1034 return error; 1035 } 1036 1037 /* 1038 * This waits/tests physio completion. 1039 */ 1040 static int 1041 aio_fphysio(struct aiocblist *iocb) 1042 { 1043 struct buf *bp; 1044 int error; 1045 1046 bp = iocb->bp; 1047 1048 crit_enter(); 1049 while ((bp->b_flags & B_DONE) == 0) { 1050 if (tsleep(bp, 0, "physstr", aiod_timeout)) { 1051 if ((bp->b_flags & B_DONE) == 0) { 1052 crit_exit(); 1053 return EINPROGRESS; 1054 } else 1055 break; 1056 } 1057 } 1058 crit_exit(); 1059 1060 /* Release mapping into kernel space. */ 1061 vunmapbuf(bp); 1062 iocb->bp = 0; 1063 1064 error = 0; 1065 1066 /* Check for an error. */ 1067 if (bp->b_flags & B_ERROR) 1068 error = bp->b_error; 1069 1070 relpbuf(bp, NULL); 1071 return (error); 1072 } 1073 #endif /* VFS_AIO */ 1074 1075 /* 1076 * Wake up aio requests that may be serviceable now. 1077 */ 1078 void 1079 aio_swake(struct socket *so, struct sockbuf *sb) 1080 { 1081 #ifndef VFS_AIO 1082 return; 1083 #else 1084 struct aiocblist *cb,*cbn; 1085 struct proc *p; 1086 struct kaioinfo *ki = NULL; 1087 int opcode, wakecount = 0; 1088 struct aioproclist *aiop; 1089 1090 if (sb == &so->so_snd) { 1091 opcode = LIO_WRITE; 1092 so->so_snd.sb_flags &= ~SB_AIO; 1093 } else { 1094 opcode = LIO_READ; 1095 so->so_rcv.sb_flags &= ~SB_AIO; 1096 } 1097 1098 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) { 1099 cbn = TAILQ_NEXT(cb, list); 1100 if (opcode == cb->uaiocb.aio_lio_opcode) { 1101 p = cb->userproc; 1102 ki = p->p_aioinfo; 1103 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1104 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist); 1105 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1106 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist); 1107 wakecount++; 1108 if (cb->jobstate != JOBST_JOBQGLOBAL) 1109 panic("invalid queue value"); 1110 } 1111 } 1112 1113 while (wakecount--) { 1114 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) { 1115 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1116 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1117 aiop->aioprocflags &= ~AIOP_FREE; 1118 wakeup(aiop->aioproc); 1119 } 1120 } 1121 #endif /* VFS_AIO */ 1122 } 1123 1124 #ifdef VFS_AIO 1125 /* 1126 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1127 * technique is done in this code. 1128 */ 1129 static int 1130 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type) 1131 { 1132 struct proc *p = curproc; 1133 struct filedesc *fdp; 1134 struct file *fp; 1135 unsigned int fd; 1136 struct socket *so; 1137 int error; 1138 int opcode, user_opcode; 1139 struct aiocblist *aiocbe; 1140 struct aioproclist *aiop; 1141 struct kaioinfo *ki; 1142 struct kevent kev; 1143 struct kqueue *kq; 1144 struct file *kq_fp; 1145 1146 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL) 1147 TAILQ_REMOVE(&aio_freejobs, aiocbe, list); 1148 else 1149 aiocbe = zalloc (aiocb_zone); 1150 1151 aiocbe->inputcharge = 0; 1152 aiocbe->outputcharge = 0; 1153 callout_init(&aiocbe->timeout); 1154 SLIST_INIT(&aiocbe->klist); 1155 1156 suword(&job->_aiocb_private.status, -1); 1157 suword(&job->_aiocb_private.error, 0); 1158 suword(&job->_aiocb_private.kernelinfo, -1); 1159 1160 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb)); 1161 if (error) { 1162 suword(&job->_aiocb_private.error, error); 1163 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1164 return error; 1165 } 1166 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL && 1167 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1168 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1169 return EINVAL; 1170 } 1171 1172 /* Save userspace address of the job info. */ 1173 aiocbe->uuaiocb = job; 1174 1175 /* Get the opcode. */ 1176 user_opcode = aiocbe->uaiocb.aio_lio_opcode; 1177 if (type != LIO_NOP) 1178 aiocbe->uaiocb.aio_lio_opcode = type; 1179 opcode = aiocbe->uaiocb.aio_lio_opcode; 1180 1181 /* Get the fd info for process. */ 1182 fdp = p->p_fd; 1183 1184 /* 1185 * Range check file descriptor. 