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 1.40 2007/04/30 07:18:54 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 *curcp, *mycp, *userp; 623 struct vmspace *myvm; 624 struct ucred *cr; 625 626 /* 627 * Local copies of curproc (cp) and vmspace (myvm). Get extra 628 * reference on myvm so we can use pmap_replacevm() 629 */ 630 mycp = curproc; 631 myvm = mycp->p_vmspace; 632 sysref_get(&myvm->vm_sysref); 633 634 if (mycp->p_textvp) { 635 vrele(mycp->p_textvp); 636 mycp->p_textvp = NULL; 637 } 638 639 /* 640 * Allocate and ready the aio control info. There is one aiop structure 641 * per daemon. 642 */ 643 aiop = zalloc(aiop_zone); 644 aiop->aioproc = mycp; 645 aiop->aioprocflags |= AIOP_FREE; 646 647 crit_enter(); 648 649 /* 650 * Place thread (lightweight process) onto the AIO free thread list. 651 */ 652 if (TAILQ_EMPTY(&aio_freeproc)) 653 wakeup(&aio_freeproc); 654 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 655 656 crit_exit(); 657 658 /* Make up a name for the daemon. */ 659 strcpy(mycp->p_comm, "aiod"); 660 661 /* 662 * Get rid of our current filedescriptors. AIOD's don't need any 663 * filedescriptors, except as temporarily inherited from the client. 664 * Credentials are also cloned, and made equivalent to "root". 665 */ 666 fdfree(mycp); 667 mycp->p_fd = NULL; 668 cr = cratom(&mycp->p_ucred); 669 cr->cr_uid = 0; 670 uireplace(&cr->cr_uidinfo, uifind(0)); 671 cr->cr_ngroups = 1; 672 cr->cr_groups[0] = 1; 673 674 /* The daemon resides in its own pgrp. */ 675 enterpgrp(mycp, mycp->p_pid, 1); 676 677 /* Mark special process type. */ 678 mycp->p_flag |= P_SYSTEM | P_KTHREADP; 679 680 /* 681 * Wakeup parent process. (Parent sleeps to keep from blasting away 682 * and creating too many daemons.) 683 */ 684 wakeup(mycp); 685 686 for (;;) { 687 /* 688 * curcp is the current daemon process context. 689 * userp is the current user process context. 690 */ 691 curcp = mycp; 692 693 /* 694 * Take daemon off of free queue 695 */ 696 if (aiop->aioprocflags & AIOP_FREE) { 697 crit_enter(); 698 TAILQ_REMOVE(&aio_freeproc, aiop, list); 699 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 700 aiop->aioprocflags &= ~AIOP_FREE; 701 crit_exit(); 702 } 703 aiop->aioprocflags &= ~AIOP_SCHED; 704 705 /* 706 * Check for jobs. 707 */ 708 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 709 cb = &aiocbe->uaiocb; 710 userp = aiocbe->userproc; 711 712 aiocbe->jobstate = JOBST_JOBRUNNING; 713 714 /* 715 * Connect to process address space for user program. 716 */ 717 if (userp != curcp) { 718 pmap_replacevm(mycp, userp->p_vmspace, 1); 719 curcp = userp; 720 } 721 722 ki = userp->p_aioinfo; 723 lj = aiocbe->lio; 724 725 /* Account for currently active jobs. */ 726 ki->kaio_active_count++; 727 728 /* Do the I/O function. */ 729 aio_process(aiocbe); 730 731 /* Decrement the active job count. */ 732 ki->kaio_active_count--; 733 734 /* 735 * Increment the completion count for wakeup/signal 736 * comparisons. 737 */ 738 aiocbe->jobflags |= AIOCBLIST_DONE; 739 ki->kaio_queue_finished_count++; 740 if (lj) 741 lj->lioj_queue_finished_count++; 742 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags 743 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) { 744 ki->kaio_flags &= ~KAIO_WAKEUP; 745 wakeup(userp); 746 } 747 748 crit_enter(); 749 if (lj && (lj->lioj_flags & 750 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) { 751 if ((lj->lioj_queue_finished_count == 752 lj->lioj_queue_count) && 753 (lj->lioj_buffer_finished_count == 754 lj->lioj_buffer_count)) { 755 ksignal(userp, 756 lj->lioj_signal.sigev_signo); 757 lj->lioj_flags |= 758 LIOJ_SIGNAL_POSTED; 759 } 760 } 761 crit_exit(); 762 763 aiocbe->jobstate = JOBST_JOBFINISHED; 764 765 crit_enter(); 766 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 767 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist); 768 crit_exit(); 769 KNOTE(&aiocbe->klist, 0); 770 771 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) { 772 wakeup(aiocbe); 773 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN; 774 } 775 776 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 777 ksignal(userp, cb->aio_sigevent.sigev_signo); 778 } 779 } 780 781 /* 782 * Disconnect from user address space. 783 */ 784 if (curcp != mycp) { 785 /* swap our original address space back in */ 786 pmap_replacevm(mycp, myvm, 1); 787 curcp = mycp; 788 } 789 790 /* 791 * If we are the first to be put onto the free queue, wakeup 792 * anyone waiting for a daemon. 793 */ 794 crit_enter(); 795 TAILQ_REMOVE(&aio_activeproc, aiop, list); 796 if (TAILQ_EMPTY(&aio_freeproc)) 797 wakeup(&aio_freeproc); 798 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 799 aiop->aioprocflags |= AIOP_FREE; 800 crit_exit(); 801 802 /* 803 * If daemon is inactive for a long time, allow it to exit, 804 * thereby freeing resources. 