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/capsicum.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/syscallsubr.h> 49 #include <sys/protosw.h> 50 #include <sys/rwlock.h> 51 #include <sys/sema.h> 52 #include <sys/socket.h> 53 #include <sys/socketvar.h> 54 #include <sys/syscall.h> 55 #include <sys/sysent.h> 56 #include <sys/sysctl.h> 57 #include <sys/syslog.h> 58 #include <sys/sx.h> 59 #include <sys/taskqueue.h> 60 #include <sys/vnode.h> 61 #include <sys/conf.h> 62 #include <sys/event.h> 63 #include <sys/mount.h> 64 #include <geom/geom.h> 65 66 #include <machine/atomic.h> 67 68 #include <vm/vm.h> 69 #include <vm/vm_page.h> 70 #include <vm/vm_extern.h> 71 #include <vm/pmap.h> 72 #include <vm/vm_map.h> 73 #include <vm/vm_object.h> 74 #include <vm/uma.h> 75 #include <sys/aio.h> 76 77 /* 78 * Counter for allocating reference ids to new jobs. Wrapped to 1 on 79 * overflow. (XXX will be removed soon.) 80 */ 81 static u_long jobrefid; 82 83 /* 84 * Counter for aio_fsync. 85 */ 86 static uint64_t jobseqno; 87 88 #ifndef MAX_AIO_PER_PROC 89 #define MAX_AIO_PER_PROC 32 90 #endif 91 92 #ifndef MAX_AIO_QUEUE_PER_PROC 93 #define MAX_AIO_QUEUE_PER_PROC 256 94 #endif 95 96 #ifndef MAX_AIO_QUEUE 97 #define MAX_AIO_QUEUE 1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */ 98 #endif 99 100 #ifndef MAX_BUF_AIO 101 #define MAX_BUF_AIO 16 102 #endif 103 104 FEATURE(aio, "Asynchronous I/O"); 105 SYSCTL_DECL(_p1003_1b); 106 107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list"); 108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list"); 109 110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, 111 "Async IO management"); 112 113 static int enable_aio_unsafe = 0; 114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0, 115 "Permit asynchronous IO on all file types, not just known-safe types"); 116 117 static unsigned int unsafe_warningcnt = 1; 118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW, 119 &unsafe_warningcnt, 0, 120 "Warnings that will be triggered upon failed IO requests on unsafe files"); 121 122 static int max_aio_procs = MAX_AIO_PROCS; 123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0, 124 "Maximum number of kernel processes to use for handling async IO "); 125 126 static int num_aio_procs = 0; 127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0, 128 "Number of presently active kernel processes for async IO"); 129 130 /* 131 * The code will adjust the actual number of AIO processes towards this 132 * number when it gets a chance. 133 */ 134 static int target_aio_procs = TARGET_AIO_PROCS; 135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 136 0, 137 "Preferred number of ready kernel processes for async IO"); 138 139 static int max_queue_count = MAX_AIO_QUEUE; 140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 141 "Maximum number of aio requests to queue, globally"); 142 143 static int num_queue_count = 0; 144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 145 "Number of queued aio requests"); 146 147 static int num_buf_aio = 0; 148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 149 "Number of aio requests presently handled by the buf subsystem"); 150 151 /* Number of async I/O processes in the process of being started */ 152 /* XXX This should be local to aio_aqueue() */ 153 static int num_aio_resv_start = 0; 154 155 static int aiod_lifetime; 156 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 157 "Maximum lifetime for idle aiod"); 158 159 static int max_aio_per_proc = MAX_AIO_PER_PROC; 160 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 161 0, 162 "Maximum active aio requests per process (stored in the process)"); 163 164 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 166 &max_aio_queue_per_proc, 0, 167 "Maximum queued aio requests per process (stored in the process)"); 168 169 static int max_buf_aio = MAX_BUF_AIO; 170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 171 "Maximum buf aio requests per process (stored in the process)"); 172 173 /* 174 * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires 175 * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with 176 * vfs.aio.aio_listio_max. 177 */ 178 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max, 179 CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc, 180 0, "Maximum aio requests for a single lio_listio call"); 181 182 #ifdef COMPAT_FREEBSD6 183 typedef struct oaiocb { 184 int aio_fildes; /* File descriptor */ 185 off_t aio_offset; /* File offset for I/O */ 186 volatile void *aio_buf; /* I/O buffer in process space */ 187 size_t aio_nbytes; /* Number of bytes for I/O */ 188 struct osigevent aio_sigevent; /* Signal to deliver */ 189 int aio_lio_opcode; /* LIO opcode */ 190 int aio_reqprio; /* Request priority -- ignored */ 191 struct __aiocb_private _aiocb_private; 192 } oaiocb_t; 193 #endif 194 195 /* 196 * Below is a key of locks used to protect each member of struct kaiocb 197 * aioliojob and kaioinfo and any backends. 198 * 199 * * - need not protected 200 * a - locked by kaioinfo lock 201 * b - locked by backend lock, the backend lock can be null in some cases, 202 * for example, BIO belongs to this type, in this case, proc lock is 203 * reused. 204 * c - locked by aio_job_mtx, the lock for the generic file I/O backend. 205 */ 206 207 /* 208 * If the routine that services an AIO request blocks while running in an 209 * AIO kernel process it can starve other I/O requests. BIO requests 210 * queued via aio_qphysio() complete in GEOM and do not use AIO kernel 211 * processes at all. Socket I/O requests use a separate pool of 212 * kprocs and also force non-blocking I/O. Other file I/O requests 213 * use the generic fo_read/fo_write operations which can block. The 214 * fsync and mlock operations can also block while executing. Ideally 215 * none of these requests would block while executing. 