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 17 /* 18 * This file contains support for the POSIX 1003.1B AIO/LIO facility. 19 */ 20 21 #include <sys/cdefs.h> 22 __FBSDID("$FreeBSD$"); 23 24 #include "opt_compat.h" 25 26 #include <sys/param.h> 27 #include <sys/systm.h> 28 #include <sys/malloc.h> 29 #include <sys/bio.h> 30 #include <sys/buf.h> 31 #include <sys/capability.h> 32 #include <sys/eventhandler.h> 33 #include <sys/sysproto.h> 34 #include <sys/filedesc.h> 35 #include <sys/kernel.h> 36 #include <sys/module.h> 37 #include <sys/kthread.h> 38 #include <sys/fcntl.h> 39 #include <sys/file.h> 40 #include <sys/limits.h> 41 #include <sys/lock.h> 42 #include <sys/mutex.h> 43 #include <sys/unistd.h> 44 #include <sys/posix4.h> 45 #include <sys/proc.h> 46 #include <sys/resourcevar.h> 47 #include <sys/signalvar.h> 48 #include <sys/protosw.h> 49 #include <sys/sema.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/syscall.h> 53 #include <sys/sysent.h> 54 #include <sys/sysctl.h> 55 #include <sys/sx.h> 56 #include <sys/taskqueue.h> 57 #include <sys/vnode.h> 58 #include <sys/conf.h> 59 #include <sys/event.h> 60 #include <sys/mount.h> 61 62 #include <machine/atomic.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_extern.h> 66 #include <vm/pmap.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_object.h> 69 #include <vm/uma.h> 70 #include <sys/aio.h> 71 72 #include "opt_vfs_aio.h" 73 74 /* 75 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 76 * overflow. (XXX will be removed soon.) 77 */ 78 static u_long jobrefid; 79 80 /* 81 * Counter for aio_fsync. 82 */ 83 static uint64_t jobseqno; 84 85 #define JOBST_NULL 0 86 #define JOBST_JOBQSOCK 1 87 #define JOBST_JOBQGLOBAL 2 88 #define JOBST_JOBRUNNING 3 89 #define JOBST_JOBFINISHED 4 90 #define JOBST_JOBQBUF 5 91 #define JOBST_JOBQSYNC 6 92 93 #ifndef MAX_AIO_PER_PROC 94 #define MAX_AIO_PER_PROC 32 95 #endif 96 97 #ifndef MAX_AIO_QUEUE_PER_PROC 98 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */ 99 #endif 100 101 #ifndef MAX_AIO_PROCS 102 #define MAX_AIO_PROCS 32 103 #endif 104 105 #ifndef MAX_AIO_QUEUE 106 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */ 107 #endif 108 109 #ifndef TARGET_AIO_PROCS 110 #define TARGET_AIO_PROCS 4 111 #endif 112 113 #ifndef MAX_BUF_AIO 114 #define MAX_BUF_AIO 16 115 #endif 116 117 #ifndef AIOD_TIMEOUT_DEFAULT 118 #define AIOD_TIMEOUT_DEFAULT (10 * hz) 119 #endif 120 121 #ifndef AIOD_LIFETIME_DEFAULT 122 #define AIOD_LIFETIME_DEFAULT (30 * hz) 123 #endif 124 125 FEATURE(aio, "Asynchronous I/O"); 126 127 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list"); 128 129 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management"); 130 131 static int max_aio_procs = MAX_AIO_PROCS; 132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, 133 CTLFLAG_RW, &max_aio_procs, 0, 134 "Maximum number of kernel threads to use for handling async IO "); 135 136 static int num_aio_procs = 0; 137 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, 138 CTLFLAG_RD, &num_aio_procs, 0, 139 "Number of presently active kernel threads for async IO"); 140 141 /* 142 * The code will adjust the actual number of AIO processes towards this 143 * number when it gets a chance. 144 */ 145 static int target_aio_procs = TARGET_AIO_PROCS; 146 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 147 0, "Preferred number of ready kernel threads for async IO"); 148 149 static int max_queue_count = MAX_AIO_QUEUE; 150 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 151 "Maximum number of aio requests to queue, globally"); 152 153 static int num_queue_count = 0; 154 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 155 "Number of queued aio requests"); 156 157 static int num_buf_aio = 0; 158 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 159 "Number of aio requests presently handled by the buf subsystem"); 160 161 /* Number of async I/O thread in the process of being started */ 162 /* XXX This should be local to aio_aqueue() */ 163 static int num_aio_resv_start = 0; 164 165 static int aiod_timeout; 166 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0, 167 "Timeout value for synchronous aio operations"); 168 169 static int aiod_lifetime; 170 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 171 "Maximum lifetime for idle aiod"); 172 173 static int unloadable = 0; 174 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0, 175 "Allow unload of aio (not recommended)"); 176 177 178 static int max_aio_per_proc = MAX_AIO_PER_PROC; 179 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 180 0, "Maximum active aio requests per process (stored in the process)"); 181 182 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 183 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 184 &max_aio_queue_per_proc, 0, 185 "Maximum queued aio requests per process (stored in the process)"); 186 187 static int max_buf_aio = MAX_BUF_AIO; 188 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 189 "Maximum buf aio requests per process (stored in the process)"); 190 191 typedef struct oaiocb { 192 int aio_fildes; /* File descriptor */ 193 off_t aio_offset; /* File offset for I/O */ 194 volatile void *aio_buf; /* I/O buffer in process space */ 195 size_t aio_nbytes; /* Number of bytes for I/O */ 196 struct osigevent aio_sigevent; /* Signal to deliver */ 197 int aio_lio_opcode; /* LIO opcode */ 198 int aio_reqprio; /* Request priority -- ignored */ 199 struct __aiocb_private _aiocb_private; 200 } oaiocb_t; 201 202 /* 203 * Below is a key of locks used to protect each member of struct aiocblist 204 * aioliojob and kaioinfo and any backends. 205 * 206 * * - need not protected 207 * a - locked by kaioinfo lock 208 * b - locked by backend lock, the backend lock can be null in some cases, 209 * for example, BIO belongs to this type, in this case, proc lock is 210 * reused. 211 * c - locked by aio_job_mtx, the lock for the generic file I/O backend. 212 */ 213 214 /* 215 * Current, there is only two backends: BIO and generic file I/O. 216 * socket I/O is served by generic file I/O, this is not a good idea, since 217 * disk file I/O and any other types without O_NONBLOCK flag can block daemon 218 * threads, if there is no thread to serve socket I/O, the socket I/O will be 219 * delayed too long or starved, we should create some threads dedicated to 220 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O 221 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file 222 * structure is not safe because there is race between userland and aio 223 * daemons. 224 */ 225 226 struct aiocblist { 227 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */ 228 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */ 229 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */ 230 int jobflags; /* (a) job flags */ 231 int jobstate; /* (b) job state */ 232 int inputcharge; /* (*) input blockes */ 233 int outputcharge; /* (*) output blockes */ 234 struct buf *bp; /* (*) private to BIO backend, 235 * buffer pointer 236 */ 237 struct proc *userproc; /* (*) user process */ 238 struct ucred *cred; /* (*) active credential when created */ 239 struct file *fd_file; /* (*) pointer to file structure */ 240 struct aioliojob *lio; /* (*) optional lio job */ 241 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */ 242 struct knlist klist; /* (a) list of knotes */ 243 struct aiocb uaiocb; /* (*) kernel I/O control block */ 244 ksiginfo_t ksi; /* (a) realtime signal info */ 245 struct task biotask; /* (*) private to BIO backend */ 246 uint64_t seqno; /* (*) job number */ 247 int pending; /* (a) number of pending I/O, aio_fsync only */ 248 }; 249 250 /* jobflags */ 251 #define AIOCBLIST_DONE 0x01 252 #define AIOCBLIST_BUFDONE 0x02 253 #define AIOCBLIST_RUNDOWN 0x04 254 #define AIOCBLIST_CHECKSYNC 0x08 255 256 /* 257 * AIO process info 258 */ 259 #define AIOP_FREE 0x1 /* proc on free queue */ 260 261 struct aiothreadlist { 262 int aiothreadflags; /* (c) AIO proc flags */ 263 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */ 264 struct thread *aiothread; /* (*) the AIO thread */ 265 }; 266 267 /* 268 * data-structure for lio signal management 269 */ 270 struct aioliojob { 271 int lioj_flags; /* (a) listio flags */ 272 int lioj_count; /* (a) listio flags */ 273 int lioj_finished_count; /* (a) listio flags */ 274 struct sigevent lioj_signal; /* (a) signal on all I/O done */ 275 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */ 276 struct knlist klist; /* (a) list of knotes */ 277 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */ 278 }; 279 280 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 281 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 282 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */ 283 284 /* 285 * per process aio data structure 286 */ 287 struct kaioinfo { 288 struct mtx kaio_mtx; /* the lock to protect this struct */ 289 int kaio_flags; /* (a) per process kaio flags */ 290 int kaio_maxactive_count; /* (*) maximum number of AIOs */ 291 int kaio_active_count; /* (c) number of currently used AIOs */ 292 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */ 293 int kaio_count; /* (a) size of AIO queue */ 294 int kaio_ballowed_count; /* (*) maximum number of buffers */ 295 int kaio_buffer_count; /* (a) number of physio buffers */ 296 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */ 297 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */ 298 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */ 299 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */ 300 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */ 301 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets, 302 * NOT USED YET. 303 */ 304 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */ 305 struct task kaio_task; /* (*) task to kick aio threads */ 306 }; 307 308 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx) 309 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx) 310 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f)) 311 #define AIO_MTX(ki) (&(ki)->kaio_mtx) 312 313 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 314 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */ 315 316 /* 317 * Operations used to interact with userland aio control blocks. 318 * Different ABIs provide their own operations. 