805 */ 806 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp, 807 0, "aiordy", aiod_lifetime)) { 808 crit_enter(); 809 if (TAILQ_EMPTY(&aio_jobs)) { 810 if ((aiop->aioprocflags & AIOP_FREE) && 811 (num_aio_procs > target_aio_procs)) { 812 TAILQ_REMOVE(&aio_freeproc, aiop, list); 813 crit_exit(); 814 zfree(aiop_zone, aiop); 815 num_aio_procs--; 816 #ifdef DIAGNOSTIC 817 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) { 818 kprintf("AIOD: bad vm refcnt for" 819 " exiting daemon: %d\n", 820 mycp->p_vmspace->vm_sysref.refcnt); 821 } 822 #endif 823 sysref_put(&myvm->vm_sysref); 824 exit1(0); 825 } 826 } 827 crit_exit(); 828 } 829 } 830 } 831 832 /* 833 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 834 * AIO daemon modifies its environment itself. 835 */ 836 static int 837 aio_newproc(void) 838 { 839 int error; 840 struct lwp *lp, *nlp; 841 struct proc *np; 842 843 lp = &lwp0; 844 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np); 845 if (error) 846 return error; 847 nlp = ONLY_LWP_IN_PROC(np); 848 cpu_set_fork_handler(nlp, aio_daemon, curproc); 849 start_forked_proc(lp, np); 850 851 /* 852 * Wait until daemon is started, but continue on just in case to 853 * handle error conditions. 854 */ 855 error = tsleep(np, 0, "aiosta", aiod_timeout); 856 num_aio_procs++; 857 858 return error; 859 } 860 861 /* 862 * Try the high-performance, low-overhead physio method for eligible 863 * VCHR devices. This method doesn't use an aio helper thread, and 864 * thus has very low overhead. 865 * 866 * Assumes that the caller, _aio_aqueue(), has incremented the file 867 * structure's reference count, preventing its deallocation for the 868 * duration of this call. 869 */ 870 static int 871 aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 872 { 873 int error; 874 struct aiocb *cb; 875 struct file *fp; 876 struct buf *bp; 877 struct vnode *vp; 878 struct kaioinfo *ki; 879 struct aio_liojob *lj; 880 int notify; 881 882 cb = &aiocbe->uaiocb; 883 fp = aiocbe->fd_file; 884 885 if (fp->f_type != DTYPE_VNODE) 886 return (-1); 887 888 vp = (struct vnode *)fp->f_data; 889 890 /* 891 * If its not a disk, we don't want to return a positive error. 892 * It causes the aio code to not fall through to try the thread 893 * way when you're talking to a regular file. 894 */ 895 if (!vn_isdisk(vp, &error)) { 896 if (error == ENOTBLK) 897 return (-1); 898 else 899 return (error); 900 } 901 902 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys) 903 return (-1); 904 905 if (cb->aio_nbytes > 906 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 907 return (-1); 908 909 ki = p->p_aioinfo; 910 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 911 return (-1); 912 913 ki->kaio_buffer_count++; 914 915 lj = aiocbe->lio; 916 if (lj) 917 lj->lioj_buffer_count++; 918 919 /* Create and build a buffer header for a transfer. */ 920 bp = getpbuf(NULL); 921 BUF_KERNPROC(bp); 922 923 /* 924 * Get a copy of the kva from the physical buffer. 925 */ 926 bp->b_bio1.bio_caller_info1.ptr = p; 927 error = 0; 928 929 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ? 930 BUF_CMD_WRITE : BUF_CMD_READ; 931 bp->b_bio1.bio_done = aio_physwakeup; 932 bp->b_bio1.bio_offset = cb->aio_offset; 933 934 /* Bring buffer into kernel space. */ 935 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) { 936 error = EFAULT; 937 goto doerror; 938 } 939 940 crit_enter(); 941 942 aiocbe->bp = bp; 943 bp->b_bio1.bio_caller_info2.ptr = aiocbe; 944 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list); 945 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 946 aiocbe->jobstate = JOBST_JOBQBUF; 947 cb->_aiocb_private.status = cb->aio_nbytes; 948 num_buf_aio++; 949 bp->b_error = 0; 950 951 crit_exit(); 952 953 /* Perform transfer. */ 954 dev_dstrategy(vp->v_rdev, &bp->b_bio1); 955 956 notify = 0; 957 crit_enter(); 958 959 /* 960 * If we had an error invoking the request, or an error in processing 961 * the request before we have returned, we process it as an error in 962 * transfer. Note that such an I/O error is not indicated immediately, 963 * but is returned using the aio_error mechanism. In this case, 964 * aio_suspend will return immediately. 965 */ 966 if (bp->b_error || (bp->b_flags & B_ERROR)) { 967 struct aiocb *job = aiocbe->uuaiocb; 968 969 aiocbe->uaiocb._aiocb_private.status = 0; 970 suword(&job->_aiocb_private.status, 0); 971 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 972 suword(&job->_aiocb_private.error, bp->b_error); 973 974 ki->kaio_buffer_finished_count++; 975 976 if (aiocbe->jobstate != JOBST_JOBBFINISHED) { 977 aiocbe->jobstate = JOBST_JOBBFINISHED; 978 aiocbe->jobflags |= AIOCBLIST_DONE; 979 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 980 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 981 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 982 notify = 1; 983 } 984 } 985 crit_exit(); 986 if (notify) 987 KNOTE(&aiocbe->klist, 0); 988 return 0; 989 990 doerror: 991 ki->kaio_buffer_count--; 992 if (lj) 993 lj->lioj_buffer_count--; 994 aiocbe->bp = NULL; 995 relpbuf(bp, NULL); 996 return error; 997 } 998 999 /* 1000 * This waits/tests physio completion. 