216 * 217 * Note that the service routines cannot toggle O_NONBLOCK in the file 218 * structure directly while handling a request due to races with 219 * userland threads. 220 */ 221 222 /* jobflags */ 223 #define KAIOCB_QUEUEING 0x01 224 #define KAIOCB_CANCELLED 0x02 225 #define KAIOCB_CANCELLING 0x04 226 #define KAIOCB_CHECKSYNC 0x08 227 #define KAIOCB_CLEARED 0x10 228 #define KAIOCB_FINISHED 0x20 229 230 /* 231 * AIO process info 232 */ 233 #define AIOP_FREE 0x1 /* proc on free queue */ 234 235 struct aioproc { 236 int aioprocflags; /* (c) AIO proc flags */ 237 TAILQ_ENTRY(aioproc) list; /* (c) list of processes */ 238 struct proc *aioproc; /* (*) the AIO proc */ 239 }; 240 241 /* 242 * data-structure for lio signal management 243 */ 244 struct aioliojob { 245 int lioj_flags; /* (a) listio flags */ 246 int lioj_count; /* (a) listio flags */ 247 int lioj_finished_count; /* (a) listio flags */ 248 struct sigevent lioj_signal; /* (a) signal on all I/O done */ 249 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */ 250 struct knlist klist; /* (a) list of knotes */ 251 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */ 252 }; 253 254 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 255 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 256 #define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */ 257 258 /* 259 * per process aio data structure 260 */ 261 struct kaioinfo { 262 struct mtx kaio_mtx; /* the lock to protect this struct */ 263 int kaio_flags; /* (a) per process kaio flags */ 264 int kaio_maxactive_count; /* (*) maximum number of AIOs */ 265 int kaio_active_count; /* (c) number of currently used AIOs */ 266 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */ 267 int kaio_count; /* (a) size of AIO queue */ 268 int kaio_ballowed_count; /* (*) maximum number of buffers */ 269 int kaio_buffer_count; /* (a) number of physio buffers */ 270 TAILQ_HEAD(,kaiocb) kaio_all; /* (a) all AIOs in a process */ 271 TAILQ_HEAD(,kaiocb) kaio_done; /* (a) done queue for process */ 272 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */ 273 TAILQ_HEAD(,kaiocb) kaio_jobqueue; /* (a) job queue for process */ 274 TAILQ_HEAD(,kaiocb) kaio_syncqueue; /* (a) queue for aio_fsync */ 275 TAILQ_HEAD(,kaiocb) kaio_syncready; /* (a) second q for aio_fsync */ 276 struct task kaio_task; /* (*) task to kick aio processes */ 277 struct task kaio_sync_task; /* (*) task to schedule fsync jobs */ 278 }; 279 280 #define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx) 281 #define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx) 282 #define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f)) 283 #define AIO_MTX(ki) (&(ki)->kaio_mtx) 284 285 #define KAIO_RUNDOWN 0x1 /* process is being run down */ 286 #define KAIO_WAKEUP 0x2 /* wakeup process when AIO completes */ 287 288 /* 289 * Operations used to interact with userland aio control blocks. 290 * Different ABIs provide their own operations. 291 */ 292 struct aiocb_ops { 293 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob); 294 long (*fetch_status)(struct aiocb *ujob); 295 long (*fetch_error)(struct aiocb *ujob); 296 int (*store_status)(struct aiocb *ujob, long status); 297 int (*store_error)(struct aiocb *ujob, long error); 298 int (*store_kernelinfo)(struct aiocb *ujob, long jobref); 299 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob); 300 }; 301 302 static TAILQ_HEAD(,aioproc) aio_freeproc; /* (c) Idle daemons */ 303 static struct sema aio_newproc_sem; 304 static struct mtx aio_job_mtx; 305 static TAILQ_HEAD(,kaiocb) aio_jobs; /* (c) Async job list */ 306 static struct unrhdr *aiod_unr; 307 308 void aio_init_aioinfo(struct proc *p); 309 static int aio_onceonly(void); 310 static int aio_free_entry(struct kaiocb *job); 311 static void aio_process_rw(struct kaiocb *job); 312 static void aio_process_sync(struct kaiocb *job); 313 static void aio_process_mlock(struct kaiocb *job); 314 static void aio_schedule_fsync(void *context, int pending); 315 static int aio_newproc(int *); 316 int aio_aqueue(struct thread *td, struct aiocb *ujob, 317 struct aioliojob *lio, int type, struct aiocb_ops *ops); 318 static int aio_queue_file(struct file *fp, struct kaiocb *job); 319 static void aio_physwakeup(struct bio *bp); 320 static void aio_proc_rundown(void *arg, struct proc *p); 321 static void aio_proc_rundown_exec(void *arg, struct proc *p, 322 struct image_params *imgp); 323 static int aio_qphysio(struct proc *p, struct kaiocb *job); 324 static void aio_daemon(void *param); 325 static void aio_bio_done_notify(struct proc *userp, struct kaiocb *job); 326 static bool aio_clear_cancel_function_locked(struct kaiocb *job); 327 static int aio_kick(struct proc *userp); 328 static void aio_kick_nowait(struct proc *userp); 329 static void aio_kick_helper(void *context, int pending); 330 static int filt_aioattach(struct knote *kn); 331 static void filt_aiodetach(struct knote *kn); 332 static int filt_aio(struct knote *kn, long hint); 333 static int filt_lioattach(struct knote *kn); 334 static void filt_liodetach(struct knote *kn); 335 static int filt_lio(struct knote *kn, long hint); 336 337 /* 338 * Zones for: 339 * kaio Per process async io info 340 * aiop async io process data 341 * aiocb async io jobs 342 * aiolio list io jobs 343 */ 344 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone; 345 346 /* kqueue filters for aio */ 347 static struct filterops aio_filtops = { 348 .f_isfd = 0, 349 .f_attach = filt_aioattach, 350 .f_detach = filt_aiodetach, 351 .f_event = filt_aio, 352 }; 353 static struct filterops lio_filtops = { 354 .f_isfd = 0, 355 .f_attach = filt_lioattach, 356 .f_detach = filt_liodetach, 357 .f_event = filt_lio 358 }; 359 360 static eventhandler_tag exit_tag, exec_tag; 361 362 TASKQUEUE_DEFINE_THREAD(aiod_kick); 363 364 /* 365 * Main operations function for use as a kernel module. 