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 1.42 2007/07/20 17:21:52 dillon 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/sysref2.h> 53 #include <sys/thread2.h> 54 55 #include <machine/limits.h> 56 #include "opt_vfs_aio.h" 57 58 #ifdef VFS_AIO 59 60 /* 61 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 62 * overflow. 63 */ 64 static long jobrefid; 65 66 #define JOBST_NULL 0x0 67 #define JOBST_JOBQGLOBAL 0x2 68 #define JOBST_JOBRUNNING 0x3 69 #define JOBST_JOBFINISHED 0x4 70 #define JOBST_JOBQBUF 0x5 71 #define JOBST_JOBBFINISHED 0x6 72 73 #ifndef MAX_AIO_PER_PROC 74 #define MAX_AIO_PER_PROC 32 75 #endif 76 77 #ifndef MAX_AIO_QUEUE_PER_PROC 78 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */ 79 #endif 80 81 #ifndef MAX_AIO_PROCS 82 #define MAX_AIO_PROCS 32 83 #endif 84 85 #ifndef MAX_AIO_QUEUE 86 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */ 87 #endif 88 89 #ifndef TARGET_AIO_PROCS 90 #define TARGET_AIO_PROCS 4 91 #endif 92 93 #ifndef MAX_BUF_AIO 94 #define MAX_BUF_AIO 16 95 #endif 96 97 #ifndef AIOD_TIMEOUT_DEFAULT 98 #define AIOD_TIMEOUT_DEFAULT (10 * hz) 99 #endif 100 101 #ifndef AIOD_LIFETIME_DEFAULT 102 #define AIOD_LIFETIME_DEFAULT (30 * hz) 103 #endif 104 105 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management"); 106 107 static int max_aio_procs = MAX_AIO_PROCS; 108 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, 109 CTLFLAG_RW, &max_aio_procs, 0, 110 "Maximum number of kernel threads to use for handling async IO"); 111 112 static int num_aio_procs = 0; 113 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, 114 CTLFLAG_RD, &num_aio_procs, 0, 115 "Number of presently active kernel threads for async IO"); 116 117 /* 118 * The code will adjust the actual number of AIO processes towards this 119 * number when it gets a chance. 120 */ 121 static int target_aio_procs = TARGET_AIO_PROCS; 122 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 123 0, "Preferred number of ready kernel threads for async IO"); 124 125 static int max_queue_count = MAX_AIO_QUEUE; 126 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 127 "Maximum number of aio requests to queue, globally"); 128 129 static int num_queue_count = 0; 130 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 131 "Number of queued aio requests"); 132 133 static int num_buf_aio = 0; 134 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 135 "Number of aio requests presently handled by the buf subsystem"); 136 137 /* Number of async I/O thread in the process of being started */ 138 /* XXX This should be local to _aio_aqueue() */ 139 static int num_aio_resv_start = 0; 140 141 static int aiod_timeout; 142 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0, 143 "Timeout value for synchronous aio operations"); 144 145 static int aiod_lifetime; 146 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 147 "Maximum lifetime for idle aiod"); 148 149 static int max_aio_per_proc = MAX_AIO_PER_PROC; 150 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 151 0, "Maximum active aio requests per process (stored in the process)"); 152 153 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 154 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 155 &max_aio_queue_per_proc, 0, 156 "Maximum queued aio requests per process (stored in the process)"); 157 158 static int max_buf_aio = MAX_BUF_AIO; 159 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 160 "Maximum buf aio requests per process (stored in the process)"); 161 162 /* 163 * AIO process info 164 */ 165 #define AIOP_FREE 0x1 /* proc on free queue */ 166 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */ 167 168 struct aioproclist { 169 int aioprocflags; /* AIO proc flags */ 170 TAILQ_ENTRY(aioproclist) list; /* List of processes */ 171 struct proc *aioproc; /* The AIO thread */ 172 }; 173 174 /* 175 * data-structure for lio signal management 176 */ 177 struct aio_liojob { 178 int lioj_flags; 179 int lioj_buffer_count; 180 int lioj_buffer_finished_count; 181 int lioj_queue_count; 182 int lioj_queue_finished_count; 183 struct sigevent lioj_signal; /* signal on all I/O done */ 184 TAILQ_ENTRY(aio_liojob) lioj_list; 185 struct kaioinfo *lioj_ki; 186 }; 187 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 188 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 189 190 /* 191 * per process aio data structure 192 */ 193 struct kaioinfo { 194 int kaio_flags; /* per process kaio flags */ 195 int kaio_maxactive_count; /* maximum number of AIOs */ 196 int kaio_active_count; /* number of currently used AIOs */ 197 int kaio_qallowed_count; /* maxiumu size of AIO queue */ 198 int kaio_queue_count; /* size of AIO queue */ 199 int kaio_ballowed_count; /* maximum number of buffers */ 200 int kaio_queue_finished_count; /* number of daemon jobs finished */ 201 int kaio_buffer_count; /* number of physio buffers */ 202 int kaio_buffer_finished_count; /* count of I/O done */ 203 struct proc *kaio_p; /* process that uses this kaio block */ 204 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */ 205 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */ 206 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */ 207 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */ 208 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */ 209 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */ 210 }; 211 212 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 213 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */ 214 215 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc; 216 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */ 217 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */ 218 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */ 219 220 static void aio_init_aioinfo(struct proc *p); 221 static void aio_onceonly(void *); 222 static int aio_free_entry(struct aiocblist *aiocbe); 223 static void aio_process(struct aiocblist *aiocbe); 224 static int aio_newproc(void); 225 static int aio_aqueue(struct aiocb *job, int type); 226 static void aio_physwakeup(struct bio *bio); 227 static int aio_fphysio(struct aiocblist *aiocbe); 228 static int aio_qphysio(struct proc *p, struct aiocblist *iocb); 229 static void aio_daemon(void *uproc, struct trapframe *frame); 230 static void process_signal(void *aioj); 231 232 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL); 233 234 /* 235 * Zones for: 236 * kaio Per process async io info 237 * aiop async io thread data 238 * aiocb async io jobs 239 * aiol list io job pointer - internal to aio_suspend XXX 240 * aiolio list io jobs 241 */ 242 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone; 243 244 /* 245 * Startup initialization 246 */ 247 static void 248 aio_onceonly(void *na) 249 { 250 TAILQ_INIT(&aio_freeproc); 251 TAILQ_INIT(&aio_activeproc); 252 TAILQ_INIT(&aio_jobs); 253 TAILQ_INIT(&aio_bufjobs); 254 TAILQ_INIT(&aio_freejobs); 255 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1); 256 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1); 257 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1); 258 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1); 259 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1); 260 aiod_timeout = AIOD_TIMEOUT_DEFAULT; 261 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 262 jobrefid = 1; 263 } 264 265 /* 266 * Init the per-process aioinfo structure. The aioinfo limits are set 267 * per-process for user limit (resource) management. 