1001 */ 1002 static int 1003 aio_fphysio(struct aiocblist *iocb) 1004 { 1005 struct buf *bp; 1006 int error; 1007 1008 bp = iocb->bp; 1009 1010 crit_enter(); 1011 while (bp->b_cmd != BUF_CMD_DONE) { 1012 if (tsleep(bp, 0, "physstr", aiod_timeout)) { 1013 if (bp->b_cmd != BUF_CMD_DONE) { 1014 crit_exit(); 1015 return EINPROGRESS; 1016 } else { 1017 break; 1018 } 1019 } 1020 } 1021 crit_exit(); 1022 1023 /* Release mapping into kernel space. */ 1024 vunmapbuf(bp); 1025 iocb->bp = 0; 1026 1027 error = 0; 1028 1029 /* Check for an error. */ 1030 if (bp->b_flags & B_ERROR) 1031 error = bp->b_error; 1032 1033 relpbuf(bp, NULL); 1034 return (error); 1035 } 1036 #endif /* VFS_AIO */ 1037 1038 /* 1039 * Wake up aio requests that may be serviceable now. 1040 */ 1041 void 1042 aio_swake(struct socket *so, struct signalsockbuf *ssb) 1043 { 1044 #ifndef VFS_AIO 1045 return; 1046 #else 1047 struct aiocblist *cb,*cbn; 1048 struct proc *p; 1049 struct kaioinfo *ki = NULL; 1050 int opcode, wakecount = 0; 1051 struct aioproclist *aiop; 1052 1053 if (ssb == &so->so_snd) { 1054 opcode = LIO_WRITE; 1055 so->so_snd.ssb_flags &= ~SSB_AIO; 1056 } else { 1057 opcode = LIO_READ; 1058 so->so_rcv.ssb_flags &= ~SSB_AIO; 1059 } 1060 1061 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) { 1062 cbn = TAILQ_NEXT(cb, list); 1063 if (opcode == cb->uaiocb.aio_lio_opcode) { 1064 p = cb->userproc; 1065 ki = p->p_aioinfo; 1066 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1067 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist); 1068 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1069 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist); 1070 wakecount++; 1071 if (cb->jobstate != JOBST_JOBQGLOBAL) 1072 panic("invalid queue value"); 1073 } 1074 } 1075 1076 while (wakecount--) { 1077 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) { 1078 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1079 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1080 aiop->aioprocflags &= ~AIOP_FREE; 1081 wakeup(aiop->aioproc); 1082 } 1083 } 1084 #endif /* VFS_AIO */ 1085 } 1086 1087 #ifdef VFS_AIO 1088 /* 1089 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1090 * technique is done in this code. 1091 */ 1092 static int 1093 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type) 1094 { 1095 struct proc *p = curproc; 1096 struct filedesc *fdp; 1097 struct file *fp; 1098 unsigned int fd; 1099 struct socket *so; 1100 int error; 1101 int opcode, user_opcode; 1102 struct aiocblist *aiocbe; 1103 struct aioproclist *aiop; 1104 struct kaioinfo *ki; 1105 struct kevent kev; 1106 struct kqueue *kq; 1107 struct file *kq_fp; 1108 1109 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL) 1110 TAILQ_REMOVE(&aio_freejobs, aiocbe, list); 1111 else 1112 aiocbe = zalloc (aiocb_zone); 1113 1114 aiocbe->inputcharge = 0; 1115 aiocbe->outputcharge = 0; 1116 callout_init(&aiocbe->timeout); 1117 SLIST_INIT(&aiocbe->klist); 1118 1119 suword(&job->_aiocb_private.status, -1); 1120 suword(&job->_aiocb_private.error, 0); 1121 suword(&job->_aiocb_private.kernelinfo, -1); 1122 1123 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb)); 1124 if (error) { 1125 suword(&job->_aiocb_private.error, error); 1126 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1127 return error; 1128 } 1129 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL && 1130 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1131 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1132 return EINVAL; 1133 } 1134 1135 /* Save userspace address of the job info. */ 1136 aiocbe->uuaiocb = job; 1137 1138 /* Get the opcode. */ 1139 user_opcode = aiocbe->uaiocb.aio_lio_opcode; 1140 if (type != LIO_NOP) 1141 aiocbe->uaiocb.aio_lio_opcode = type; 1142 opcode = aiocbe->uaiocb.aio_lio_opcode; 1143 1144 /* Get the fd info for process. */ 1145 fdp = p->p_fd; 1146 1147 /* 1148 * Range check file descriptor. 1149 */ 1150 fd = aiocbe->uaiocb.aio_fildes; 1151 if (fd >= fdp->fd_nfiles) { 1152 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1153 if (type == 0) 1154 suword(&job->_aiocb_private.error, EBADF); 1155 return EBADF; 1156 } 1157 1158 fp = aiocbe->fd_file = fdp->fd_files[fd].fp; 1159 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 1160 0))) { 1161 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1162 if (type == 0) 1163 suword(&job->_aiocb_private.error, EBADF); 1164 return EBADF; 1165 } 1166 fhold(fp); 1167 1168 if (aiocbe->uaiocb.aio_offset == -1LL) { 1169 error = EINVAL; 1170 goto aqueue_fail; 1171 } 1172 error = suword(&job->_aiocb_private.kernelinfo, jobrefid); 1173 if (error) { 1174 error = EINVAL; 1175 goto aqueue_fail; 1176 } 1177 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid; 1178 if (jobrefid == LONG_MAX) 1179 jobrefid = 1; 1180 else 1181 jobrefid++; 1182 1183 if (opcode == LIO_NOP) { 1184 fdrop(fp); 1185 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1186 if (type == 0) { 1187 suword(&job->_aiocb_private.error, 0); 1188 suword(&job->_aiocb_private.status, 0); 1189 suword(&job->_aiocb_private.kernelinfo, 0); 1190 } 1191 return 0; 1192 } 1193 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) { 1194 if (type == 0) 1195 suword(&job->_aiocb_private.status, 0); 1196 error = EINVAL; 1197 goto aqueue_fail; 1198 } 1199 1200 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) { 1201 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1202 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr; 1203 } 1204 else { 1205 /* 1206 * This method for requesting kevent-based notification won't 1207 * work on the alpha, since we're passing in a pointer 1208 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT- 1209 * based method instead. 