366 */ 367 static int 368 aio_modload(struct module *module, int cmd, void *arg) 369 { 370 int error = 0; 371 372 switch (cmd) { 373 case MOD_LOAD: 374 aio_onceonly(); 375 break; 376 case MOD_SHUTDOWN: 377 break; 378 default: 379 error = EOPNOTSUPP; 380 break; 381 } 382 return (error); 383 } 384 385 static moduledata_t aio_mod = { 386 "aio", 387 &aio_modload, 388 NULL 389 }; 390 391 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY); 392 MODULE_VERSION(aio, 1); 393 394 /* 395 * Startup initialization 396 */ 397 static int 398 aio_onceonly(void) 399 { 400 401 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 402 EVENTHANDLER_PRI_ANY); 403 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, 404 NULL, EVENTHANDLER_PRI_ANY); 405 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 406 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 407 TAILQ_INIT(&aio_freeproc); 408 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 409 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 410 TAILQ_INIT(&aio_jobs); 411 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 412 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 413 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 414 aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL, 415 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 416 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL, 417 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 418 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 419 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 420 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 421 jobrefid = 1; 422 p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO); 423 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 424 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 425 426 return (0); 427 } 428 429 /* 430 * Init the per-process aioinfo structure. The aioinfo limits are set 431 * per-process for user limit (resource) management. 432 */ 433 void 434 aio_init_aioinfo(struct proc *p) 435 { 436 struct kaioinfo *ki; 437 438 ki = uma_zalloc(kaio_zone, M_WAITOK); 439 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW); 440 ki->kaio_flags = 0; 441 ki->kaio_maxactive_count = max_aio_per_proc; 442 ki->kaio_active_count = 0; 443 ki->kaio_qallowed_count = max_aio_queue_per_proc; 444 ki->kaio_count = 0; 445 ki->kaio_ballowed_count = max_buf_aio; 446 ki->kaio_buffer_count = 0; 447 TAILQ_INIT(&ki->kaio_all); 448 TAILQ_INIT(&ki->kaio_done); 449 TAILQ_INIT(&ki->kaio_jobqueue); 450 TAILQ_INIT(&ki->kaio_liojoblist); 451 TAILQ_INIT(&ki->kaio_syncqueue); 452 TAILQ_INIT(&ki->kaio_syncready); 453 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 454 TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki); 455 PROC_LOCK(p); 456 if (p->p_aioinfo == NULL) { 457 p->p_aioinfo = ki; 458 PROC_UNLOCK(p); 459 } else { 460 PROC_UNLOCK(p); 461 mtx_destroy(&ki->kaio_mtx); 462 uma_zfree(kaio_zone, ki); 463 } 464 465 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs)) 466 aio_newproc(NULL); 467 } 468 469 static int 470 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi) 471 { 472 struct thread *td; 473 int error; 474 475 error = sigev_findtd(p, sigev, &td); 476 if (error) 477 return (error); 478 if (!KSI_ONQ(ksi)) { 479 ksiginfo_set_sigev(ksi, sigev); 480 ksi->ksi_code = SI_ASYNCIO; 481 ksi->ksi_flags |= KSI_EXT | KSI_INS; 482 tdsendsignal(p, td, ksi->ksi_signo, ksi); 483 } 484 PROC_UNLOCK(p); 485 return (error); 486 } 487 488 /* 489 * Free a job entry. Wait for completion if it is currently active, but don't 490 * delay forever. If we delay, we return a flag that says that we have to 491 * restart the queue scan. 492 */ 493 static int 494 aio_free_entry(struct kaiocb *job) 495 { 496 struct kaioinfo *ki; 497 struct aioliojob *lj; 498 struct proc *p; 499 500 p = job->userproc; 501 MPASS(curproc == p); 502 ki = p->p_aioinfo; 503 MPASS(ki != NULL); 504 505 AIO_LOCK_ASSERT(ki, MA_OWNED); 506 MPASS(job->jobflags & KAIOCB_FINISHED); 507 508 atomic_subtract_int(&num_queue_count, 1); 509 510 ki->kaio_count--; 511 MPASS(ki->kaio_count >= 0); 512 513 TAILQ_REMOVE(&ki->kaio_done, job, plist); 514 TAILQ_REMOVE(&ki->kaio_all, job, allist); 515 516 lj = job->lio; 517 if (lj) { 518 lj->lioj_count--; 519 lj->lioj_finished_count--; 520 521 if (lj->lioj_count == 0) { 522 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 523 /* lio is going away, we need to destroy any knotes */ 524 knlist_delete(&lj->klist, curthread, 1); 525 PROC_LOCK(p); 526 sigqueue_take(&lj->lioj_ksi); 527 PROC_UNLOCK(p); 528 uma_zfree(aiolio_zone, lj); 529 } 530 } 531 532 /* job is going away, we need to destroy any knotes */ 533 knlist_delete(&job->klist, curthread, 1); 534 PROC_LOCK(p); 535 sigqueue_take(&job->ksi); 536 PROC_UNLOCK(p); 537 538 AIO_UNLOCK(ki); 539 540 /* 541 * The thread argument here is used to find the owning process 542 * and is also passed to fo_close() which may pass it to various 543 * places such as devsw close() routines. Because of that, we 544 * need a thread pointer from the process owning the job that is 545 * persistent and won't disappear out from under us or move to 546 * another process. 547 * 548 * Currently, all the callers of this function call it to remove 549 * a kaiocb from the current process' job list either via a 550 * syscall or due to the current process calling exit() or 551 * execve(). Thus, we know that p == curproc. We also know that 552 * curthread can't exit since we are curthread. 553 * 554 * Therefore, we use curthread as the thread to pass to 555 * knlist_delete(). This does mean that it is possible for the 556 * thread pointer at close time to differ from the thread pointer 557 * at open time, but this is already true of file descriptors in 558 * a multithreaded process. 