vfs_aio.c revision 106998
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: head/sys/kern/vfs_aio.c 106998 2002-11-17 04:15:34Z alfred $ 17 */ 18 19/* 20 * This file contains support for the POSIX 1003.1B AIO/LIO facility. 21 */ 22 23#include <sys/param.h> 24#include <sys/systm.h> 25#include <sys/malloc.h> 26#include <sys/bio.h> 27#include <sys/buf.h> 28#include <sys/sysproto.h> 29#include <sys/filedesc.h> 30#include <sys/kernel.h> 31#include <sys/kthread.h> 32#include <sys/fcntl.h> 33#include <sys/file.h> 34#include <sys/lock.h> 35#include <sys/mutex.h> 36#include <sys/unistd.h> 37#include <sys/proc.h> 38#include <sys/resourcevar.h> 39#include <sys/signalvar.h> 40#include <sys/protosw.h> 41#include <sys/socketvar.h> 42#include <sys/syscall.h> 43#include <sys/sysent.h> 44#include <sys/sysctl.h> 45#include <sys/sx.h> 46#include <sys/vnode.h> 47#include <sys/conf.h> 48#include <sys/event.h> 49 50#include <posix4/posix4.h> 51#include <vm/vm.h> 52#include <vm/vm_extern.h> 53#include <vm/pmap.h> 54#include <vm/vm_map.h> 55#include <vm/uma.h> 56#include <sys/aio.h> 57 58#include <machine/limits.h> 59 60#include "opt_vfs_aio.h" 61 62/* 63 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 64 * overflow. 65 */ 66static long jobrefid; 67 68#define JOBST_NULL 0x0 69#define JOBST_JOBQGLOBAL 0x2 70#define JOBST_JOBRUNNING 0x3 71#define JOBST_JOBFINISHED 0x4 72#define JOBST_JOBQBUF 0x5 73#define JOBST_JOBBFINISHED 0x6 74 75#ifndef MAX_AIO_PER_PROC 76#define MAX_AIO_PER_PROC 32 77#endif 78 79#ifndef MAX_AIO_QUEUE_PER_PROC 80#define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */ 81#endif 82 83#ifndef MAX_AIO_PROCS 84#define MAX_AIO_PROCS 32 85#endif 86 87#ifndef MAX_AIO_QUEUE 88#define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */ 89#endif 90 91#ifndef TARGET_AIO_PROCS 92#define TARGET_AIO_PROCS 4 93#endif 94 95#ifndef MAX_BUF_AIO 96#define MAX_BUF_AIO 16 97#endif 98 99#ifndef AIOD_TIMEOUT_DEFAULT 100#define AIOD_TIMEOUT_DEFAULT (10 * hz) 101#endif 102 103#ifndef AIOD_LIFETIME_DEFAULT 104#define AIOD_LIFETIME_DEFAULT (30 * hz) 105#endif 106 107SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management"); 108 109static int max_aio_procs = MAX_AIO_PROCS; 110SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, 111 CTLFLAG_RW, &max_aio_procs, 0, 112 "Maximum number of kernel threads to use for handling async IO "); 113 114static int num_aio_procs = 0; 115SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, 116 CTLFLAG_RD, &num_aio_procs, 0, 117 "Number of presently active kernel threads for async IO"); 118 119/* 120 * The code will adjust the actual number of AIO processes towards this 121 * number when it gets a chance. 122 */ 123static int target_aio_procs = TARGET_AIO_PROCS; 124SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 125 0, "Preferred number of ready kernel threads for async IO"); 126 127static int max_queue_count = MAX_AIO_QUEUE; 128SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 129 "Maximum number of aio requests to queue, globally"); 130 131static int num_queue_count = 0; 132SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 133 "Number of queued aio requests"); 134 135static int num_buf_aio = 0; 136SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 137 "Number of aio requests presently handled by the buf subsystem"); 138 139/* Number of async I/O thread in the process of being started */ 140/* XXX This should be local to _aio_aqueue() */ 141static int num_aio_resv_start = 0; 142 143static int aiod_timeout; 144SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0, 145 "Timeout value for synchronous aio operations"); 146 147static int aiod_lifetime; 148SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 149 "Maximum lifetime for idle aiod"); 150 151static int unloadable = 0; 152SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0, 153 "Allow unload of aio (not recommended)"); 154 155 156static int max_aio_per_proc = MAX_AIO_PER_PROC; 157SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 158 0, "Maximum active aio requests per process (stored in the process)"); 159 160static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 161SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 162 &max_aio_queue_per_proc, 0, 163 "Maximum queued aio requests per process (stored in the process)"); 164 165static int max_buf_aio = MAX_BUF_AIO; 166SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 167 "Maximum buf aio requests per process (stored in the process)"); 168 169struct aiocblist { 170 TAILQ_ENTRY(aiocblist) list; /* List of jobs */ 171 TAILQ_ENTRY(aiocblist) plist; /* List of jobs for proc */ 172 int jobflags; 173 int jobstate; 174 int inputcharge; 175 int outputcharge; 176 struct callout_handle timeouthandle; 177 struct buf *bp; /* Buffer pointer */ 178 struct proc *userproc; /* User process */ /* Not td! */ 179 struct ucred *cred; /* Active credential when created */ 180 struct file *fd_file; /* Pointer to file structure */ 181 struct aio_liojob *lio; /* Optional lio job */ 182 struct aiocb *uuaiocb; /* Pointer in userspace of aiocb */ 183 struct klist klist; /* list of knotes */ 184 struct aiocb uaiocb; /* Kernel I/O control block */ 185}; 186 187/* jobflags */ 188#define AIOCBLIST_RUNDOWN 0x4 189#define AIOCBLIST_DONE 0x10 190 191/* 192 * AIO process info 193 */ 194#define AIOP_FREE 0x1 /* proc on free queue */ 195#define AIOP_SCHED 0x2 /* proc explicitly scheduled */ 196 197struct aiothreadlist { 198 int aiothreadflags; /* AIO proc flags */ 199 TAILQ_ENTRY(aiothreadlist) list; /* List of processes */ 200 struct thread *aiothread; /* The AIO thread */ 201}; 202 203/* 204 * data-structure for lio signal management 205 */ 206struct aio_liojob { 207 int lioj_flags; 208 int lioj_buffer_count; 209 int lioj_buffer_finished_count; 210 int lioj_queue_count; 211 int lioj_queue_finished_count; 212 struct sigevent lioj_signal; /* signal on all I/O done */ 213 TAILQ_ENTRY(aio_liojob) lioj_list; 214 struct kaioinfo *lioj_ki; 215}; 216#define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 217#define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 218 219/* 220 * per process aio data structure 221 */ 222struct kaioinfo { 223 int kaio_flags; /* per process kaio flags */ 224 int kaio_maxactive_count; /* maximum number of AIOs */ 225 int kaio_active_count; /* number of currently used AIOs */ 226 int kaio_qallowed_count; /* maxiumu size of AIO queue */ 227 int kaio_queue_count; /* size of AIO queue */ 228 int kaio_ballowed_count; /* maximum number of buffers */ 229 int kaio_queue_finished_count; /* number of daemon jobs finished */ 230 int kaio_buffer_count; /* number of physio buffers */ 231 int kaio_buffer_finished_count; /* count of I/O done */ 232 struct proc *kaio_p; /* process that uses this kaio block */ 233 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */ 234 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */ 235 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */ 236 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */ 237 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */ 238 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */ 239}; 240 241#define KAIO_RUNDOWN 0x1 /* process is being run down */ 242#define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */ 243 244static TAILQ_HEAD(,aiothreadlist) aio_activeproc; /* Active daemons */ 245static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* Idle daemons */ 246static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */ 247static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */ 248 249static void aio_init_aioinfo(struct proc *p); 250static void aio_onceonly(void); 251static int aio_free_entry(struct aiocblist *aiocbe); 252static void aio_process(struct aiocblist *aiocbe); 253static int aio_newproc(void); 254static int aio_aqueue(struct thread *td, struct aiocb *job, int type); 255static void aio_physwakeup(struct buf *bp); 256static void aio_proc_rundown(struct proc *p); 257static int aio_fphysio(struct aiocblist *aiocbe); 258static int aio_qphysio(struct proc *p, struct aiocblist *iocb); 259static void aio_daemon(void *uproc); 260static void aio_swake_cb(struct socket *, struct sockbuf *); 261static int aio_unload(void); 262static void process_signal(void *aioj); 263static int filt_aioattach(struct knote *kn); 264static void filt_aiodetach(struct knote *kn); 265static int filt_aio(struct knote *kn, long hint); 266 267/* 268 * Zones for: 269 * kaio Per process async io info 270 * aiop async io thread data 271 * aiocb async io jobs 272 * aiol list io job pointer - internal to aio_suspend XXX 273 * aiolio list io jobs 274 */ 275static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone; 276 277/* kqueue filters for aio */ 278static struct filterops aio_filtops = 279 { 0, filt_aioattach, filt_aiodetach, filt_aio }; 280 281/* 282 * Main operations function for use as a kernel module. 