1210 */ 1211 if (user_opcode == LIO_NOP || user_opcode == LIO_READ || 1212 user_opcode == LIO_WRITE) 1213 goto no_kqueue; 1214 1215 error = copyin((struct kevent *)(uintptr_t)user_opcode, 1216 &kev, sizeof(kev)); 1217 if (error) 1218 goto aqueue_fail; 1219 } 1220 if ((u_int)kev.ident >= fdp->fd_nfiles || 1221 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL || 1222 (kq_fp->f_type != DTYPE_KQUEUE)) { 1223 error = EBADF; 1224 goto aqueue_fail; 1225 } 1226 kq = (struct kqueue *)kq_fp->f_data; 1227 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1228 kev.filter = EVFILT_AIO; 1229 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 1230 kev.data = (intptr_t)aiocbe; 1231 /* XXX lwp kqueue_register takes a thread, but only uses its proc */ 1232 error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread); 1233 aqueue_fail: 1234 if (error) { 1235 fdrop(fp); 1236 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list); 1237 if (type == 0) 1238 suword(&job->_aiocb_private.error, error); 1239 goto done; 1240 } 1241 no_kqueue: 1242 1243 suword(&job->_aiocb_private.error, EINPROGRESS); 1244 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1245 aiocbe->userproc = p; 1246 aiocbe->jobflags = 0; 1247 aiocbe->lio = lj; 1248 ki = p->p_aioinfo; 1249 1250 if (fp->f_type == DTYPE_SOCKET) { 1251 /* 1252 * Alternate queueing for socket ops: Reach down into the 1253 * descriptor to get the socket data. Then check to see if the 1254 * socket is ready to be read or written (based on the requested 1255 * operation). 1256 * 1257 * If it is not ready for io, then queue the aiocbe on the 1258 * socket, and set the flags so we get a call when ssb_notify() 1259 * happens. 1260 */ 1261 so = (struct socket *)fp->f_data; 1262 crit_enter(); 1263 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1264 LIO_WRITE) && (!sowriteable(so)))) { 1265 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1266 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist); 1267 if (opcode == LIO_READ) 1268 so->so_rcv.ssb_flags |= SSB_AIO; 1269 else 1270 so->so_snd.ssb_flags |= SSB_AIO; 1271 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */ 1272 ki->kaio_queue_count++; 1273 num_queue_count++; 1274 crit_exit(); 1275 error = 0; 1276 goto done; 1277 } 1278 crit_exit(); 1279 } 1280 1281 if ((error = aio_qphysio(p, aiocbe)) == 0) 1282 goto done; 1283 if (error > 0) { 1284 suword(&job->_aiocb_private.status, 0); 1285 aiocbe->uaiocb._aiocb_private.error = error; 1286 suword(&job->_aiocb_private.error, error); 1287 goto done; 1288 } 1289 1290 /* No buffer for daemon I/O. */ 1291 aiocbe->bp = NULL; 1292 1293 ki->kaio_queue_count++; 1294 if (lj) 1295 lj->lioj_queue_count++; 1296 crit_enter(); 1297 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1298 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1299 crit_exit(); 1300 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1301 1302 num_queue_count++; 1303 error = 0; 1304 1305 /* 1306 * If we don't have a free AIO process, and we are below our quota, then 1307 * start one. Otherwise, depend on the subsequent I/O completions to 1308 * pick-up this job. If we don't successfully create the new process 1309 * (thread) due to resource issues, we return an error for now (EAGAIN), 1310 * which is likely not the correct thing to do. 1311 */ 1312 crit_enter(); 1313 retryproc: 1314 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1315 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1316 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1317 aiop->aioprocflags &= ~AIOP_FREE; 1318 wakeup(aiop->aioproc); 1319 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1320 ((ki->kaio_active_count + num_aio_resv_start) < 1321 ki->kaio_maxactive_count)) { 1322 num_aio_resv_start++; 1323 if ((error = aio_newproc()) == 0) { 1324 num_aio_resv_start--; 1325 goto retryproc; 1326 } 1327 num_aio_resv_start--; 1328 } 1329 crit_exit(); 1330 done: 1331 return error; 1332 } 1333 1334 /* 1335 * This routine queues an AIO request, checking for quotas. 1336 */ 1337 static int 1338 aio_aqueue(struct aiocb *job, int type) 1339 { 1340 struct proc *p = curproc; 1341 struct kaioinfo *ki; 1342 1343 if (p->p_aioinfo == NULL) 1344 aio_init_aioinfo(p); 1345 1346 if (num_queue_count >= max_queue_count) 1347 return EAGAIN; 1348 1349 ki = p->p_aioinfo; 1350 if (ki->kaio_queue_count >= ki->kaio_qallowed_count) 1351 return EAGAIN; 1352 1353 return _aio_aqueue(job, NULL, type); 1354 } 1355 #endif /* VFS_AIO */ 1356 1357 /* 1358 * Support the aio_return system call, as a side-effect, kernel resources are 1359 * released. 