559 */ 560 if (job->fd_file) 561 fdrop(job->fd_file, curthread); 562 crfree(job->cred); 563 uma_zfree(aiocb_zone, job); 564 AIO_LOCK(ki); 565 566 return (0); 567 } 568 569 static void 570 aio_proc_rundown_exec(void *arg, struct proc *p, 571 struct image_params *imgp __unused) 572 { 573 aio_proc_rundown(arg, p); 574 } 575 576 static int 577 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job) 578 { 579 aio_cancel_fn_t *func; 580 int cancelled; 581 582 AIO_LOCK_ASSERT(ki, MA_OWNED); 583 if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED)) 584 return (0); 585 MPASS((job->jobflags & KAIOCB_CANCELLING) == 0); 586 job->jobflags |= KAIOCB_CANCELLED; 587 588 func = job->cancel_fn; 589 590 /* 591 * If there is no cancel routine, just leave the job marked as 592 * cancelled. The job should be in active use by a caller who 593 * should complete it normally or when it fails to install a 594 * cancel routine. 595 */ 596 if (func == NULL) 597 return (0); 598 599 /* 600 * Set the CANCELLING flag so that aio_complete() will defer 601 * completions of this job. This prevents the job from being 602 * freed out from under the cancel callback. After the 603 * callback any deferred completion (whether from the callback 604 * or any other source) will be completed. 605 */ 606 job->jobflags |= KAIOCB_CANCELLING; 607 AIO_UNLOCK(ki); 608 func(job); 609 AIO_LOCK(ki); 610 job->jobflags &= ~KAIOCB_CANCELLING; 611 if (job->jobflags & KAIOCB_FINISHED) { 612 cancelled = job->uaiocb._aiocb_private.error == ECANCELED; 613 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 614 aio_bio_done_notify(p, job); 615 } else { 616 /* 617 * The cancel callback might have scheduled an 618 * operation to cancel this request, but it is 619 * only counted as cancelled if the request is 620 * cancelled when the callback returns. 621 */ 622 cancelled = 0; 623 } 624 return (cancelled); 625 } 626 627 /* 628 * Rundown the jobs for a given process. 629 */ 630 static void 631 aio_proc_rundown(void *arg, struct proc *p) 632 { 633 struct kaioinfo *ki; 634 struct aioliojob *lj; 635 struct kaiocb *job, *jobn; 636 637 KASSERT(curthread->td_proc == p, 638 ("%s: called on non-curproc", __func__)); 639 ki = p->p_aioinfo; 640 if (ki == NULL) 641 return; 642 643 AIO_LOCK(ki); 644 ki->kaio_flags |= KAIO_RUNDOWN; 645 646 restart: 647 648 /* 649 * Try to cancel all pending requests. This code simulates 650 * aio_cancel on all pending I/O requests. 651 */ 652 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 653 aio_cancel_job(p, ki, job); 654 } 655 656 /* Wait for all running I/O to be finished */ 657 if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) { 658 ki->kaio_flags |= KAIO_WAKEUP; 659 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 660 goto restart; 661 } 662 663 /* Free all completed I/O requests. */ 664 while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL) 665 aio_free_entry(job); 666 667 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 668 if (lj->lioj_count == 0) { 669 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 670 knlist_delete(&lj->klist, curthread, 1); 671 PROC_LOCK(p); 672 sigqueue_take(&lj->lioj_ksi); 673 PROC_UNLOCK(p); 674 uma_zfree(aiolio_zone, lj); 675 } else { 676 panic("LIO job not cleaned up: C:%d, FC:%d\n", 677 lj->lioj_count, lj->lioj_finished_count); 678 } 679 } 680 AIO_UNLOCK(ki); 681 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task); 682 taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task); 683 mtx_destroy(&ki->kaio_mtx); 684 uma_zfree(kaio_zone, ki); 685 p->p_aioinfo = NULL; 686 } 687 688 /* 689 * Select a job to run (called by an AIO daemon). 690 */ 691 static struct kaiocb * 692 aio_selectjob(struct aioproc *aiop) 693 { 694 struct kaiocb *job; 695 struct kaioinfo *ki; 696 struct proc *userp; 697 698 mtx_assert(&aio_job_mtx, MA_OWNED); 699 restart: 700 TAILQ_FOREACH(job, &aio_jobs, list) { 701 userp = job->userproc; 702 ki = userp->p_aioinfo; 703 704 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 705 TAILQ_REMOVE(&aio_jobs, job, list); 706 if (!aio_clear_cancel_function(job)) 707 goto restart; 708 709 /* Account for currently active jobs. */ 710 ki->kaio_active_count++; 711 break; 712 } 713 } 714 return (job); 715 } 716 717 /* 718 * Move all data to a permanent storage device. This code 719 * simulates the fsync syscall. 720 */ 721 static int 722 aio_fsync_vnode(struct thread *td, struct vnode *vp) 723 { 724 struct mount *mp; 725 int error; 726 727 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 728 goto drop; 729 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 730 if (vp->v_object != NULL) { 731 VM_OBJECT_WLOCK(vp->v_object); 732 vm_object_page_clean(vp->v_object, 0, 0, 0); 733 VM_OBJECT_WUNLOCK(vp->v_object); 734 } 735 error = VOP_FSYNC(vp, MNT_WAIT, td); 736 737 VOP_UNLOCK(vp, 0); 738 vn_finished_write(mp); 739 drop: 740 return (error); 741 } 742 743 /* 744 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that 745 * does the I/O request for the non-physio version of the operations. The 746 * normal vn operations are used, and this code should work in all instances 747 * for every type of file, including pipes, sockets, fifos, and regular files. 748 * 749 * XXX I don't think it works well for socket, pipe, and fifo. 750 */ 751 static void 752 aio_process_rw(struct kaiocb *job) 753 { 754 struct ucred *td_savedcred; 755 struct thread *td; 756 struct aiocb *cb; 757 struct file *fp; 758 struct uio auio; 759 struct iovec aiov; 760 ssize_t cnt; 761 long msgsnd_st, msgsnd_end; 762 long msgrcv_st, msgrcv_end; 763 long oublock_st, oublock_end; 764 long inblock_st, inblock_end; 765 int error; 766 767 KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ || 768 job->uaiocb.aio_lio_opcode == LIO_WRITE, 769 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 770 771 aio_switch_vmspace(job); 772 td = curthread; 773 td_savedcred = td->td_ucred; 774 td->td_ucred = job->cred; 775 cb = &job->uaiocb; 776 fp = job->fd_file; 777 778 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 779 aiov.iov_len = cb->aio_nbytes; 780 781 auio.uio_iov = &aiov; 782 auio.uio_iovcnt = 1; 783 auio.uio_offset = cb->aio_offset; 784 auio.uio_resid = cb->aio_nbytes; 785 cnt = cb->aio_nbytes; 786 auio.uio_segflg = UIO_USERSPACE; 787 auio.uio_td = td; 788 789 msgrcv_st = td->td_ru.ru_msgrcv; 790 msgsnd_st = td->td_ru.ru_msgsnd; 791 inblock_st = td->td_ru.ru_inblock; 792 oublock_st = td->td_ru.ru_oublock; 793 794 /* 795 * aio_aqueue() acquires a reference to the file that is 796 * released in aio_free_entry(). 