283 */ 284static int 285aio_modload(struct module *module, int cmd, void *arg) 286{ 287 int error = 0; 288 289 switch (cmd) { 290 case MOD_LOAD: 291 aio_onceonly(); 292 break; 293 case MOD_UNLOAD: 294 error = aio_unload(); 295 break; 296 case MOD_SHUTDOWN: 297 break; 298 default: 299 error = EINVAL; 300 break; 301 } 302 return (error); 303} 304 305static moduledata_t aio_mod = { 306 "aio", 307 &aio_modload, 308 NULL 309}; 310 311SYSCALL_MODULE_HELPER(aio_return); 312SYSCALL_MODULE_HELPER(aio_suspend); 313SYSCALL_MODULE_HELPER(aio_cancel); 314SYSCALL_MODULE_HELPER(aio_error); 315SYSCALL_MODULE_HELPER(aio_read); 316SYSCALL_MODULE_HELPER(aio_write); 317SYSCALL_MODULE_HELPER(aio_waitcomplete); 318SYSCALL_MODULE_HELPER(lio_listio); 319 320DECLARE_MODULE(aio, aio_mod, 321 SI_SUB_VFS, SI_ORDER_ANY); 322MODULE_VERSION(aio, 1); 323 324/* 325 * Startup initialization 326 */ 327static void 328aio_onceonly(void) 329{ 330 331 /* XXX: should probably just use so->callback */ 332 aio_swake = &aio_swake_cb; 333 at_exit(aio_proc_rundown); 334 at_exec(aio_proc_rundown); 335 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 336 TAILQ_INIT(&aio_freeproc); 337 TAILQ_INIT(&aio_activeproc); 338 TAILQ_INIT(&aio_jobs); 339 TAILQ_INIT(&aio_bufjobs); 340 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 341 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 342 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL, 343 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 344 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL, 345 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 346 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL, 347 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 348 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL, 349 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 350 aiod_timeout = AIOD_TIMEOUT_DEFAULT; 351 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 352 jobrefid = 1; 353 async_io_version = _POSIX_VERSION; 354 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX); 355 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 356 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 357} 358 359/* 360 * Callback for unload of AIO when used as a module. 361 */ 362static int 363aio_unload(void) 364{ 365 366 /* 367 * XXX: no unloads by default, it's too dangerous. 368 * perhaps we could do it if locked out callers and then 369 * did an aio_proc_rundown() on each process. 370 */ 371 if (!unloadable) 372 return (EOPNOTSUPP); 373 374 async_io_version = 0; 375 aio_swake = NULL; 376 rm_at_exit(aio_proc_rundown); 377 rm_at_exec(aio_proc_rundown); 378 kqueue_del_filteropts(EVFILT_AIO); 379 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1); 380 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1); 381 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1); 382 return (0); 383} 384 385/* 386 * Init the per-process aioinfo structure. The aioinfo limits are set 387 * per-process for user limit (resource) management. 388 */ 389static void 390aio_init_aioinfo(struct proc *p) 391{ 392 struct kaioinfo *ki; 393 if (p->p_aioinfo == NULL) { 394 ki = uma_zalloc(kaio_zone, M_WAITOK); 395 p->p_aioinfo = ki; 396 ki->kaio_flags = 0; 397 ki->kaio_maxactive_count = max_aio_per_proc; 398 ki->kaio_active_count = 0; 399 ki->kaio_qallowed_count = max_aio_queue_per_proc; 400 ki->kaio_queue_count = 0; 401 ki->kaio_ballowed_count = max_buf_aio; 402 ki->kaio_buffer_count = 0; 403 ki->kaio_buffer_finished_count = 0; 404 ki->kaio_p = p; 405 TAILQ_INIT(&ki->kaio_jobdone); 406 TAILQ_INIT(&ki->kaio_jobqueue); 407 TAILQ_INIT(&ki->kaio_bufdone); 408 TAILQ_INIT(&ki->kaio_bufqueue); 409 TAILQ_INIT(&ki->kaio_liojoblist); 410 TAILQ_INIT(&ki->kaio_sockqueue); 411 } 412 413 while (num_aio_procs < target_aio_procs) 414 aio_newproc(); 415} 416 417/* 418 * Free a job entry. Wait for completion if it is currently active, but don't 419 * delay forever. If we delay, we return a flag that says that we have to 420 * restart the queue scan. 421 */ 422static int 423aio_free_entry(struct aiocblist *aiocbe) 424{ 425 struct kaioinfo *ki; 426 struct aio_liojob *lj; 427 struct proc *p; 428 int error; 429 int s; 430 431 if (aiocbe->jobstate == JOBST_NULL) 432 panic("aio_free_entry: freeing already free job"); 433 434 p = aiocbe->userproc; 435 ki = p->p_aioinfo; 436 lj = aiocbe->lio; 437 if (ki == NULL) 438 panic("aio_free_entry: missing p->p_aioinfo"); 439 440 while (aiocbe->jobstate == JOBST_JOBRUNNING) { 441 aiocbe->jobflags |= AIOCBLIST_RUNDOWN; 442 tsleep(aiocbe, PRIBIO, "jobwai", 0); 443 } 444 if (aiocbe->bp == NULL) { 445 if (ki->kaio_queue_count <= 0) 446 panic("aio_free_entry: process queue size <= 0"); 447 if (num_queue_count <= 0) 448 panic("aio_free_entry: system wide queue size <= 0"); 449 450 if (lj) { 451 lj->lioj_queue_count--; 452 if (aiocbe->jobflags & AIOCBLIST_DONE) 453 lj->lioj_queue_finished_count--; 454 } 455 ki->kaio_queue_count--; 456 if (aiocbe->jobflags & AIOCBLIST_DONE) 457 ki->kaio_queue_finished_count--; 458 num_queue_count--; 459 } else { 460 if (lj) { 461 lj->lioj_buffer_count--; 462 if (aiocbe->jobflags & AIOCBLIST_DONE) 463 lj->lioj_buffer_finished_count--; 464 } 465 if (aiocbe->jobflags & AIOCBLIST_DONE) 466 ki->kaio_buffer_finished_count--; 467 ki->kaio_buffer_count--; 468 num_buf_aio--; 469 } 470 471 /* aiocbe is going away, we need to destroy any knotes */ 472 /* XXXKSE Note the thread here is used to eventually find the 473 * owning process again, but it is also used to do a fo_close 474 * and that requires the thread. (but does it require the 475 * OWNING thread? (or maybe the running thread?) 476 * There is a semantic problem here... 477 */ 478 knote_remove(FIRST_THREAD_IN_PROC(p), &aiocbe->klist); /* XXXKSE */ 479 480 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN) 481 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) { 482 ki->kaio_flags &= ~KAIO_WAKEUP; 483 wakeup(p); 484 } 485 486 if (aiocbe->jobstate == JOBST_JOBQBUF) { 487 if ((error = aio_fphysio(aiocbe)) != 0) 488 return error; 489 if (aiocbe->jobstate != JOBST_JOBBFINISHED) 490 panic("aio_free_entry: invalid physio finish-up state"); 491 s = splbio(); 492 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist); 493 splx(s); 494 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) { 495 s = splnet(); 496 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 497 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 498 splx(s); 499 } else if (aiocbe->jobstate == JOBST_JOBFINISHED) 500 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist); 501 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) { 502 s = splbio(); 503 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist); 504 splx(s); 505 if (aiocbe->bp) { 506 vunmapbuf(aiocbe->bp); 507 relpbuf(aiocbe->bp, NULL); 508 aiocbe->bp = NULL; 509 } 510 } 511 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) { 512 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 513 uma_zfree(aiolio_zone, lj); 514 } 515 aiocbe->jobstate = JOBST_NULL; 516 untimeout(process_signal, aiocbe, aiocbe->timeouthandle); 517 fdrop(aiocbe->fd_file, curthread); 518 crfree(aiocbe->cred); 519 uma_zfree(aiocb_zone, aiocbe); 520 return 0; 521} 522 523/* 524 * Rundown the jobs for a given process. 