268 */ 269 static void 270 aio_init_aioinfo(struct proc *p) 271 { 272 struct kaioinfo *ki; 273 if (p->p_aioinfo == NULL) { 274 ki = zalloc(kaio_zone); 275 p->p_aioinfo = ki; 276 ki->kaio_flags = 0; 277 ki->kaio_maxactive_count = max_aio_per_proc; 278 ki->kaio_active_count = 0; 279 ki->kaio_qallowed_count = max_aio_queue_per_proc; 280 ki->kaio_queue_count = 0; 281 ki->kaio_ballowed_count = max_buf_aio; 282 ki->kaio_buffer_count = 0; 283 ki->kaio_buffer_finished_count = 0; 284 ki->kaio_p = p; 285 TAILQ_INIT(&ki->kaio_jobdone); 286 TAILQ_INIT(&ki->kaio_jobqueue); 287 TAILQ_INIT(&ki->kaio_bufdone); 288 TAILQ_INIT(&ki->kaio_bufqueue); 289 TAILQ_INIT(&ki->kaio_liojoblist); 290 TAILQ_INIT(&ki->kaio_sockqueue); 291 } 292 293 while (num_aio_procs < target_aio_procs) 294 aio_newproc(); 295 } 296 297 /* 298 * Free a job entry. Wait for completion if it is currently active, but don't 299 * delay forever. If we delay, we return a flag that says that we have to 300 * restart the queue scan. 301 */ 302 static int 303 aio_free_entry(struct aiocblist *aiocbe) 304 { 305 struct kaioinfo *ki; 306 struct aio_liojob *lj; 307 struct proc *p; 308 int error; 309 310 if (aiocbe->jobstate == JOBST_NULL) 311 panic("aio_free_entry: freeing already free job"); 312 313 p = aiocbe->userproc; 314 ki = p->p_aioinfo; 315 lj = aiocbe->lio; 316 if (ki == NULL) 317 panic("aio_free_entry: missing p->p_aioinfo"); 318 319 while (aiocbe->jobstate == JOBST_JOBRUNNING) { 320 aiocbe->jobflags |= AIOCBLIST_RUNDOWN; 321 tsleep(aiocbe, 0, "jobwai", 0); 322 } 323 if (aiocbe->bp == NULL) { 324 if (ki->kaio_queue_count <= 0) 325 panic("aio_free_entry: process queue size <= 0"); 326 if (num_queue_count <= 0) 327 panic("aio_free_entry: system wide queue size <= 0"); 328 329 if (lj) { 330 lj->lioj_queue_count--; 331 if (aiocbe->jobflags & AIOCBLIST_DONE) 332 lj->lioj_queue_finished_count--; 333 } 334 ki->kaio_queue_count--; 335 if (aiocbe->jobflags & AIOCBLIST_DONE) 336 ki->kaio_queue_finished_count--; 337 num_queue_count--; 338 } else { 339 if (lj) { 340 lj->lioj_buffer_count--; 341 if (aiocbe->jobflags & AIOCBLIST_DONE) 342 lj->lioj_buffer_finished_count--; 343 } 344 if (aiocbe->jobflags & AIOCBLIST_DONE) 345 ki->kaio_buffer_finished_count--; 346 ki->kaio_buffer_count--; 347 num_buf_aio--; 348 } 349 350 /* aiocbe is going away, we need to destroy any knotes */ 351 /* XXX lwp knote wants a thread, but only cares about the process */ 352 knote_remove(FIRST_LWP_IN_PROC(p)->lwp_thread, &aiocbe->klist); 353 354 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN) 355 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) { 356 ki->kaio_flags &= ~KAIO_WAKEUP; 357 wakeup(p); 358 } 359 360 if (aiocbe->jobstate == JOBST_JOBQBUF) { 361 if ((error = aio_fphysio(aiocbe)) != 0) 362 return error; 363 if (aiocbe->jobstate != JOBST_JOBBFINISHED) 364 panic("aio_free_entry: invalid physio finish-up state"); 365 crit_enter(); 366 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist); 367 crit_exit(); 368 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) { 369 crit_enter(); 370 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 371 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 372 crit_exit(); 373 } else if (aiocbe->jobstate == JOBST_JOBFINISHED) 374 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist); 375 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) { 376 crit_enter(); 377 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist); 378 crit_exit(); 379 if (aiocbe->bp) { 380 vunmapbuf(aiocbe->bp); 381 relpbuf(aiocbe->bp, NULL); 382 aiocbe->bp = NULL; 383 } 384 } 385 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) { 386 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 387 zfree(aiolio_zone, lj); 388 } 389 aiocbe->jobstate = JOBST_NULL; 390 callout_stop(&aiocbe->timeout); 391 fdrop(aiocbe->fd_file); 392 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 393 return 0; 394 } 395 #endif /* VFS_AIO */ 396 397 /* 398 * Rundown the jobs for a given process. 399 */ 400 void 401 aio_proc_rundown(struct proc *p) 402 { 403 #ifndef VFS_AIO 404 return; 405 #else 406 struct kaioinfo *ki; 407 struct aio_liojob *lj, *ljn; 408 struct aiocblist *aiocbe, *aiocbn; 409 struct file *fp; 410 struct socket *so; 411 412 ki = p->p_aioinfo; 413 if (ki == NULL) 414 return; 415 416 ki->kaio_flags |= LIOJ_SIGNAL_POSTED; 417 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count > 418 ki->kaio_buffer_finished_count)) { 419 ki->kaio_flags |= KAIO_RUNDOWN; 420 if (tsleep(p, 0, "kaiowt", aiod_timeout)) 421 break; 422 } 423 424 /* 425 * Move any aio ops that are waiting on socket I/O to the normal job 426 * queues so they are cleaned up with any others. 427 */ 428 crit_enter(); 429 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe = 430 aiocbn) { 431 aiocbn = TAILQ_NEXT(aiocbe, plist); 432 fp = aiocbe->fd_file; 433 if (fp != NULL) { 434 so = (struct socket *)fp->f_data; 435 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list); 436 if (TAILQ_EMPTY(&so->so_aiojobq)) { 437 so->so_snd.ssb_flags &= ~SSB_AIO; 438 so->so_rcv.ssb_flags &= ~SSB_AIO; 439 } 440 } 441 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist); 442 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list); 443 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist); 444 } 445 crit_exit(); 446 447 restart1: 448 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) { 449 aiocbn = TAILQ_NEXT(aiocbe, plist); 450 if (aio_free_entry(aiocbe)) 451 goto restart1; 452 } 453 454 restart2: 455 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe = 456 aiocbn) { 457 aiocbn = TAILQ_NEXT(aiocbe, plist); 458 if (aio_free_entry(aiocbe)) 459 goto restart2; 460 } 461 462 restart3: 463 crit_enter(); 464 while (TAILQ_FIRST(&ki->kaio_bufqueue)) { 465 ki->kaio_flags |= KAIO_WAKEUP; 466 tsleep(p, 0, "aioprn", 0); 467 crit_exit(); 468 goto restart3; 469 } 470 crit_exit(); 471 472 restart4: 473 crit_enter(); 474 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) { 475 aiocbn = TAILQ_NEXT(aiocbe, plist); 476 if (aio_free_entry(aiocbe)) { 477 crit_exit(); 478 goto restart4; 479 } 480 } 481 crit_exit(); 482 483 /* 484 * If we've slept, jobs might have moved from one queue to another. 485 * Retry rundown if we didn't manage to empty the queues. 486 */ 487 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL || 488 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL || 489 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL || 490 TAILQ_FIRST(&ki->kaio_bufdone) != NULL) 491 goto restart1; 492 493 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) { 494 ljn = TAILQ_NEXT(lj, lioj_list); 495 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 496 0)) { 497 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 498 zfree(aiolio_zone, lj); 499 } else { 500 #ifdef DIAGNOSTIC 501 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, " 502 "QF:%d\n", lj->lioj_buffer_count, 503 lj->lioj_buffer_finished_count, 504 lj->lioj_queue_count, 505 lj->lioj_queue_finished_count); 506 #endif 507 } 508 } 509 510 zfree(kaio_zone, ki); 511 p->p_aioinfo = NULL; 512 #endif /* VFS_AIO */ 513 } 514 515 #ifdef VFS_AIO 516 /* 517 * Select a job to run (called by an AIO daemon). 