797 */ 798 if (cb->aio_lio_opcode == LIO_READ) { 799 auio.uio_rw = UIO_READ; 800 if (auio.uio_resid == 0) 801 error = 0; 802 else 803 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td); 804 } else { 805 if (fp->f_type == DTYPE_VNODE) 806 bwillwrite(); 807 auio.uio_rw = UIO_WRITE; 808 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td); 809 } 810 msgrcv_end = td->td_ru.ru_msgrcv; 811 msgsnd_end = td->td_ru.ru_msgsnd; 812 inblock_end = td->td_ru.ru_inblock; 813 oublock_end = td->td_ru.ru_oublock; 814 815 job->msgrcv = msgrcv_end - msgrcv_st; 816 job->msgsnd = msgsnd_end - msgsnd_st; 817 job->inblock = inblock_end - inblock_st; 818 job->outblock = oublock_end - oublock_st; 819 820 if ((error) && (auio.uio_resid != cnt)) { 821 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 822 error = 0; 823 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) { 824 PROC_LOCK(job->userproc); 825 kern_psignal(job->userproc, SIGPIPE); 826 PROC_UNLOCK(job->userproc); 827 } 828 } 829 830 cnt -= auio.uio_resid; 831 td->td_ucred = td_savedcred; 832 if (error) 833 aio_complete(job, -1, error); 834 else 835 aio_complete(job, cnt, 0); 836 } 837 838 static void 839 aio_process_sync(struct kaiocb *job) 840 { 841 struct thread *td = curthread; 842 struct ucred *td_savedcred = td->td_ucred; 843 struct file *fp = job->fd_file; 844 int error = 0; 845 846 KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC, 847 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 848 849 td->td_ucred = job->cred; 850 if (fp->f_vnode != NULL) 851 error = aio_fsync_vnode(td, fp->f_vnode); 852 td->td_ucred = td_savedcred; 853 if (error) 854 aio_complete(job, -1, error); 855 else 856 aio_complete(job, 0, 0); 857 } 858 859 static void 860 aio_process_mlock(struct kaiocb *job) 861 { 862 struct aiocb *cb = &job->uaiocb; 863 int error; 864 865 KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK, 866 ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode)); 867 868 aio_switch_vmspace(job); 869 error = kern_mlock(job->userproc, job->cred, 870 __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes); 871 aio_complete(job, error != 0 ? -1 : 0, error); 872 } 873 874 static void 875 aio_bio_done_notify(struct proc *userp, struct kaiocb *job) 876 { 877 struct aioliojob *lj; 878 struct kaioinfo *ki; 879 struct kaiocb *sjob, *sjobn; 880 int lj_done; 881 bool schedule_fsync; 882 883 ki = userp->p_aioinfo; 884 AIO_LOCK_ASSERT(ki, MA_OWNED); 885 lj = job->lio; 886 lj_done = 0; 887 if (lj) { 888 lj->lioj_finished_count++; 889 if (lj->lioj_count == lj->lioj_finished_count) 890 lj_done = 1; 891 } 892 TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist); 893 MPASS(job->jobflags & KAIOCB_FINISHED); 894 895 if (ki->kaio_flags & KAIO_RUNDOWN) 896 goto notification_done; 897 898 if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 899 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 900 aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi); 901 902 KNOTE_LOCKED(&job->klist, 1); 903 904 if (lj_done) { 905 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 906 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 907 KNOTE_LOCKED(&lj->klist, 1); 908 } 909 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 910 == LIOJ_SIGNAL 911 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 912 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 913 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi); 914 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 915 } 916 } 917 918 notification_done: 919 if (job->jobflags & KAIOCB_CHECKSYNC) { 920 schedule_fsync = false; 921 TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) { 922 if (job->fd_file != sjob->fd_file || 923 job->seqno >= sjob->seqno) 924 continue; 925 if (--sjob->pending > 0) 926 continue; 927 TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list); 928 if (!aio_clear_cancel_function_locked(sjob)) 929 continue; 930 TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list); 931 schedule_fsync = true; 932 } 933 if (schedule_fsync) 934 taskqueue_enqueue(taskqueue_aiod_kick, 935 &ki->kaio_sync_task); 936 } 937 if (ki->kaio_flags & KAIO_WAKEUP) { 938 ki->kaio_flags &= ~KAIO_WAKEUP; 939 wakeup(&userp->p_aioinfo); 940 } 941 } 942 943 static void 944 aio_schedule_fsync(void *context, int pending) 945 { 946 struct kaioinfo *ki; 947 struct kaiocb *job; 948 949 ki = context; 950 AIO_LOCK(ki); 951 while (!TAILQ_EMPTY(&ki->kaio_syncready)) { 952 job = TAILQ_FIRST(&ki->kaio_syncready); 953 TAILQ_REMOVE(&ki->kaio_syncready, job, list); 954 AIO_UNLOCK(ki); 955 aio_schedule(job, aio_process_sync); 956 AIO_LOCK(ki); 957 } 958 AIO_UNLOCK(ki); 959 } 960 961 bool 962 aio_cancel_cleared(struct kaiocb *job) 963 { 964 struct kaioinfo *ki; 965 966 /* 967 * The caller should hold the same queue lock held when 968 * aio_clear_cancel_function() was called and set this flag 969 * ensuring this check sees an up-to-date value. However, 970 * there is no way to assert that. 971 */ 972 ki = job->userproc->p_aioinfo; 973 return ((job->jobflags & KAIOCB_CLEARED) != 0); 974 } 975 976 static bool 977 aio_clear_cancel_function_locked(struct kaiocb *job) 978 { 979 980 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 981 MPASS(job->cancel_fn != NULL); 982 if (job->jobflags & KAIOCB_CANCELLING) { 983 job->jobflags |= KAIOCB_CLEARED; 