319 */ 320 struct aiocb_ops { 321 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob); 322 long (*fetch_status)(struct aiocb *ujob); 323 long (*fetch_error)(struct aiocb *ujob); 324 int (*store_status)(struct aiocb *ujob, long status); 325 int (*store_error)(struct aiocb *ujob, long error); 326 int (*store_kernelinfo)(struct aiocb *ujob, long jobref); 327 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob); 328 }; 329 330 static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */ 331 static struct sema aio_newproc_sem; 332 static struct mtx aio_job_mtx; 333 static struct mtx aio_sock_mtx; 334 static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */ 335 static struct unrhdr *aiod_unr; 336 337 void aio_init_aioinfo(struct proc *p); 338 static int aio_onceonly(void); 339 static int aio_free_entry(struct aiocblist *aiocbe); 340 static void aio_process(struct aiocblist *aiocbe); 341 static int aio_newproc(int *); 342 int aio_aqueue(struct thread *td, struct aiocb *job, 343 struct aioliojob *lio, int type, struct aiocb_ops *ops); 344 static void aio_physwakeup(struct buf *bp); 345 static void aio_proc_rundown(void *arg, struct proc *p); 346 static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp); 347 static int aio_qphysio(struct proc *p, struct aiocblist *iocb); 348 static void biohelper(void *, int); 349 static void aio_daemon(void *param); 350 static void aio_swake_cb(struct socket *, struct sockbuf *); 351 static int aio_unload(void); 352 static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type); 353 #define DONE_BUF 1 354 #define DONE_QUEUE 2 355 static int aio_kick(struct proc *userp); 356 static void aio_kick_nowait(struct proc *userp); 357 static void aio_kick_helper(void *context, int pending); 358 static int filt_aioattach(struct knote *kn); 359 static void filt_aiodetach(struct knote *kn); 360 static int filt_aio(struct knote *kn, long hint); 361 static int filt_lioattach(struct knote *kn); 362 static void filt_liodetach(struct knote *kn); 363 static int filt_lio(struct knote *kn, long hint); 364 365 /* 366 * Zones for: 367 * kaio Per process async io info 368 * aiop async io thread data 369 * aiocb async io jobs 370 * aiol list io job pointer - internal to aio_suspend XXX 371 * aiolio list io jobs 372 */ 373 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone; 374 375 /* kqueue filters for aio */ 376 static struct filterops aio_filtops = { 377 .f_isfd = 0, 378 .f_attach = filt_aioattach, 379 .f_detach = filt_aiodetach, 380 .f_event = filt_aio, 381 }; 382 static struct filterops lio_filtops = { 383 .f_isfd = 0, 384 .f_attach = filt_lioattach, 385 .f_detach = filt_liodetach, 386 .f_event = filt_lio 387 }; 388 389 static eventhandler_tag exit_tag, exec_tag; 390 391 TASKQUEUE_DEFINE_THREAD(aiod_bio); 392 393 /* 394 * Main operations function for use as a kernel module. 395 */ 396 static int 397 aio_modload(struct module *module, int cmd, void *arg) 398 { 399 int error = 0; 400 401 switch (cmd) { 402 case MOD_LOAD: 403 aio_onceonly(); 404 break; 405 case MOD_UNLOAD: 406 error = aio_unload(); 407 break; 408 case MOD_SHUTDOWN: 409 break; 410 default: 411 error = EINVAL; 412 break; 413 } 414 return (error); 415 } 416 417 static moduledata_t aio_mod = { 418 "aio", 419 &aio_modload, 420 NULL 421 }; 422 423 static struct syscall_helper_data aio_syscalls[] = { 424 SYSCALL_INIT_HELPER(aio_cancel), 425 SYSCALL_INIT_HELPER(aio_error), 426 SYSCALL_INIT_HELPER(aio_fsync), 427 SYSCALL_INIT_HELPER(aio_read), 428 SYSCALL_INIT_HELPER(aio_return), 429 SYSCALL_INIT_HELPER(aio_suspend), 430 SYSCALL_INIT_HELPER(aio_waitcomplete), 431 SYSCALL_INIT_HELPER(aio_write), 432 SYSCALL_INIT_HELPER(lio_listio), 433 SYSCALL_INIT_HELPER(oaio_read), 434 SYSCALL_INIT_HELPER(oaio_write), 435 SYSCALL_INIT_HELPER(olio_listio), 436 SYSCALL_INIT_LAST 437 }; 438 439 #ifdef COMPAT_FREEBSD32 440 #include <sys/mount.h> 441 #include <sys/socket.h> 442 #include <compat/freebsd32/freebsd32.h> 443 #include <compat/freebsd32/freebsd32_proto.h> 444 #include <compat/freebsd32/freebsd32_signal.h> 445 #include <compat/freebsd32/freebsd32_syscall.h> 446 #include <compat/freebsd32/freebsd32_util.h> 447 448 static struct syscall_helper_data aio32_syscalls[] = { 449 SYSCALL32_INIT_HELPER(freebsd32_aio_return), 450 SYSCALL32_INIT_HELPER(freebsd32_aio_suspend), 451 SYSCALL32_INIT_HELPER(freebsd32_aio_cancel), 452 SYSCALL32_INIT_HELPER(freebsd32_aio_error), 453 SYSCALL32_INIT_HELPER(freebsd32_aio_fsync), 454 SYSCALL32_INIT_HELPER(freebsd32_aio_read), 455 SYSCALL32_INIT_HELPER(freebsd32_aio_write), 456 SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete), 457 SYSCALL32_INIT_HELPER(freebsd32_lio_listio), 458 SYSCALL32_INIT_HELPER(freebsd32_oaio_read), 459 SYSCALL32_INIT_HELPER(freebsd32_oaio_write), 460 SYSCALL32_INIT_HELPER(freebsd32_olio_listio), 461 SYSCALL_INIT_LAST 462 }; 463 #endif 464 465 DECLARE_MODULE(aio, aio_mod, 466 SI_SUB_VFS, SI_ORDER_ANY); 467 MODULE_VERSION(aio, 1); 468 469 /* 470 * Startup initialization 471 */ 472 static int 473 aio_onceonly(void) 474 { 475 int error; 476 477 /* XXX: should probably just use so->callback */ 478 aio_swake = &aio_swake_cb; 479 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 480 EVENTHANDLER_PRI_ANY); 481 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL, 482 EVENTHANDLER_PRI_ANY); 483 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 484 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 485 TAILQ_INIT(&aio_freeproc); 486 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 487 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 488 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF); 489 TAILQ_INIT(&aio_jobs); 490 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 491 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 492 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 493 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL, 494 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 495 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL, 496 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 497 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL, 498 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 499 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 500 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 501 aiod_timeout = AIOD_TIMEOUT_DEFAULT; 502 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 503 jobrefid = 1; 504 async_io_version = _POSIX_VERSION; 505 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX); 506 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 507 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 508 509 error = syscall_helper_register(aio_syscalls); 510 if (error) 511 return (error); 512 #ifdef COMPAT_FREEBSD32 513 error = syscall32_helper_register(aio32_syscalls); 514 if (error) 515 return (error); 516 #endif 517 return (0); 518 } 519 520 /* 521 * Callback for unload of AIO when used as a module. 522 */ 523 static int 524 aio_unload(void) 525 { 526 int error; 527 528 /* 529 * XXX: no unloads by default, it's too dangerous. 530 * perhaps we could do it if locked out callers and then 531 * did an aio_proc_rundown() on each process. 532 * 533 * jhb: aio_proc_rundown() needs to run on curproc though, 534 * so I don't think that would fly. 535 */ 536 if (!unloadable) 537 return (EOPNOTSUPP); 538 539 #ifdef COMPAT_FREEBSD32 540 syscall32_helper_unregister(aio32_syscalls); 541 #endif 542 syscall_helper_unregister(aio_syscalls); 543 544 error = kqueue_del_filteropts(EVFILT_AIO); 545 if (error) 546 return error; 547 error = kqueue_del_filteropts(EVFILT_LIO); 548 if (error) 549 return error; 550 async_io_version = 0; 551 aio_swake = NULL; 552 taskqueue_free(taskqueue_aiod_bio); 553 delete_unrhdr(aiod_unr); 554 uma_zdestroy(kaio_zone); 555 uma_zdestroy(aiop_zone); 556 uma_zdestroy(aiocb_zone); 557 uma_zdestroy(aiol_zone); 558 uma_zdestroy(aiolio_zone); 559 EVENTHANDLER_DEREGISTER(process_exit, exit_tag); 560 EVENTHANDLER_DEREGISTER(process_exec, exec_tag); 561 mtx_destroy(&aio_job_mtx); 562 mtx_destroy(&aio_sock_mtx); 563 sema_destroy(&aio_newproc_sem); 564 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1); 565 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1); 566 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1); 567 return (0); 568 } 569 570 /* 571 * Init the per-process aioinfo structure. The aioinfo limits are set 572 * per-process for user limit (resource) management. 573 */ 574 void 575 aio_init_aioinfo(struct proc *p) 576 { 577 struct kaioinfo *ki; 578 579 ki = uma_zalloc(kaio_zone, M_WAITOK); 580 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF); 581 ki->kaio_flags = 0; 582 ki->kaio_maxactive_count = max_aio_per_proc; 583 ki->kaio_active_count = 0; 584 ki->kaio_qallowed_count = max_aio_queue_per_proc; 585 ki->kaio_count = 0; 586 ki->kaio_ballowed_count = max_buf_aio; 587 ki->kaio_buffer_count = 0; 588 TAILQ_INIT(&ki->kaio_all); 589 TAILQ_INIT(&ki->kaio_done); 590 TAILQ_INIT(&ki->kaio_jobqueue); 591 TAILQ_INIT(&ki->kaio_bufqueue); 592 TAILQ_INIT(&ki->kaio_liojoblist); 593 TAILQ_INIT(&ki->kaio_sockqueue); 594 TAILQ_INIT(&ki->kaio_syncqueue); 595 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 596 PROC_LOCK(p); 597 if (p->p_aioinfo == NULL) { 598 p->p_aioinfo = ki; 599 PROC_UNLOCK(p); 600 } else { 601 PROC_UNLOCK(p); 602 mtx_destroy(&ki->kaio_mtx); 603 uma_zfree(kaio_zone, ki); 604 } 605 606 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs)) 607 aio_newproc(NULL); 608 } 609 610 static int 611 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi) 612 { 613 struct thread *td; 614 int error; 615 616 error = sigev_findtd(p, sigev, &td); 617 if (error) 618 return (error); 619 if (!KSI_ONQ(ksi)) { 620 ksiginfo_set_sigev(ksi, sigev); 621 ksi->ksi_code = SI_ASYNCIO; 622 ksi->ksi_flags |= KSI_EXT | KSI_INS; 623 tdsendsignal(p, td, ksi->ksi_signo, ksi); 624 } 625 PROC_UNLOCK(p); 626 return (error); 627 } 628 629 /* 630 * Free a job entry. Wait for completion if it is currently active, but don't 631 * delay forever. If we delay, we return a flag that says that we have to 632 * restart the queue scan. 633 */ 634 static int 635 aio_free_entry(struct aiocblist *aiocbe) 636 { 637 struct kaioinfo *ki; 638 struct aioliojob *lj; 639 struct proc *p; 640 641 p = aiocbe->userproc; 642 MPASS(curproc == p); 643 ki = p->p_aioinfo; 644 MPASS(ki != NULL); 645 646 AIO_LOCK_ASSERT(ki, MA_OWNED); 647 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED); 648 649 atomic_subtract_int(&num_queue_count, 1); 650 651 ki->kaio_count--; 652 MPASS(ki->kaio_count >= 0); 653 654 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist); 655 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist); 656 657 lj = aiocbe->lio; 658 if (lj) { 659 lj->lioj_count--; 660 lj->lioj_finished_count--; 661 662 if (lj->lioj_count == 0) { 663 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 664 /* lio is going away, we need to destroy any knotes */ 665 knlist_delete(&lj->klist, curthread, 1); 666 PROC_LOCK(p); 667 sigqueue_take(&lj->lioj_ksi); 668 PROC_UNLOCK(p); 669 uma_zfree(aiolio_zone, lj); 670 } 671 } 672 673 /* aiocbe is going away, we need to destroy any knotes */ 674 knlist_delete(&aiocbe->klist, curthread, 1); 675 PROC_LOCK(p); 676 sigqueue_take(&aiocbe->ksi); 677 PROC_UNLOCK(p); 678 679 MPASS(aiocbe->bp == NULL); 680 aiocbe->jobstate = JOBST_NULL; 681 AIO_UNLOCK(ki); 682 683 /* 684 * The thread argument here is used to find the owning process 685 * and is also passed to fo_close() which may pass it to various 686 * places such as devsw close() routines. Because of that, we 687 * need a thread pointer from the process owning the job that is 688 * persistent and won't disappear out from under us or move to 689 * another process. 