518 */ 519 static struct aiocblist * 520 aio_selectjob(struct aioproclist *aiop) 521 { 522 struct aiocblist *aiocbe; 523 struct kaioinfo *ki; 524 struct proc *userp; 525 526 crit_enter(); 527 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe = 528 TAILQ_NEXT(aiocbe, list)) { 529 userp = aiocbe->userproc; 530 ki = userp->p_aioinfo; 531 532 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 533 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 534 crit_exit(); 535 return aiocbe; 536 } 537 } 538 crit_exit(); 539 540 return NULL; 541 } 542 543 /* 544 * The AIO processing activity. This is the code that does the I/O request for 545 * the non-physio version of the operations. The normal vn operations are used, 546 * and this code should work in all instances for every type of file, including 547 * pipes, sockets, fifos, and regular files. 548 */ 549 static void 550 aio_process(struct aiocblist *aiocbe) 551 { 552 struct thread *mytd; 553 struct aiocb *cb; 554 struct file *fp; 555 struct uio auio; 556 struct iovec aiov; 557 int cnt; 558 int error; 559 int oublock_st, oublock_end; 560 int inblock_st, inblock_end; 561 562 mytd = curthread; 563 cb = &aiocbe->uaiocb; 564 fp = aiocbe->fd_file; 565 566 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 567 aiov.iov_len = cb->aio_nbytes; 568 569 auio.uio_iov = &aiov; 570 auio.uio_iovcnt = 1; 571 auio.uio_offset = cb->aio_offset; 572 auio.uio_resid = cb->aio_nbytes; 573 cnt = cb->aio_nbytes; 574 auio.uio_segflg = UIO_USERSPACE; 575 auio.uio_td = mytd; 576 577 inblock_st = mytd->td_lwp->lwp_ru.ru_inblock; 578 oublock_st = mytd->td_lwp->lwp_ru.ru_oublock; 579 /* 580 * _aio_aqueue() acquires a reference to the file that is 581 * released in aio_free_entry(). 582 */ 583 if (cb->aio_lio_opcode == LIO_READ) { 584 auio.uio_rw = UIO_READ; 585 error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET); 586 } else { 587 auio.uio_rw = UIO_WRITE; 588 error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET); 589 } 590 inblock_end = mytd->td_lwp->lwp_ru.ru_inblock; 591 oublock_end = mytd->td_lwp->lwp_ru.ru_oublock; 592 593 aiocbe->inputcharge = inblock_end - inblock_st; 594 aiocbe->outputcharge = oublock_end - oublock_st; 595 596 if ((error) && (auio.uio_resid != cnt)) { 597 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 598 error = 0; 599 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) 600 ksignal(aiocbe->userproc, SIGPIPE); 601 } 602 603 cnt -= auio.uio_resid; 604 cb->_aiocb_private.error = error; 605 cb->_aiocb_private.status = cnt; 606 } 607 608 /* 609 * The AIO daemon, most of the actual work is done in aio_process, 610 * but the setup (and address space mgmt) is done in this routine. 611 * 612 * The MP lock is held on entry. 613 */ 614 static void 615 aio_daemon(void *uproc, struct trapframe *frame) 616 { 617 struct aio_liojob *lj; 618 struct aiocb *cb; 619 struct aiocblist *aiocbe; 620 struct aioproclist *aiop; 621 struct kaioinfo *ki; 622 struct proc *mycp, *userp; 623 struct vmspace *curvm; 624 struct lwp *mylwp; 625 struct ucred *cr; 626 627 mylwp = curthread->td_lwp; 628 mycp = mylwp->lwp_proc; 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 curvm = NULL; 682 683 for (;;) { 684 /* 685 * Take daemon off of free queue 686 */ 687 if (aiop->aioprocflags & AIOP_FREE) { 688 crit_enter(); 689 TAILQ_REMOVE(&aio_freeproc, aiop, list); 690 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 691 aiop->aioprocflags &= ~AIOP_FREE; 692 crit_exit(); 693 } 694 aiop->aioprocflags &= ~AIOP_SCHED; 695 696 /* 697 * Check for jobs. 698 */ 699 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 700 cb = &aiocbe->uaiocb; 701 userp = aiocbe->userproc; 702 703 aiocbe->jobstate = JOBST_JOBRUNNING; 704 705 /* 706 * Connect to process address space for user program. 707 */ 708 if (curvm != userp->p_vmspace) { 709 pmap_setlwpvm(mylwp, userp->p_vmspace); 710 if (curvm) 711 sysref_put(&curvm->vm_sysref); 712 curvm = userp->p_vmspace; 713 sysref_get(&curvm->vm_sysref); 714 } 715 716 ki = userp->p_aioinfo; 717 lj = aiocbe->lio; 718 719 /* Account for currently active jobs. */ 720 ki->kaio_active_count++; 721 722 /* Do the I/O function. */ 723 aio_process(aiocbe); 724 725 /* Decrement the active job count. */ 726 ki->kaio_active_count--; 727 728 /* 729 * Increment the completion count for wakeup/signal 730 * comparisons. 731 */ 732 aiocbe->jobflags |= AIOCBLIST_DONE; 733 ki->kaio_queue_finished_count++; 734 if (lj) 735 lj->lioj_queue_finished_count++; 736 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags 737 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) { 738 ki->kaio_flags &= ~KAIO_WAKEUP; 739 wakeup(userp); 740 } 741 742 crit_enter(); 743 if (lj && (lj->lioj_flags & 744 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) { 745 if ((lj->lioj_queue_finished_count == 746 lj->lioj_queue_count) && 747 (lj->lioj_buffer_finished_count == 748 lj->lioj_buffer_count)) { 749 ksignal(userp, 750 lj->lioj_signal.sigev_signo); 751 lj->lioj_flags |= 752 LIOJ_SIGNAL_POSTED; 753 } 754 } 755 crit_exit(); 756 757 aiocbe->jobstate = JOBST_JOBFINISHED; 758 759 crit_enter(); 760 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 761 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist); 762 crit_exit(); 763 KNOTE(&aiocbe->klist, 0); 764 765 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) { 766 wakeup(aiocbe); 767 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN; 768 } 769 770 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 771 ksignal(userp, cb->aio_sigevent.sigev_signo); 772 } 773 } 774 775 /* 776 * Disconnect from user address space. 777 */ 778 if (curvm) { 779 /* swap our original address space back in */ 780 pmap_setlwpvm(mylwp, mycp->p_vmspace); 781 sysref_put(&curvm->vm_sysref); 782 curvm = NULL; 783 } 784 785 /* 786 * If we are the first to be put onto the free queue, wakeup 787 * anyone waiting for a daemon. 788 */ 789 crit_enter(); 790 TAILQ_REMOVE(&aio_activeproc, aiop, list); 791 if (TAILQ_EMPTY(&aio_freeproc)) 792 wakeup(&aio_freeproc); 793 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 794 aiop->aioprocflags |= AIOP_FREE; 795 crit_exit(); 796 797 /* 798 * If daemon is inactive for a long time, allow it to exit, 799 * thereby freeing resources. 800 */ 801 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp, 802 0, "aiordy", aiod_lifetime)) { 803 crit_enter(); 804 if (TAILQ_EMPTY(&aio_jobs)) { 805 if ((aiop->aioprocflags & AIOP_FREE) && 806 (num_aio_procs > target_aio_procs)) { 807 TAILQ_REMOVE(&aio_freeproc, aiop, list); 808 crit_exit(); 809 zfree(aiop_zone, aiop); 810 num_aio_procs--; 811 #ifdef DIAGNOSTIC 812 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) { 813 kprintf("AIOD: bad vm refcnt for" 814 " exiting daemon: %d\n", 815 mycp->p_vmspace->vm_sysref.refcnt); 816 } 817 #endif 818 exit1(0); 819 } 820 } 821 crit_exit(); 822 } 823 } 824 } 825 826 /* 827 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 828 * AIO daemon modifies its environment itself. 