984 return (false); 985 } 986 job->cancel_fn = NULL; 987 return (true); 988 } 989 990 bool 991 aio_clear_cancel_function(struct kaiocb *job) 992 { 993 struct kaioinfo *ki; 994 bool ret; 995 996 ki = job->userproc->p_aioinfo; 997 AIO_LOCK(ki); 998 ret = aio_clear_cancel_function_locked(job); 999 AIO_UNLOCK(ki); 1000 return (ret); 1001 } 1002 1003 static bool 1004 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func) 1005 { 1006 1007 AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED); 1008 if (job->jobflags & KAIOCB_CANCELLED) 1009 return (false); 1010 job->cancel_fn = func; 1011 return (true); 1012 } 1013 1014 bool 1015 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func) 1016 { 1017 struct kaioinfo *ki; 1018 bool ret; 1019 1020 ki = job->userproc->p_aioinfo; 1021 AIO_LOCK(ki); 1022 ret = aio_set_cancel_function_locked(job, func); 1023 AIO_UNLOCK(ki); 1024 return (ret); 1025 } 1026 1027 void 1028 aio_complete(struct kaiocb *job, long status, int error) 1029 { 1030 struct kaioinfo *ki; 1031 struct proc *userp; 1032 1033 job->uaiocb._aiocb_private.error = error; 1034 job->uaiocb._aiocb_private.status = status; 1035 1036 userp = job->userproc; 1037 ki = userp->p_aioinfo; 1038 1039 AIO_LOCK(ki); 1040 KASSERT(!(job->jobflags & KAIOCB_FINISHED), 1041 ("duplicate aio_complete")); 1042 job->jobflags |= KAIOCB_FINISHED; 1043 if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) { 1044 TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist); 1045 aio_bio_done_notify(userp, job); 1046 } 1047 AIO_UNLOCK(ki); 1048 } 1049 1050 void 1051 aio_cancel(struct kaiocb *job) 1052 { 1053 1054 aio_complete(job, -1, ECANCELED); 1055 } 1056 1057 void 1058 aio_switch_vmspace(struct kaiocb *job) 1059 { 1060 1061 vmspace_switch_aio(job->userproc->p_vmspace); 1062 } 1063 1064 /* 1065 * The AIO daemon, most of the actual work is done in aio_process_*, 1066 * but the setup (and address space mgmt) is done in this routine. 1067 */ 1068 static void 1069 aio_daemon(void *_id) 1070 { 1071 struct kaiocb *job; 1072 struct aioproc *aiop; 1073 struct kaioinfo *ki; 1074 struct proc *p; 1075 struct vmspace *myvm; 1076 struct thread *td = curthread; 1077 int id = (intptr_t)_id; 1078 1079 /* 1080 * Grab an extra reference on the daemon's vmspace so that it 1081 * doesn't get freed by jobs that switch to a different 1082 * vmspace. 1083 */ 1084 p = td->td_proc; 1085 myvm = vmspace_acquire_ref(p); 1086 1087 KASSERT(p->p_textvp == NULL, ("kthread has a textvp")); 1088 1089 /* 1090 * Allocate and ready the aio control info. There is one aiop structure 1091 * per daemon. 1092 */ 1093 aiop = uma_zalloc(aiop_zone, M_WAITOK); 1094 aiop->aioproc = p; 1095 aiop->aioprocflags = 0; 1096 1097 /* 1098 * Wakeup parent process. (Parent sleeps to keep from blasting away 1099 * and creating too many daemons.) 1100 */ 1101 sema_post(&aio_newproc_sem); 1102 1103 mtx_lock(&aio_job_mtx); 1104 for (;;) { 1105 /* 1106 * Take daemon off of free queue 1107 */ 1108 if (aiop->aioprocflags & AIOP_FREE) { 1109 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1110 aiop->aioprocflags &= ~AIOP_FREE; 1111 } 1112 1113 /* 1114 * Check for jobs. 1115 */ 1116 while ((job = aio_selectjob(aiop)) != NULL) { 1117 mtx_unlock(&aio_job_mtx); 1118 1119 ki = job->userproc->p_aioinfo; 1120 job->handle_fn(job); 1121 1122 mtx_lock(&aio_job_mtx); 1123 /* Decrement the active job count. */ 1124 ki->kaio_active_count--; 1125 } 1126 1127 /* 1128 * Disconnect from user address space. 1129 */ 1130 if (p->p_vmspace != myvm) { 1131 mtx_unlock(&aio_job_mtx); 1132 vmspace_switch_aio(myvm); 1133 mtx_lock(&aio_job_mtx); 1134 /* 1135 * We have to restart to avoid race, we only sleep if 1136 * no job can be selected. 1137 */ 1138 continue; 1139 } 1140 1141 mtx_assert(&aio_job_mtx, MA_OWNED); 1142 1143 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1144 aiop->aioprocflags |= AIOP_FREE; 1145 1146 /* 1147 * If daemon is inactive for a long time, allow it to exit, 1148 * thereby freeing resources. 1149 */ 1150 if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy", 1151 aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) && 1152 (aiop->aioprocflags & AIOP_FREE) && 1153 num_aio_procs > target_aio_procs) 1154 break; 1155 } 1156 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1157 num_aio_procs--; 1158 mtx_unlock(&aio_job_mtx); 1159 uma_zfree(aiop_zone, aiop); 1160 free_unr(aiod_unr, id); 1161 vmspace_free(myvm); 1162 1163 KASSERT(p->p_vmspace == myvm, 1164 ("AIOD: bad vmspace for exiting daemon")); 1165 KASSERT(myvm->vm_refcnt > 1, 1166 ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt)); 1167 kproc_exit(0); 1168 } 1169 1170 /* 1171 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1172 * AIO daemon modifies its environment itself. 1173 */ 1174 static int 1175 aio_newproc(int *start) 1176 { 1177 int error; 1178 struct proc *p; 1179 int id; 1180 1181 id = alloc_unr(aiod_unr); 1182 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p, 1183 RFNOWAIT, 0, "aiod%d", id); 1184 if (error == 0) { 1185 /* 1186 * Wait until daemon is started. 1187 */ 1188 sema_wait(&aio_newproc_sem); 1189 mtx_lock(&aio_job_mtx); 1190 num_aio_procs++; 1191 if (start != NULL) 1192 (*start)--; 1193 mtx_unlock(&aio_job_mtx); 1194 } else { 1195 free_unr(aiod_unr, id); 1196 } 1197 return (error); 1198 } 1199 1200 /* 1201 * Try the high-performance, low-overhead physio method for eligible 1202 * VCHR devices. This method doesn't use an aio helper thread, and 1203 * thus has very low overhead. 