1186 */ 1187 fd = aiocbe->uaiocb.aio_fildes; 1188 if (fd >= fdp->fd_nfiles) { 1189 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1190 if (type == 0) 1191 suword(&job->_aiocb_private.error, EBADF); 1192 return EBADF; 1193 } 1194 1195 fp = aiocbe->fd_file = fdp->fd_files[fd].fp; 1196 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 1197 0))) { 1198 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1199 if (type == 0) 1200 suword(&job->_aiocb_private.error, EBADF); 1201 return EBADF; 1202 } 1203 fhold(fp); 1204 1205 if (aiocbe->uaiocb.aio_offset == -1LL) { 1206 error = EINVAL; 1207 goto aqueue_fail; 1208 } 1209 error = suword(&job->_aiocb_private.kernelinfo, jobrefid); 1210 if (error) { 1211 error = EINVAL; 1212 goto aqueue_fail; 1213 } 1214 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid; 1215 if (jobrefid == LONG_MAX) 1216 jobrefid = 1; 1217 else 1218 jobrefid++; 1219 1220 if (opcode == LIO_NOP) { 1221 fdrop(fp, p->p_thread); 1222 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1223 if (type == 0) { 1224 suword(&job->_aiocb_private.error, 0); 1225 suword(&job->_aiocb_private.status, 0); 1226 suword(&job->_aiocb_private.kernelinfo, 0); 1227 } 1228 return 0; 1229 } 1230 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) { 1231 if (type == 0) 1232 suword(&job->_aiocb_private.status, 0); 1233 error = EINVAL; 1234 goto aqueue_fail; 1235 } 1236 1237 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) { 1238 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1239 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr; 1240 } 1241 else { 1242 /* 1243 * This method for requesting kevent-based notification won't 1244 * work on the alpha, since we're passing in a pointer 1245 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT- 1246 * based method instead. 1247 */ 1248 if (user_opcode == LIO_NOP || user_opcode == LIO_READ || 1249 user_opcode == LIO_WRITE) 1250 goto no_kqueue; 1251 1252 error = copyin((struct kevent *)(uintptr_t)user_opcode, 1253 &kev, sizeof(kev)); 1254 if (error) 1255 goto aqueue_fail; 1256 } 1257 if ((u_int)kev.ident >= fdp->fd_nfiles || 1258 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL || 1259 (kq_fp->f_type != DTYPE_KQUEUE)) { 1260 error = EBADF; 1261 goto aqueue_fail; 1262 } 1263 kq = (struct kqueue *)kq_fp->f_data; 1264 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1265 kev.filter = EVFILT_AIO; 1266 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 1267 kev.data = (intptr_t)aiocbe; 1268 error = kqueue_register(kq, &kev, p->p_thread); 1269 aqueue_fail: 1270 if (error) { 1271 fdrop(fp, p->p_thread); 1272 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1273 if (type == 0) 1274 suword(&job->_aiocb_private.error, error); 1275 goto done; 1276 } 1277 no_kqueue: 1278 1279 suword(&job->_aiocb_private.error, EINPROGRESS); 1280 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1281 aiocbe->userproc = p; 1282 aiocbe->jobflags = 0; 1283 aiocbe->lio = lj; 1284 ki = p->p_aioinfo; 1285 1286 if (fp->f_type == DTYPE_SOCKET) { 1287 /* 1288 * Alternate queueing for socket ops: Reach down into the 1289 * descriptor to get the socket data. Then check to see if the 1290 * socket is ready to be read or written (based on the requested 1291 * operation). 1292 * 1293 * If it is not ready for io, then queue the aiocbe on the 1294 * socket, and set the flags so we get a call when sbnotify() 1295 * happens. 