829 */ 830 static int 831 aio_newproc(void) 832 { 833 int error; 834 struct lwp *lp, *nlp; 835 struct proc *np; 836 837 lp = &lwp0; 838 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np); 839 if (error) 840 return error; 841 nlp = ONLY_LWP_IN_PROC(np); 842 cpu_set_fork_handler(nlp, aio_daemon, curproc); 843 start_forked_proc(lp, np); 844 845 /* 846 * Wait until daemon is started, but continue on just in case to 847 * handle error conditions. 848 */ 849 error = tsleep(np, 0, "aiosta", aiod_timeout); 850 num_aio_procs++; 851 852 return error; 853 } 854 855 /* 856 * Try the high-performance, low-overhead physio method for eligible 857 * VCHR devices. This method doesn't use an aio helper thread, and 858 * thus has very low overhead. 859 * 860 * Assumes that the caller, _aio_aqueue(), has incremented the file 861 * structure's reference count, preventing its deallocation for the 862 * duration of this call. 863 */ 864 static int 865 aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 866 { 867 int error; 868 struct aiocb *cb; 869 struct file *fp; 870 struct buf *bp; 871 struct vnode *vp; 872 struct kaioinfo *ki; 873 struct aio_liojob *lj; 874 int notify; 875 876 cb = &aiocbe->uaiocb; 877 fp = aiocbe->fd_file; 878 879 if (fp->f_type != DTYPE_VNODE) 880 return (-1); 881 882 vp = (struct vnode *)fp->f_data; 883 884 /* 885 * If its not a disk, we don't want to return a positive error. 886 * It causes the aio code to not fall through to try the thread 887 * way when you're talking to a regular file. 888 */ 889 if (!vn_isdisk(vp, &error)) { 890 if (error == ENOTBLK) 891 return (-1); 892 else 893 return (error); 894 } 895 896 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys) 897 return (-1); 898 899 if (cb->aio_nbytes > 900 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 901 return (-1); 902 903 ki = p->p_aioinfo; 904 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 905 return (-1); 906 907 ki->kaio_buffer_count++; 908 909 lj = aiocbe->lio; 910 if (lj) 911 lj->lioj_buffer_count++; 912 913 /* Create and build a buffer header for a transfer. */ 914 bp = getpbuf(NULL); 915 BUF_KERNPROC(bp); 916 917 /* 918 * Get a copy of the kva from the physical buffer. 919 */ 920 bp->b_bio1.bio_caller_info1.ptr = p; 921 error = 0; 922 923 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ? 924 BUF_CMD_WRITE : BUF_CMD_READ; 925 bp->b_bio1.bio_done = aio_physwakeup; 926 bp->b_bio1.bio_offset = cb->aio_offset; 927 928 /* Bring buffer into kernel space. */ 929 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) { 930 error = EFAULT; 931 goto doerror; 932 } 933 934 crit_enter(); 935 936 aiocbe->bp = bp; 937 bp->b_bio1.bio_caller_info2.ptr = aiocbe; 938 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list); 939 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 940 aiocbe->jobstate = JOBST_JOBQBUF; 941 cb->_aiocb_private.status = cb->aio_nbytes; 942 num_buf_aio++; 943 bp->b_error = 0; 944 945 crit_exit(); 946 947 /* 948 * Perform the transfer. vn_strategy must be used even though we 949 * know we have a device in order to deal with requests which exceed 950 * device DMA limitations. 951 */ 952 vn_strategy(vp, &bp->b_bio1); 953 954 notify = 0; 955 crit_enter(); 956 957 /* 958 * If we had an error invoking the request, or an error in processing 959 * the request before we have returned, we process it as an error in 960 * transfer. Note that such an I/O error is not indicated immediately, 961 * but is returned using the aio_error mechanism. In this case, 962 * aio_suspend will return immediately. 963 */ 964 if (bp->b_error || (bp->b_flags & B_ERROR)) { 965 struct aiocb *job = aiocbe->uuaiocb; 966 967 aiocbe->uaiocb._aiocb_private.status = 0; 968 suword(&job->_aiocb_private.status, 0); 969 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 970 suword(&job->_aiocb_private.error, bp->b_error); 971 972 ki->kaio_buffer_finished_count++; 973 974 if (aiocbe->jobstate != JOBST_JOBBFINISHED) { 975 aiocbe->jobstate = JOBST_JOBBFINISHED; 976 aiocbe->jobflags |= AIOCBLIST_DONE; 977 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 978 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 979 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 980 notify = 1; 981 } 982 } 983 crit_exit(); 984 if (notify) 985 KNOTE(&aiocbe->klist, 0); 986 return 0; 987 988 doerror: 989 ki->kaio_buffer_count--; 990 if (lj) 991 lj->lioj_buffer_count--; 992 aiocbe->bp = NULL; 993 relpbuf(bp, NULL); 994 return error; 995 } 996 997 /* 998 * This waits/tests physio completion. 999 */ 1000 static int 1001 aio_fphysio(struct aiocblist *iocb) 1002 { 1003 struct buf *bp; 1004 int error; 1005 1006 bp = iocb->bp; 1007 1008 crit_enter(); 1009 while (bp->b_cmd != BUF_CMD_DONE) { 1010 if (tsleep(bp, 0, "physstr", aiod_timeout)) { 1011 if (bp->b_cmd != BUF_CMD_DONE) { 1012 crit_exit(); 1013 return EINPROGRESS; 1014 } else { 1015 break; 1016 } 1017 } 1018 } 1019 crit_exit(); 1020 1021 /* Release mapping into kernel space. */ 1022 vunmapbuf(bp); 1023 iocb->bp = 0; 1024 1025 error = 0; 1026 1027 /* Check for an error. */ 1028 if (bp->b_flags & B_ERROR) 1029 error = bp->b_error; 1030 1031 relpbuf(bp, NULL); 1032 return (error); 1033 } 1034 #endif /* VFS_AIO */ 1035 1036 /* 1037 * Wake up aio requests that may be serviceable now. 1038 */ 1039 void 1040 aio_swake(struct socket *so, struct signalsockbuf *ssb) 1041 { 1042 #ifndef VFS_AIO 1043 return; 1044 #else 1045 struct aiocblist *cb,*cbn; 1046 struct proc *p; 1047 struct kaioinfo *ki = NULL; 1048 int opcode, wakecount = 0; 1049 struct aioproclist *aiop; 1050 1051 if (ssb == &so->so_snd) { 1052 opcode = LIO_WRITE; 1053 so->so_snd.ssb_flags &= ~SSB_AIO; 1054 } else { 1055 opcode = LIO_READ; 1056 so->so_rcv.ssb_flags &= ~SSB_AIO; 1057 } 1058 1059 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) { 1060 cbn = TAILQ_NEXT(cb, list); 1061 if (opcode == cb->uaiocb.aio_lio_opcode) { 1062 p = cb->userproc; 1063 ki = p->p_aioinfo; 1064 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1065 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist); 1066 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1067 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist); 1068 wakecount++; 1069 if (cb->jobstate != JOBST_JOBQGLOBAL) 1070 panic("invalid queue value"); 1071 } 1072 } 1073 1074 while (wakecount--) { 1075 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) { 1076 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1077 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1078 aiop->aioprocflags &= ~AIOP_FREE; 1079 wakeup(aiop->aioproc); 1080 } 1081 } 1082 #endif /* VFS_AIO */ 1083 } 1084 1085 #ifdef VFS_AIO 1086 /* 1087 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1088 * technique is done in this code. 1089 */ 1090 static int 1091 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type) 1092 { 1093 struct proc *p = curproc; 