1204 * 1205 * Assumes that the caller, aio_aqueue(), has incremented the file 1206 * structure's reference count, preventing its deallocation for the 1207 * duration of this call. 1208 */ 1209 static int 1210 aio_qphysio(struct proc *p, struct kaiocb *job) 1211 { 1212 struct aiocb *cb; 1213 struct file *fp; 1214 struct bio *bp; 1215 struct buf *pbuf; 1216 struct vnode *vp; 1217 struct cdevsw *csw; 1218 struct cdev *dev; 1219 struct kaioinfo *ki; 1220 int error, ref, poff; 1221 vm_prot_t prot; 1222 1223 cb = &job->uaiocb; 1224 fp = job->fd_file; 1225 1226 if (fp == NULL || fp->f_type != DTYPE_VNODE) 1227 return (-1); 1228 1229 vp = fp->f_vnode; 1230 if (vp->v_type != VCHR) 1231 return (-1); 1232 if (vp->v_bufobj.bo_bsize == 0) 1233 return (-1); 1234 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize) 1235 return (-1); 1236 1237 ref = 0; 1238 csw = devvn_refthread(vp, &dev, &ref); 1239 if (csw == NULL) 1240 return (ENXIO); 1241 1242 if ((csw->d_flags & D_DISK) == 0) { 1243 error = -1; 1244 goto unref; 1245 } 1246 if (cb->aio_nbytes > dev->si_iosize_max) { 1247 error = -1; 1248 goto unref; 1249 } 1250 1251 ki = p->p_aioinfo; 1252 poff = (vm_offset_t)cb->aio_buf & PAGE_MASK; 1253 if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) { 1254 if (cb->aio_nbytes > MAXPHYS) { 1255 error = -1; 1256 goto unref; 1257 } 1258 1259 pbuf = NULL; 1260 } else { 1261 if (cb->aio_nbytes > MAXPHYS - poff) { 1262 error = -1; 1263 goto unref; 1264 } 1265 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) { 1266 error = -1; 1267 goto unref; 1268 } 1269 1270 job->pbuf = pbuf = (struct buf *)getpbuf(NULL); 1271 BUF_KERNPROC(pbuf); 1272 AIO_LOCK(ki); 1273 ki->kaio_buffer_count++; 1274 AIO_UNLOCK(ki); 1275 } 1276 job->bp = bp = g_alloc_bio(); 1277 1278 bp->bio_length = cb->aio_nbytes; 1279 bp->bio_bcount = cb->aio_nbytes; 1280 bp->bio_done = aio_physwakeup; 1281 bp->bio_data = (void *)(uintptr_t)cb->aio_buf; 1282 bp->bio_offset = cb->aio_offset; 1283 bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ; 1284 bp->bio_dev = dev; 1285 bp->bio_caller1 = (void *)job; 1286 1287 prot = VM_PROT_READ; 1288 if (cb->aio_lio_opcode == LIO_READ) 1289 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 1290 job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 1291 (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages, 1292 nitems(job->pages)); 1293 if (job->npages < 0) { 1294 error = EFAULT; 1295 goto doerror; 1296 } 1297 if (pbuf != NULL) { 1298 pmap_qenter((vm_offset_t)pbuf->b_data, 1299 job->pages, job->npages); 1300 bp->bio_data = pbuf->b_data + poff; 1301 atomic_add_int(&num_buf_aio, 1); 1302 } else { 1303 bp->bio_ma = job->pages; 1304 bp->bio_ma_n = job->npages; 1305 bp->bio_ma_offset = poff; 1306 bp->bio_data = unmapped_buf; 1307 bp->bio_flags |= BIO_UNMAPPED; 1308 } 1309 1310 /* Perform transfer. */ 1311 csw->d_strategy(bp); 1312 dev_relthread(dev, ref); 1313 return (0); 1314 1315 doerror: 1316 if (pbuf != NULL) { 1317 AIO_LOCK(ki); 1318 ki->kaio_buffer_count--; 1319 AIO_UNLOCK(ki); 1320 relpbuf(pbuf, NULL); 1321 job->pbuf = NULL; 1322 } 1323 g_destroy_bio(bp); 1324 job->bp = NULL; 1325 unref: 1326 dev_relthread(dev, ref); 1327 return (error); 1328 } 1329 1330 #ifdef COMPAT_FREEBSD6 1331 static int 1332 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig) 1333 { 1334 1335 /* 1336 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 1337 * supported by AIO with the old sigevent structure. 1338 */ 1339 nsig->sigev_notify = osig->sigev_notify; 1340 switch (nsig->sigev_notify) { 1341 case SIGEV_NONE: 1342 break; 1343 case SIGEV_SIGNAL: 1344 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 1345 break; 1346 case SIGEV_KEVENT: 1347 nsig->sigev_notify_kqueue = 1348 osig->__sigev_u.__sigev_notify_kqueue; 1349 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr; 1350 break; 1351 default: 1352 return (EINVAL); 1353 } 1354 return (0); 1355 } 1356 1357 static int 1358 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob) 1359 { 1360 struct oaiocb *ojob; 1361 int error; 1362 1363 bzero(kjob, sizeof(struct aiocb)); 1364 error = copyin(ujob, kjob, sizeof(struct oaiocb)); 1365 if (error) 1366 return (error); 1367 ojob = (struct oaiocb *)kjob; 1368 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent)); 1369 } 1370 #endif 1371 1372 static int 1373 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob) 1374 { 1375 1376 return (copyin(ujob, kjob, sizeof(struct aiocb))); 1377 } 1378 1379 static long 1380 aiocb_fetch_status(struct aiocb *ujob) 1381 { 1382 1383 return (fuword(&ujob->_aiocb_private.status)); 1384 } 1385 1386 static long 1387 aiocb_fetch_error(struct aiocb *ujob) 1388 { 1389 1390 return (fuword(&ujob->_aiocb_private.error)); 1391 } 1392 1393 static int 1394 aiocb_store_status(struct aiocb *ujob, long status) 1395 { 1396 1397 return (suword(&ujob->_aiocb_private.status, status)); 1398 } 1399 1400 static int 1401 aiocb_store_error(struct aiocb *ujob, long error) 1402 { 1403 1404 return (suword(&ujob->_aiocb_private.error, error)); 1405 } 1406 1407 static int 1408 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref) 1409 { 1410 1411 return (suword(&ujob->_aiocb_private.kernelinfo, jobref)); 1412 } 1413 1414 static int 1415 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 1416 { 1417 1418 return (suword(ujobp, (long)ujob)); 1419 } 1420 1421 static struct aiocb_ops aiocb_ops = { 1422 .copyin = aiocb_copyin, 1423 .fetch_status = aiocb_fetch_status, 1424 .fetch_error = aiocb_fetch_error, 1425 .store_status = aiocb_store_status, 1426 .store_error = aiocb_store_error, 1427 .store_kernelinfo = aiocb_store_kernelinfo, 1428 .store_aiocb = aiocb_store_aiocb, 1429 }; 1430 1431 #ifdef COMPAT_FREEBSD6 1432 static struct aiocb_ops aiocb_ops_osigevent = { 1433 .copyin = aiocb_copyin_old_sigevent, 1434 .fetch_status = aiocb_fetch_status, 1435 .fetch_error = aiocb_fetch_error, 1436 .store_status = aiocb_store_status, 1437 .store_error = aiocb_store_error, 1438 .store_kernelinfo = aiocb_store_kernelinfo, 1439 .store_aiocb = aiocb_store_aiocb, 1440 }; 1441 #endif 1442 1443 /* 1444 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1445 * technique is done in this code. 1446 */ 1447 int 1448 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj, 1449 