525 */ 526static void 527aio_proc_rundown(struct proc *p) 528{ 529 int s; 530 struct kaioinfo *ki; 531 struct aio_liojob *lj, *ljn; 532 struct aiocblist *aiocbe, *aiocbn; 533 struct file *fp; 534 struct socket *so; 535 536 ki = p->p_aioinfo; 537 if (ki == NULL) 538 return; 539 540 ki->kaio_flags |= LIOJ_SIGNAL_POSTED; 541 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count > 542 ki->kaio_buffer_finished_count)) { 543 ki->kaio_flags |= KAIO_RUNDOWN; 544 if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout)) 545 break; 546 } 547 548 /* 549 * Move any aio ops that are waiting on socket I/O to the normal job 550 * queues so they are cleaned up with any others. 551 */ 552 s = splnet(); 553 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe = 554 aiocbn) { 555 aiocbn = TAILQ_NEXT(aiocbe, plist); 556 fp = aiocbe->fd_file; 557 if (fp != NULL) { 558 so = (struct socket *)fp->f_data; 559 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list); 560 if (TAILQ_EMPTY(&so->so_aiojobq)) { 561 so->so_snd.sb_flags &= ~SB_AIO; 562 so->so_rcv.sb_flags &= ~SB_AIO; 563 } 564 } 565 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist); 566 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list); 567 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist); 568 } 569 splx(s); 570 571restart1: 572 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) { 573 aiocbn = TAILQ_NEXT(aiocbe, plist); 574 if (aio_free_entry(aiocbe)) 575 goto restart1; 576 } 577 578restart2: 579 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe = 580 aiocbn) { 581 aiocbn = TAILQ_NEXT(aiocbe, plist); 582 if (aio_free_entry(aiocbe)) 583 goto restart2; 584 } 585 586/* 587 * Note the use of lots of splbio here, trying to avoid splbio for long chains 588 * of I/O. Probably unnecessary. 589 */ 590restart3: 591 s = splbio(); 592 while (TAILQ_FIRST(&ki->kaio_bufqueue)) { 593 ki->kaio_flags |= KAIO_WAKEUP; 594 tsleep(p, PRIBIO, "aioprn", 0); 595 splx(s); 596 goto restart3; 597 } 598 splx(s); 599 600restart4: 601 s = splbio(); 602 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) { 603 aiocbn = TAILQ_NEXT(aiocbe, plist); 604 if (aio_free_entry(aiocbe)) { 605 splx(s); 606 goto restart4; 607 } 608 } 609 splx(s); 610 611 /* 612 * If we've slept, jobs might have moved from one queue to another. 613 * Retry rundown if we didn't manage to empty the queues. 614 */ 615 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL || 616 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL || 617 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL || 618 TAILQ_FIRST(&ki->kaio_bufdone) != NULL) 619 goto restart1; 620 621 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) { 622 ljn = TAILQ_NEXT(lj, lioj_list); 623 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 624 0)) { 625 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 626 uma_zfree(aiolio_zone, lj); 627 } else { 628#ifdef DIAGNOSTIC 629 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, " 630 "QF:%d\n", lj->lioj_buffer_count, 631 lj->lioj_buffer_finished_count, 632 lj->lioj_queue_count, 633 lj->lioj_queue_finished_count); 634#endif 635 } 636 } 637 638 uma_zfree(kaio_zone, ki); 639 p->p_aioinfo = NULL; 640} 641 642/* 643 * Select a job to run (called by an AIO daemon). 644 */ 645static struct aiocblist * 646aio_selectjob(struct aiothreadlist *aiop) 647{ 648 int s; 649 struct aiocblist *aiocbe; 650 struct kaioinfo *ki; 651 struct proc *userp; 652 653 s = splnet(); 654 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe = 655 TAILQ_NEXT(aiocbe, list)) { 656 userp = aiocbe->userproc; 657 ki = userp->p_aioinfo; 658 659 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 660 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 661 splx(s); 662 return aiocbe; 663 } 664 } 665 splx(s); 666 667 return NULL; 668} 669 670/* 671 * The AIO processing activity. This is the code that does the I/O request for 672 * the non-physio version of the operations. The normal vn operations are used, 673 * and this code should work in all instances for every type of file, including 674 * pipes, sockets, fifos, and regular files. 675 */ 676static void 677aio_process(struct aiocblist *aiocbe) 678{ 679 struct ucred *td_savedcred; 680 struct thread *td; 681 struct proc *mycp; 682 struct aiocb *cb; 683 struct file *fp; 684 struct uio auio; 685 struct iovec aiov; 686 int cnt; 687 int error; 688 int oublock_st, oublock_end; 689 int inblock_st, inblock_end; 690 691 td = curthread; 692 td_savedcred = td->td_ucred; 693 td->td_ucred = aiocbe->cred; 694 mycp = td->td_proc; 695 cb = &aiocbe->uaiocb; 696 fp = aiocbe->fd_file; 697 698 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 699 aiov.iov_len = cb->aio_nbytes; 700 701 auio.uio_iov = &aiov; 702 auio.uio_iovcnt = 1; 703 auio.uio_offset = cb->aio_offset; 704 auio.uio_resid = cb->aio_nbytes; 705 cnt = cb->aio_nbytes; 706 auio.uio_segflg = UIO_USERSPACE; 707 auio.uio_td = td; 708 709 inblock_st = mycp->p_stats->p_ru.ru_inblock; 710 oublock_st = mycp->p_stats->p_ru.ru_oublock; 711 /* 712 * _aio_aqueue() acquires a reference to the file that is 713 * released in aio_free_entry(). 714 */ 715 if (cb->aio_lio_opcode == LIO_READ) { 716 auio.uio_rw = UIO_READ; 717 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td); 718 } else { 719 auio.uio_rw = UIO_WRITE; 720 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td); 721 } 722 inblock_end = mycp->p_stats->p_ru.ru_inblock; 723 oublock_end = mycp->p_stats->p_ru.ru_oublock; 724 725 aiocbe->inputcharge = inblock_end - inblock_st; 726 aiocbe->outputcharge = oublock_end - oublock_st; 727 728 if ((error) && (auio.uio_resid != cnt)) { 729 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 730 error = 0; 731 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) { 732 PROC_LOCK(aiocbe->userproc); 733 psignal(aiocbe->userproc, SIGPIPE); 734 PROC_UNLOCK(aiocbe->userproc); 735 } 736 } 737 738 cnt -= auio.uio_resid; 739 cb->_aiocb_private.error = error; 740 cb->_aiocb_private.status = cnt; 741 td->td_ucred = td_savedcred; 742} 743 744/* 745 * The AIO daemon, most of the actual work is done in aio_process, 746 * but the setup (and address space mgmt) is done in this routine. 747 */ 748static void 749aio_daemon(void *uproc) 750{ 751 int s; 752 struct aio_liojob *lj; 753 struct aiocb *cb; 754 struct aiocblist *aiocbe; 755 struct aiothreadlist *aiop; 756 struct kaioinfo *ki; 757 struct proc *curcp, *mycp, *userp; 758 struct vmspace *myvm, *tmpvm; 759 struct thread *td = curthread; 760 struct pgrp *newpgrp; 761 struct session *newsess; 762 763 mtx_lock(&Giant); 764 /* 765 * Local copies of curproc (cp) and vmspace (myvm) 766 */ 767 mycp = td->td_proc; 768 myvm = mycp->p_vmspace; 769 770 if (mycp->p_textvp) { 771 vrele(mycp->p_textvp); 772 mycp->p_textvp = NULL; 773 } 774 775 /* 776 * Allocate and ready the aio control info. There is one aiop structure 777 * per daemon. 778 */ 779 aiop = uma_zalloc(aiop_zone, M_WAITOK); 780 aiop->aiothread = td; 781 aiop->aiothreadflags |= AIOP_FREE; 782 783 s = splnet(); 784 785 /* 786 * Place thread (lightweight process) onto the AIO free thread list. 787 */ 788 if (TAILQ_EMPTY(&aio_freeproc)) 789 wakeup(&aio_freeproc); 790 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 791 792 splx(s); 793 794 /* 795 * Get rid of our current filedescriptors. AIOD's don't need any 796 * filedescriptors, except as temporarily inherited from the client. 