1094 struct filedesc *fdp; 1095 struct file *fp; 1096 unsigned int fd; 1097 struct socket *so; 1098 int error; 1099 int opcode, user_opcode; 1100 struct aiocblist *aiocbe; 1101 struct aioproclist *aiop; 1102 struct kaioinfo *ki; 1103 struct kevent kev; 1104 struct kqueue *kq; 1105 struct file *kq_fp; 1106 1107 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL) 1108 TAILQ_REMOVE(&aio_freejobs, aiocbe, list); 1109 else 1110 aiocbe = zalloc (aiocb_zone); 1111 1112 aiocbe->inputcharge = 0; 1113 aiocbe->outputcharge = 0; 1114 callout_init(&aiocbe->timeout); 1115 SLIST_INIT(&aiocbe->klist); 1116 1117 suword(&job->_aiocb_private.status, -1); 1118 suword(&job->_aiocb_private.error, 0); 1119 suword(&job->_aiocb_private.kernelinfo, -1); 1120 1121 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb)); 1122 if (error) { 1123 suword(&job->_aiocb_private.error, error); 1124 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1125 return error; 1126 } 1127 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL && 1128 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1129 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1130 return EINVAL; 1131 } 1132 1133 /* Save userspace address of the job info. */ 1134 aiocbe->uuaiocb = job; 1135 1136 /* Get the opcode. */ 1137 user_opcode = aiocbe->uaiocb.aio_lio_opcode; 1138 if (type != LIO_NOP) 1139 aiocbe->uaiocb.aio_lio_opcode = type; 1140 opcode = aiocbe->uaiocb.aio_lio_opcode; 1141 1142 /* Get the fd info for process. */ 1143 fdp = p->p_fd; 1144 1145 /* 1146 * Range check file descriptor. 1147 */ 1148 fd = aiocbe->uaiocb.aio_fildes; 1149 if (fd >= fdp->fd_nfiles) { 1150 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1151 if (type == 0) 1152 suword(&job->_aiocb_private.error, EBADF); 1153 return EBADF; 1154 } 1155 1156 fp = aiocbe->fd_file = fdp->fd_files[fd].fp; 1157 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 1158 0))) { 1159 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1160 if (type == 0) 1161 suword(&job->_aiocb_private.error, EBADF); 1162 return EBADF; 1163 } 1164 fhold(fp); 1165 1166 if (aiocbe->uaiocb.aio_offset == -1LL) { 1167 error = EINVAL; 1168 goto aqueue_fail; 1169 } 1170 error = suword(&job->_aiocb_private.kernelinfo, jobrefid); 1171 if (error) { 1172 error = EINVAL; 1173 goto aqueue_fail; 1174 } 1175 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid; 1176 if (jobrefid == LONG_MAX) 1177 jobrefid = 1; 1178 else 1179 jobrefid++; 1180 1181 if (opcode == LIO_NOP) { 1182 fdrop(fp); 1183 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1184 if (type == 0) { 1185 suword(&job->_aiocb_private.error, 0); 1186 suword(&job->_aiocb_private.status, 0); 1187 suword(&job->_aiocb_private.kernelinfo, 0); 1188 } 1189 return 0; 1190 } 1191 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) { 1192 if (type == 0) 1193 suword(&job->_aiocb_private.status, 0); 1194 error = EINVAL; 1195 goto aqueue_fail; 1196 } 1197 1198 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) { 1199 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1200 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr; 1201 } 1202 else { 1203 /* 1204 * This method for requesting kevent-based notification won't 1205 * work on the alpha, since we're passing in a pointer 1206 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT- 1207 * based method instead. 1208 */ 1209 if (user_opcode == LIO_NOP || user_opcode == LIO_READ || 1210 user_opcode == LIO_WRITE) 1211 goto no_kqueue; 1212 1213 error = copyin((struct kevent *)(uintptr_t)user_opcode, 1214 &kev, sizeof(kev)); 1215 if (error) 1216 goto aqueue_fail; 1217 } 1218 if ((u_int)kev.ident >= fdp->fd_nfiles || 1219 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL || 1220 (kq_fp->f_type != DTYPE_KQUEUE)) { 1221 error = EBADF; 1222 goto aqueue_fail; 1223 } 1224 kq = (struct kqueue *)kq_fp->f_data; 1225 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1226 kev.filter = EVFILT_AIO; 1227 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 1228 kev.data = (intptr_t)aiocbe; 1229 /* XXX lwp kqueue_register takes a thread, but only uses its proc */ 1230 error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread); 1231 aqueue_fail: 1232 if (error) { 1233 fdrop(fp); 1234 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1235 if (type == 0) 1236 suword(&job->_aiocb_private.error, error); 1237 goto done; 1238 } 1239 no_kqueue: 1240 1241 suword(&job->_aiocb_private.error, EINPROGRESS); 1242 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1243 aiocbe->userproc = p; 1244 aiocbe->jobflags = 0; 1245 aiocbe->lio = lj; 1246 ki = p->p_aioinfo; 1247 1248 if (fp->f_type == DTYPE_SOCKET) { 1249 /* 1250 * Alternate queueing for socket ops: Reach down into the 1251 * descriptor to get the socket data. Then check to see if the 1252 * socket is ready to be read or written (based on the requested 1253 * operation). 1254 * 1255 * If it is not ready for io, then queue the aiocbe on the 1256 * socket, and set the flags so we get a call when ssb_notify() 1257 * happens. 1258 */ 1259 so = (struct socket *)fp->f_data; 1260 crit_enter(); 1261 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1262 LIO_WRITE) && (!sowriteable(so)))) { 1263 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1264 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist); 1265 if (opcode == LIO_READ) 1266 so->so_rcv.ssb_flags |= SSB_AIO; 1267 else 1268 so->so_snd.ssb_flags |= SSB_AIO; 1269 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */ 1270 ki->kaio_queue_count++; 1271 num_queue_count++; 1272 crit_exit(); 1273 error = 0; 1274 goto done; 1275 } 1276 crit_exit(); 1277 } 1278 1279 if ((error = aio_qphysio(p, aiocbe)) == 0) 1280 goto done; 1281 if (error > 0) { 1282 suword(&job->_aiocb_private.status, 0); 1283 aiocbe->uaiocb._aiocb_private.error = error; 1284 suword(&job->_aiocb_private.error, error); 1285 goto done; 1286 } 1287 1288 /* No buffer for daemon I/O. */ 1289 aiocbe->bp = NULL; 1290 1291 ki->kaio_queue_count++; 1292 if (lj) 1293 lj->lioj_queue_count++; 1294 crit_enter(); 1295 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1296 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1297 crit_exit(); 1298 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1299 1300 num_queue_count++; 1301 error = 0; 1302 1303 /* 1304 * If we don't have a free AIO process, and we are below our quota, then 1305 * start one. Otherwise, depend on the subsequent I/O completions to 1306 * pick-up this job. If we don't successfully create the new process 1307 * (thread) due to resource issues, we return an error for now (EAGAIN), 1308 * which is likely not the correct thing to do. 1309 */ 1310 crit_enter(); 1311 retryproc: 1312 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1313 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1314 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1315 aiop->aioprocflags &= ~AIOP_FREE; 1316 wakeup(aiop->aioproc); 1317 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1318 ((ki->kaio_active_count + num_aio_resv_start) < 1319 ki->kaio_maxactive_count)) { 1320 num_aio_resv_start++; 1321 if ((error = aio_newproc()) == 0) { 1322 num_aio_resv_start--; 1323 goto retryproc; 1324 } 1325 num_aio_resv_start--; 1326 } 1327 crit_exit(); 1328 done: 1329 return error; 1330 } 1331 1332 /* 1333 * This routine queues an AIO request, checking for quotas. 