690 * 691 * Currently, all the callers of this function call it to remove 692 * an aiocblist from the current process' job list either via a 693 * syscall or due to the current process calling exit() or 694 * execve(). Thus, we know that p == curproc. We also know that 695 * curthread can't exit since we are curthread. 696 * 697 * Therefore, we use curthread as the thread to pass to 698 * knlist_delete(). This does mean that it is possible for the 699 * thread pointer at close time to differ from the thread pointer 700 * at open time, but this is already true of file descriptors in 701 * a multithreaded process. 702 */ 703 fdrop(aiocbe->fd_file, curthread); 704 crfree(aiocbe->cred); 705 uma_zfree(aiocb_zone, aiocbe); 706 AIO_LOCK(ki); 707 708 return (0); 709 } 710 711 static void 712 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused) 713 { 714 aio_proc_rundown(arg, p); 715 } 716 717 /* 718 * Rundown the jobs for a given process. 719 */ 720 static void 721 aio_proc_rundown(void *arg, struct proc *p) 722 { 723 struct kaioinfo *ki; 724 struct aioliojob *lj; 725 struct aiocblist *cbe, *cbn; 726 struct file *fp; 727 struct socket *so; 728 int remove; 729 730 KASSERT(curthread->td_proc == p, 731 ("%s: called on non-curproc", __func__)); 732 ki = p->p_aioinfo; 733 if (ki == NULL) 734 return; 735 736 AIO_LOCK(ki); 737 ki->kaio_flags |= KAIO_RUNDOWN; 738 739 restart: 740 741 /* 742 * Try to cancel all pending requests. This code simulates 743 * aio_cancel on all pending I/O requests. 744 */ 745 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) { 746 remove = 0; 747 mtx_lock(&aio_job_mtx); 748 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 749 TAILQ_REMOVE(&aio_jobs, cbe, list); 750 remove = 1; 751 } else if (cbe->jobstate == JOBST_JOBQSOCK) { 752 fp = cbe->fd_file; 753 MPASS(fp->f_type == DTYPE_SOCKET); 754 so = fp->f_data; 755 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 756 remove = 1; 757 } else if (cbe->jobstate == JOBST_JOBQSYNC) { 758 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list); 759 remove = 1; 760 } 761 mtx_unlock(&aio_job_mtx); 762 763 if (remove) { 764 cbe->jobstate = JOBST_JOBFINISHED; 765 cbe->uaiocb._aiocb_private.status = -1; 766 cbe->uaiocb._aiocb_private.error = ECANCELED; 767 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 768 aio_bio_done_notify(p, cbe, DONE_QUEUE); 769 } 770 } 771 772 /* Wait for all running I/O to be finished */ 773 if (TAILQ_FIRST(&ki->kaio_bufqueue) || 774 TAILQ_FIRST(&ki->kaio_jobqueue)) { 775 ki->kaio_flags |= KAIO_WAKEUP; 776 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 777 goto restart; 778 } 779 780 /* Free all completed I/O requests. */ 781 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL) 782 aio_free_entry(cbe); 783 784 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 785 if (lj->lioj_count == 0) { 786 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 787 knlist_delete(&lj->klist, curthread, 1); 788 PROC_LOCK(p); 789 sigqueue_take(&lj->lioj_ksi); 790 PROC_UNLOCK(p); 791 uma_zfree(aiolio_zone, lj); 792 } else { 793 panic("LIO job not cleaned up: C:%d, FC:%d\n", 794 lj->lioj_count, lj->lioj_finished_count); 795 } 796 } 797 AIO_UNLOCK(ki); 798 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task); 799 mtx_destroy(&ki->kaio_mtx); 800 uma_zfree(kaio_zone, ki); 801 p->p_aioinfo = NULL; 802 } 803 804 /* 805 * Select a job to run (called by an AIO daemon). 806 */ 807 static struct aiocblist * 808 aio_selectjob(struct aiothreadlist *aiop) 809 { 810 struct aiocblist *aiocbe; 811 struct kaioinfo *ki; 812 struct proc *userp; 813 814 mtx_assert(&aio_job_mtx, MA_OWNED); 815 TAILQ_FOREACH(aiocbe, &aio_jobs, list) { 816 userp = aiocbe->userproc; 817 ki = userp->p_aioinfo; 818 819 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 820 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 821 /* Account for currently active jobs. */ 822 ki->kaio_active_count++; 823 aiocbe->jobstate = JOBST_JOBRUNNING; 824 break; 825 } 826 } 827 return (aiocbe); 828 } 829 830 /* 831 * Move all data to a permanent storage device, this code 832 * simulates fsync syscall. 833 */ 834 static int 835 aio_fsync_vnode(struct thread *td, struct vnode *vp) 836 { 837 struct mount *mp; 838 int vfslocked; 839 int error; 840 841 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 842 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 843 goto drop; 844 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 845 if (vp->v_object != NULL) { 846 VM_OBJECT_LOCK(vp->v_object); 847 vm_object_page_clean(vp->v_object, 0, 0, 0); 848 VM_OBJECT_UNLOCK(vp->v_object); 849 } 850 error = VOP_FSYNC(vp, MNT_WAIT, td); 851 852 VOP_UNLOCK(vp, 0); 853 vn_finished_write(mp); 854 drop: 855 VFS_UNLOCK_GIANT(vfslocked); 856 return (error); 857 } 858 859 /* 860 * The AIO processing activity. This is the code that does the I/O request for 861 * the non-physio version of the operations. The normal vn operations are used, 862 * and this code should work in all instances for every type of file, including 863 * pipes, sockets, fifos, and regular files. 864 * 865 * XXX I don't think it works well for socket, pipe, and fifo. 866 */ 867 static void 868 aio_process(struct aiocblist *aiocbe) 869 { 870 struct ucred *td_savedcred; 871 struct thread *td; 872 struct aiocb *cb; 873 struct file *fp; 874 struct socket *so; 875 struct uio auio; 876 struct iovec aiov; 877 int cnt; 878 int error; 879 int oublock_st, oublock_end; 880 int inblock_st, inblock_end; 881 882 td = curthread; 883 td_savedcred = td->td_ucred; 884 td->td_ucred = aiocbe->cred; 885 cb = &aiocbe->uaiocb; 886 fp = aiocbe->fd_file; 887 888 if (cb->aio_lio_opcode == LIO_SYNC) { 889 error = 0; 890 cnt = 0; 891 if (fp->f_vnode != NULL) 892 error = aio_fsync_vnode(td, fp->f_vnode); 893 cb->_aiocb_private.error = error; 894 cb->_aiocb_private.status = 0; 895 td->td_ucred = td_savedcred; 896 return; 897 } 898 899 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 900 aiov.iov_len = cb->aio_nbytes; 901 902 auio.uio_iov = &aiov; 903 auio.uio_iovcnt = 1; 904 auio.uio_offset = cb->aio_offset; 905 auio.uio_resid = cb->aio_nbytes; 906 cnt = cb->aio_nbytes; 907 auio.uio_segflg = UIO_USERSPACE; 908 auio.uio_td = td; 909 910 inblock_st = td->td_ru.ru_inblock; 911 oublock_st = td->td_ru.ru_oublock; 912 /* 913 * aio_aqueue() acquires a reference to the file that is 914 * released in aio_free_entry(). 915 */ 916 if (cb->aio_lio_opcode == LIO_READ) { 917 auio.uio_rw = UIO_READ; 918 if (auio.uio_resid == 0) 919 error = 0; 920 else 921 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td); 922 } else { 923 if (fp->f_type == DTYPE_VNODE) 924 bwillwrite(); 925 auio.uio_rw = UIO_WRITE; 926 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td); 927 } 928 inblock_end = td->td_ru.ru_inblock; 929 oublock_end = td->td_ru.ru_oublock; 930 931 aiocbe->inputcharge = inblock_end - inblock_st; 932 aiocbe->outputcharge = oublock_end - oublock_st; 933 934 if ((error) && (auio.uio_resid != cnt)) { 935 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 936 error = 0; 937 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) { 938 int sigpipe = 1; 939 if (fp->f_type == DTYPE_SOCKET) { 940 so = fp->f_data; 941 if (so->so_options & SO_NOSIGPIPE) 942 sigpipe = 0; 943 } 944 if (sigpipe) { 945 PROC_LOCK(aiocbe->userproc); 946 kern_psignal(aiocbe->userproc, SIGPIPE); 947 PROC_UNLOCK(aiocbe->userproc); 948 } 949 } 950 } 951 952 cnt -= auio.uio_resid; 953 cb->_aiocb_private.error = error; 954 cb->_aiocb_private.status = cnt; 955 td->td_ucred = td_savedcred; 956 } 957 958 static void 959 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type) 960 { 961 struct aioliojob *lj; 962 struct kaioinfo *ki; 963 struct aiocblist *scb, *scbn; 964 int lj_done; 965 966 ki = userp->p_aioinfo; 967 AIO_LOCK_ASSERT(ki, MA_OWNED); 968 lj = aiocbe->lio; 969 lj_done = 0; 970 if (lj) { 971 lj->lioj_finished_count++; 972 if (lj->lioj_count == lj->lioj_finished_count) 973 lj_done = 1; 974 } 975 if (type == DONE_QUEUE) { 976 aiocbe->jobflags |= AIOCBLIST_DONE; 977 } else { 978 aiocbe->jobflags |= AIOCBLIST_BUFDONE; 979 } 980 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist); 981 aiocbe->jobstate = JOBST_JOBFINISHED; 982 983 if (ki->kaio_flags & KAIO_RUNDOWN) 984 goto notification_done; 985 986 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 987 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 988 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi); 989 990 KNOTE_LOCKED(&aiocbe->klist, 1); 991 992 if (lj_done) { 993 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 994 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 995 KNOTE_LOCKED(&lj->klist, 1); 996 } 997 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 998 == LIOJ_SIGNAL 999 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 1000 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 1001 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi); 1002 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 1003 } 1004 } 1005 1006 notification_done: 1007 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) { 1008 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) { 1009 if (aiocbe->fd_file == scb->fd_file && 1010 aiocbe->seqno < scb->seqno) { 1011 if (--scb->pending == 0) { 1012 mtx_lock(&aio_job_mtx); 1013 scb->jobstate = JOBST_JOBQGLOBAL; 1014 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list); 1015 TAILQ_INSERT_TAIL(&aio_jobs, scb, list); 1016 aio_kick_nowait(userp); 1017 mtx_unlock(&aio_job_mtx); 1018 } 1019 } 1020 } 1021 } 1022 if (ki->kaio_flags & KAIO_WAKEUP) { 1023 ki->kaio_flags &= ~KAIO_WAKEUP; 1024 wakeup(&userp->p_aioinfo); 1025 } 1026 } 1027 1028 /* 1029 * The AIO daemon, most of the actual work is done in aio_process, 1030 * but the setup (and address space mgmt) is done in this routine. 