1296 */ 1297 so = (struct socket *)fp->f_data; 1298 crit_enter(); 1299 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1300 LIO_WRITE) && (!sowriteable(so)))) { 1301 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1302 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist); 1303 if (opcode == LIO_READ) 1304 so->so_rcv.sb_flags |= SB_AIO; 1305 else 1306 so->so_snd.sb_flags |= SB_AIO; 1307 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */ 1308 ki->kaio_queue_count++; 1309 num_queue_count++; 1310 crit_exit(); 1311 error = 0; 1312 goto done; 1313 } 1314 crit_exit(); 1315 } 1316 1317 if ((error = aio_qphysio(p, aiocbe)) == 0) 1318 goto done; 1319 if (error > 0) { 1320 suword(&job->_aiocb_private.status, 0); 1321 aiocbe->uaiocb._aiocb_private.error = error; 1322 suword(&job->_aiocb_private.error, error); 1323 goto done; 1324 } 1325 1326 /* No buffer for daemon I/O. */ 1327 aiocbe->bp = NULL; 1328 1329 ki->kaio_queue_count++; 1330 if (lj) 1331 lj->lioj_queue_count++; 1332 crit_enter(); 1333 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1334 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1335 crit_exit(); 1336 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1337 1338 num_queue_count++; 1339 error = 0; 1340 1341 /* 1342 * If we don't have a free AIO process, and we are below our quota, then 1343 * start one. Otherwise, depend on the subsequent I/O completions to 1344 * pick-up this job. If we don't successfully create the new process 1345 * (thread) due to resource issues, we return an error for now (EAGAIN), 1346 * which is likely not the correct thing to do. 1347 */ 1348 crit_enter(); 1349 retryproc: 1350 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1351 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1352 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1353 aiop->aioprocflags &= ~AIOP_FREE; 1354 wakeup(aiop->aioproc); 1355 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1356 ((ki->kaio_active_count + num_aio_resv_start) < 1357 ki->kaio_maxactive_count)) { 1358 num_aio_resv_start++; 1359 if ((error = aio_newproc()) == 0) { 1360 num_aio_resv_start--; 1361 goto retryproc; 1362 } 1363 num_aio_resv_start--; 1364 } 1365 crit_exit(); 1366 done: 1367 return error; 1368 } 1369 1370 /* 1371 * This routine queues an AIO request, checking for quotas. 1372 */ 1373 static int 1374 aio_aqueue(struct aiocb *job, int type) 1375 { 1376 struct proc *p = curproc; 1377 struct kaioinfo *ki; 1378 1379 if (p->p_aioinfo == NULL) 1380 aio_init_aioinfo(p); 1381 1382 if (num_queue_count >= max_queue_count) 1383 return EAGAIN; 1384 1385 ki = p->p_aioinfo; 1386 if (ki->kaio_queue_count >= ki->kaio_qallowed_count) 1387 return EAGAIN; 1388 1389 return _aio_aqueue(job, NULL, type); 1390 } 1391 #endif /* VFS_AIO */ 1392 1393 /* 1394 * Support the aio_return system call, as a side-effect, kernel resources are 1395 * released. 1396 */ 1397 int 1398 aio_return(struct aio_return_args *uap) 1399 { 1400 #ifndef VFS_AIO 1401 return ENOSYS; 1402 #else 1403 struct proc *p = curproc; 1404 long jobref; 1405 struct aiocblist *cb, *ncb; 1406 struct aiocb *ujob; 1407 struct kaioinfo *ki; 1408 1409 ki = p->p_aioinfo; 1410 if (ki == NULL) 1411 return EINVAL; 1412 1413 ujob = uap->aiocbp; 1414 1415 jobref = fuword(&ujob->_aiocb_private.kernelinfo); 1416 if (jobref == -1 || jobref == 0) 1417 return EINVAL; 1418 1419 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1420 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == 1421 jobref) { 1422 if (ujob == cb->uuaiocb) { 1423 uap->sysmsg_result = 1424 cb->uaiocb._aiocb_private.status; 1425 } else 1426 uap->sysmsg_result = EFAULT; 1427 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1428 p->p_stats->p_ru.ru_oublock += 1429 cb->outputcharge; 1430 cb->outputcharge = 0; 1431 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1432 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 1433 cb->inputcharge = 0; 1434 } 1435 aio_free_entry(cb); 1436 return 0; 1437 } 1438 } 1439 crit_enter(); 1440 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) { 1441 ncb = TAILQ_NEXT(cb, plist); 1442 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) 1443 == jobref) { 1444 crit_exit(); 1445 if (ujob == cb->uuaiocb) { 1446 uap->sysmsg_result = 1447 cb->uaiocb._aiocb_private.status; 1448 } else 1449 uap->sysmsg_result = EFAULT; 1450 aio_free_entry(cb); 1451 return 0; 1452 } 1453 } 1454 crit_exit(); 1455 1456 return (EINVAL); 1457 #endif /* VFS_AIO */ 1458 } 1459 1460 /* 1461 * Allow a process to wakeup when any of the I/O requests are completed. 