797 */ 798 fdfree(td); 799 800 mtx_unlock(&Giant); 801 /* The daemon resides in its own pgrp. */ 802 MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP, 803 M_WAITOK | M_ZERO); 804 MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION, 805 M_WAITOK | M_ZERO); 806 807 sx_xlock(&proctree_lock); 808 enterpgrp(mycp, mycp->p_pid, newpgrp, newsess); 809 sx_xunlock(&proctree_lock); 810 mtx_lock(&Giant); 811 812 /* Mark special process type. */ 813 mycp->p_flag |= P_SYSTEM; 814 815 /* 816 * Wakeup parent process. (Parent sleeps to keep from blasting away 817 * and creating too many daemons.) 818 */ 819 wakeup(mycp); 820 821 for (;;) { 822 /* 823 * curcp is the current daemon process context. 824 * userp is the current user process context. 825 */ 826 curcp = mycp; 827 828 /* 829 * Take daemon off of free queue 830 */ 831 if (aiop->aiothreadflags & AIOP_FREE) { 832 s = splnet(); 833 TAILQ_REMOVE(&aio_freeproc, aiop, list); 834 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 835 aiop->aiothreadflags &= ~AIOP_FREE; 836 splx(s); 837 } 838 aiop->aiothreadflags &= ~AIOP_SCHED; 839 840 /* 841 * Check for jobs. 842 */ 843 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 844 cb = &aiocbe->uaiocb; 845 userp = aiocbe->userproc; 846 847 aiocbe->jobstate = JOBST_JOBRUNNING; 848 849 /* 850 * Connect to process address space for user program. 851 */ 852 if (userp != curcp) { 853 /* 854 * Save the current address space that we are 855 * connected to. 856 */ 857 tmpvm = mycp->p_vmspace; 858 859 /* 860 * Point to the new user address space, and 861 * refer to it. 862 */ 863 mycp->p_vmspace = userp->p_vmspace; 864 mycp->p_vmspace->vm_refcnt++; 865 866 /* Activate the new mapping. */ 867 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 868 869 /* 870 * If the old address space wasn't the daemons 871 * own address space, then we need to remove the 872 * daemon's reference from the other process 873 * that it was acting on behalf of. 874 */ 875 if (tmpvm != myvm) { 876 vmspace_free(tmpvm); 877 } 878 curcp = userp; 879 } 880 881 ki = userp->p_aioinfo; 882 lj = aiocbe->lio; 883 884 /* Account for currently active jobs. */ 885 ki->kaio_active_count++; 886 887 /* Do the I/O function. */ 888 aio_process(aiocbe); 889 890 /* Decrement the active job count. */ 891 ki->kaio_active_count--; 892 893 /* 894 * Increment the completion count for wakeup/signal 895 * comparisons. 896 */ 897 aiocbe->jobflags |= AIOCBLIST_DONE; 898 ki->kaio_queue_finished_count++; 899 if (lj) 900 lj->lioj_queue_finished_count++; 901 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags 902 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) { 903 ki->kaio_flags &= ~KAIO_WAKEUP; 904 wakeup(userp); 905 } 906 907 s = splbio(); 908 if (lj && (lj->lioj_flags & 909 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) { 910 if ((lj->lioj_queue_finished_count == 911 lj->lioj_queue_count) && 912 (lj->lioj_buffer_finished_count == 913 lj->lioj_buffer_count)) { 914 PROC_LOCK(userp); 915 psignal(userp, 916 lj->lioj_signal.sigev_signo); 917 PROC_UNLOCK(userp); 918 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 919 } 920 } 921 splx(s); 922 923 aiocbe->jobstate = JOBST_JOBFINISHED; 924 925 s = splnet(); 926 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 927 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist); 928 splx(s); 929 KNOTE(&aiocbe->klist, 0); 930 931 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) { 932 wakeup(aiocbe); 933 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN; 934 } 935 936 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 937 PROC_LOCK(userp); 938 psignal(userp, cb->aio_sigevent.sigev_signo); 939 PROC_UNLOCK(userp); 940 } 941 } 942 943 /* 944 * Disconnect from user address space. 945 */ 946 if (curcp != mycp) { 947 /* Get the user address space to disconnect from. */ 948 tmpvm = mycp->p_vmspace; 949 950 /* Get original address space for daemon. */ 951 mycp->p_vmspace = myvm; 952 953 /* Activate the daemon's address space. */ 954 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 955#ifdef DIAGNOSTIC 956 if (tmpvm == myvm) { 957 printf("AIOD: vmspace problem -- %d\n", 958 mycp->p_pid); 959 } 960#endif 961 /* Remove our vmspace reference. */ 962 vmspace_free(tmpvm); 963 964 curcp = mycp; 965 } 966 967 /* 968 * If we are the first to be put onto the free queue, wakeup 969 * anyone waiting for a daemon. 970 */ 971 s = splnet(); 972 TAILQ_REMOVE(&aio_activeproc, aiop, list); 973 if (TAILQ_EMPTY(&aio_freeproc)) 974 wakeup(&aio_freeproc); 975 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 976 aiop->aiothreadflags |= AIOP_FREE; 977 splx(s); 978 979 /* 980 * If daemon is inactive for a long time, allow it to exit, 981 * thereby freeing resources. 982 */ 983 if ((aiop->aiothreadflags & AIOP_SCHED) == 0 && 984 tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) { 985 s = splnet(); 986 if (TAILQ_EMPTY(&aio_jobs)) { 987 if ((aiop->aiothreadflags & AIOP_FREE) && 988 (num_aio_procs > target_aio_procs)) { 989 TAILQ_REMOVE(&aio_freeproc, aiop, list); 990 splx(s); 991 uma_zfree(aiop_zone, aiop); 992 num_aio_procs--; 993#ifdef DIAGNOSTIC 994 if (mycp->p_vmspace->vm_refcnt <= 1) { 995 printf("AIOD: bad vm refcnt for" 996 " exiting daemon: %d\n", 997 mycp->p_vmspace->vm_refcnt); 998 } 999#endif 1000 kthread_exit(0); 1001 } 1002 } 1003 splx(s); 1004 } 1005 } 1006} 1007 1008/* 1009 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1010 * AIO daemon modifies its environment itself. 1011 */ 1012static int 1013aio_newproc() 1014{ 1015 int error; 1016 struct proc *p; 1017 1018 error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, 0, "aiod%d", 1019 num_aio_procs); 1020 if (error) 1021 return error; 1022 1023 /* 1024 * Wait until daemon is started, but continue on just in case to 1025 * handle error conditions. 1026 */ 1027 error = tsleep(p, PZERO, "aiosta", aiod_timeout); 1028 1029 num_aio_procs++; 1030 1031 return error; 1032} 1033 1034/* 1035 * Try the high-performance, low-overhead physio method for eligible 1036 * VCHR devices. This method doesn't use an aio helper thread, and 1037 * thus has very low overhead. 1038 * 1039 * Assumes that the caller, _aio_aqueue(), has incremented the file 1040 * structure's reference count, preventing its deallocation for the 1041 * duration of this call. 1042 */ 1043static int 1044aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 1045{ 1046 int error; 1047 struct aiocb *cb; 1048 struct file *fp; 1049 struct buf *bp; 1050 struct vnode *vp; 1051 struct kaioinfo *ki; 1052 struct aio_liojob *lj; 1053 int s; 1054 int notify; 1055 1056 cb = &aiocbe->uaiocb; 1057 fp = aiocbe->fd_file; 1058 1059 if (fp->f_type != DTYPE_VNODE) 1060 return (-1); 1061 1062 vp = (struct vnode *)fp->f_data; 1063 1064 /* 1065 * If its not a disk, we don't want to return a positive error. 1066 * It causes the aio code to not fall through to try the thread 1067 * way when you're talking to a regular file. 1068 */ 1069 if (!vn_isdisk(vp, &error)) { 1070 if (error == ENOTBLK) 1071 return (-1); 1072 else 1073 return (error); 1074 } 1075 1076 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys) 1077 return (-1); 1078 1079 if (cb->aio_nbytes > 1080 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 1081 return (-1); 1082 1083 ki = p->p_aioinfo; 1084 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 1085 return (-1); 1086 1087 ki->kaio_buffer_count++; 1088 1089 lj = aiocbe->lio; 1090 if (lj) 1091 lj->lioj_buffer_count++; 1092 1093 /* Create and build a buffer header for a transfer. */ 1094 bp = (struct buf *)getpbuf(NULL); 1095 BUF_KERNPROC(bp); 1096 1097 /* 1098 * Get a copy of the kva from the physical buffer. 