1360 */ 1361 int 1362 sys_aio_return(struct aio_return_args *uap) 1363 { 1364 #ifndef VFS_AIO 1365 return ENOSYS; 1366 #else 1367 struct proc *p = curproc; 1368 struct lwp *lp = curthread->td_lwp; 1369 long jobref; 1370 struct aiocblist *cb, *ncb; 1371 struct aiocb *ujob; 1372 struct kaioinfo *ki; 1373 1374 ki = p->p_aioinfo; 1375 if (ki == NULL) 1376 return EINVAL; 1377 1378 ujob = uap->aiocbp; 1379 1380 jobref = fuword(&ujob->_aiocb_private.kernelinfo); 1381 if (jobref == -1 || jobref == 0) 1382 return EINVAL; 1383 1384 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1385 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == 1386 jobref) { 1387 if (ujob == cb->uuaiocb) { 1388 uap->sysmsg_result = 1389 cb->uaiocb._aiocb_private.status; 1390 } else 1391 uap->sysmsg_result = EFAULT; 1392 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1393 lp->lwp_ru.ru_oublock += cb->outputcharge; 1394 cb->outputcharge = 0; 1395 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1396 lp->lwp_ru.ru_inblock += cb->inputcharge; 1397 cb->inputcharge = 0; 1398 } 1399 aio_free_entry(cb); 1400 return 0; 1401 } 1402 } 1403 crit_enter(); 1404 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) { 1405 ncb = TAILQ_NEXT(cb, plist); 1406 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) 1407 == jobref) { 1408 crit_exit(); 1409 if (ujob == cb->uuaiocb) { 1410 uap->sysmsg_result = 1411 cb->uaiocb._aiocb_private.status; 1412 } else 1413 uap->sysmsg_result = EFAULT; 1414 aio_free_entry(cb); 1415 return 0; 1416 } 1417 } 1418 crit_exit(); 1419 1420 return (EINVAL); 1421 #endif /* VFS_AIO */ 1422 } 1423 1424 /* 1425 * Allow a process to wakeup when any of the I/O requests are completed. 1426 */ 1427 int 1428 sys_aio_suspend(struct aio_suspend_args *uap) 1429 { 1430 #ifndef VFS_AIO 1431 return ENOSYS; 1432 #else 1433 struct proc *p = curproc; 1434 struct timeval atv; 1435 struct timespec ts; 1436 struct aiocb *const *cbptr, *cbp; 1437 struct kaioinfo *ki; 1438 struct aiocblist *cb; 1439 int i; 1440 int njoblist; 1441 int error, timo; 1442 long *ijoblist; 1443 struct aiocb **ujoblist; 1444 1445 if (uap->nent > AIO_LISTIO_MAX) 1446 return EINVAL; 1447 1448 timo = 0; 1449 if (uap->timeout) { 1450 /* Get timespec struct. */ 1451 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1452 return error; 1453 1454 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) 1455 return (EINVAL); 1456 1457 TIMESPEC_TO_TIMEVAL(&atv, &ts); 1458 if (itimerfix(&atv)) 1459 return (EINVAL); 1460 timo = tvtohz_high(&atv); 1461 } 1462 1463 ki = p->p_aioinfo; 1464 if (ki == NULL) 1465 return EAGAIN; 1466 1467 njoblist = 0; 1468 ijoblist = zalloc(aiol_zone); 1469 ujoblist = zalloc(aiol_zone); 1470 cbptr = uap->aiocbp; 1471 1472 for (i = 0; i < uap->nent; i++) { 1473 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1474 if (cbp == 0) 1475 continue; 1476 ujoblist[njoblist] = cbp; 1477 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo); 1478 njoblist++; 1479 } 1480 1481 if (njoblist == 0) { 1482 zfree(aiol_zone, ijoblist); 1483 zfree(aiol_zone, ujoblist); 1484 return 0; 1485 } 1486 1487 error = 0; 1488 for (;;) { 1489 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1490 for (i = 0; i < njoblist; i++) { 1491 if (((intptr_t) 1492 cb->uaiocb._aiocb_private.kernelinfo) == 1493 ijoblist[i]) { 1494 if (ujoblist[i] != cb->uuaiocb) 1495 error = EINVAL; 1496 zfree(aiol_zone, ijoblist); 1497 zfree(aiol_zone, ujoblist); 1498 return error; 1499 } 1500 } 1501 } 1502 1503 crit_enter(); 1504 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = 1505 TAILQ_NEXT(cb, plist)) { 1506 for (i = 0; i < njoblist; i++) { 1507 if (((intptr_t) 1508 cb->uaiocb._aiocb_private.kernelinfo) == 1509 ijoblist[i]) { 1510 crit_exit(); 1511 if (ujoblist[i] != cb->uuaiocb) 1512 error = EINVAL; 1513 zfree(aiol_zone, ijoblist); 1514 zfree(aiol_zone, ujoblist); 1515 return error; 1516 } 1517 } 1518 } 1519 1520 ki->kaio_flags |= KAIO_WAKEUP; 1521 error = tsleep(p, PCATCH, "aiospn", timo); 1522 crit_exit(); 1523 1524 if (error == ERESTART || error == EINTR) { 1525 zfree(aiol_zone, ijoblist); 1526 zfree(aiol_zone, ujoblist); 1527 return EINTR; 1528 } else if (error == EWOULDBLOCK) { 1529 zfree(aiol_zone, ijoblist); 1530 zfree(aiol_zone, ujoblist); 1531 return EAGAIN; 1532 } 1533 } 1534 1535 /* NOTREACHED */ 1536 return EINVAL; 1537 #endif /* VFS_AIO */ 1538 } 1539 1540 /* 1541 * aio_cancel cancels any non-physio aio operations not currently in 1542 * progress. 1543 */ 1544 int 1545 sys_aio_cancel(struct aio_cancel_args *uap) 1546 { 1547 #ifndef VFS_AIO 1548 return ENOSYS; 1549 #else 1550 struct proc *p = curproc; 1551 struct kaioinfo *ki; 1552 struct aiocblist *cbe, *cbn; 1553 struct file *fp; 1554 struct filedesc *fdp; 1555 struct socket *so; 1556 struct proc *po; 1557 int error; 1558 int cancelled=0; 1559 int notcancelled=0; 1560 struct vnode *vp; 1561 1562 fdp = p->p_fd; 1563 if ((u_int)uap->fd >= fdp->fd_nfiles || 1564 (fp = fdp->fd_files[uap->fd].fp) == NULL) 1565 return (EBADF); 1566 1567 if (fp->f_type == DTYPE_VNODE) { 1568 vp = (struct vnode *)fp->f_data; 1569 1570 if (vn_isdisk(vp,&error)) { 1571 uap->sysmsg_result = AIO_NOTCANCELED; 1572 return 0; 1573 } 1574 } else if (fp->f_type == DTYPE_SOCKET) { 1575 so = (struct socket *)fp->f_data; 1576 1577 crit_enter(); 1578 1579 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) { 1580 cbn = TAILQ_NEXT(cbe, list); 1581 if ((uap->aiocbp == NULL) || 1582 (uap->aiocbp == cbe->uuaiocb) ) { 1583 po = cbe->userproc; 1584 ki = po->p_aioinfo; 1585 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 1586 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist); 1587 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist); 1588 if (ki->kaio_flags & KAIO_WAKEUP) { 1589 wakeup(po); 1590 } 1591 cbe->jobstate = JOBST_JOBFINISHED; 1592 cbe->uaiocb._aiocb_private.status=-1; 1593 cbe->uaiocb._aiocb_private.error=ECANCELED; 1594 cancelled++; 1595 /* XXX cancelled, knote? */ 1596 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1597 SIGEV_SIGNAL) 1598 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1599 if (uap->aiocbp) 1600 break; 1601 } 1602 } 1603 crit_exit(); 1604 1605 if ((cancelled) && (uap->aiocbp)) { 1606 uap->sysmsg_result = AIO_CANCELED; 1607 return 0; 1608 } 1609 } 1610 ki=p->p_aioinfo; 1611 if (ki == NULL) 1612 goto done; 1613 crit_enter(); 1614 1615 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) { 1616 cbn = TAILQ_NEXT(cbe, plist); 1617 1618 if ((uap->fd == cbe->uaiocb.aio_fildes) && 1619 ((uap->aiocbp == NULL ) || 1620 (uap->aiocbp == cbe->uuaiocb))) { 1621 1622 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 1623 TAILQ_REMOVE(&aio_jobs, cbe, list); 1624 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 1625 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, 1626 plist); 1627 cancelled++; 1628 ki->kaio_queue_finished_count++; 1629 cbe->jobstate = JOBST_JOBFINISHED; 1630 cbe->uaiocb._aiocb_private.status = -1; 1631 cbe->uaiocb._aiocb_private.error = ECANCELED; 1632 /* XXX cancelled, knote? */ 1633 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1634 SIGEV_SIGNAL) 1635 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1636 } else { 1637 notcancelled++; 1638 } 1639 } 1640 } 1641 crit_exit(); 1642 done: 1643 if (notcancelled) { 1644 uap->sysmsg_result = AIO_NOTCANCELED; 1645 return 0; 1646 } 1647 if (cancelled) { 1648 uap->sysmsg_result = AIO_CANCELED; 1649 return 0; 1650 } 1651 uap->sysmsg_result = AIO_ALLDONE; 1652 1653 return 0; 1654 #endif /* VFS_AIO */ 1655 } 1656 1657 /* 1658 * aio_error is implemented in the kernel level for compatibility purposes only. 1659 * For a user mode async implementation, it would be best to do it in a userland 1660 * subroutine. 1661 */ 1662 int 1663 sys_aio_error(struct aio_error_args *uap) 1664 { 1665 #ifndef VFS_AIO 1666 return ENOSYS; 1667 #else 1668 struct proc *p = curproc; 1669 struct aiocblist *cb; 1670 struct kaioinfo *ki; 1671 long jobref; 1672 1673 ki = p->p_aioinfo; 1674 if (ki == NULL) 1675 return EINVAL; 1676 1677 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo); 1678 if ((jobref == -1) || (jobref == 0)) 1679 return EINVAL; 1680 1681 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1682 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1683 jobref) { 1684 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1685 return 0; 1686 } 1687 } 1688 1689 crit_enter(); 1690 1691 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb, 1692 plist)) { 1693 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1694 jobref) { 1695 uap->sysmsg_result = EINPROGRESS; 1696 crit_exit(); 1697 return 0; 1698 } 1699 } 1700 1701 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb, 1702 plist)) { 1703 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1704 jobref) { 1705 uap->sysmsg_result = EINPROGRESS; 1706 crit_exit(); 1707 return 0; 1708 } 1709 } 1710 crit_exit(); 1711 1712 crit_enter(); 1713 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb, 1714 plist)) { 1715 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1716 jobref) { 1717 uap->sysmsg_result = cb->uaiocb._aiocb_private.error; 1718 crit_exit(); 1719 return 0; 1720 } 1721 } 1722 1723 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb, 1724 plist)) { 1725 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1726 jobref) { 1727 uap->sysmsg_result = EINPROGRESS; 1728 crit_exit(); 1729 return 0; 1730 } 1731 } 1732 crit_exit(); 1733 1734 #if (0) 1735 /* 1736 * Hack for lio. 1737 */ 1738 status = fuword(&uap->aiocbp->_aiocb_private.status); 1739 if (status == -1) 1740 return fuword(&uap->aiocbp->_aiocb_private.error); 1741 #endif 1742 return EINVAL; 1743 #endif /* VFS_AIO */ 1744 } 1745 1746 /* syscall - asynchronous read from a file (REALTIME) */ 1747 int 1748 sys_aio_read(struct aio_read_args *uap) 1749 { 1750 #ifndef VFS_AIO 1751 return ENOSYS; 1752 #else 1753 return aio_aqueue(uap->aiocbp, LIO_READ); 1754 #endif /* VFS_AIO */ 1755 } 1756 1757 /* syscall - asynchronous write to a file (REALTIME) */ 1758 int 1759 sys_aio_write(struct aio_write_args *uap) 1760 { 1761 #ifndef VFS_AIO 1762 return ENOSYS; 1763 #else 1764 return aio_aqueue(uap->aiocbp, LIO_WRITE); 1765 #endif /* VFS_AIO */ 1766 } 1767 1768 /* syscall - XXX undocumented */ 1769 int 1770 sys_lio_listio(struct lio_listio_args *uap) 1771 { 1772 #ifndef VFS_AIO 1773 return ENOSYS; 1774 #else 1775 