1462 */ 1463 int 1464 aio_suspend(struct aio_suspend_args *uap) 1465 { 1466 #ifndef VFS_AIO 1467 return ENOSYS; 1468 #else 1469 struct proc *p = curproc; 1470 struct timeval atv; 1471 struct timespec ts; 1472 struct aiocb *const *cbptr, *cbp; 1473 struct kaioinfo *ki; 1474 struct aiocblist *cb; 1475 int i; 1476 int njoblist; 1477 int error, timo; 1478 long *ijoblist; 1479 struct aiocb **ujoblist; 1480 1481 if (uap->nent > AIO_LISTIO_MAX) 1482 return EINVAL; 1483 1484 timo = 0; 1485 if (uap->timeout) { 1486 /* Get timespec struct. */ 1487 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1488 return error; 1489 1490 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) 1491 return (EINVAL); 1492 1493 TIMESPEC_TO_TIMEVAL(&atv, &ts); 1494 if (itimerfix(&atv)) 1495 return (EINVAL); 1496 timo = tvtohz_high(&atv); 1497 } 1498 1499 ki = p->p_aioinfo; 1500 if (ki == NULL) 1501 return EAGAIN; 1502 1503 njoblist = 0; 1504 ijoblist = zalloc(aiol_zone); 1505 ujoblist = zalloc(aiol_zone); 1506 cbptr = uap->aiocbp; 1507 1508 for (i = 0; i < uap->nent; i++) { 1509 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1510 if (cbp == 0) 1511 continue; 1512 ujoblist[njoblist] = cbp; 1513 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo); 1514 njoblist++; 1515 } 1516 1517 if (njoblist == 0) { 1518 zfree(aiol_zone, ijoblist); 1519 zfree(aiol_zone, ujoblist); 1520 return 0; 1521 } 1522 1523 error = 0; 1524 for (;;) { 1525 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1526 for (i = 0; i < njoblist; i++) { 1527 if (((intptr_t) 1528 cb->uaiocb._aiocb_private.kernelinfo) == 1529 ijoblist[i]) { 1530 if (ujoblist[i] != cb->uuaiocb) 1531 error = EINVAL; 1532 zfree(aiol_zone, ijoblist); 1533 zfree(aiol_zone, ujoblist); 1534 return error; 1535 } 1536 } 1537 } 1538 1539 crit_enter(); 1540 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = 1541 TAILQ_NEXT(cb, plist)) { 1542 for (i = 0; i < njoblist; i++) { 1543 if (((intptr_t) 1544 cb->uaiocb._aiocb_private.kernelinfo) == 1545 ijoblist[i]) { 1546 crit_exit(); 1547 if (ujoblist[i] != cb->uuaiocb) 1548 error = EINVAL; 1549 zfree(aiol_zone, ijoblist); 1550 zfree(aiol_zone, ujoblist); 1551 return error; 1552 } 1553 } 1554 } 1555 1556 ki->kaio_flags |= KAIO_WAKEUP; 1557 error = tsleep(p, PCATCH, "aiospn", timo); 1558 crit_exit(); 1559 1560 if (error == ERESTART || error == EINTR) { 1561 zfree(aiol_zone, ijoblist); 1562 zfree(aiol_zone, ujoblist); 1563 return EINTR; 1564 } else if (error == EWOULDBLOCK) { 1565 zfree(aiol_zone, ijoblist); 1566 zfree(aiol_zone, ujoblist); 1567 return EAGAIN; 1568 } 1569 } 1570 1571 /* NOTREACHED */ 1572 return EINVAL; 1573 #endif /* VFS_AIO */ 1574 } 1575 1576 /* 1577 * aio_cancel cancels any non-physio aio operations not currently in 1578 * progress. 1579 */ 1580 int 1581 aio_cancel(struct aio_cancel_args *uap) 1582 { 1583 #ifndef VFS_AIO 1584 return ENOSYS; 1585 #else 1586 struct proc *p = curproc; 1587 struct kaioinfo *ki; 1588 struct aiocblist *cbe, *cbn; 1589 struct file *fp; 1590 struct filedesc *fdp; 1591 struct socket *so; 1592 struct proc *po; 1593 int error; 1594 int cancelled=0; 1595 int notcancelled=0; 1596 struct vnode *vp; 1597 1598 fdp = p->p_fd; 1599 if ((u_int)uap->fd >= fdp->fd_nfiles || 1600 (fp = fdp->fd_files[uap->fd].fp) == NULL) 1601 return (EBADF); 1602 1603 if (fp->f_type == DTYPE_VNODE) { 1604 vp = (struct vnode *)fp->f_data; 1605 1606 