1031 */ 1032 static void 1033 aio_daemon(void *_id) 1034 { 1035 struct aiocblist *aiocbe; 1036 struct aiothreadlist *aiop; 1037 struct kaioinfo *ki; 1038 struct proc *curcp, *mycp, *userp; 1039 struct vmspace *myvm, *tmpvm; 1040 struct thread *td = curthread; 1041 int id = (intptr_t)_id; 1042 1043 /* 1044 * Local copies of curproc (cp) and vmspace (myvm) 1045 */ 1046 mycp = td->td_proc; 1047 myvm = mycp->p_vmspace; 1048 1049 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp")); 1050 1051 /* 1052 * Allocate and ready the aio control info. There is one aiop structure 1053 * per daemon. 1054 */ 1055 aiop = uma_zalloc(aiop_zone, M_WAITOK); 1056 aiop->aiothread = td; 1057 aiop->aiothreadflags = 0; 1058 1059 /* The daemon resides in its own pgrp. */ 1060 sys_setsid(td, NULL); 1061 1062 /* 1063 * Wakeup parent process. (Parent sleeps to keep from blasting away 1064 * and creating too many daemons.) 1065 */ 1066 sema_post(&aio_newproc_sem); 1067 1068 mtx_lock(&aio_job_mtx); 1069 for (;;) { 1070 /* 1071 * curcp is the current daemon process context. 1072 * userp is the current user process context. 1073 */ 1074 curcp = mycp; 1075 1076 /* 1077 * Take daemon off of free queue 1078 */ 1079 if (aiop->aiothreadflags & AIOP_FREE) { 1080 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1081 aiop->aiothreadflags &= ~AIOP_FREE; 1082 } 1083 1084 /* 1085 * Check for jobs. 1086 */ 1087 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 1088 mtx_unlock(&aio_job_mtx); 1089 userp = aiocbe->userproc; 1090 1091 /* 1092 * Connect to process address space for user program. 1093 */ 1094 if (userp != curcp) { 1095 /* 1096 * Save the current address space that we are 1097 * connected to. 1098 */ 1099 tmpvm = mycp->p_vmspace; 1100 1101 /* 1102 * Point to the new user address space, and 1103 * refer to it. 1104 */ 1105 mycp->p_vmspace = userp->p_vmspace; 1106 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1); 1107 1108 /* Activate the new mapping. */ 1109 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 1110 1111 /* 1112 * If the old address space wasn't the daemons 1113 * own address space, then we need to remove the 1114 * daemon's reference from the other process 1115 * that it was acting on behalf of. 1116 */ 1117 if (tmpvm != myvm) { 1118 vmspace_free(tmpvm); 1119 } 1120 curcp = userp; 1121 } 1122 1123 ki = userp->p_aioinfo; 1124 1125 /* Do the I/O function. */ 1126 aio_process(aiocbe); 1127 1128 mtx_lock(&aio_job_mtx); 1129 /* Decrement the active job count. */ 1130 ki->kaio_active_count--; 1131 mtx_unlock(&aio_job_mtx); 1132 1133 AIO_LOCK(ki); 1134 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 1135 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE); 1136 AIO_UNLOCK(ki); 1137 1138 mtx_lock(&aio_job_mtx); 1139 } 1140 1141 /* 1142 * Disconnect from user address space. 1143 */ 1144 if (curcp != mycp) { 1145 1146 mtx_unlock(&aio_job_mtx); 1147 1148 /* Get the user address space to disconnect from. */ 1149 tmpvm = mycp->p_vmspace; 1150 1151 /* Get original address space for daemon. */ 1152 mycp->p_vmspace = myvm; 1153 1154 /* Activate the daemon's address space. */ 1155 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 1156 #ifdef DIAGNOSTIC 1157 if (tmpvm == myvm) { 1158 printf("AIOD: vmspace problem -- %d\n", 1159 mycp->p_pid); 1160 } 1161 #endif 1162 /* Remove our vmspace reference. */ 1163 vmspace_free(tmpvm); 1164 1165 curcp = mycp; 1166 1167 mtx_lock(&aio_job_mtx); 1168 /* 1169 * We have to restart to avoid race, we only sleep if 1170 * no job can be selected, that should be 1171 * curcp == mycp. 1172 */ 1173 continue; 1174 } 1175 1176 mtx_assert(&aio_job_mtx, MA_OWNED); 1177 1178 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1179 aiop->aiothreadflags |= AIOP_FREE; 1180 1181 /* 1182 * If daemon is inactive for a long time, allow it to exit, 1183 * thereby freeing resources. 1184 */ 1185 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy", 1186 aiod_lifetime)) { 1187 if (TAILQ_EMPTY(&aio_jobs)) { 1188 if ((aiop->aiothreadflags & AIOP_FREE) && 1189 (num_aio_procs > target_aio_procs)) { 1190 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1191 num_aio_procs--; 1192 mtx_unlock(&aio_job_mtx); 1193 uma_zfree(aiop_zone, aiop); 1194 free_unr(aiod_unr, id); 1195 #ifdef DIAGNOSTIC 1196 if (mycp->p_vmspace->vm_refcnt <= 1) { 1197 printf("AIOD: bad vm refcnt for" 1198 " exiting daemon: %d\n", 1199 mycp->p_vmspace->vm_refcnt); 1200 } 1201 #endif 1202 kproc_exit(0); 1203 } 1204 } 1205 } 1206 } 1207 mtx_unlock(&aio_job_mtx); 1208 panic("shouldn't be here\n"); 1209 } 1210 1211 /* 1212 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1213 * AIO daemon modifies its environment itself. 1214 */ 1215 static int 1216 aio_newproc(int *start) 1217 { 1218 int error; 1219 struct proc *p; 1220 int id; 1221 1222 id = alloc_unr(aiod_unr); 1223 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p, 1224 RFNOWAIT, 0, "aiod%d", id); 1225 if (error == 0) { 1226 /* 1227 * Wait until daemon is started. 1228 */ 1229 sema_wait(&aio_newproc_sem); 1230 mtx_lock(&aio_job_mtx); 1231 num_aio_procs++; 1232 if (start != NULL) 1233 (*start)--; 1234 mtx_unlock(&aio_job_mtx); 1235 } else { 1236 free_unr(aiod_unr, id); 1237 } 1238 return (error); 1239 } 1240 1241 /* 1242 * Try the high-performance, low-overhead physio method for eligible 1243 * VCHR devices. This method doesn't use an aio helper thread, and 1244 * thus has very low overhead. 1245 * 1246 * Assumes that the caller, aio_aqueue(), has incremented the file 1247 * structure's reference count, preventing its deallocation for the 1248 * duration of this call. 1249 */ 1250 static int 1251 aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 1252 { 1253 struct aiocb *cb; 1254 struct file *fp; 1255 struct buf *bp; 1256 struct vnode *vp; 1257 struct kaioinfo *ki; 1258 struct aioliojob *lj; 1259 int error; 1260 1261 cb = &aiocbe->uaiocb; 1262 fp = aiocbe->fd_file; 1263 1264 if (fp->f_type != DTYPE_VNODE) 1265 return (-1); 1266 1267 vp = fp->f_vnode; 1268 1269 /* 1270 * If its not a disk, we don't want to return a positive error. 1271 * It causes the aio code to not fall through to try the thread 1272 * way when you're talking to a regular file. 1273 */ 1274 if (!vn_isdisk(vp, &error)) { 1275 if (error == ENOTBLK) 1276 return (-1); 1277 else 1278 return (error); 1279 } 1280 1281 if (vp->v_bufobj.bo_bsize == 0) 1282 return (-1); 1283 1284 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize) 1285 return (-1); 1286 1287 if (cb->aio_nbytes > vp->v_rdev->si_iosize_max) 1288 return (-1); 1289 1290 if (cb->aio_nbytes > 1291 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 1292 return (-1); 1293 1294 ki = p->p_aioinfo; 1295 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 1296 return (-1); 1297 1298 /* Create and build a buffer header for a transfer. */ 1299 bp = (struct buf *)getpbuf(NULL); 1300 BUF_KERNPROC(bp); 1301 1302 AIO_LOCK(ki); 1303 ki->kaio_count++; 1304 ki->kaio_buffer_count++; 1305 lj = aiocbe->lio; 1306 if (lj) 1307 lj->lioj_count++; 1308 AIO_UNLOCK(ki); 1309 1310 /* 1311 * Get a copy of the kva from the physical buffer. 1312 */ 1313 error = 0; 1314 1315 bp->b_bcount = cb->aio_nbytes; 1316 bp->b_bufsize = cb->aio_nbytes; 1317 bp->b_iodone = aio_physwakeup; 1318 bp->b_saveaddr = bp->b_data; 1319 bp->b_data = (void *)(uintptr_t)cb->aio_buf; 1320 bp->b_offset = cb->aio_offset; 1321 bp->b_iooffset = cb->aio_offset; 1322 bp->b_blkno = btodb(cb->aio_offset); 1323 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ; 1324 1325 /* 1326 * Bring buffer into kernel space. 1327 */ 1328 if (vmapbuf(bp) < 0) { 1329 error = EFAULT; 1330 goto doerror; 1331 } 1332 1333 AIO_LOCK(ki); 1334 aiocbe->bp = bp; 1335 bp->b_caller1 = (void *)aiocbe; 1336 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 1337 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1338 aiocbe->jobstate = JOBST_JOBQBUF; 1339 cb->_aiocb_private.status = cb->aio_nbytes; 1340 AIO_UNLOCK(ki); 1341 1342 atomic_add_int(&num_queue_count, 1); 1343 atomic_add_int(&num_buf_aio, 1); 1344 1345 bp->b_error = 0; 1346 1347 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe); 1348 1349 /* Perform transfer. */ 1350 dev_strategy(vp->v_rdev, bp); 1351 return (0); 1352 1353 doerror: 1354 AIO_LOCK(ki); 1355 ki->kaio_count--; 1356 ki->kaio_buffer_count--; 1357 if (lj) 1358 lj->lioj_count--; 1359 aiocbe->bp = NULL; 1360 AIO_UNLOCK(ki); 1361 relpbuf(bp, NULL); 1362 return (error); 1363 } 1364 1365 /* 1366 * Wake up aio requests that may be serviceable now. 1367 */ 1368 static void 1369 aio_swake_cb(struct socket *so, struct sockbuf *sb) 1370 { 1371 struct aiocblist *cb, *cbn; 1372 int opcode; 1373 1374 SOCKBUF_LOCK_ASSERT(sb); 1375 if (sb == &so->so_snd) 1376 opcode = LIO_WRITE; 1377 else 1378 opcode = LIO_READ; 1379 1380 sb->sb_flags &= ~SB_AIO; 1381 mtx_lock(&aio_job_mtx); 1382 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) { 1383 if (opcode == cb->uaiocb.aio_lio_opcode) { 1384 if (cb->jobstate != JOBST_JOBQSOCK) 1385 panic("invalid queue value"); 1386 /* XXX 1387 * We don't have actual sockets backend yet, 1388 * so we simply move the requests to the generic 1389 * file I/O backend. 1390 */ 1391 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1392 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1393 aio_kick_nowait(cb->userproc); 1394 } 1395 } 1396 mtx_unlock(&aio_job_mtx); 1397 } 1398 1399 static int 1400 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig) 1401 { 1402 1403 /* 1404 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 1405 * supported by AIO with the old sigevent structure. 1406 */ 1407 nsig->sigev_notify = osig->sigev_notify; 1408 switch (nsig->sigev_notify) { 1409 case SIGEV_NONE: 1410 break; 1411 case SIGEV_SIGNAL: 1412 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 1413 break; 1414 case SIGEV_KEVENT: 1415 nsig->sigev_notify_kqueue = 1416 osig->__sigev_u.