1099 */ 1100 bp->b_caller1 = p; 1101 bp->b_dev = vp->v_rdev; 1102 error = bp->b_error = 0; 1103 1104 bp->b_bcount = cb->aio_nbytes; 1105 bp->b_bufsize = cb->aio_nbytes; 1106 bp->b_flags = B_PHYS; 1107 bp->b_iodone = aio_physwakeup; 1108 bp->b_saveaddr = bp->b_data; 1109 bp->b_data = (void *)(uintptr_t)cb->aio_buf; 1110 bp->b_blkno = btodb(cb->aio_offset); 1111 1112 if (cb->aio_lio_opcode == LIO_WRITE) { 1113 bp->b_iocmd = BIO_WRITE; 1114 if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_READ)) { 1115 error = EFAULT; 1116 goto doerror; 1117 } 1118 } else { 1119 bp->b_iocmd = BIO_READ; 1120 if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_WRITE)) { 1121 error = EFAULT; 1122 goto doerror; 1123 } 1124 } 1125 1126 /* Bring buffer into kernel space. */ 1127 vmapbuf(bp); 1128 1129 s = splbio(); 1130 aiocbe->bp = bp; 1131 bp->b_spc = (void *)aiocbe; 1132 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list); 1133 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 1134 aiocbe->jobstate = JOBST_JOBQBUF; 1135 cb->_aiocb_private.status = cb->aio_nbytes; 1136 num_buf_aio++; 1137 bp->b_error = 0; 1138 1139 splx(s); 1140 1141 /* Perform transfer. */ 1142 DEV_STRATEGY(bp, 0); 1143 1144 notify = 0; 1145 s = splbio(); 1146 1147 /* 1148 * If we had an error invoking the request, or an error in processing 1149 * the request before we have returned, we process it as an error in 1150 * transfer. Note that such an I/O error is not indicated immediately, 1151 * but is returned using the aio_error mechanism. In this case, 1152 * aio_suspend will return immediately. 1153 */ 1154 if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) { 1155 struct aiocb *job = aiocbe->uuaiocb; 1156 1157 aiocbe->uaiocb._aiocb_private.status = 0; 1158 suword(&job->_aiocb_private.status, 0); 1159 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 1160 suword(&job->_aiocb_private.error, bp->b_error); 1161 1162 ki->kaio_buffer_finished_count++; 1163 1164 if (aiocbe->jobstate != JOBST_JOBBFINISHED) { 1165 aiocbe->jobstate = JOBST_JOBBFINISHED; 1166 aiocbe->jobflags |= AIOCBLIST_DONE; 1167 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 1168 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 1169 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 1170 notify = 1; 1171 } 1172 } 1173 splx(s); 1174 if (notify) 1175 KNOTE(&aiocbe->klist, 0); 1176 return 0; 1177 1178doerror: 1179 ki->kaio_buffer_count--; 1180 if (lj) 1181 lj->lioj_buffer_count--; 1182 aiocbe->bp = NULL; 1183 relpbuf(bp, NULL); 1184 return error; 1185} 1186 1187/* 1188 * This waits/tests physio completion. 1189 */ 1190static int 1191aio_fphysio(struct aiocblist *iocb) 1192{ 1193 int s; 1194 struct buf *bp; 1195 int error; 1196 1197 bp = iocb->bp; 1198 1199 s = splbio(); 1200 while ((bp->b_flags & B_DONE) == 0) { 1201 if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) { 1202 if ((bp->b_flags & B_DONE) == 0) { 1203 splx(s); 1204 return EINPROGRESS; 1205 } else 1206 break; 1207 } 1208 } 1209 splx(s); 1210 1211 /* Release mapping into kernel space. */ 1212 vunmapbuf(bp); 1213 iocb->bp = 0; 1214 1215 error = 0; 1216 1217 /* Check for an error. */ 1218 if (bp->b_ioflags & BIO_ERROR) 1219 error = bp->b_error; 1220 1221 relpbuf(bp, NULL); 1222 return (error); 1223} 1224 1225/* 1226 * Wake up aio requests that may be serviceable now. 1227 */ 1228static void 1229aio_swake_cb(struct socket *so, struct sockbuf *sb) 1230{ 1231 struct aiocblist *cb,*cbn; 1232 struct proc *p; 1233 struct kaioinfo *ki = NULL; 1234 int opcode, wakecount = 0; 1235 struct aiothreadlist *aiop; 1236 1237 if (sb == &so->so_snd) { 1238 opcode = LIO_WRITE; 1239 so->so_snd.sb_flags &= ~SB_AIO; 1240 } else { 1241 opcode = LIO_READ; 1242 so->so_rcv.sb_flags &= ~SB_AIO; 1243 } 1244 1245 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) { 1246 cbn = TAILQ_NEXT(cb, list); 1247 if (opcode == cb->uaiocb.aio_lio_opcode) { 1248 p = cb->userproc; 1249 ki = p->p_aioinfo; 1250 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1251 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist); 1252 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1253 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist); 1254 wakecount++; 1255 if (cb->jobstate != JOBST_JOBQGLOBAL) 1256 panic("invalid queue value"); 1257 } 1258 } 1259 1260 while (wakecount--) { 1261 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) { 1262 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1263 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1264 aiop->aiothreadflags &= ~AIOP_FREE; 1265 wakeup(aiop->aiothread); 1266 } 1267 } 1268} 1269 1270/* 1271 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1272 * technique is done in this code. 1273 */ 1274static int 1275_aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type) 1276{ 1277 struct proc *p = td->td_proc; 1278 struct filedesc *fdp; 1279 struct file *fp; 1280 unsigned int fd; 1281 struct socket *so; 1282 int s; 1283 int error; 1284 int opcode, user_opcode; 1285 struct aiocblist *aiocbe; 1286 struct aiothreadlist *aiop; 1287 struct kaioinfo *ki; 1288 struct kevent kev; 1289 struct kqueue *kq; 1290 struct file *kq_fp; 1291 1292 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK); 1293 aiocbe->inputcharge = 0; 1294 aiocbe->outputcharge = 0; 1295 callout_handle_init(&aiocbe->timeouthandle); 1296 SLIST_INIT(&aiocbe->klist); 1297 1298 suword(&job->_aiocb_private.status, -1); 1299 suword(&job->_aiocb_private.error, 0); 1300 suword(&job->_aiocb_private.kernelinfo, -1); 1301 1302 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb)); 1303 if (error) { 1304 suword(&job->_aiocb_private.error, error); 1305 uma_zfree(aiocb_zone, aiocbe); 1306 return error; 1307 } 1308 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL && 1309 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1310 uma_zfree(aiocb_zone, aiocbe); 1311 return EINVAL; 1312 } 1313 1314 /* Save userspace address of the job info. */ 1315 aiocbe->uuaiocb = job; 1316 1317 /* Get the opcode. */ 1318 user_opcode = aiocbe->uaiocb.aio_lio_opcode; 1319 if (type != LIO_NOP) 1320 aiocbe->uaiocb.aio_lio_opcode = type; 1321 opcode = aiocbe->uaiocb.aio_lio_opcode; 1322 1323 /* Get the fd info for process. */ 1324 fdp = p->p_fd; 1325 1326 /* 1327 * Range check file descriptor. 1328 */ 1329 fd = aiocbe->uaiocb.aio_fildes; 1330 if (fd >= fdp->fd_nfiles) { 1331 uma_zfree(aiocb_zone, aiocbe); 1332 if (type == 0) 1333 suword(&job->_aiocb_private.error, EBADF); 1334 return EBADF; 1335 } 1336 1337 fp = aiocbe->fd_file = fdp->fd_ofiles[fd]; 1338 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 1339 0))) { 1340 uma_zfree(aiocb_zone, aiocbe); 1341 if (type == 0) 1342 suword(&job->_aiocb_private.error, EBADF); 1343 return EBADF; 1344 } 1345 fhold(fp); 1346 1347 if (aiocbe->uaiocb.aio_offset == -1LL) { 1348 error = EINVAL; 1349 goto aqueue_fail; 1350 } 1351 error = suword(&job->_aiocb_private.kernelinfo, jobrefid); 1352 if (error) { 1353 error = EINVAL; 1354 goto aqueue_fail; 1355 } 1356 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid; 1357 if (jobrefid == LONG_MAX) 1358 jobrefid = 1; 1359 else 1360 jobrefid++; 1361 1362 if (opcode == LIO_NOP) { 1363 fdrop(fp, td); 1364 uma_zfree(aiocb_zone, aiocbe); 1365 if (type == 0) { 1366 suword(&job->_aiocb_private.error, 0); 1367 suword(&job->_aiocb_private.status, 0); 1368 suword(&job->_aiocb_private.kernelinfo, 0); 1369 } 1370 return 0; 1371 } 1372 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) { 1373 if (type == 0) 1374 suword(&job->_aiocb_private.status, 0); 1375 error = EINVAL; 1376 goto aqueue_fail; 1377 } 1378 1379 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) { 1380 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1381 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr; 1382 } 1383 else { 1384 /* 1385 * This method for requesting kevent-based notification won't 1386 * work on the alpha, since we're passing in a pointer 1387 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT- 1388 * based method instead. 