if (vn_isdisk(vp,&error)) { 1607 uap->sysmsg_result = AIO_NOTCANCELED; 1608 return 0; 1609 } 1610 } else if (fp->f_type == DTYPE_SOCKET) { 1611 so = (struct socket *)fp->f_data; 1612 1613 crit_enter(); 1614 1615 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) { 1616 cbn = TAILQ_NEXT(cbe, list); 1617 if ((uap->aiocbp == NULL) || 1618 (uap->aiocbp == cbe->uuaiocb) ) { 1619 po = cbe->userproc; 1620 ki = po->p_aioinfo; 1621 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 1622 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist); 1623 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist); 1624 if (ki->kaio_flags & KAIO_WAKEUP) { 1625 wakeup(po); 1626 } 1627 cbe->jobstate = JOBST_JOBFINISHED; 1628 cbe->uaiocb._aiocb_private.status=-1; 1629 cbe->uaiocb._aiocb_private.error=ECANCELED; 1630 cancelled++; 1631 /* XXX cancelled, knote? */ 1632 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1633 SIGEV_SIGNAL) 1634 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1635 if (uap->aiocbp) 1636 break; 1637 } 1638 } 1639 crit_exit(); 1640 1641 if ((cancelled) && (uap->aiocbp)) { 1642 uap->sysmsg_result = AIO_CANCELED; 1643 return 0; 1644 } 1645 } 1646 ki=p->p_aioinfo; 1647 if (ki == NULL) 1648 goto done; 1649 crit_enter(); 1650 1651 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) { 1652 cbn = TAILQ_NEXT(cbe, plist); 1653 1654 if ((uap->fd == cbe->uaiocb.aio_fildes) && 1655 ((uap->aiocbp == NULL ) || 1656 (uap->aiocbp == cbe->uuaiocb))) { 1657 1658 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 1659 TAILQ_REMOVE(&aio_jobs, cbe, list); 1660 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 1661 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, 1662 plist); 1663 cancelled++; 1664 ki->kaio_queue_finished_count++; 1665 cbe->jobstate = JOBST_JOBFINISHED; 1666 cbe->uaiocb._aiocb_private.status = -1; 1667 cbe->uaiocb._aiocb_private.error = ECANCELED; 1668 /* XXX cancelled, knote? */ 1669 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1670 SIGEV_SIGNAL) 1671 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1672 } else { 1673 notcancelled++; 1674 } 1675 } 1676 } 1677 crit_exit(); 1678 done: 1679 if (notcancelled) { 1680 uap->sysmsg_result = AIO_NOTCANCELED; 1681 return 0; 1682 } 1683 if (cancelled) { 1684 uap->sysmsg_result = AIO_CANCELED; 1685 return 0; 1686 } 1687 uap->sysmsg_result = AIO_ALLDONE; 1688 1689 return 0; 1690 #endif /* VFS_AIO */ 1691 } 1692 1693 /* 1694 * aio_error is implemented in the kernel level for compatibility purposes only. 1695 * For a user mode async implementation, it would be best to do it in a userland 1696 * subroutine. 1697 */ 1698 int 1699 aio_error(struct aio_error_args *uap) 1700 { 1701 #ifndef VFS_AIO 1702 return ENOSYS; 1703 #else 1704 struct proc *p = curproc; 1705 struct aiocblist *cb; 1706 struct kaioinfo *ki; 1707 long jobref; 1708 1709 ki = p->p_aioinfo; 1710 if (ki == NULL) 1711 return EINVAL; 1712 1713 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo); 1714 if ((jobref == -1) || (jobref == 0)) 1715 return EINVAL; 1716 1717 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1718 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1719 jobref) { 1720 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1721 return 0; 1722 } 1723 } 1724 1725 crit_enter(); 1726 1727 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb, 1728 plist)) { 1729 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1730 jobref) { 1731 uap->sysmsg_result = EINPROGRESS; 1732 crit_exit(); 1733 return 0; 1734 } 1735 } 1736 1737 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb, 1738 plist)) { 1739 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1740 jobref) { 1741 uap->sysmsg_result = EINPROGRESS; 1742 crit_exit(); 1743 return 0; 1744 } 1745 } 1746 crit_exit(); 1747 1748 crit_enter(); 1749 