struct proc *p = curproc; 1776 struct lwp *lp = curthread->td_lwp; 1777 int nent, nentqueued; 1778 struct aiocb *iocb, * const *cbptr; 1779 struct aiocblist *cb; 1780 struct kaioinfo *ki; 1781 struct aio_liojob *lj; 1782 int error, runningcode; 1783 int nerror; 1784 int i; 1785 1786 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 1787 return EINVAL; 1788 1789 nent = uap->nent; 1790 if (nent > AIO_LISTIO_MAX) 1791 return EINVAL; 1792 1793 if (p->p_aioinfo == NULL) 1794 aio_init_aioinfo(p); 1795 1796 if ((nent + num_queue_count) > max_queue_count) 1797 return EAGAIN; 1798 1799 ki = p->p_aioinfo; 1800 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) 1801 return EAGAIN; 1802 1803 lj = zalloc(aiolio_zone); 1804 if (!lj) 1805 return EAGAIN; 1806 1807 lj->lioj_flags = 0; 1808 lj->lioj_buffer_count = 0; 1809 lj->lioj_buffer_finished_count = 0; 1810 lj->lioj_queue_count = 0; 1811 lj->lioj_queue_finished_count = 0; 1812 lj->lioj_ki = ki; 1813 1814 /* 1815 * Setup signal. 1816 */ 1817 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 1818 error = copyin(uap->sig, &lj->lioj_signal, 1819 sizeof(lj->lioj_signal)); 1820 if (error) { 1821 zfree(aiolio_zone, lj); 1822 return error; 1823 } 1824 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 1825 zfree(aiolio_zone, lj); 1826 return EINVAL; 1827 } 1828 lj->lioj_flags |= LIOJ_SIGNAL; 1829 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED; 1830 } else 1831 lj->lioj_flags &= ~LIOJ_SIGNAL; 1832 1833 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 1834 /* 1835 * Get pointers to the list of I/O requests. 1836 */ 1837 nerror = 0; 1838 nentqueued = 0; 1839 cbptr = uap->acb_list; 1840 for (i = 0; i < uap->nent; i++) { 1841 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1842 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) { 1843 error = _aio_aqueue(iocb, lj, 0); 1844 if (error == 0) 1845 nentqueued++; 1846 else 1847 nerror++; 1848 } 1849 } 1850 1851 /* 1852 * If we haven't queued any, then just return error. 1853 */ 1854 if (nentqueued == 0) 1855 return 0; 1856 1857 /* 1858 * Calculate the appropriate error return. 1859 */ 1860 runningcode = 0; 1861 if (nerror) 1862 runningcode = EIO; 1863 1864 if (uap->mode == LIO_WAIT) { 1865 int command, found, jobref; 1866 1867 for (;;) { 1868 found = 0; 1869 for (i = 0; i < uap->nent; i++) { 1870 /* 1871 * Fetch address of the control buf pointer in 1872 * user space. 1873 */ 1874 iocb = (struct aiocb *) 1875 (intptr_t)fuword(&cbptr[i]); 1876 if (((intptr_t)iocb == -1) || ((intptr_t)iocb 1877 == 0)) 1878 continue; 1879 1880 /* 1881 * Fetch the associated command from user space. 1882 */ 1883 command = fuword(&iocb->aio_lio_opcode); 1884 if (command == LIO_NOP) { 1885 found++; 1886 continue; 1887 } 1888 1889 jobref = fuword(&iocb->_aiocb_private.kernelinfo); 1890 1891 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1892 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1893 == jobref) { 1894 if (cb->uaiocb.aio_lio_opcode 1895 == LIO_WRITE) { 1896 lp->lwp_ru.ru_oublock += 1897 cb->outputcharge; 1898 cb->outputcharge = 0; 1899 } else if (cb->uaiocb.aio_lio_opcode 1900 == LIO_READ) { 1901 lp->lwp_ru.ru_inblock += 1902 cb->inputcharge; 1903 cb->inputcharge = 0; 1904 } 1905 found++; 1906 break; 1907 } 1908 } 1909 1910 crit_enter(); 1911 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) { 1912 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 1913 == jobref) { 1914 found++; 1915 break; 1916 } 1917 } 1918 crit_exit(); 1919 } 1920 1921 /* 1922 * If all I/Os have been disposed of, then we can 1923 * return. 1924 */ 1925 if (found == nentqueued) 1926 return runningcode; 1927 1928 ki->kaio_flags |= KAIO_WAKEUP; 1929 error = tsleep(p, PCATCH, "aiospn", 0); 1930 1931 if (error == EINTR) 1932 return EINTR; 1933 else if (error == EWOULDBLOCK) 1934 return EAGAIN; 1935 } 1936 } 1937 1938 return runningcode; 1939 #endif /* VFS_AIO */ 1940 } 1941 1942 #ifdef VFS_AIO 1943 /* 1944 * This is a weird hack so that we can post a signal. It is safe to do so from 1945 * a timeout routine, but *not* from an interrupt routine. 1946 */ 1947 static void 1948 process_signal(void *aioj) 1949 { 1950 struct aiocblist *aiocbe = aioj; 1951 struct aio_liojob *lj = aiocbe->lio; 1952 struct aiocb *cb = &aiocbe->uaiocb; 1953 1954 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) && 1955 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) { 1956 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo); 1957 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 1958 } 1959 1960 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) 1961 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo); 1962 } 1963 1964 /* 1965 * Interrupt handler for physio, performs the necessary process wakeups, and 1966 * signals. 