1389 */ 1390 if (user_opcode == LIO_NOP || user_opcode == LIO_READ || 1391 user_opcode == LIO_WRITE) 1392 goto no_kqueue; 1393 1394 error = copyin((struct kevent *)(uintptr_t)user_opcode, 1395 &kev, sizeof(kev)); 1396 if (error) 1397 goto aqueue_fail; 1398 } 1399 if ((u_int)kev.ident >= fdp->fd_nfiles || 1400 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL || 1401 (kq_fp->f_type != DTYPE_KQUEUE)) { 1402 error = EBADF; 1403 goto aqueue_fail; 1404 } 1405 kq = (struct kqueue *)kq_fp->f_data; 1406 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1407 kev.filter = EVFILT_AIO; 1408 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 1409 kev.data = (intptr_t)aiocbe; 1410 error = kqueue_register(kq, &kev, td); 1411aqueue_fail: 1412 if (error) { 1413 fdrop(fp, td); 1414 uma_zfree(aiocb_zone, aiocbe); 1415 if (type == 0) 1416 suword(&job->_aiocb_private.error, error); 1417 goto done; 1418 } 1419no_kqueue: 1420 1421 suword(&job->_aiocb_private.error, EINPROGRESS); 1422 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1423 aiocbe->userproc = p; 1424 aiocbe->cred = crhold(td->td_ucred); 1425 aiocbe->jobflags = 0; 1426 aiocbe->lio = lj; 1427 ki = p->p_aioinfo; 1428 1429 if (fp->f_type == DTYPE_SOCKET) { 1430 /* 1431 * Alternate queueing for socket ops: Reach down into the 1432 * descriptor to get the socket data. Then check to see if the 1433 * socket is ready to be read or written (based on the requested 1434 * operation). 1435 * 1436 * If it is not ready for io, then queue the aiocbe on the 1437 * socket, and set the flags so we get a call when sbnotify() 1438 * happens. 1439 */ 1440 so = (struct socket *)fp->f_data; 1441 s = splnet(); 1442 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1443 LIO_WRITE) && (!sowriteable(so)))) { 1444 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1445 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist); 1446 if (opcode == LIO_READ) 1447 so->so_rcv.sb_flags |= SB_AIO; 1448 else 1449 so->so_snd.sb_flags |= SB_AIO; 1450 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */ 1451 ki->kaio_queue_count++; 1452 num_queue_count++; 1453 splx(s); 1454 error = 0; 1455 goto done; 1456 } 1457 splx(s); 1458 } 1459 1460 if ((error = aio_qphysio(p, aiocbe)) == 0) 1461 goto done; 1462 if (error > 0) { 1463 suword(&job->_aiocb_private.status, 0); 1464 aiocbe->uaiocb._aiocb_private.error = error; 1465 suword(&job->_aiocb_private.error, error); 1466 goto done; 1467 } 1468 1469 /* No buffer for daemon I/O. */ 1470 aiocbe->bp = NULL; 1471 1472 ki->kaio_queue_count++; 1473 if (lj) 1474 lj->lioj_queue_count++; 1475 s = splnet(); 1476 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1477 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1478 splx(s); 1479 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1480 1481 num_queue_count++; 1482 error = 0; 1483 1484 /* 1485 * If we don't have a free AIO process, and we are below our quota, then 1486 * start one. Otherwise, depend on the subsequent I/O completions to 1487 * pick-up this job. If we don't sucessfully create the new process 1488 * (thread) due to resource issues, we return an error for now (EAGAIN), 1489 * which is likely not the correct thing to do. 1490 */ 1491 s = splnet(); 1492retryproc: 1493 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1494 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1495 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list); 1496 aiop->aiothreadflags &= ~AIOP_FREE; 1497 wakeup(aiop->aiothread); 1498 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1499 ((ki->kaio_active_count + num_aio_resv_start) < 1500 ki->kaio_maxactive_count)) { 1501 num_aio_resv_start++; 1502 if ((error = aio_newproc()) == 0) { 1503 num_aio_resv_start--; 1504 goto retryproc; 1505 } 1506 num_aio_resv_start--; 1507 } 1508 splx(s); 1509done: 1510 return error; 1511} 1512 1513/* 1514 * This routine queues an AIO request, checking for quotas. 1515 */ 1516static int 1517aio_aqueue(struct thread *td, struct aiocb *job, int type) 1518{ 1519 struct proc *p = td->td_proc; 1520 struct kaioinfo *ki; 1521 1522 if (p->p_aioinfo == NULL) 1523 aio_init_aioinfo(p); 1524 1525 if (num_queue_count >= max_queue_count) 1526 return EAGAIN; 1527 1528 ki = p->p_aioinfo; 1529 if (ki->kaio_queue_count >= ki->kaio_qallowed_count) 1530 return EAGAIN; 1531 1532 return _aio_aqueue(td, job, NULL, type); 1533} 1534 1535/* 1536 * Support the aio_return system call, as a side-effect, kernel resources are 1537 * released. 1538 */ 1539int 1540aio_return(struct thread *td, struct aio_return_args *uap) 1541{ 1542 struct proc *p = td->td_proc; 1543 int s; 1544 long jobref; 1545 struct aiocblist *cb, *ncb; 1546 struct aiocb *ujob; 1547 struct kaioinfo *ki; 1548 1549 ujob = uap->aiocbp; 1550 jobref = fuword(&ujob->_aiocb_private.kernelinfo); 1551 if (jobref == -1 || jobref == 0) 1552 return EINVAL; 1553 1554 ki = p->p_aioinfo; 1555 if (ki == NULL) 1556 return EINVAL; 1557 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1558 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == 1559 jobref) { 1560 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1561 p->p_stats->p_ru.ru_oublock += 1562 cb->outputcharge; 1563 cb->outputcharge = 0; 1564 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1565 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 1566 cb->inputcharge = 0; 1567 } 1568 goto done; 1569 } 1570 } 1571 s = splbio(); 1572 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) { 1573 ncb = TAILQ_NEXT(cb, plist); 1574 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) 1575 == jobref) { 1576 break; 1577 } 1578 } 1579 splx(s); 1580 done: 1581 if (cb != NULL) { 1582 if (ujob == cb->uuaiocb) { 1583 td->td_retval[0] = 1584 cb->uaiocb._aiocb_private.status; 1585 } else 1586 td->td_retval[0] = EFAULT; 1587 aio_free_entry(cb); 1588 return (0); 1589 } 1590 return (EINVAL); 1591} 1592 1593/* 1594 * Allow a process to wakeup when any of the I/O requests are completed. 1595 */ 1596int 1597aio_suspend(struct thread *td, struct aio_suspend_args *uap) 1598{ 1599 struct proc *p = td->td_proc; 1600 struct timeval atv; 1601 struct timespec ts; 1602 struct aiocb *const *cbptr, *cbp; 1603 struct kaioinfo *ki; 1604 struct aiocblist *cb; 1605 int i; 1606 int njoblist; 1607 int error, s, timo; 1608 long *ijoblist; 1609 struct aiocb **ujoblist; 1610 1611 if (uap->nent > AIO_LISTIO_MAX) 1612 return EINVAL; 1613 1614 timo = 0; 1615 if (uap->timeout) { 1616 /* Get timespec struct. */ 1617 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1618 return error; 1619 1620 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) 1621 return (EINVAL); 1622 1623 TIMESPEC_TO_TIMEVAL(&atv, &ts); 1624 if (itimerfix(&atv)) 1625 return (EINVAL); 1626 timo = tvtohz(&atv); 1627 } 1628 1629 ki = p->p_aioinfo; 1630 if (ki == NULL) 1631 return EAGAIN; 1632 1633 njoblist = 0; 1634 ijoblist = uma_zalloc(aiol_zone, M_WAITOK); 1635 ujoblist = uma_zalloc(aiol_zone, M_WAITOK); 1636 cbptr = uap->aiocbp; 1637 1638 for (i = 0; i < uap->nent; i++) { 1639 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1640 if (cbp == 0) 1641 continue; 1642 ujoblist[njoblist] = cbp; 1643 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo); 1644 njoblist++; 1645 } 1646 1647 if (njoblist == 0) { 1648 uma_zfree(aiol_zone, ijoblist); 1649 uma_zfree(aiol_zone, ujoblist); 1650 return 0; 1651 } 1652 1653 error = 0; 1654 for (;;) { 1655 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1656 for (i = 0; i < njoblist; i++) { 1657 if (((intptr_t) 1658 cb->uaiocb._aiocb_private.kernelinfo) == 1659 ijoblist[i]) { 1660 if (ujoblist[i] != cb->uuaiocb) 1661 error = EINVAL; 1662 uma_zfree(aiol_zone, ijoblist); 1663 uma_zfree(aiol_zone, ujoblist); 1664 return error; 1665 } 1666 } 1667 } 1668 1669 s = splbio(); 1670 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = 1671 TAILQ_NEXT(cb, plist)) { 1672 for (i = 0; i < njoblist; i++) { 1673 if (((intptr_t) 1674 cb->uaiocb._aiocb_private.kernelinfo) == 1675 ijoblist[i]) { 1676 splx(s); 1677 if (ujoblist[i] != cb->uuaiocb) 1678 error = EINVAL; 1679 uma_zfree(aiol_zone, ijoblist); 1680 uma_zfree(aiol_zone, ujoblist); 1681 return error; 1682 } 1683 } 1684 } 1685 1686 ki->kaio_flags |= KAIO_WAKEUP; 1687 error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo); 1688 splx(s); 1689 1690 if (error == ERESTART || error == EINTR) { 1691 uma_zfree(aiol_zone, ijoblist); 1692 uma_zfree(aiol_zone, ujoblist); 1693 return EINTR; 1694 } else if (error == EWOULDBLOCK) { 1695 uma_zfree(aiol_zone, ijoblist); 1696 uma_zfree(aiol_zone, ujoblist); 1697 return EAGAIN; 1698 } 1699 } 1700 1701/* NOTREACHED */ 1702 return EINVAL; 1703} 1704 1705/* 1706 * aio_cancel cancels any non-physio aio operations not currently in 1707 * progress. 