1334 */ 1335 static int 1336 aio_aqueue(struct aiocb *job, int type) 1337 { 1338 struct proc *p = curproc; 1339 struct kaioinfo *ki; 1340 1341 if (p->p_aioinfo == NULL) 1342 aio_init_aioinfo(p); 1343 1344 if (num_queue_count >= max_queue_count) 1345 return EAGAIN; 1346 1347 ki = p->p_aioinfo; 1348 if (ki->kaio_queue_count >= ki->kaio_qallowed_count) 1349 return EAGAIN; 1350 1351 return _aio_aqueue(job, NULL, type); 1352 } 1353 #endif /* VFS_AIO */ 1354 1355 /* 1356 * Support the aio_return system call, as a side-effect, kernel resources are 1357 * released. 1358 */ 1359 int 1360 sys_aio_return(struct aio_return_args *uap) 1361 { 1362 #ifndef VFS_AIO 1363 return ENOSYS; 1364 #else 1365 struct proc *p = curproc; 1366 struct lwp *lp = curthread->td_lwp; 1367 long jobref; 1368 struct aiocblist *cb, *ncb; 1369 struct aiocb *ujob; 1370 struct kaioinfo *ki; 1371 1372 ki = p->p_aioinfo; 1373 if (ki == NULL) 1374 return EINVAL; 1375 1376 ujob = uap->aiocbp; 1377 1378 jobref = fuword(&ujob->_aiocb_private.kernelinfo); 1379 if (jobref == -1 || jobref == 0) 1380 return EINVAL; 1381 1382 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1383 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == 1384 jobref) { 1385 if (ujob == cb->uuaiocb) { 1386 uap->sysmsg_result = 1387 cb->uaiocb._aiocb_private.status; 1388 } else 1389 uap->sysmsg_result = EFAULT; 1390 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1391 lp->lwp_ru.ru_oublock += cb->outputcharge; 1392 cb->outputcharge = 0; 1393 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1394 lp->lwp_ru.ru_inblock += cb->inputcharge; 1395 cb->inputcharge = 0; 1396 } 1397 aio_free_entry(cb); 1398 return 0; 1399 } 1400 } 1401 crit_enter(); 1402 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) { 1403 ncb = TAILQ_NEXT(cb, plist); 1404 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) 1405 == jobref) { 1406 crit_exit(); 1407 if (ujob == cb->uuaiocb) { 1408 uap->sysmsg_result = 1409 cb->uaiocb._aiocb_private.status; 1410 } else 1411 uap->sysmsg_result = EFAULT; 1412 aio_free_entry(cb); 1413 return 0; 1414 } 1415 } 1416 crit_exit(); 1417 1418 return (EINVAL); 1419 #endif /* VFS_AIO */ 1420 } 1421 1422 /* 1423 * Allow a process to wakeup when any of the I/O requests are completed. 1424 */ 1425 int 1426 sys_aio_suspend(struct aio_suspend_args *uap) 1427 { 1428 #ifndef VFS_AIO 1429 return ENOSYS; 1430 #else 1431 struct proc *p = curproc; 1432 struct timeval atv; 1433 struct timespec ts; 1434 struct aiocb *const *cbptr, *cbp; 1435 struct kaioinfo *ki; 1436 struct aiocblist *cb; 1437 int i; 1438 int njoblist; 1439 int error, timo; 1440 long *ijoblist; 1441 struct aiocb **ujoblist; 1442 1443 if (uap->nent > AIO_LISTIO_MAX) 1444 return EINVAL; 1445 1446 timo = 0; 1447 if (uap->timeout) { 1448 /* Get timespec struct. */ 1449 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1450 return error; 1451 1452 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) 1453 return (EINVAL); 1454 1455 TIMESPEC_TO_TIMEVAL(&atv, &ts); 1456 if (itimerfix(&atv)) 1457 return (EINVAL); 1458 timo = tvtohz_high(&atv); 1459 } 1460 1461 ki = p->p_aioinfo; 1462 if (ki == NULL) 1463 return EAGAIN; 1464 1465 njoblist = 0; 1466 ijoblist = zalloc(aiol_zone); 1467 ujoblist = zalloc(aiol_zone); 1468 cbptr = uap->aiocbp; 1469 1470 for (i = 0; i < uap->nent; i++) { 1471 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1472 if (cbp == 0) 1473 continue; 1474 ujoblist[njoblist] = cbp; 1475 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo); 1476 njoblist++; 1477 } 1478 1479 if (njoblist == 0) { 1480 zfree(aiol_zone, ijoblist); 1481 zfree(aiol_zone, ujoblist); 1482 return 0; 1483 } 1484 1485 error = 0; 1486 for (;;) { 1487 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1488 for (i = 0; i < njoblist; i++) { 1489 if (((intptr_t) 1490 cb->uaiocb._aiocb_private.kernelinfo) == 1491 ijoblist[i]) { 1492 if (ujoblist[i] != cb->uuaiocb) 1493 error = EINVAL; 1494 zfree(aiol_zone, ijoblist); 1495 zfree(aiol_zone, ujoblist); 1496 return error; 1497 } 1498 } 1499 } 1500 1501 crit_enter(); 1502 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = 1503 TAILQ_NEXT(cb, plist)) { 1504 for (i = 0; i < njoblist; i++) { 1505 if (((intptr_t) 1506 cb->uaiocb._aiocb_private.kernelinfo) == 1507 ijoblist[i]) { 1508 crit_exit(); 1509 if (ujoblist[i] != cb->uuaiocb) 1510 error = EINVAL; 1511 zfree(aiol_zone, ijoblist); 1512 zfree(aiol_zone, ujoblist); 1513 return error; 1514 } 1515 } 1516 } 1517 1518 ki->kaio_flags |= KAIO_WAKEUP; 1519 error = tsleep(p, PCATCH, "aiospn", timo); 1520 crit_exit(); 1521 1522 if (error == ERESTART || error == EINTR) { 1523 zfree(aiol_zone, ijoblist); 1524 zfree(aiol_zone, ujoblist); 1525 return EINTR; 1526 } else if (error == EWOULDBLOCK) { 1527 zfree(aiol_zone, ijoblist); 1528 zfree(aiol_zone, ujoblist); 1529 return EAGAIN; 1530 } 1531 } 1532 1533 /* NOTREACHED */ 1534 return EINVAL; 1535 #endif /* VFS_AIO */ 1536 } 1537 1538 /* 1539 * aio_cancel cancels any non-physio aio operations not currently in 1540 * progress. 1541 */ 1542 int 1543 sys_aio_cancel(struct aio_cancel_args *uap) 1544 { 1545 #ifndef VFS_AIO 1546 return ENOSYS; 1547 #else 1548 struct proc *p = curproc; 1549 struct kaioinfo *ki; 1550 struct aiocblist *cbe, *cbn; 1551 struct file *fp; 1552 struct filedesc *fdp; 1553 struct socket *so; 1554 struct proc *po; 1555 int error; 1556 int cancelled=0; 1557 int notcancelled=0; 1558 struct vnode *vp; 1559 1560 fdp = p->p_fd; 1561 if ((u_int)uap->fd >= fdp->fd_nfiles || 1562 (fp = fdp->fd_files[uap->fd].fp) == NULL) 1563 return (EBADF); 1564 1565 if (fp->f_type == DTYPE_VNODE) { 1566 vp = (struct vnode *)fp->f_data; 1567 1568 if (vn_isdisk(vp,&error)) { 1569 uap->sysmsg_result = AIO_NOTCANCELED; 1570 return 0; 1571 } 1572 } else if (fp->f_type == DTYPE_SOCKET) { 1573 so = (struct socket *)fp->f_data; 1574 1575 crit_enter(); 1576 1577 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) { 1578 cbn = TAILQ_NEXT(cbe, list); 1579 if ((uap->aiocbp == NULL) || 1580 (uap->aiocbp == cbe->uuaiocb) ) { 1581 po = cbe->userproc; 1582 ki = po->p_aioinfo; 1583 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 1584 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist); 1585 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist); 1586 if (ki->kaio_flags & KAIO_WAKEUP) { 1587 wakeup(po); 1588 } 1589 cbe->jobstate = JOBST_JOBFINISHED; 1590 cbe->uaiocb._aiocb_private.status=-1; 1591 cbe->uaiocb._aiocb_private.error=ECANCELED; 1592 cancelled++; 1593 /* XXX cancelled, knote? */ 1594 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1595 SIGEV_SIGNAL) 1596 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1597 if (uap->aiocbp) 1598 break; 1599 } 1600 } 1601 crit_exit(); 1602 1603 if ((cancelled) && (uap->aiocbp)) { 1604 uap->sysmsg_result = AIO_CANCELED; 1605 return 0; 1606 } 1607 } 1608 ki=p->p_aioinfo; 1609 if (ki == NULL) 1610 goto