1967 */ 1968 static void 1969 aio_physwakeup(struct bio *bio) 1970 { 1971 struct buf *bp = bio->bio_buf; 1972 struct aiocblist *aiocbe; 1973 struct proc *p; 1974 struct kaioinfo *ki; 1975 struct aio_liojob *lj; 1976 1977 aiocbe = bio->bio_caller_info2.ptr; 1978 1979 if (aiocbe) { 1980 p = bio->bio_caller_info1.ptr; 1981 1982 aiocbe->jobstate = JOBST_JOBBFINISHED; 1983 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 1984 aiocbe->uaiocb._aiocb_private.error = 0; 1985 aiocbe->jobflags |= AIOCBLIST_DONE; 1986 1987 if (bp->b_flags & B_ERROR) 1988 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 1989 1990 lj = aiocbe->lio; 1991 if (lj) { 1992 lj->lioj_buffer_finished_count++; 1993 1994 /* 1995 * wakeup/signal if all of the interrupt jobs are done. 1996 */ 1997 if (lj->lioj_buffer_finished_count == 1998 lj->lioj_buffer_count) { 1999 /* 2000 * Post a signal if it is called for. 2001 */ 2002 if ((lj->lioj_flags & 2003 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == 2004 LIOJ_SIGNAL) { 2005 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2006 callout_reset(&aiocbe->timeout, 0, 2007 process_signal, aiocbe); 2008 } 2009 } 2010 } 2011 2012 ki = p->p_aioinfo; 2013 if (ki) { 2014 ki->kaio_buffer_finished_count++; 2015 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 2016 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 2017 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 2018 2019 KNOTE(&aiocbe->klist, 0); 2020 /* Do the wakeup. */ 2021 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) { 2022 ki->kaio_flags &= ~KAIO_WAKEUP; 2023 wakeup(p); 2024 } 2025 } 2026 2027 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 2028 callout_reset(&aiocbe->timeout, 0, 2029 process_signal, aiocbe); 2030 } 2031 } 2032 bp->b_cmd = BUF_CMD_DONE; 2033 wakeup(bp); 2034 } 2035 #endif /* VFS_AIO */ 2036 2037 /* syscall - wait for the next completion of an aio request */ 2038 int 2039 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap) 2040 { 2041 #ifndef VFS_AIO 2042 return ENOSYS; 2043 #else 2044 struct proc *p = curproc; 2045 struct lwp *lp = curthread->td_lwp; 2046 struct timeval atv; 2047 struct timespec ts; 2048 struct kaioinfo *ki; 2049 struct aiocblist *cb = NULL; 2050 int error, timo; 2051 2052 suword(uap->aiocbp, (int)NULL); 2053 2054 timo = 0; 2055 if (uap->timeout) { 2056 /* Get timespec struct. */ 2057 error = copyin(uap->timeout, &ts, sizeof(ts)); 2058 if (error) 2059 return error; 2060 2061 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000)) 2062 return (EINVAL); 2063 2064 TIMESPEC_TO_TIMEVAL(&atv, &ts); 2065 if (itimerfix(&atv)) 2066 return (EINVAL); 2067 timo = tvtohz_high(&atv); 2068 } 2069 2070 ki = p->p_aioinfo; 2071 if (ki == NULL) 2072 return EAGAIN; 2073 2074 for (;;) { 2075 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) { 2076 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2077 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2078 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2079 lp->lwp_ru.ru_oublock += 2080 cb->outputcharge; 2081 cb->outputcharge = 0; 2082 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2083 lp->lwp_ru.ru_inblock += cb->inputcharge; 2084 cb->inputcharge = 0; 2085 } 2086 aio_free_entry(cb); 2087 return cb->uaiocb._aiocb_private.error; 2088 } 2089 2090 crit_enter(); 2091 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) { 2092 crit_exit(); 2093 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2094 uap->sysmsg_result = cb->uaiocb._aiocb_private.status; 2095 aio_free_entry(cb); 2096 return cb->uaiocb._aiocb_private.error; 2097 } 2098 2099 ki->kaio_flags |= KAIO_WAKEUP; 2100 error = tsleep(p, PCATCH, "aiowc", timo); 2101 crit_exit(); 2102 2103 if (error == ERESTART) 2104 return EINTR; 2105 else if (error < 0) 2106 return error; 2107 else if (error == EINTR) 2108 return EINTR; 2109 else if (error == EWOULDBLOCK) 2110 return EAGAIN; 2111 } 2112 #endif /* VFS_AIO */ 2113 } 2114 2115 #ifndef VFS_AIO 2116 static int 2117 filt_aioattach(struct knote *kn) 2118 { 2119 2120 return (ENXIO); 2121 } 2122 2123 struct filterops aio_filtops = 2124 { 0, filt_aioattach, NULL, NULL }; 2125 2126 #else 2127 /* kqueue attach function */ 2128 static int 2129 filt_aioattach(struct knote *kn) 2130 { 2131 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2132 2133 /* 2134 * The aiocbe pointer must be validated before using it, so 2135 * registration is restricted to the kernel; the user cannot 2136 * set EV_FLAG1. 2137 */ 2138 if ((kn->kn_flags & EV_FLAG1) == 0) 2139 return (EPERM); 2140 kn->kn_flags &= ~EV_FLAG1; 2141 2142 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext); 2143 2144 return (0); 2145 } 2146 2147 /* kqueue detach function */ 2148 static void 2149 filt_aiodetach(struct knote *kn) 2150 { 2151 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2152 2153 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext); 2154 } 2155 2156 /* kqueue filter function */ 2157 /*ARGSUSED*/ 2158 static int 2159 filt_aio(struct knote *kn, long hint) 2160 { 2161 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2162 2163 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2164 if (aiocbe->jobstate != JOBST_JOBFINISHED && 2165 aiocbe->jobstate != JOBST_JOBBFINISHED) 2166 return (0); 2167 kn->kn_flags |= EV_EOF; 2168 return (1); 2169 } 2170 2171 struct filterops aio_filtops = 2172 { 0, filt_aioattach, filt_aiodetach, filt_aio }; 2173 #endif /* VFS_AIO */ 2174