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb, 1750 plist)) { 1751 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1752 jobref) { 1753 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1754 crit_exit(); 1755 return 0; 1756 } 1757 } 1758 1759 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb, 1760 plist)) { 1761 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1762 jobref) { 1763 uap->sysmsg_result = EINPROGRESS; 1764 crit_exit(); 1765 return 0; 1766 } 1767 } 1768 crit_exit(); 1769 1770 #if (0) 1771 /* 1772 * Hack for lio. 1773 */ 1774 status = fuword(&uap->aiocbp->_aiocb_private.status); 1775 if (status == -1) 1776 return fuword(&uap->aiocbp->_aiocb_private.error); 1777 #endif 1778 return EINVAL; 1779 #endif /* VFS_AIO */ 1780 } 1781 1782 /* syscall - asynchronous read from a file (REALTIME) */ 1783 int 1784 aio_read(struct aio_read_args *uap) 1785 { 1786 #ifndef VFS_AIO 1787 return ENOSYS; 1788 #else 1789 return aio_aqueue(uap->aiocbp, LIO_READ); 1790 #endif /* VFS_AIO */ 1791 } 1792 1793 /* syscall - asynchronous write to a file (REALTIME) */ 1794 int 1795 aio_write(struct aio_write_args *uap) 1796 { 1797 #ifndef VFS_AIO 1798 return ENOSYS; 1799 #else 1800 return aio_aqueue(uap->aiocbp, LIO_WRITE); 1801 #endif /* VFS_AIO */ 1802 } 1803 1804 /* syscall - XXX undocumented */ 1805 int 1806 lio_listio(struct lio_listio_args *uap) 1807 { 1808 #ifndef VFS_AIO 1809 return ENOSYS; 1810 #else 1811 struct proc *p = curproc; 1812 int nent, nentqueued; 1813 struct aiocb *iocb, * const *cbptr; 1814 struct aiocblist *cb; 1815 struct kaioinfo *ki; 1816 struct aio_liojob *lj; 1817 int error, runningcode; 1818 int nerror; 1819 int i; 1820 1821 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 1822 return EINVAL; 1823 1824 nent = uap->nent; 1825 if (nent > AIO_LISTIO_MAX) 1826 return EINVAL; 1827 1828 if (p->p_aioinfo == NULL) 1829 aio_init_aioinfo(p); 1830 1831 if ((nent + num_queue_count) > max_queue_count) 1832 return EAGAIN; 1833 1834 ki = p->p_aioinfo; 1835 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) 1836 return EAGAIN; 1837 1838 lj = zalloc(aiolio_zone); 1839 if (!lj) 1840 return EAGAIN; 1841 1842 lj->lioj_flags = 0; 1843 lj->lioj_buffer_count = 0; 1844 lj->lioj_buffer_finished_count = 0; 1845 lj->lioj_queue_count = 0; 1846 lj->lioj_queue_finished_count = 0; 1847 lj->lioj_ki = ki; 1848 1849 /* 1850 * Setup signal. 1851 */ 1852 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 1853 error = copyin(uap->sig, &lj->lioj_signal, 1854 sizeof(lj->lioj_signal)); 1855 if (error) { 1856 zfree(aiolio_zone, lj); 1857 return error; 1858 } 1859 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 1860 zfree(aiolio_zone, lj); 1861 return EINVAL; 1862 } 1863 lj->lioj_flags |= LIOJ_SIGNAL; 1864 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED; 1865 } else 1866 lj->lioj_flags &= ~LIOJ_SIGNAL; 1867 1868 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 1869 /* 1870 * Get pointers to the list of I/O requests. 1871 */ 1872 nerror = 0; 1873 nentqueued = 0; 1874 cbptr = uap->acb_list; 1875 for (i = 0; i < uap->nent; i++) { 1876 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1877 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) { 1878 error = _aio_aqueue(iocb, lj, 0); 1879 if (error == 0) 1880 nentqueued++; 1881 else 1882 nerror++; 1883 } 1884 } 1885 1886 /* 1887 * If we haven't queued any, then just return error. 1888 */ 1889 if (nentqueued == 0) 1890 return 0; 1891 1892 /* 1893 * Calculate the appropriate error return. 1894 */ 1895 runningcode = 0; 1896 if (nerror) 1897 runningcode = EIO; 1898 1899 if (uap->mode == LIO_WAIT) { 1900 int command, found, jobref; 1901 1902 for (;;) { 1903 found = 0; 1904 for (i = 0; i < uap->nent; i++) { 1905 /* 1906 * Fetch address of the control buf pointer in 1907 * user space. 