done; 1611 crit_enter(); 1612 1613 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) { 1614 cbn = TAILQ_NEXT(cbe, plist); 1615 1616 if ((uap->fd == cbe->uaiocb.aio_fildes) && 1617 ((uap->aiocbp == NULL ) || 1618 (uap->aiocbp == cbe->uuaiocb))) { 1619 1620 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 1621 TAILQ_REMOVE(&aio_jobs, cbe, list); 1622 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 1623 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, 1624 plist); 1625 cancelled++; 1626 ki->kaio_queue_finished_count++; 1627 cbe->jobstate = JOBST_JOBFINISHED; 1628 cbe->uaiocb._aiocb_private.status = -1; 1629 cbe->uaiocb._aiocb_private.error = ECANCELED; 1630 /* XXX cancelled, knote? */ 1631 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1632 SIGEV_SIGNAL) 1633 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1634 } else { 1635 notcancelled++; 1636 } 1637 } 1638 } 1639 crit_exit(); 1640 done: 1641 if (notcancelled) { 1642 uap->sysmsg_result = AIO_NOTCANCELED; 1643 return 0; 1644 } 1645 if (cancelled) { 1646 uap->sysmsg_result = AIO_CANCELED; 1647 return 0; 1648 } 1649 uap->sysmsg_result = AIO_ALLDONE; 1650 1651 return 0; 1652 #endif /* VFS_AIO */ 1653 } 1654 1655 /* 1656 * aio_error is implemented in the kernel level for compatibility purposes only. 1657 * For a user mode async implementation, it would be best to do it in a userland 1658 * subroutine. 1659 */ 1660 int 1661 sys_aio_error(struct aio_error_args *uap) 1662 { 1663 #ifndef VFS_AIO 1664 return ENOSYS; 1665 #else 1666 struct proc *p = curproc; 1667 struct aiocblist *cb; 1668 struct kaioinfo *ki; 1669 long jobref; 1670 1671 ki = p->p_aioinfo; 1672 if (ki == NULL) 1673 return EINVAL; 1674 1675 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo); 1676 if ((jobref == -1) || (jobref == 0)) 1677 return EINVAL; 1678 1679 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1680 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1681 jobref) { 1682 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1683 return 0; 1684 } 1685 } 1686 1687 crit_enter(); 1688 1689 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb, 1690 plist)) { 1691 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1692 jobref) { 1693 uap->sysmsg_result = EINPROGRESS; 1694 crit_exit(); 1695 return 0; 1696 } 1697 } 1698 1699 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb, 1700 plist)) { 1701 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1702 jobref) { 1703 uap->sysmsg_result = EINPROGRESS; 1704 crit_exit(); 1705 return 0; 1706 } 1707 } 1708 crit_exit(); 1709 1710 crit_enter(); 1711 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb, 1712 plist)) { 1713 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1714 jobref) { 1715 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1716 crit_exit(); 1717 return 0; 1718 } 1719 } 1720 1721 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb, 1722 plist)) { 1723 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1724 jobref) { 1725 uap->sysmsg_result = EINPROGRESS; 1726 crit_exit(); 1727 return 0; 1728 } 1729 } 1730 crit_exit(); 1731 1732 #if (0) 1733 /* 1734 * Hack for lio. 1735 */ 1736 status = fuword(&uap->aiocbp->_aiocb_private.status); 1737 if (status == -1) 1738 return fuword(&uap->aiocbp->_aiocb_private.error); 1739 #endif 1740 return EINVAL; 1741 #endif /* VFS_AIO */ 1742 } 1743 1744 /* syscall - asynchronous read from a file (REALTIME) */ 1745 int 1746 sys_aio_read(struct aio_read_args *uap) 1747 { 1748 #ifndef VFS_AIO 1749 return ENOSYS; 1750 #else 1751 return aio_aqueue(uap->aiocbp, LIO_READ); 1752 #endif /* VFS_AIO */ 1753 } 1754 1755 /* syscall - asynchronous write to a file (REALTIME) */ 1756 int 1757 sys_aio_write(struct aio_write_args *uap) 1758 { 1759 #ifndef VFS_AIO 1760 return ENOSYS; 1761 #else 1762 return aio_aqueue(uap->aiocbp, LIO_WRITE); 1763 #endif /* VFS_AIO */ 1764 } 1765 1766 /* syscall - XXX undocumented */ 1767 int 1768 sys_lio_listio(struct lio_listio_args *uap) 1769 { 1770 #ifndef VFS_AIO 1771 return ENOSYS; 1772 #else 1773 struct proc *p = curproc; 1774 struct lwp *lp = curthread->td_lwp; 1775 int nent, nentqueued; 1776 struct aiocb *iocb, * const *cbptr; 1777 struct aiocblist *cb; 1778 struct kaioinfo *ki; 1779 struct aio_liojob *lj; 1780 int error, runningcode; 1781 int nerror; 1782 int i; 1783 1784 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 1785 return EINVAL; 1786 1787 nent = uap->nent; 1788 if (nent > AIO_LISTIO_MAX) 1789 return EINVAL; 1790 1791 if (p->p_aioinfo == NULL) 1792 aio_init_aioinfo(p); 1793 1794 if ((nent + num_queue_count) > max_queue_count) 1795 return EAGAIN; 1796 1797 ki = p->p_aioinfo; 1798 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) 1799 return EAGAIN; 1800 1801 lj = zalloc(aiolio_zone); 1802 if (!lj) 1803 return EAGAIN; 1804 1805 lj->lioj_flags = 0; 1806 lj->lioj_buffer_count = 0; 1807 lj->lioj_buffer_finished_count = 0; 1808 lj->lioj_queue_count = 0; 1809 lj->lioj_queue_finished_count = 0; 1810 lj->lioj_ki = ki; 1811 1812 /* 1813 * Setup signal. 1814 */ 1815 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 1816 error = copyin(uap->sig, &lj->lioj_signal, 1817 sizeof(lj->lioj_signal)); 1818 if (error) { 1819 zfree(aiolio_zone, lj); 1820 return error; 1821 } 1822 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 1823 zfree(aiolio_zone, lj); 1824 return EINVAL; 1825 } 1826 lj->lioj_flags |= LIOJ_SIGNAL; 1827 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED; 1828 } else 1829 lj->lioj_flags &= ~LIOJ_SIGNAL; 1830 1831 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 1832 /* 1833 * Get pointers to the list of I/O requests. 1834 */ 1835 nerror = 0; 1836 nentqueued = 0; 1837 cbptr = uap->acb_list; 1838 for (i = 0; i < uap->nent; i++) { 1839 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1840 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) { 1841 error = _aio_aqueue(iocb, lj, 0); 1842 if (error == 0) 1843 nentqueued++; 1844 else 1845 nerror++; 1846 } 1847 } 1848 1849 /* 1850 * If we haven't queued any, then just return error. 1851 */ 1852 if (nentqueued == 0) 1853 return 0; 1854 1855 /* 1856 * Calculate the appropriate error return. 1857 */ 1858 runningcode = 0; 1859 if (nerror) 1860 runningcode = EIO; 1861 1862 if (uap->mode == LIO_WAIT) { 1863 int command, found, jobref; 1864 1865 for (;;) { 1866 found = 0; 1867 for (i = 0; i < uap->nent; i++) { 1868 /* 1869 * Fetch address of the control buf pointer in 1870 * user space. 1871 */ 1872 iocb = (struct aiocb *) 1873 (intptr_t)fuword(&cbptr[i]); 1874 if (((intptr_t)iocb == -1) || ((intptr_t)iocb 1875 == 0)) 1876 continue; 1877 1878 /* 1879 * Fetch the associated command from user space. 