__sigev_notify_kqueue; 1417 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr; 1418 break; 1419 default: 1420 return (EINVAL); 1421 } 1422 return (0); 1423 } 1424 1425 static int 1426 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob) 1427 { 1428 struct oaiocb *ojob; 1429 int error; 1430 1431 bzero(kjob, sizeof(struct aiocb)); 1432 error = copyin(ujob, kjob, sizeof(struct oaiocb)); 1433 if (error) 1434 return (error); 1435 ojob = (struct oaiocb *)kjob; 1436 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent)); 1437 } 1438 1439 static int 1440 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob) 1441 { 1442 1443 return (copyin(ujob, kjob, sizeof(struct aiocb))); 1444 } 1445 1446 static long 1447 aiocb_fetch_status(struct aiocb *ujob) 1448 { 1449 1450 return (fuword(&ujob->_aiocb_private.status)); 1451 } 1452 1453 static long 1454 aiocb_fetch_error(struct aiocb *ujob) 1455 { 1456 1457 return (fuword(&ujob->_aiocb_private.error)); 1458 } 1459 1460 static int 1461 aiocb_store_status(struct aiocb *ujob, long status) 1462 { 1463 1464 return (suword(&ujob->_aiocb_private.status, status)); 1465 } 1466 1467 static int 1468 aiocb_store_error(struct aiocb *ujob, long error) 1469 { 1470 1471 return (suword(&ujob->_aiocb_private.error, error)); 1472 } 1473 1474 static int 1475 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref) 1476 { 1477 1478 return (suword(&ujob->_aiocb_private.kernelinfo, jobref)); 1479 } 1480 1481 static int 1482 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 1483 { 1484 1485 return (suword(ujobp, (long)ujob)); 1486 } 1487 1488 static struct aiocb_ops aiocb_ops = { 1489 .copyin = aiocb_copyin, 1490 .fetch_status = aiocb_fetch_status, 1491 .fetch_error = aiocb_fetch_error, 1492 .store_status = aiocb_store_status, 1493 .store_error = aiocb_store_error, 1494 .store_kernelinfo = aiocb_store_kernelinfo, 1495 .store_aiocb = aiocb_store_aiocb, 1496 }; 1497 1498 static struct aiocb_ops aiocb_ops_osigevent = { 1499 .copyin = aiocb_copyin_old_sigevent, 1500 .fetch_status = aiocb_fetch_status, 1501 .fetch_error = aiocb_fetch_error, 1502 .store_status = aiocb_store_status, 1503 .store_error = aiocb_store_error, 1504 .store_kernelinfo = aiocb_store_kernelinfo, 1505 .store_aiocb = aiocb_store_aiocb, 1506 }; 1507 1508 /* 1509 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1510 * technique is done in this code. 1511 */ 1512 int 1513 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj, 1514 int type, struct aiocb_ops *ops) 1515 { 1516 struct proc *p = td->td_proc; 1517 struct file *fp; 1518 struct socket *so; 1519 struct aiocblist *aiocbe, *cb; 1520 struct kaioinfo *ki; 1521 struct kevent kev; 1522 struct sockbuf *sb; 1523 int opcode; 1524 int error; 1525 int fd, kqfd; 1526 int jid; 1527 u_short evflags; 1528 1529 if (p->p_aioinfo == NULL) 1530 aio_init_aioinfo(p); 1531 1532 ki = p->p_aioinfo; 1533 1534 ops->store_status(job, -1); 1535 ops->store_error(job, 0); 1536 ops->store_kernelinfo(job, -1); 1537 1538 if (num_queue_count >= max_queue_count || 1539 ki->kaio_count >= ki->kaio_qallowed_count) { 1540 ops->store_error(job, EAGAIN); 1541 return (EAGAIN); 1542 } 1543 1544 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1545 aiocbe->inputcharge = 0; 1546 aiocbe->outputcharge = 0; 1547 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki)); 1548 1549 error = ops->copyin(job, &aiocbe->uaiocb); 1550 if (error) { 1551 ops->store_error(job, error); 1552 uma_zfree(aiocb_zone, aiocbe); 1553 return (error); 1554 } 1555 1556 /* XXX: aio_nbytes is later casted to signed types. */ 1557 if (aiocbe->uaiocb.aio_nbytes > INT_MAX) { 1558 uma_zfree(aiocb_zone, aiocbe); 1559 return (EINVAL); 1560 } 1561 1562 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1563 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1564 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1565 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1566 ops->store_error(job, EINVAL); 1567 uma_zfree(aiocb_zone, aiocbe); 1568 return (EINVAL); 1569 } 1570 1571 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1572 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1573 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1574 uma_zfree(aiocb_zone, aiocbe); 1575 return (EINVAL); 1576 } 1577 1578 ksiginfo_init(&aiocbe->ksi); 1579 1580 /* Save userspace address of the job info. */ 1581 aiocbe->uuaiocb = job; 1582 1583 /* Get the opcode. */ 1584 if (type != LIO_NOP) 1585 aiocbe->uaiocb.aio_lio_opcode = type; 1586 opcode = aiocbe->uaiocb.aio_lio_opcode; 1587 1588 /* 1589 * Validate the opcode and fetch the file object for the specified 1590 * file descriptor. 1591 * 1592 * XXXRW: Moved the opcode validation up here so that we don't 1593 * retrieve a file descriptor without knowing what the capabiltity 1594 * should be. 1595 */ 1596 fd = aiocbe->uaiocb.aio_fildes; 1597 switch (opcode) { 1598 case LIO_WRITE: 1599 error = fget_write(td, fd, CAP_WRITE | CAP_SEEK, &fp); 1600 break; 1601 case LIO_READ: 1602 error = fget_read(td, fd, CAP_READ | CAP_SEEK, &fp); 1603 break; 1604 case LIO_SYNC: 1605 error = fget(td, fd, CAP_FSYNC, &fp); 1606 break; 1607 case LIO_NOP: 1608 error = fget(td, fd, 0, &fp); 1609 break; 1610 default: 1611 error = EINVAL; 1612 } 1613 if (error) { 1614 uma_zfree(aiocb_zone, aiocbe); 1615 ops->store_error(job, error); 1616 return (error); 1617 } 1618 1619 if (opcode == LIO_SYNC && fp->f_vnode == NULL) { 1620 error = EINVAL; 1621 goto aqueue_fail; 1622 } 1623 1624 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) { 1625 error = EINVAL; 1626 goto aqueue_fail; 1627 } 1628 1629 aiocbe->fd_file = fp; 1630 1631 mtx_lock(&aio_job_mtx); 1632 jid = jobrefid++; 1633 aiocbe->seqno = jobseqno++; 1634 mtx_unlock(&aio_job_mtx); 1635 error = ops->store_kernelinfo(job, jid); 1636 if (error) { 1637 error = EINVAL; 1638 goto aqueue_fail; 1639 } 1640 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid; 1641 1642 if (opcode == LIO_NOP) { 1643 fdrop(fp, td); 1644 uma_zfree(aiocb_zone, aiocbe); 1645 return (0); 1646 } 1647 1648 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT) 1649 goto no_kqueue; 1650 evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags; 1651 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) { 1652 error = EINVAL; 1653 goto aqueue_fail; 1654 } 1655 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1656 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1657 kev.filter = EVFILT_AIO; 1658 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags; 1659 kev.data = (intptr_t)aiocbe; 1660 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1661 error = kqfd_register(kqfd, &kev, td, 1); 1662 aqueue_fail: 1663 if (error) { 1664 fdrop(fp, td); 1665 uma_zfree(aiocb_zone, aiocbe); 1666 ops->store_error(job, error); 1667 goto done; 1668 } 1669 no_kqueue: 1670 1671 ops->store_error(job, EINPROGRESS); 1672 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1673 aiocbe->userproc = p; 1674 aiocbe->cred = crhold(td->td_ucred); 1675 aiocbe->jobflags = 0; 1676 aiocbe->lio = lj; 1677 1678 if (opcode == LIO_SYNC) 1679 goto queueit; 1680 1681 if (fp->f_type == DTYPE_SOCKET) { 1682 /* 1683 * Alternate queueing for socket ops: Reach down into the 1684 * descriptor to get the socket data. Then check to see if the 1685 * socket is ready to be read or written (based on the requested 1686 * operation). 1687 * 1688 * If it is not ready for io, then queue the aiocbe on the 1689 * socket, and set the flags so we get a call when sbnotify() 1690 * happens. 1691 * 1692 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock 1693 * and unlock the snd sockbuf for no reason. 1694 */ 1695 so = fp->f_data; 1696 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd; 1697 SOCKBUF_LOCK(sb); 1698 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1699 LIO_WRITE) && (!sowriteable(so)))) { 1700 sb->sb_flags |= SB_AIO; 1701 1702 mtx_lock(&aio_job_mtx); 1703 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1704 mtx_unlock(&aio_job_mtx); 1705 1706 AIO_LOCK(ki); 1707 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1708 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1709 aiocbe->jobstate = JOBST_JOBQSOCK; 1710 ki->kaio_count++; 1711 if (lj) 1712 lj->lioj_count++; 1713 AIO_UNLOCK(ki); 1714 SOCKBUF_UNLOCK(sb); 1715 atomic_add_int(&num_queue_count, 1); 1716 error = 0; 1717 goto done; 1718 } 1719 SOCKBUF_UNLOCK(sb); 1720 } 1721 1722 if ((error = aio_qphysio(p, aiocbe)) == 0) 1723 goto done; 1724 #if 0 1725 if (error > 0) { 1726 aiocbe->uaiocb._aiocb_private.error = error; 1727 ops->store_error(job, error); 1728 goto done; 1729 } 1730 #endif 1731 queueit: 1732 /* No buffer for daemon I/O. */ 1733 aiocbe->bp = NULL; 1734 atomic_add_int(&num_queue_count, 1); 1735 1736 AIO_LOCK(ki); 1737 ki->kaio_count++; 1738 if (lj) 1739 lj->lioj_count++; 1740 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1741 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1742 if (opcode == LIO_SYNC) { 1743 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) { 1744 if (cb->fd_file == aiocbe->fd_file && 1745 cb->uaiocb.aio_lio_opcode != LIO_SYNC && 1746 cb->seqno < aiocbe->seqno) { 1747 cb->jobflags |= AIOCBLIST_CHECKSYNC; 1748 aiocbe->pending++; 1749 } 1750 } 1751 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) { 1752 if (cb->fd_file == aiocbe->fd_file && 1753 cb->uaiocb.aio_lio_opcode != LIO_SYNC && 1754 cb->seqno < aiocbe->seqno) { 1755 cb->jobflags |= AIOCBLIST_CHECKSYNC; 1756 aiocbe->pending++; 1757 } 1758 } 1759 if (aiocbe->pending != 0) { 1760 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list); 1761 aiocbe->jobstate = JOBST_JOBQSYNC; 1762 AIO_UNLOCK(ki); 1763 goto done; 1764 } 1765 } 1766 mtx_lock(&aio_job_mtx); 1767 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1768 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1769 aio_kick_nowait(p); 1770 mtx_unlock(&aio_job_mtx); 1771 AIO_UNLOCK(ki); 1772 error = 0; 1773 done: 1774 return (error); 1775 } 1776 1777 static void 1778 aio_kick_nowait(struct proc *userp) 1779 { 1780 struct kaioinfo *ki = userp->p_aioinfo; 1781 struct aiothreadlist *aiop; 1782 1783 mtx_assert(&aio_job_mtx, MA_OWNED); 1784 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1785 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1786 aiop->aiothreadflags &= ~AIOP_FREE; 1787 wakeup(aiop->aiothread); 1788 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1789 ((ki->kaio_active_count + num_aio_resv_start) < 1790 ki->kaio_maxactive_count)) { 1791 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task); 1792 } 1793 } 1794 1795 static int 1796 aio_kick(struct proc *userp) 1797 { 1798 struct kaioinfo *ki = userp->p_aioinfo; 1799 struct aiothreadlist *aiop; 1800 int error, ret = 0; 1801 1802 mtx_assert(&aio_job_mtx, MA_OWNED); 1803 retryproc: 1804 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1805 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1806 aiop->aiothreadflags &= ~AIOP_FREE; 1807 wakeup(aiop->aiothread); 1808 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1809 ((ki->kaio_active_count + num_aio_resv_start) < 1810 ki->kaio_maxactive_count)) { 1811 num_aio_resv_start++; 1812 mtx_unlock(&aio_job_mtx); 1813 error = aio_newproc(&num_aio_resv_start); 1814 mtx_lock(&aio_job_mtx); 1815 if (error) { 1816 num_aio_resv_start--; 1817 goto retryproc; 1818 } 1819 } else { 1820 ret = -1; 1821 } 1822 return (ret); 1823 } 1824 1825 static void 1826 aio_kick_helper(void *context, int pending) 1827 { 1828 struct proc *userp = context; 1829 1830 mtx_lock(&aio_job_mtx); 1831 while (--pending >= 0) { 1832 if (aio_kick(userp)) 1833 break; 1834 } 1835 mtx_unlock(&aio_job_mtx); 1836 } 1837 1838 /* 1839 * Support the aio_return system call, as a side-effect, kernel resources are 1840 * released. 