1908 */ 1909 iocb = (struct aiocb *) 1910 (intptr_t)fuword(&cbptr[i]); 1911 if (((intptr_t)iocb == -1) || ((intptr_t)iocb 1912 == 0)) 1913 continue; 1914 1915 /* 1916 * Fetch the associated command from user space. 1917 */ 1918 command = fuword(&iocb->aio_lio_opcode); 1919 if (command == LIO_NOP) { 1920 found++; 1921 continue; 1922 } 1923 1924 jobref = fuword(&iocb->_aiocb_private.kernelinfo); 1925 1926 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1927 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1928 == jobref) { 1929 if (cb->uaiocb.aio_lio_opcode 1930 == LIO_WRITE) { 1931 p->p_stats->p_ru.ru_oublock 1932 += 1933 cb->outputcharge; 1934 cb->outputcharge = 0; 1935 } else if (cb->uaiocb.aio_lio_opcode 1936 == LIO_READ) { 1937 p->p_stats->p_ru.ru_inblock 1938 += cb->inputcharge; 1939 cb->inputcharge = 0; 1940 } 1941 found++; 1942 break; 1943 } 1944 } 1945 1946 crit_enter(); 1947 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) { 1948 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1949 == jobref) { 1950 found++; 1951 break; 1952 } 1953 } 1954 crit_exit(); 1955 } 1956 1957 /* 1958 * If all I/Os have been disposed of, then we can 1959 * return. 1960 */ 1961 if (found == nentqueued) 1962 return runningcode; 1963 1964 ki->kaio_flags |= KAIO_WAKEUP; 1965 error = tsleep(p, PCATCH, "aiospn", 0); 1966 1967 if (error == EINTR) 1968 return EINTR; 1969 else if (error == EWOULDBLOCK) 1970 return EAGAIN; 1971 } 1972 } 1973 1974 return runningcode; 1975 #endif /* VFS_AIO */ 1976 } 1977 1978 #ifdef VFS_AIO 1979 /* 1980 * This is a weird hack so that we can post a signal. It is safe to do so from 1981 * a timeout routine, but *not* from an interrupt routine. 1982 */ 1983 static void 1984 process_signal(void *aioj) 1985 { 1986 struct aiocblist *aiocbe = aioj; 1987 struct aio_liojob *lj = aiocbe->lio; 1988 struct aiocb *cb = &aiocbe->uaiocb; 1989 1990 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) && 1991 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) { 1992 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo); 1993 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 1994 } 1995 1996 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) 1997 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo); 1998 } 1999 2000 /* 2001 * Interrupt handler for physio, performs the necessary process wakeups, and 2002 * signals. 2003 */ 2004 static void 2005 aio_physwakeup(struct buf *bp) 2006 { 2007 struct aiocblist *aiocbe; 2008 struct proc *p; 2009 struct kaioinfo *ki; 2010 struct aio_liojob *lj; 2011 2012 wakeup(bp); 2013 2014 aiocbe = (struct aiocblist *)bp->b_spc; 2015 if (aiocbe) { 2016 p = bp->b_caller1; 2017 2018 aiocbe->jobstate = JOBST_JOBBFINISHED; 2019 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 2020 aiocbe->uaiocb._aiocb_private.error = 0; 2021 aiocbe->jobflags |= AIOCBLIST_DONE; 2022 2023 if (bp->b_flags & B_ERROR) 2024 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 2025 2026 lj = aiocbe->lio; 2027 if (lj) { 2028 lj->lioj_buffer_finished_count++; 2029 2030 /* 2031 * wakeup/signal if all of the interrupt jobs are done. 2032 */ 2033 if (lj->lioj_buffer_finished_count == 2034 lj->lioj_buffer_count) { 2035 /* 2036 * Post a signal if it is called for. 2037 */ 2038 if ((lj->lioj_flags & 2039 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == 2040 LIOJ_SIGNAL) { 2041 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2042 callout_reset(&aiocbe->timeout, 0, 2043 process_signal, aiocbe); 2044 } 2045 } 2046 } 2047 2048 ki = p->p_aioinfo; 2049 if (ki) { 2050 ki->kaio_buffer_finished_count++; 2051 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 2052 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 2053 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 2054 2055 KNOTE(&aiocbe->klist, 0); 2056 /* Do the wakeup. */ 2057 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) { 