1880 */ 1881 command = fuword(&iocb->aio_lio_opcode); 1882 if (command == LIO_NOP) { 1883 found++; 1884 continue; 1885 } 1886 1887 jobref = fuword(&iocb->_aiocb_private.kernelinfo); 1888 1889 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1890 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1891 == jobref) { 1892 if (cb->uaiocb.aio_lio_opcode 1893 == LIO_WRITE) { 1894 lp->lwp_ru.ru_oublock += 1895 cb->outputcharge; 1896 cb->outputcharge = 0; 1897 } else if (cb->uaiocb.aio_lio_opcode 1898 == LIO_READ) { 1899 lp->lwp_ru.ru_inblock += 1900 cb->inputcharge; 1901 cb->inputcharge = 0; 1902 } 1903 found++; 1904 break; 1905 } 1906 } 1907 1908 crit_enter(); 1909 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) { 1910 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1911 == jobref) { 1912 found++; 1913 break; 1914 } 1915 } 1916 crit_exit(); 1917 } 1918 1919 /* 1920 * If all I/Os have been disposed of, then we can 1921 * return. 1922 */ 1923 if (found == nentqueued) 1924 return runningcode; 1925 1926 ki->kaio_flags |= KAIO_WAKEUP; 1927 error = tsleep(p, PCATCH, "aiospn", 0); 1928 1929 if (error == EINTR) 1930 return EINTR; 1931 else if (error == EWOULDBLOCK) 1932 return EAGAIN; 1933 } 1934 } 1935 1936 return runningcode; 1937 #endif /* VFS_AIO */ 1938 } 1939 1940 #ifdef VFS_AIO 1941 /* 1942 * This is a weird hack so that we can post a signal. It is safe to do so from 1943 * a timeout routine, but *not* from an interrupt routine. 1944 */ 1945 static void 1946 process_signal(void *aioj) 1947 { 1948 struct aiocblist *aiocbe = aioj; 1949 struct aio_liojob *lj = aiocbe->lio; 1950 struct aiocb *cb = &aiocbe->uaiocb; 1951 1952 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) && 1953 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) { 1954 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo); 1955 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 1956 } 1957 1958 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) 1959 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo); 1960 } 1961 1962 /* 1963 * Interrupt handler for physio, performs the necessary process wakeups, and 1964 * signals. 1965 */ 1966 static void 1967 aio_physwakeup(struct bio *bio) 1968 { 1969 struct buf *bp = bio->bio_buf; 1970 struct aiocblist *aiocbe; 1971 struct proc *p; 1972 struct kaioinfo *ki; 1973 struct aio_liojob *lj; 1974 1975 aiocbe = bio->bio_caller_info2.ptr; 1976 1977 if (aiocbe) { 1978 p = bio->bio_caller_info1.ptr; 1979 1980 aiocbe->jobstate = JOBST_JOBBFINISHED; 1981 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 1982 aiocbe->uaiocb._aiocb_private.error = 0; 1983 aiocbe->jobflags |= AIOCBLIST_DONE; 1984 1985 if (bp->b_flags & B_ERROR) 1986 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 1987 1988 lj = aiocbe->lio; 1989 if (lj) { 1990 lj->lioj_buffer_finished_count++; 1991 1992 /* 1993 * wakeup/signal if all of the interrupt jobs are done. 1994 */ 1995 if (lj->lioj_buffer_finished_count == 1996 lj->lioj_buffer_count) { 1997 /* 1998 * Post a signal if it is called for. 1999 */ 2000 if ((lj->lioj_flags & 2001 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == 2002 LIOJ_SIGNAL) { 2003 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2004 callout_reset(&aiocbe->timeout, 0, 2005 process_signal, aiocbe); 2006 } 2007 } 2008 } 2009 2010 ki = p->p_aioinfo; 2011 if (ki) { 2012 ki->kaio_buffer_finished_count++; 2013 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 2014 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 2015 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 2016 2017 KNOTE(&aiocbe->klist, 0); 2018 /* Do the wakeup. */ 2019 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) { 2020 ki->kaio_flags &= ~KAIO_WAKEUP; 2021 wakeup(p); 2022 } 2023 } 2024 2025 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 2026 callout_reset(&aiocbe->timeout, 0, 2027 process_signal, aiocbe); 2028 } 2029 } 2030 bp->b_cmd = BUF_CMD_DONE; 2031 wakeup(bp); 2032 } 2033 #endif /* VFS_AIO */ 2034 2035 /* syscall - wait for the next completion of an aio request */ 2036 int 2037 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap) 2038 { 2039 #ifndef VFS_AIO 2040 return ENOSYS; 2041 #else 2042 struct proc *p = curproc; 2043 struct lwp *lp = curthread->td_lwp; 2044 struct timeval atv; 2045 struct timespec ts; 2046 struct kaioinfo *ki; 2047 struct aiocblist *cb = NULL; 2048 int error, timo; 2049 2050 suword(uap->aiocbp, (int)NULL); 2051 2052 timo = 0; 2053 if (uap->timeout) { 2054 /* Get timespec struct. */ 2055 error = copyin(uap->timeout, &ts, sizeof(ts)); 2056 if (error) 2057 return error; 2058 2059 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000)) 2060 return (EINVAL); 2061 2062 TIMESPEC_TO_TIMEVAL(&atv, &ts); 2063 if (itimerfix(&atv)) 2064 return (EINVAL); 2065 timo = tvtohz_high(&atv); 2066 } 2067 2068 ki = p->p_aioinfo; 2069 if (ki == NULL) 2070 return EAGAIN; 2071 2072 for (;;) { 2073 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) { 2074 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2075 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2076 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2077 lp->lwp_ru.ru_oublock += 2078 cb->outputcharge; 2079 cb->outputcharge = 0; 2080 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2081 lp->lwp_ru.ru_inblock += cb->inputcharge; 2082 cb->inputcharge = 0; 2083 } 2084 aio_free_entry(cb); 2085 return cb->uaiocb._aiocb_private.error; 2086 } 2087 2088 crit_enter(); 2089 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) { 2090 crit_exit(); 2091 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2092 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2093 aio_free_entry(cb); 2094 return cb->uaiocb._aiocb_private.error; 2095 } 2096 2097 ki->kaio_flags |= KAIO_WAKEUP; 2098 error = tsleep(p, PCATCH, "aiowc", timo); 2099 crit_exit(); 2100 2101 if (error == ERESTART) 2102 return EINTR; 2103 else if (error < 0) 2104 return error; 2105 else if (error == EINTR) 2106 return EINTR; 2107 else if (error == EWOULDBLOCK) 2108 return EAGAIN; 2109 } 2110 #endif /* VFS_AIO */ 2111 } 2112 2113 #ifndef VFS_AIO 2114 static int 2115 filt_aioattach(struct knote *kn) 2116 { 2117 2118 return (ENXIO); 2119 } 2120 2121 struct filterops aio_filtops = 2122 { 0, filt_aioattach, NULL, NULL }; 2123 2124 #else 2125 /* kqueue attach function */ 2126 static int 2127 filt_aioattach(struct knote *kn) 2128 { 2129 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2130 2131 /* 2132 * The aiocbe pointer must be validated before using it, so 2133 * registration is restricted to the kernel; the user cannot 2134 * set EV_FLAG1. 2135 */ 2136 if ((kn->kn_flags & EV_FLAG1) == 0) 2137 return (EPERM); 2138 kn->kn_flags &= ~EV_FLAG1; 2139 2140 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext); 2141 2142 return (0); 2143 } 2144 2145 /* kqueue detach function */ 2146 static void 2147 filt_aiodetach(struct knote *kn) 2148 { 2149 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2150 2151 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext); 2152 } 2153 2154 /* kqueue filter function */ 2155 /*ARGSUSED*/ 2156 static int 2157 filt_aio(struct knote *kn, long hint) 2158 { 2159 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2160 2161 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2162 if (aiocbe->jobstate != JOBST_JOBFINISHED && 2163 aiocbe->jobstate != JOBST_JOBBFINISHED) 2164 return (0); 2165 kn->kn_flags |= EV_EOF; 2166 return (1); 2167 } 2168 2169 struct filterops aio_filtops = 2170 { 0, filt_aioattach, filt_aiodetach, filt_aio }; 2171 #endif /* VFS_AIO */ 2172