1841 */ 1842 static int 1843 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops) 1844 { 1845 struct proc *p = td->td_proc; 1846 struct aiocblist *cb; 1847 struct kaioinfo *ki; 1848 int status, error; 1849 1850 ki = p->p_aioinfo; 1851 if (ki == NULL) 1852 return (EINVAL); 1853 AIO_LOCK(ki); 1854 TAILQ_FOREACH(cb, &ki->kaio_done, plist) { 1855 if (cb->uuaiocb == uaiocb) 1856 break; 1857 } 1858 if (cb != NULL) { 1859 MPASS(cb->jobstate == JOBST_JOBFINISHED); 1860 status = cb->uaiocb._aiocb_private.status; 1861 error = cb->uaiocb._aiocb_private.error; 1862 td->td_retval[0] = status; 1863 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1864 td->td_ru.ru_oublock += cb->outputcharge; 1865 cb->outputcharge = 0; 1866 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1867 td->td_ru.ru_inblock += cb->inputcharge; 1868 cb->inputcharge = 0; 1869 } 1870 aio_free_entry(cb); 1871 AIO_UNLOCK(ki); 1872 ops->store_error(uaiocb, error); 1873 ops->store_status(uaiocb, status); 1874 } else { 1875 error = EINVAL; 1876 AIO_UNLOCK(ki); 1877 } 1878 return (error); 1879 } 1880 1881 int 1882 sys_aio_return(struct thread *td, struct aio_return_args *uap) 1883 { 1884 1885 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops)); 1886 } 1887 1888 /* 1889 * Allow a process to wakeup when any of the I/O requests are completed. 1890 */ 1891 static int 1892 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist, 1893 struct timespec *ts) 1894 { 1895 struct proc *p = td->td_proc; 1896 struct timeval atv; 1897 struct kaioinfo *ki; 1898 struct aiocblist *cb, *cbfirst; 1899 int error, i, timo; 1900 1901 timo = 0; 1902 if (ts) { 1903 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000) 1904 return (EINVAL); 1905 1906 TIMESPEC_TO_TIMEVAL(&atv, ts); 1907 if (itimerfix(&atv)) 1908 return (EINVAL); 1909 timo = tvtohz(&atv); 1910 } 1911 1912 ki = p->p_aioinfo; 1913 if (ki == NULL) 1914 return (EAGAIN); 1915 1916 if (njoblist == 0) 1917 return (0); 1918 1919 AIO_LOCK(ki); 1920 for (;;) { 1921 cbfirst = NULL; 1922 error = 0; 1923 TAILQ_FOREACH(cb, &ki->kaio_all, allist) { 1924 for (i = 0; i < njoblist; i++) { 1925 if (cb->uuaiocb == ujoblist[i]) { 1926 if (cbfirst == NULL) 1927 cbfirst = cb; 1928 if (cb->jobstate == JOBST_JOBFINISHED) 1929 goto RETURN; 1930 } 1931 } 1932 } 1933 /* All tasks were finished. */ 1934 if (cbfirst == NULL) 1935 break; 1936 1937 ki->kaio_flags |= KAIO_WAKEUP; 1938 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 1939 "aiospn", timo); 1940 if (error == ERESTART) 1941 error = EINTR; 1942 if (error) 1943 break; 1944 } 1945 RETURN: 1946 AIO_UNLOCK(ki); 1947 return (error); 1948 } 1949 1950 int 1951 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap) 1952 { 1953 struct timespec ts, *tsp; 1954 struct aiocb **ujoblist; 1955 int error; 1956 1957 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX) 1958 return (EINVAL); 1959 1960 if (uap->timeout) { 1961 /* Get timespec struct. */ 1962 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1963 return (error); 1964 tsp = &ts; 1965 } else 1966 tsp = NULL; 1967 1968 ujoblist = uma_zalloc(aiol_zone, M_WAITOK); 1969 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0])); 1970 if (error == 0) 1971 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 1972 uma_zfree(aiol_zone, ujoblist); 1973 return (error); 1974 } 1975 1976 /* 1977 * aio_cancel cancels any non-physio aio operations not currently in 1978 * progress. 1979 */ 1980 int 1981 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap) 1982 { 1983 struct proc *p = td->td_proc; 1984 struct kaioinfo *ki; 1985 struct aiocblist *cbe, *cbn; 1986 struct file *fp; 1987 struct socket *so; 1988 int error; 1989 int remove; 1990 int cancelled = 0; 1991 int notcancelled = 0; 1992 struct vnode *vp; 1993 1994 /* Lookup file object. */ 1995 error = fget(td, uap->fd, 0, &fp); 1996 if (error) 1997 return (error); 1998 1999 ki = p->p_aioinfo; 2000 if (ki == NULL) 2001 goto done; 2002 2003 if (fp->f_type == DTYPE_VNODE) { 2004 vp = fp->f_vnode; 2005 if (vn_isdisk(vp, &error)) { 2006 fdrop(fp, td); 2007 td->td_retval[0] = AIO_NOTCANCELED; 2008 return (0); 2009 } 2010 } 2011 2012 AIO_LOCK(ki); 2013 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) { 2014 if ((uap->fd == cbe->uaiocb.aio_fildes) && 2015 ((uap->aiocbp == NULL) || 2016 (uap->aiocbp == cbe->uuaiocb))) { 2017 remove = 0; 2018 2019 mtx_lock(&aio_job_mtx); 2020 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 2021 TAILQ_REMOVE(&aio_jobs, cbe, list); 2022 remove = 1; 2023 } else if (cbe->jobstate == JOBST_JOBQSOCK) { 2024 MPASS(fp->f_type == DTYPE_SOCKET); 2025 so = fp->f_data; 2026 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 2027 remove = 1; 2028 } else if (cbe->jobstate == JOBST_JOBQSYNC) { 2029 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list); 2030 remove = 1; 2031 } 2032 mtx_unlock(&aio_job_mtx); 2033 2034 if (remove) { 2035 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 2036 cbe->uaiocb._aiocb_private.status = -1; 2037 cbe->uaiocb._aiocb_private.error = ECANCELED; 2038 aio_bio_done_notify(p, cbe, DONE_QUEUE); 2039 cancelled++; 2040 } else { 2041 notcancelled++; 2042 } 2043 if (uap->aiocbp != NULL) 2044 break; 2045 } 2046 } 2047 AIO_UNLOCK(ki); 2048 2049 done: 2050 fdrop(fp, td); 2051 2052 if (uap->aiocbp != NULL) { 2053 if (cancelled) { 2054 td->td_retval[0] = AIO_CANCELED; 2055 return (0); 2056 } 2057 } 2058 2059 if (notcancelled) { 2060 td->td_retval[0] = AIO_NOTCANCELED; 2061 return (0); 2062 } 2063 2064 if (cancelled) { 2065 td->td_retval[0] = AIO_CANCELED; 2066 return (0); 2067 } 2068 2069 td->td_retval[0] = AIO_ALLDONE; 2070 2071 return (0); 2072 } 2073 2074 /* 2075 * aio_error is implemented in the kernel level for compatibility purposes 2076 * only. For a user mode async implementation, it would be best to do it in 2077 * a userland subroutine. 2078 */ 2079 static int 2080 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops) 2081 { 2082 struct proc *p = td->td_proc; 2083 struct aiocblist *cb; 2084 struct kaioinfo *ki; 2085 int status; 2086 2087 ki = p->p_aioinfo; 2088 if (ki == NULL) { 2089 td->td_retval[0] = EINVAL; 2090 return (0); 2091 } 2092 2093 AIO_LOCK(ki); 2094 TAILQ_FOREACH(cb, &ki->kaio_all, allist) { 2095 if (cb->uuaiocb == aiocbp) { 2096 if (cb->jobstate == JOBST_JOBFINISHED) 2097 td->td_retval[0] = 2098 cb->uaiocb._aiocb_private.error; 2099 else 2100 td->td_retval[0] = EINPROGRESS; 2101 AIO_UNLOCK(ki); 2102 return (0); 2103 } 2104 } 2105 AIO_UNLOCK(ki); 2106 2107 /* 2108 * Hack for failure of aio_aqueue. 2109 */ 2110 status = ops->fetch_status(aiocbp); 2111 if (status == -1) { 2112 td->td_retval[0] = ops->fetch_error(aiocbp); 2113 return (0); 2114 } 2115 2116 td->td_retval[0] = EINVAL; 2117 return (0); 2118 } 2119 2120 int 2121 sys_aio_error(struct thread *td, struct aio_error_args *uap) 2122 { 2123 2124 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops)); 2125 } 2126 2127 /* syscall - asynchronous read from a file (REALTIME) */ 2128 int 2129 sys_oaio_read(struct thread *td, struct oaio_read_args *uap) 2130 { 2131 2132 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2133 &aiocb_ops_osigevent)); 2134 } 2135 2136 int 2137 sys_aio_read(struct thread *td, struct aio_read_args *uap) 2138 { 2139 2140 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops)); 2141 } 2142 2143 /* syscall - asynchronous write to a file (REALTIME) */ 2144 int 2145 sys_oaio_write(struct thread *td, struct oaio_write_args *uap) 2146 { 2147 2148 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2149 &aiocb_ops_osigevent)); 2150 } 2151 2152 int 2153 sys_aio_write(struct thread *td, struct aio_write_args *uap) 2154 { 2155 2156 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops)); 2157 } 2158 2159 static int 2160 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list, 2161 struct aiocb **acb_list, int nent, struct sigevent *sig, 2162 struct aiocb_ops *ops) 2163 { 2164 struct proc *p = td->td_proc; 2165 struct aiocb *iocb; 2166 struct kaioinfo *ki; 2167 struct aioliojob *lj; 2168 struct kevent kev; 2169 int error; 2170 int nerror; 2171 int i; 2172 2173 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT)) 2174 return (EINVAL); 2175 2176 if (nent < 0 || nent > AIO_LISTIO_MAX) 2177 return (EINVAL); 2178 2179 if (p->p_aioinfo == NULL) 2180 aio_init_aioinfo(p); 2181 2182 ki = p->p_aioinfo; 2183 2184 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2185 lj->lioj_flags = 0; 2186 lj->lioj_count = 0; 2187 lj->lioj_finished_count = 0; 2188 knlist_init_mtx(&lj->klist, AIO_MTX(ki)); 2189 ksiginfo_init(&lj->lioj_ksi); 2190 2191 /* 2192 * Setup signal. 2193 */ 2194 if (sig && (mode == LIO_NOWAIT)) { 2195 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal)); 2196 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2197 /* Assume only new style KEVENT */ 2198 kev.filter = EVFILT_LIO; 2199 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2200 kev.ident = (uintptr_t)uacb_list; /* something unique */ 2201 kev.data = (intptr_t)lj; 2202 /* pass user defined sigval data */ 2203 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2204 error = kqfd_register( 2205 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1); 2206 if (error) { 2207 uma_zfree(aiolio_zone, lj); 2208 return (error); 2209 } 2210 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2211 ; 2212 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2213 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2214 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2215 uma_zfree(aiolio_zone, lj); 2216 return EINVAL; 2217 } 2218 lj->lioj_flags |= LIOJ_SIGNAL; 2219 } else { 2220 uma_zfree(aiolio_zone, lj); 2221 return EINVAL; 2222 } 2223 } 2224 2225 AIO_LOCK(ki); 2226 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2227 /* 2228 * Add extra aiocb count to avoid the lio to be freed 2229 * by other threads doing aio_waitcomplete or aio_return, 2230 * and prevent event from being sent until we have queued 2231 * all tasks. 2232 */ 2233 lj->lioj_count = 1; 2234 AIO_UNLOCK(ki); 2235 2236 /* 2237 * Get pointers to the list of I/O requests. 