2058 ki->kaio_flags &= ~KAIO_WAKEUP; 2059 wakeup(p); 2060 } 2061 } 2062 2063 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 2064 callout_reset(&aiocbe->timeout, 0, 2065 process_signal, aiocbe); 2066 } 2067 } 2068 } 2069 #endif /* VFS_AIO */ 2070 2071 /* syscall - wait for the next completion of an aio request */ 2072 int 2073 aio_waitcomplete(struct aio_waitcomplete_args *uap) 2074 { 2075 #ifndef VFS_AIO 2076 return ENOSYS; 2077 #else 2078 struct proc *p = curproc; 2079 struct timeval atv; 2080 struct timespec ts; 2081 struct kaioinfo *ki; 2082 struct aiocblist *cb = NULL; 2083 int error, timo; 2084 2085 suword(uap->aiocbp, (int)NULL); 2086 2087 timo = 0; 2088 if (uap->timeout) { 2089 /* Get timespec struct. */ 2090 error = copyin(uap->timeout, &ts, sizeof(ts)); 2091 if (error) 2092 return error; 2093 2094 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000)) 2095 return (EINVAL); 2096 2097 TIMESPEC_TO_TIMEVAL(&atv, &ts); 2098 if (itimerfix(&atv)) 2099 return (EINVAL); 2100 timo = tvtohz_high(&atv); 2101 } 2102 2103 ki = p->p_aioinfo; 2104 if (ki == NULL) 2105 return EAGAIN; 2106 2107 for (;;) { 2108 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) { 2109 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2110 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2111 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2112 p->p_stats->p_ru.ru_oublock += 2113 cb->outputcharge; 2114 cb->outputcharge = 0; 2115 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2116 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 2117 cb->inputcharge = 0; 2118 } 2119 aio_free_entry(cb); 2120 return cb->uaiocb._aiocb_private.error; 2121 } 2122 2123 crit_enter(); 2124 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) { 2125 crit_exit(); 2126 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2127 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2128 aio_free_entry(cb); 2129 return cb->uaiocb._aiocb_private.error; 2130 } 2131 2132 ki->kaio_flags |= KAIO_WAKEUP; 2133 error = tsleep(p, PCATCH, "aiowc", timo); 2134 crit_exit(); 2135 2136 if (error == ERESTART) 2137 return EINTR; 2138 else if (error < 0) 2139 return error; 2140 else if (error == EINTR) 2141 return EINTR; 2142 else if (error == EWOULDBLOCK) 2143 return EAGAIN; 2144 } 2145 #endif /* VFS_AIO */ 2146 } 2147 2148 #ifndef VFS_AIO 2149 static int 2150 filt_aioattach(struct knote *kn) 2151 { 2152 2153 return (ENXIO); 2154 } 2155 2156 struct filterops aio_filtops = 2157 { 0, filt_aioattach, NULL, NULL }; 2158 2159 #else 2160 /* kqueue attach function */ 2161 static int 2162 filt_aioattach(struct knote *kn) 2163 { 2164 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2165 2166 /* 2167 * The aiocbe pointer must be validated before using it, so 2168 * registration is restricted to the kernel; the user cannot 2169 * set EV_FLAG1. 2170 */ 2171 if ((kn->kn_flags & EV_FLAG1) == 0) 2172 return (EPERM); 2173 kn->kn_flags &= ~EV_FLAG1; 2174 2175 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext); 2176 2177 return (0); 2178 } 2179 2180 /* kqueue detach function */ 2181 static void 2182 filt_aiodetach(struct knote *kn) 2183 { 2184 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2185 2186 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext); 2187 } 2188 2189 /* kqueue filter function */ 2190 /*ARGSUSED*/ 2191 static int 2192 filt_aio(struct knote *kn, long hint) 2193 { 2194 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2195 2196 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2197 if (aiocbe->jobstate != JOBST_JOBFINISHED && 2198 aiocbe->jobstate != JOBST_JOBBFINISHED) 2199 return (0); 2200 kn->kn_flags |= EV_EOF; 2201 return (1); 2202 } 2203 2204 struct filterops aio_filtops = 2205 { 0, filt_aioattach, filt_aiodetach, filt_aio }; 2206 #endif /* VFS_AIO */ 2207