1708 */ 1709int 1710aio_cancel(struct thread *td, struct aio_cancel_args *uap) 1711{ 1712 struct proc *p = td->td_proc; 1713 struct kaioinfo *ki; 1714 struct aiocblist *cbe, *cbn; 1715 struct file *fp; 1716 struct filedesc *fdp; 1717 struct socket *so; 1718 struct proc *po; 1719 int s,error; 1720 int cancelled=0; 1721 int notcancelled=0; 1722 struct vnode *vp; 1723 1724 fdp = p->p_fd; 1725 if ((u_int)uap->fd >= fdp->fd_nfiles || 1726 (fp = fdp->fd_ofiles[uap->fd]) == NULL) 1727 return (EBADF); 1728 1729 if (fp->f_type == DTYPE_VNODE) { 1730 vp = (struct vnode *)fp->f_data; 1731 1732 if (vn_isdisk(vp,&error)) { 1733 td->td_retval[0] = AIO_NOTCANCELED; 1734 return 0; 1735 } 1736 } else if (fp->f_type == DTYPE_SOCKET) { 1737 so = (struct socket *)fp->f_data; 1738 1739 s = splnet(); 1740 1741 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) { 1742 cbn = TAILQ_NEXT(cbe, list); 1743 if ((uap->aiocbp == NULL) || 1744 (uap->aiocbp == cbe->uuaiocb) ) { 1745 po = cbe->userproc; 1746 ki = po->p_aioinfo; 1747 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 1748 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist); 1749 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist); 1750 if (ki->kaio_flags & KAIO_WAKEUP) { 1751 wakeup(po); 1752 } 1753 cbe->jobstate = JOBST_JOBFINISHED; 1754 cbe->uaiocb._aiocb_private.status=-1; 1755 cbe->uaiocb._aiocb_private.error=ECANCELED; 1756 cancelled++; 1757/* XXX cancelled, knote? */ 1758 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1759 SIGEV_SIGNAL) { 1760 PROC_LOCK(cbe->userproc); 1761 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1762 PROC_UNLOCK(cbe->userproc); 1763 } 1764 if (uap->aiocbp) 1765 break; 1766 } 1767 } 1768 splx(s); 1769 1770 if ((cancelled) && (uap->aiocbp)) { 1771 td->td_retval[0] = AIO_CANCELED; 1772 return 0; 1773 } 1774 } 1775 ki=p->p_aioinfo; 1776 if (ki == NULL) 1777 goto done; 1778 s = splnet(); 1779 1780 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) { 1781 cbn = TAILQ_NEXT(cbe, plist); 1782 1783 if ((uap->fd == cbe->uaiocb.aio_fildes) && 1784 ((uap->aiocbp == NULL ) || 1785 (uap->aiocbp == cbe->uuaiocb))) { 1786 1787 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 1788 TAILQ_REMOVE(&aio_jobs, cbe, list); 1789 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 1790 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, 1791 plist); 1792 cancelled++; 1793 ki->kaio_queue_finished_count++; 1794 cbe->jobstate = JOBST_JOBFINISHED; 1795 cbe->uaiocb._aiocb_private.status = -1; 1796 cbe->uaiocb._aiocb_private.error = ECANCELED; 1797/* XXX cancelled, knote? */ 1798 if (cbe->uaiocb.aio_sigevent.sigev_notify == 1799 SIGEV_SIGNAL) { 1800 PROC_LOCK(cbe->userproc); 1801 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo); 1802 PROC_UNLOCK(cbe->userproc); 1803 } 1804 } else { 1805 notcancelled++; 1806 } 1807 } 1808 } 1809 splx(s); 1810done: 1811 if (notcancelled) { 1812 td->td_retval[0] = AIO_NOTCANCELED; 1813 return 0; 1814 } 1815 if (cancelled) { 1816 td->td_retval[0] = AIO_CANCELED; 1817 return 0; 1818 } 1819 td->td_retval[0] = AIO_ALLDONE; 1820 1821 return 0; 1822} 1823 1824/* 1825 * aio_error is implemented in the kernel level for compatibility purposes only. 1826 * For a user mode async implementation, it would be best to do it in a userland 1827 * subroutine. 1828 */ 1829int 1830aio_error(struct thread *td, struct aio_error_args *uap) 1831{ 1832 struct proc *p = td->td_proc; 1833 int s; 1834 struct aiocblist *cb; 1835 struct kaioinfo *ki; 1836 long jobref; 1837 1838 ki = p->p_aioinfo; 1839 if (ki == NULL) 1840 return EINVAL; 1841 1842 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo); 1843 if ((jobref == -1) || (jobref == 0)) 1844 return EINVAL; 1845 1846 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 1847 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1848 jobref) { 1849 td->td_retval[0] = cb->uaiocb._aiocb_private.error; 1850 return 0; 1851 } 1852 } 1853 1854 s = splnet(); 1855 1856 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb, 1857 plist)) { 1858 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1859 jobref) { 1860 td->td_retval[0] = EINPROGRESS; 1861 splx(s); 1862 return 0; 1863 } 1864 } 1865 1866 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb, 1867 plist)) { 1868 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1869 jobref) { 1870 td->td_retval[0] = EINPROGRESS; 1871 splx(s); 1872 return 0; 1873 } 1874 } 1875 splx(s); 1876 1877 s = splbio(); 1878 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb, 1879 plist)) { 1880 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1881 jobref) { 1882 td->td_retval[0] = cb->uaiocb._aiocb_private.error; 1883 splx(s); 1884 return 0; 1885 } 1886 } 1887 1888 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb, 1889 plist)) { 1890 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) == 1891 jobref) { 1892 td->td_retval[0] = EINPROGRESS; 1893 splx(s); 1894 return 0; 1895 } 1896 } 1897 splx(s); 1898 1899#if (0) 1900 /* 1901 * Hack for lio. 1902 */ 1903 status = fuword(&uap->aiocbp->_aiocb_private.status); 1904 if (status == -1) 1905 return fuword(&uap->aiocbp->_aiocb_private.error); 1906#endif 1907 return EINVAL; 1908} 1909 1910/* syscall - asynchronous read from a file (REALTIME) */ 1911int 1912aio_read(struct thread *td, struct aio_read_args *uap) 1913{ 1914 1915 return aio_aqueue(td, uap->aiocbp, LIO_READ); 1916} 1917 1918/* syscall - asynchronous write to a file (REALTIME) */ 1919int 1920aio_write(struct thread *td, struct aio_write_args *uap) 1921{ 1922 1923 return aio_aqueue(td, uap->aiocbp, LIO_WRITE); 1924} 1925 1926/* syscall - XXX undocumented */ 1927int 1928lio_listio(struct thread *td, struct lio_listio_args *uap) 1929{ 1930 struct proc *p = td->td_proc; 1931 int nent, nentqueued; 1932 struct aiocb *iocb, * const *cbptr; 1933 struct aiocblist *cb; 1934 struct kaioinfo *ki; 1935 struct aio_liojob *lj; 1936 int error, runningcode; 1937 int nerror; 1938 int i; 1939 int s; 1940 1941 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 1942 return EINVAL; 1943 1944 nent = uap->nent; 1945 if (nent > AIO_LISTIO_MAX) 1946 return EINVAL; 1947 1948 if (p->p_aioinfo == NULL) 1949 aio_init_aioinfo(p); 1950 1951 if ((nent + num_queue_count) > max_queue_count) 1952 return EAGAIN; 1953 1954 ki = p->p_aioinfo; 1955 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) 1956 return EAGAIN; 1957 1958 lj = uma_zalloc(aiolio_zone, M_WAITOK); 1959 if (!lj) 1960 return EAGAIN; 1961 1962 lj->lioj_flags = 0; 1963 lj->lioj_buffer_count = 0; 1964 lj->lioj_buffer_finished_count = 0; 1965 lj->lioj_queue_count = 0; 1966 lj->lioj_queue_finished_count = 0; 1967 lj->lioj_ki = ki; 1968 1969 /* 1970 * Setup signal. 1971 */ 1972 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 1973 error = copyin(uap->sig, &lj->lioj_signal, 1974 sizeof(lj->lioj_signal)); 1975 if (error) { 1976 uma_zfree(aiolio_zone, lj); 1977 return error; 1978 } 1979 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 1980 uma_zfree(aiolio_zone, lj); 1981 return EINVAL; 1982 } 1983 lj->lioj_flags |= LIOJ_SIGNAL; 1984 } 1985 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 1986 /* 1987 * Get pointers to the list of I/O requests. 1988 */ 1989 nerror = 0; 1990 nentqueued = 0; 1991 cbptr = uap->acb_list; 1992 for (i = 0; i < uap->nent; i++) { 1993 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]); 1994 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) { 1995 error = _aio_aqueue(td, iocb, lj, 0); 1996 if (error == 0) 1997 nentqueued++; 1998 else 1999 nerror++; 2000 } 2001 } 2002 2003 /* 2004 * If we haven't queued any, then just return error. 