2238 */ 2239 nerror = 0; 2240 for (i = 0; i < nent; i++) { 2241 iocb = acb_list[i]; 2242 if (iocb != NULL) { 2243 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops); 2244 if (error != 0) 2245 nerror++; 2246 } 2247 } 2248 2249 error = 0; 2250 AIO_LOCK(ki); 2251 if (mode == LIO_WAIT) { 2252 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2253 ki->kaio_flags |= KAIO_WAKEUP; 2254 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2255 PRIBIO | PCATCH, "aiospn", 0); 2256 if (error == ERESTART) 2257 error = EINTR; 2258 if (error) 2259 break; 2260 } 2261 } else { 2262 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2263 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2264 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2265 KNOTE_LOCKED(&lj->klist, 1); 2266 } 2267 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 2268 == LIOJ_SIGNAL 2269 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2270 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2271 aio_sendsig(p, &lj->lioj_signal, 2272 &lj->lioj_ksi); 2273 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2274 } 2275 } 2276 } 2277 lj->lioj_count--; 2278 if (lj->lioj_count == 0) { 2279 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2280 knlist_delete(&lj->klist, curthread, 1); 2281 PROC_LOCK(p); 2282 sigqueue_take(&lj->lioj_ksi); 2283 PROC_UNLOCK(p); 2284 AIO_UNLOCK(ki); 2285 uma_zfree(aiolio_zone, lj); 2286 } else 2287 AIO_UNLOCK(ki); 2288 2289 if (nerror) 2290 return (EIO); 2291 return (error); 2292 } 2293 2294 /* syscall - list directed I/O (REALTIME) */ 2295 int 2296 sys_olio_listio(struct thread *td, struct olio_listio_args *uap) 2297 { 2298 struct aiocb **acb_list; 2299 struct sigevent *sigp, sig; 2300 struct osigevent osig; 2301 int error, nent; 2302 2303 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2304 return (EINVAL); 2305 2306 nent = uap->nent; 2307 if (nent < 0 || nent > AIO_LISTIO_MAX) 2308 return (EINVAL); 2309 2310 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2311 error = copyin(uap->sig, &osig, sizeof(osig)); 2312 if (error) 2313 return (error); 2314 error = convert_old_sigevent(&osig, &sig); 2315 if (error) 2316 return (error); 2317 sigp = &sig; 2318 } else 2319 sigp = NULL; 2320 2321 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2322 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2323 if (error == 0) 2324 error = kern_lio_listio(td, uap->mode, 2325 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2326 &aiocb_ops_osigevent); 2327 free(acb_list, M_LIO); 2328 return (error); 2329 } 2330 2331 /* syscall - list directed I/O (REALTIME) */ 2332 int 2333 sys_lio_listio(struct thread *td, struct lio_listio_args *uap) 2334 { 2335 struct aiocb **acb_list; 2336 struct sigevent *sigp, sig; 2337 int error, nent; 2338 2339 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2340 return (EINVAL); 2341 2342 nent = uap->nent; 2343 if (nent < 0 || nent > AIO_LISTIO_MAX) 2344 return (EINVAL); 2345 2346 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2347 error = copyin(uap->sig, &sig, sizeof(sig)); 2348 if (error) 2349 return (error); 2350 sigp = &sig; 2351 } else 2352 sigp = NULL; 2353 2354 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2355 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2356 if (error == 0) 2357 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list, 2358 nent, sigp, &aiocb_ops); 2359 free(acb_list, M_LIO); 2360 return (error); 2361 } 2362 2363 /* 2364 * Called from interrupt thread for physio, we should return as fast 2365 * as possible, so we schedule a biohelper task. 2366 */ 2367 static void 2368 aio_physwakeup(struct buf *bp) 2369 { 2370 struct aiocblist *aiocbe; 2371 2372 aiocbe = (struct aiocblist *)bp->b_caller1; 2373 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask); 2374 } 2375 2376 /* 2377 * Task routine to perform heavy tasks, process wakeup, and signals. 2378 */ 2379 static void 2380 biohelper(void *context, int pending) 2381 { 2382 struct aiocblist *aiocbe = context; 2383 struct buf *bp; 2384 struct proc *userp; 2385 struct kaioinfo *ki; 2386 int nblks; 2387 2388 bp = aiocbe->bp; 2389 userp = aiocbe->userproc; 2390 ki = userp->p_aioinfo; 2391 AIO_LOCK(ki); 2392 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 2393 aiocbe->uaiocb._aiocb_private.error = 0; 2394 if (bp->b_ioflags & BIO_ERROR) 2395 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 2396 nblks = btodb(aiocbe->uaiocb.aio_nbytes); 2397 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE) 2398 aiocbe->outputcharge += nblks; 2399 else 2400 aiocbe->inputcharge += nblks; 2401 aiocbe->bp = NULL; 2402 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist); 2403 ki->kaio_buffer_count--; 2404 aio_bio_done_notify(userp, aiocbe, DONE_BUF); 2405 AIO_UNLOCK(ki); 2406 2407 /* Release mapping into kernel space. */ 2408 vunmapbuf(bp); 2409 relpbuf(bp, NULL); 2410 atomic_subtract_int(&num_buf_aio, 1); 2411 } 2412 2413 /* syscall - wait for the next completion of an aio request */ 2414 static int 2415 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp, 2416 struct timespec *ts, struct aiocb_ops *ops) 2417 { 2418 struct proc *p = td->td_proc; 2419 struct timeval atv; 2420 struct kaioinfo *ki; 2421 struct aiocblist *cb; 2422 struct aiocb *uuaiocb; 2423 int error, status, timo; 2424 2425 ops->store_aiocb(aiocbp, NULL); 2426 2427 timo = 0; 2428 if (ts) { 2429 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000)) 2430 return (EINVAL); 2431 2432 TIMESPEC_TO_TIMEVAL(&atv, ts); 2433 if (itimerfix(&atv)) 2434 return (EINVAL); 2435 timo = tvtohz(&atv); 2436 } 2437 2438 if (p->p_aioinfo == NULL) 2439 aio_init_aioinfo(p); 2440 ki = p->p_aioinfo; 2441 2442 error = 0; 2443 cb = NULL; 2444 AIO_LOCK(ki); 2445 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2446 ki->kaio_flags |= KAIO_WAKEUP; 2447 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2448 "aiowc", timo); 2449 if (timo && error == ERESTART) 2450 error = EINTR; 2451 if (error) 2452 break; 2453 } 2454 2455 if (cb != NULL) { 2456 MPASS(cb->jobstate == JOBST_JOBFINISHED); 2457 uuaiocb = cb->uuaiocb; 2458 status = cb->uaiocb._aiocb_private.status; 2459 error = cb->uaiocb._aiocb_private.error; 2460 td->td_retval[0] = status; 2461 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2462 td->td_ru.ru_oublock += cb->outputcharge; 2463 cb->outputcharge = 0; 2464 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2465 td->td_ru.ru_inblock += cb->inputcharge; 2466 cb->inputcharge = 0; 2467 } 2468 aio_free_entry(cb); 2469 AIO_UNLOCK(ki); 2470 ops->store_aiocb(aiocbp, uuaiocb); 2471 ops->store_error(uuaiocb, error); 2472 ops->store_status(uuaiocb, status); 2473 } else 2474 AIO_UNLOCK(ki); 2475 2476 return (error); 2477 } 2478 2479 int 2480 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2481 { 2482 struct timespec ts, *tsp; 2483 int error; 2484 2485 if (uap->timeout) { 2486 /* Get timespec struct. */ 2487 error = copyin(uap->timeout, &ts, sizeof(ts)); 2488 if (error) 2489 return (error); 2490 tsp = &ts; 2491 } else 2492 tsp = NULL; 2493 2494 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops)); 2495 } 2496 2497 static int 2498 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp, 2499 struct aiocb_ops *ops) 2500 { 2501 struct proc *p = td->td_proc; 2502 struct kaioinfo *ki; 2503 2504 if (op != O_SYNC) /* XXX lack of O_DSYNC */ 2505 return (EINVAL); 2506 ki = p->p_aioinfo; 2507 if (ki == NULL) 2508 aio_init_aioinfo(p); 2509 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops)); 2510 } 2511 2512 int 2513 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2514 { 2515 2516 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops)); 2517 } 2518 2519 /* kqueue attach function */ 2520 static int 2521 filt_aioattach(struct knote *kn) 2522 { 2523 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2524 2525 /* 2526 * The aiocbe pointer must be validated before using it, so 2527 * registration is restricted to the kernel; the user cannot 2528 * set EV_FLAG1. 2529 */ 2530 if ((kn->kn_flags & EV_FLAG1) == 0) 2531 return (EPERM); 2532 kn->kn_ptr.p_aio = aiocbe; 2533 kn->kn_flags &= ~EV_FLAG1; 2534 2535 knlist_add(&aiocbe->klist, kn, 0); 2536 2537 return (0); 2538 } 2539 2540 /* kqueue detach function */ 2541 static void 2542 filt_aiodetach(struct knote *kn) 2543 { 2544 struct knlist *knl; 2545 2546 knl = &kn->kn_ptr.p_aio->klist; 2547 knl->kl_lock(knl->kl_lockarg); 2548 if (!knlist_empty(knl)) 2549 knlist_remove(knl, kn, 1); 2550 knl->kl_unlock(knl->kl_lockarg); 2551 } 2552 2553 /* kqueue filter function */ 2554 /*ARGSUSED*/ 2555 static int 2556 filt_aio(struct knote *kn, long hint) 2557 { 2558 struct aiocblist *aiocbe = kn->kn_ptr.p_aio; 2559 2560 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2561 if (aiocbe->jobstate != JOBST_JOBFINISHED) 2562 return (0); 2563 kn->kn_flags |= EV_EOF; 2564 return (1); 2565 } 2566 2567 /* kqueue attach function */ 2568 static int 2569 filt_lioattach(struct knote *kn) 2570 { 2571 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata; 2572 2573 /* 2574 * The aioliojob pointer must be validated before using it, so 2575 * registration is restricted to the kernel; the user cannot 2576 * set EV_FLAG1. 2577 */ 2578 if ((kn->kn_flags & EV_FLAG1) == 0) 2579 return (EPERM); 2580 kn->kn_ptr.p_lio = lj; 2581 kn->kn_flags &= ~EV_FLAG1; 2582 2583 knlist_add(&lj->klist, kn, 0); 2584 2585 return (0); 2586 } 2587 2588 /* kqueue detach function */ 2589 static void 2590 filt_liodetach(struct knote *kn) 2591 { 2592 struct knlist *knl; 2593 2594 knl = &kn->kn_ptr.p_lio->klist; 2595 knl->kl_lock(knl->kl_lockarg); 2596 if (!knlist_empty(knl)) 2597 knlist_remove(knl, kn, 1); 2598 knl->kl_unlock(knl->kl_lockarg); 2599 } 2600 2601 /* kqueue filter function */ 2602 /*ARGSUSED*/ 2603 static int 2604 filt_lio(struct knote *kn, long hint) 2605 { 2606 struct aioliojob * lj = kn->kn_ptr.p_lio; 2607 2608 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2609 } 2610 2611 #ifdef COMPAT_FREEBSD32 2612 2613 struct __aiocb_private32 { 2614 int32_t status; 2615 int32_t error; 2616 uint32_t kernelinfo; 2617 }; 2618 2619 typedef struct oaiocb32 { 2620 int aio_fildes; /* File descriptor */ 2621 uint64_t aio_offset __packed; /* File offset for I/O */ 2622 uint32_t aio_buf; /* I/O buffer in process space */ 2623 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2624 struct osigevent32 aio_sigevent; /* Signal to deliver */ 2625 int aio_lio_opcode; /* LIO opcode */ 2626 int aio_reqprio; /* Request priority -- ignored */ 2627 struct __aiocb_private32 _aiocb_private; 2628 } oaiocb32_t; 2629 2630 typedef struct aiocb32 { 2631 int32_t aio_fildes; /* File descriptor */ 2632 uint64_t aio_offset __packed; /* File offset for I/O */ 2633 uint32_t aio_buf; /* I/O buffer in process space */ 2634 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2635 int __spare__[2]; 2636 uint32_t __spare2__; 2637 int aio_lio_opcode; /* LIO opcode */ 2638 int aio_reqprio; /* Request priority -- ignored */ 2639 struct __aiocb_private32 _aiocb_private; 2640 struct sigevent32 aio_sigevent; /* Signal to deliver */ 2641 } aiocb32_t; 2642 2643 static int 2644 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig) 2645 { 2646 2647 /* 2648 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 2649 * supported by AIO with the old sigevent structure. 