2005 */ 2006 if (nentqueued == 0) 2007 return 0; 2008 2009 /* 2010 * Calculate the appropriate error return. 2011 */ 2012 runningcode = 0; 2013 if (nerror) 2014 runningcode = EIO; 2015 2016 if (uap->mode == LIO_WAIT) { 2017 int command, found, jobref; 2018 2019 for (;;) { 2020 found = 0; 2021 for (i = 0; i < uap->nent; i++) { 2022 /* 2023 * Fetch address of the control buf pointer in 2024 * user space. 2025 */ 2026 iocb = (struct aiocb *) 2027 (intptr_t)fuword(&cbptr[i]); 2028 if (((intptr_t)iocb == -1) || ((intptr_t)iocb 2029 == 0)) 2030 continue; 2031 2032 /* 2033 * Fetch the associated command from user space. 2034 */ 2035 command = fuword(&iocb->aio_lio_opcode); 2036 if (command == LIO_NOP) { 2037 found++; 2038 continue; 2039 } 2040 2041 jobref = fuword(&iocb->_aiocb_private.kernelinfo); 2042 2043 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) { 2044 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 2045 == jobref) { 2046 if (cb->uaiocb.aio_lio_opcode 2047 == LIO_WRITE) { 2048 p->p_stats->p_ru.ru_oublock 2049 += 2050 cb->outputcharge; 2051 cb->outputcharge = 0; 2052 } else if (cb->uaiocb.aio_lio_opcode 2053 == LIO_READ) { 2054 p->p_stats->p_ru.ru_inblock 2055 += cb->inputcharge; 2056 cb->inputcharge = 0; 2057 } 2058 found++; 2059 break; 2060 } 2061 } 2062 2063 s = splbio(); 2064 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) { 2065 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) 2066 == jobref) { 2067 found++; 2068 break; 2069 } 2070 } 2071 splx(s); 2072 } 2073 2074 /* 2075 * If all I/Os have been disposed of, then we can 2076 * return. 2077 */ 2078 if (found == nentqueued) 2079 return runningcode; 2080 2081 ki->kaio_flags |= KAIO_WAKEUP; 2082 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0); 2083 2084 if (error == EINTR) 2085 return EINTR; 2086 else if (error == EWOULDBLOCK) 2087 return EAGAIN; 2088 } 2089 } 2090 2091 return runningcode; 2092} 2093 2094/* 2095 * This is a weird hack so that we can post a signal. It is safe to do so from 2096 * a timeout routine, but *not* from an interrupt routine. 2097 */ 2098static void 2099process_signal(void *aioj) 2100{ 2101 struct aiocblist *aiocbe = aioj; 2102 struct aio_liojob *lj = aiocbe->lio; 2103 struct aiocb *cb = &aiocbe->uaiocb; 2104 2105 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) && 2106 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) { 2107 PROC_LOCK(lj->lioj_ki->kaio_p); 2108 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo); 2109 PROC_UNLOCK(lj->lioj_ki->kaio_p); 2110 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2111 } 2112 2113 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) { 2114 PROC_LOCK(aiocbe->userproc); 2115 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo); 2116 PROC_UNLOCK(aiocbe->userproc); 2117 } 2118} 2119 2120/* 2121 * Interrupt handler for physio, performs the necessary process wakeups, and 2122 * signals. 2123 */ 2124static void 2125aio_physwakeup(struct buf *bp) 2126{ 2127 struct aiocblist *aiocbe; 2128 struct proc *p; 2129 struct kaioinfo *ki; 2130 struct aio_liojob *lj; 2131 2132 wakeup(bp); 2133 2134 aiocbe = (struct aiocblist *)bp->b_spc; 2135 if (aiocbe) { 2136 p = bp->b_caller1; 2137 2138 aiocbe->jobstate = JOBST_JOBBFINISHED; 2139 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 2140 aiocbe->uaiocb._aiocb_private.error = 0; 2141 aiocbe->jobflags |= AIOCBLIST_DONE; 2142 2143 if (bp->b_ioflags & BIO_ERROR) 2144 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 2145 2146 lj = aiocbe->lio; 2147 if (lj) { 2148 lj->lioj_buffer_finished_count++; 2149 2150 /* 2151 * wakeup/signal if all of the interrupt jobs are done. 2152 */ 2153 if (lj->lioj_buffer_finished_count == 2154 lj->lioj_buffer_count) { 2155 /* 2156 * Post a signal if it is called for. 2157 */ 2158 if ((lj->lioj_flags & 2159 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == 2160 LIOJ_SIGNAL) { 2161 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2162 aiocbe->timeouthandle = 2163 timeout(process_signal, 2164 aiocbe, 0); 2165 } 2166 } 2167 } 2168 2169 ki = p->p_aioinfo; 2170 if (ki) { 2171 ki->kaio_buffer_finished_count++; 2172 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list); 2173 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist); 2174 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist); 2175 2176 KNOTE(&aiocbe->klist, 0); 2177 /* Do the wakeup. */ 2178 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) { 2179 ki->kaio_flags &= ~KAIO_WAKEUP; 2180 wakeup(p); 2181 } 2182 } 2183 2184 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) 2185 aiocbe->timeouthandle = 2186 timeout(process_signal, aiocbe, 0); 2187 } 2188} 2189 2190/* syscall - wait for the next completion of an aio request */ 2191int 2192aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2193{ 2194 struct proc *p = td->td_proc; 2195 struct timeval atv; 2196 struct timespec ts; 2197 struct kaioinfo *ki; 2198 struct aiocblist *cb = NULL; 2199 int error, s, timo; 2200 2201 suword(uap->aiocbp, (int)NULL); 2202 2203 timo = 0; 2204 if (uap->timeout) { 2205 /* Get timespec struct. */ 2206 error = copyin(uap->timeout, &ts, sizeof(ts)); 2207 if (error) 2208 return error; 2209 2210 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000)) 2211 return (EINVAL); 2212 2213 TIMESPEC_TO_TIMEVAL(&atv, &ts); 2214 if (itimerfix(&atv)) 2215 return (EINVAL); 2216 timo = tvtohz(&atv); 2217 } 2218 2219 ki = p->p_aioinfo; 2220 if (ki == NULL) 2221 return EAGAIN; 2222 2223 for (;;) { 2224 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) { 2225 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2226 td->td_retval[0] = cb->uaiocb._aiocb_private.status; 2227 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2228 p->p_stats->p_ru.ru_oublock += 2229 cb->outputcharge; 2230 cb->outputcharge = 0; 2231 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2232 p->p_stats->p_ru.ru_inblock += cb->inputcharge; 2233 cb->inputcharge = 0; 2234 } 2235 aio_free_entry(cb); 2236 return cb->uaiocb._aiocb_private.error; 2237 } 2238 2239 s = splbio(); 2240 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) { 2241 splx(s); 2242 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb); 2243 td->td_retval[0] = cb->uaiocb._aiocb_private.status; 2244 aio_free_entry(cb); 2245 return cb->uaiocb._aiocb_private.error; 2246 } 2247 2248 ki->kaio_flags |= KAIO_WAKEUP; 2249 error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo); 2250 splx(s); 2251 2252 if (error == ERESTART) 2253 return EINTR; 2254 else if (error < 0) 2255 return error; 2256 else if (error == EINTR) 2257 return EINTR; 2258 else if (error == EWOULDBLOCK) 2259 return EAGAIN; 2260 } 2261} 2262 2263/* kqueue attach function */ 2264static int 2265filt_aioattach(struct knote *kn) 2266{ 2267 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2268 2269 /* 2270 * The aiocbe pointer must be validated before using it, so 2271 * registration is restricted to the kernel; the user cannot 2272 * set EV_FLAG1. 2273 */ 2274 if ((kn->kn_flags & EV_FLAG1) == 0) 2275 return (EPERM); 2276 kn->kn_flags &= ~EV_FLAG1; 2277 2278 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext); 2279 2280 return (0); 2281} 2282 2283/* kqueue detach function */ 2284static void 2285filt_aiodetach(struct knote *kn) 2286{ 2287 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2288 2289 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext); 2290} 2291 2292/* kqueue filter function */ 2293/*ARGSUSED*/ 2294static int 2295filt_aio(struct knote *kn, long hint) 2296{ 2297 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2298 2299 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2300 if (aiocbe->jobstate != JOBST_JOBFINISHED && 2301 aiocbe->jobstate != JOBST_JOBBFINISHED) 2302 return (0); 2303 kn->kn_flags |= EV_EOF; 2304 return (1); 2305} 2306