2650 */ 2651 CP(*osig, *nsig, sigev_notify); 2652 switch (nsig->sigev_notify) { 2653 case SIGEV_NONE: 2654 break; 2655 case SIGEV_SIGNAL: 2656 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 2657 break; 2658 case SIGEV_KEVENT: 2659 nsig->sigev_notify_kqueue = 2660 osig->__sigev_u.__sigev_notify_kqueue; 2661 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr); 2662 break; 2663 default: 2664 return (EINVAL); 2665 } 2666 return (0); 2667 } 2668 2669 static int 2670 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob) 2671 { 2672 struct oaiocb32 job32; 2673 int error; 2674 2675 bzero(kjob, sizeof(struct aiocb)); 2676 error = copyin(ujob, &job32, sizeof(job32)); 2677 if (error) 2678 return (error); 2679 2680 CP(job32, *kjob, aio_fildes); 2681 CP(job32, *kjob, aio_offset); 2682 PTRIN_CP(job32, *kjob, aio_buf); 2683 CP(job32, *kjob, aio_nbytes); 2684 CP(job32, *kjob, aio_lio_opcode); 2685 CP(job32, *kjob, aio_reqprio); 2686 CP(job32, *kjob, _aiocb_private.status); 2687 CP(job32, *kjob, _aiocb_private.error); 2688 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo); 2689 return (convert_old_sigevent32(&job32.aio_sigevent, 2690 &kjob->aio_sigevent)); 2691 } 2692 2693 static int 2694 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig) 2695 { 2696 2697 CP(*sig32, *sig, sigev_notify); 2698 switch (sig->sigev_notify) { 2699 case SIGEV_NONE: 2700 break; 2701 case SIGEV_THREAD_ID: 2702 CP(*sig32, *sig, sigev_notify_thread_id); 2703 /* FALLTHROUGH */ 2704 case SIGEV_SIGNAL: 2705 CP(*sig32, *sig, sigev_signo); 2706 break; 2707 case SIGEV_KEVENT: 2708 CP(*sig32, *sig, sigev_notify_kqueue); 2709 CP(*sig32, *sig, sigev_notify_kevent_flags); 2710 PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr); 2711 break; 2712 default: 2713 return (EINVAL); 2714 } 2715 return (0); 2716 } 2717 2718 static int 2719 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob) 2720 { 2721 struct aiocb32 job32; 2722 int error; 2723 2724 error = copyin(ujob, &job32, sizeof(job32)); 2725 if (error) 2726 return (error); 2727 CP(job32, *kjob, aio_fildes); 2728 CP(job32, *kjob, aio_offset); 2729 PTRIN_CP(job32, *kjob, aio_buf); 2730 CP(job32, *kjob, aio_nbytes); 2731 CP(job32, *kjob, aio_lio_opcode); 2732 CP(job32, *kjob, aio_reqprio); 2733 CP(job32, *kjob, _aiocb_private.status); 2734 CP(job32, *kjob, _aiocb_private.error); 2735 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo); 2736 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent)); 2737 } 2738 2739 static long 2740 aiocb32_fetch_status(struct aiocb *ujob) 2741 { 2742 struct aiocb32 *ujob32; 2743 2744 ujob32 = (struct aiocb32 *)ujob; 2745 return (fuword32(&ujob32->_aiocb_private.status)); 2746 } 2747 2748 static long 2749 aiocb32_fetch_error(struct aiocb *ujob) 2750 { 2751 struct aiocb32 *ujob32; 2752 2753 ujob32 = (struct aiocb32 *)ujob; 2754 return (fuword32(&ujob32->_aiocb_private.error)); 2755 } 2756 2757 static int 2758 aiocb32_store_status(struct aiocb *ujob, long status) 2759 { 2760 struct aiocb32 *ujob32; 2761 2762 ujob32 = (struct aiocb32 *)ujob; 2763 return (suword32(&ujob32->_aiocb_private.status, status)); 2764 } 2765 2766 static int 2767 aiocb32_store_error(struct aiocb *ujob, long error) 2768 { 2769 struct aiocb32 *ujob32; 2770 2771 ujob32 = (struct aiocb32 *)ujob; 2772 return (suword32(&ujob32->_aiocb_private.error, error)); 2773 } 2774 2775 static int 2776 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref) 2777 { 2778 struct aiocb32 *ujob32; 2779 2780 ujob32 = (struct aiocb32 *)ujob; 2781 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref)); 2782 } 2783 2784 static int 2785 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 2786 { 2787 2788 return (suword32(ujobp, (long)ujob)); 2789 } 2790 2791 static struct aiocb_ops aiocb32_ops = { 2792 .copyin = aiocb32_copyin, 2793 .fetch_status = aiocb32_fetch_status, 2794 .fetch_error = aiocb32_fetch_error, 2795 .store_status = aiocb32_store_status, 2796 .store_error = aiocb32_store_error, 2797 .store_kernelinfo = aiocb32_store_kernelinfo, 2798 .store_aiocb = aiocb32_store_aiocb, 2799 }; 2800 2801 static struct aiocb_ops aiocb32_ops_osigevent = { 2802 .copyin = aiocb32_copyin_old_sigevent, 2803 .fetch_status = aiocb32_fetch_status, 2804 .fetch_error = aiocb32_fetch_error, 2805 .store_status = aiocb32_store_status, 2806 .store_error = aiocb32_store_error, 2807 .store_kernelinfo = aiocb32_store_kernelinfo, 2808 .store_aiocb = aiocb32_store_aiocb, 2809 }; 2810 2811 int 2812 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap) 2813 { 2814 2815 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2816 } 2817 2818 int 2819 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap) 2820 { 2821 struct timespec32 ts32; 2822 struct timespec ts, *tsp; 2823 struct aiocb **ujoblist; 2824 uint32_t *ujoblist32; 2825 int error, i; 2826 2827 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX) 2828 return (EINVAL); 2829 2830 if (uap->timeout) { 2831 /* Get timespec struct. */ 2832 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0) 2833 return (error); 2834 CP(ts32, ts, tv_sec); 2835 CP(ts32, ts, tv_nsec); 2836 tsp = &ts; 2837 } else 2838 tsp = NULL; 2839 2840 ujoblist = uma_zalloc(aiol_zone, M_WAITOK); 2841 ujoblist32 = (uint32_t *)ujoblist; 2842 error = copyin(uap->aiocbp, ujoblist32, uap->nent * 2843 sizeof(ujoblist32[0])); 2844 if (error == 0) { 2845 for (i = uap->nent; i > 0; i--) 2846 ujoblist[i] = PTRIN(ujoblist32[i]); 2847 2848 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2849 } 2850 uma_zfree(aiol_zone, ujoblist); 2851 return (error); 2852 } 2853 2854 int 2855 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap) 2856 { 2857 2858 return (sys_aio_cancel(td, (struct aio_cancel_args *)uap)); 2859 } 2860 2861 int 2862 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap) 2863 { 2864 2865 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2866 } 2867 2868 int 2869 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap) 2870 { 2871 2872 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2873 &aiocb32_ops_osigevent)); 2874 } 2875 2876 int 2877 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap) 2878 { 2879 2880 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2881 &aiocb32_ops)); 2882 } 2883 2884 int 2885 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap) 2886 { 2887 2888 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2889 &aiocb32_ops_osigevent)); 2890 } 2891 2892 int 2893 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap) 2894 { 2895 2896 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2897 &aiocb32_ops)); 2898 } 2899 2900 int 2901 freebsd32_aio_waitcomplete(struct thread *td, 2902 struct freebsd32_aio_waitcomplete_args *uap) 2903 { 2904 struct timespec32 ts32; 2905 struct timespec ts, *tsp; 2906 int error; 2907 2908 if (uap->timeout) { 2909 /* Get timespec struct. */ 2910 error = copyin(uap->timeout, &ts32, sizeof(ts32)); 2911 if (error) 2912 return (error); 2913 CP(ts32, ts, tv_sec); 2914 CP(ts32, ts, tv_nsec); 2915 tsp = &ts; 2916 } else 2917 tsp = NULL; 2918 2919 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp, 2920 &aiocb32_ops)); 2921 } 2922 2923 int 2924 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap) 2925 { 2926 2927 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp, 2928 &aiocb32_ops)); 2929 } 2930 2931 int 2932 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap) 2933 { 2934 struct aiocb **acb_list; 2935 struct sigevent *sigp, sig; 2936 struct osigevent32 osig; 2937 uint32_t *acb_list32; 2938 int error, i, nent; 2939 2940 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2941 return (EINVAL); 2942 2943 nent = uap->nent; 2944 if (nent < 0 || nent > AIO_LISTIO_MAX) 2945 return (EINVAL); 2946 2947 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2948 error = copyin(uap->sig, &osig, sizeof(osig)); 2949 if (error) 2950 return (error); 2951 error = convert_old_sigevent32(&osig, &sig); 2952 if (error) 2953 return (error); 2954 sigp = &sig; 2955 } else 2956 sigp = NULL; 2957 2958 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 2959 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 2960 if (error) { 2961 free(acb_list32, M_LIO); 2962 return (error); 2963 } 2964 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2965 for (i = 0; i < nent; i++) 2966 acb_list[i] = PTRIN(acb_list32[i]); 2967 free(acb_list32, M_LIO); 2968 2969 error = kern_lio_listio(td, uap->mode, 2970 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2971 &aiocb32_ops_osigevent); 2972 free(acb_list, M_LIO); 2973 return (error); 2974 } 2975 2976 int 2977 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap) 2978 { 2979 struct aiocb **acb_list; 2980 struct sigevent *sigp, sig; 2981 struct sigevent32 sig32; 2982 uint32_t *acb_list32; 2983 int error, i, nent; 2984 2985 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2986 return (EINVAL); 2987 2988 nent = uap->nent; 2989 if (nent < 0 || nent > AIO_LISTIO_MAX) 2990 return (EINVAL); 2991 2992 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2993 error = copyin(uap->sig, &sig32, sizeof(sig32)); 2994 if (error) 2995 return (error); 2996 error = convert_sigevent32(&sig32, &sig); 2997 if (error) 2998 return (error); 2999 sigp = &sig; 3000 } else 3001 sigp = NULL; 3002 3003 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3004 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3005 if (error) { 3006 free(acb_list32, M_LIO); 3007 return (error); 3008 } 3009 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3010 for (i = 0; i < nent; i++) 3011 acb_list[i] = PTRIN(acb_list32[i]); 3012 free(acb_list32, M_LIO); 3013 3014 error = kern_lio_listio(td, uap->mode, 3015 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3016 &aiocb32_ops); 3017 free(acb_list, M_LIO); 3018 return (error); 3019 } 3020 3021 #endif 3022