int type, struct aiocb_ops *ops) 1450 { 1451 struct proc *p = td->td_proc; 1452 cap_rights_t rights; 1453 struct file *fp; 1454 struct kaiocb *job; 1455 struct kaioinfo *ki; 1456 struct kevent kev; 1457 int opcode; 1458 int error; 1459 int fd, kqfd; 1460 int jid; 1461 u_short evflags; 1462 1463 if (p->p_aioinfo == NULL) 1464 aio_init_aioinfo(p); 1465 1466 ki = p->p_aioinfo; 1467 1468 ops->store_status(ujob, -1); 1469 ops->store_error(ujob, 0); 1470 ops->store_kernelinfo(ujob, -1); 1471 1472 if (num_queue_count >= max_queue_count || 1473 ki->kaio_count >= ki->kaio_qallowed_count) { 1474 ops->store_error(ujob, EAGAIN); 1475 return (EAGAIN); 1476 } 1477 1478 job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1479 knlist_init_mtx(&job->klist, AIO_MTX(ki)); 1480 1481 error = ops->copyin(ujob, &job->uaiocb); 1482 if (error) { 1483 ops->store_error(ujob, error); 1484 uma_zfree(aiocb_zone, job); 1485 return (error); 1486 } 1487 1488 if (job->uaiocb.aio_nbytes > IOSIZE_MAX) { 1489 uma_zfree(aiocb_zone, job); 1490 return (EINVAL); 1491 } 1492 1493 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1494 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1495 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1496 job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1497 ops->store_error(ujob, EINVAL); 1498 uma_zfree(aiocb_zone, job); 1499 return (EINVAL); 1500 } 1501 1502 if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1503 job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1504 !_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) { 1505 uma_zfree(aiocb_zone, job); 1506 return (EINVAL); 1507 } 1508 1509 ksiginfo_init(&job->ksi); 1510 1511 /* Save userspace address of the job info. */ 1512 job->ujob = ujob; 1513 1514 /* Get the opcode. */ 1515 if (type != LIO_NOP) 1516 job->uaiocb.aio_lio_opcode = type; 1517 opcode = job->uaiocb.aio_lio_opcode; 1518 1519 /* 1520 * Validate the opcode and fetch the file object for the specified 1521 * file descriptor. 1522 * 1523 * XXXRW: Moved the opcode validation up here so that we don't 1524 * retrieve a file descriptor without knowing what the capabiltity 1525 * should be. 1526 */ 1527 fd = job->uaiocb.aio_fildes; 1528 switch (opcode) { 1529 case LIO_WRITE: 1530 error = fget_write(td, fd, 1531 cap_rights_init(&rights, CAP_PWRITE), &fp); 1532 break; 1533 case LIO_READ: 1534 error = fget_read(td, fd, 1535 cap_rights_init(&rights, CAP_PREAD), &fp); 1536 break; 1537 case LIO_SYNC: 1538 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp); 1539 break; 1540 case LIO_MLOCK: 1541 fp = NULL; 1542 break; 1543 case LIO_NOP: 1544 error = fget(td, fd, cap_rights_init(&rights), &fp); 1545 break; 1546 default: 1547 error = EINVAL; 1548 } 1549 if (error) { 1550 uma_zfree(aiocb_zone, job); 1551 ops->store_error(ujob, error); 1552 return (error); 1553 } 1554 1555 if (opcode == LIO_SYNC && fp->f_vnode == NULL) { 1556 error = EINVAL; 1557 goto aqueue_fail; 1558 } 1559 1560 if ((opcode == LIO_READ || opcode == LIO_WRITE) && 1561 job->uaiocb.aio_offset < 0 && 1562 (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) { 1563 error = EINVAL; 1564 goto aqueue_fail; 1565 } 1566 1567 job->fd_file = fp; 1568 1569 mtx_lock(&aio_job_mtx); 1570 jid = jobrefid++; 1571 job->seqno = jobseqno++; 1572 mtx_unlock(&aio_job_mtx); 1573 error = ops->store_kernelinfo(ujob, jid); 1574 if (error) { 1575 error = EINVAL; 1576 goto aqueue_fail; 1577 } 1578 job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid; 1579 1580 if (opcode == LIO_NOP) { 1581 fdrop(fp, td); 1582 uma_zfree(aiocb_zone, job); 1583 return (0); 1584 } 1585 1586 if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT) 1587 goto no_kqueue; 1588 evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags; 1589 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) { 1590 error = EINVAL; 1591 goto aqueue_fail; 1592 } 1593 kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue; 1594 kev.ident = (uintptr_t)job->ujob; 1595 kev.filter = EVFILT_AIO; 1596 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags; 1597 kev.data = (intptr_t)job; 1598 kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1599 error = kqfd_register(kqfd, &kev, td, 1); 1600 if (error) 1601 goto aqueue_fail; 1602 1603 no_kqueue: 1604 1605 ops->store_error(ujob, EINPROGRESS); 1606 job->uaiocb._aiocb_private.error = EINPROGRESS; 1607 job->userproc = p; 1608 job->cred = crhold(td->td_ucred); 1609 job->jobflags = KAIOCB_QUEUEING; 1610 job->lio = lj; 1611 1612 if (opcode == LIO_MLOCK) { 1613 aio_schedule(job, aio_process_mlock); 1614 error = 0; 1615 } else if (fp->f_ops->fo_aio_queue == NULL) 1616 error = aio_queue_file(fp, job); 1617 else 1618 error = fo_aio_queue(fp, job); 1619 if (error) 1620 goto aqueue_fail; 1621 1622 AIO_LOCK(ki); 1623 job->jobflags &= ~KAIOCB_QUEUEING; 1624 TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist); 1625 ki->kaio_count++; 1626 if (lj) 1627 lj->lioj_count++; 1628 atomic_add_int(&num_queue_count, 1); 1629 if (job->jobflags & KAIOCB_FINISHED) { 1630 /* 1631 * The queue callback completed the request synchronously. 1632 * The bulk of the completion is deferred in that case 1633 * until this point. 1634 */ 1635 aio_bio_done_notify(p, job); 1636 } else 1637 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist); 1638 AIO_UNLOCK(ki); 1639 return (0); 1640 1641 aqueue_fail: 1642 knlist_delete(&job->klist, curthread, 0); 1643 if (fp) 1644 fdrop(fp, td); 1645 uma_zfree(aiocb_zone, job); 1646 ops->store_error(ujob, error); 1647 return (error); 1648 } 1649 1650 static void 1651 aio_cancel_daemon_job(struct kaiocb *job) 1652 { 1653 1654 mtx_lock(&aio_job_mtx); 1655 if (!aio_cancel_cleared(job)) 1656 TAILQ_REMOVE(&aio_jobs, job, list); 1657 mtx_unlock(&aio_job_mtx); 1658 aio_cancel(job); 1659 } 1660 1661 void 1662 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func) 1663 { 1664 1665 mtx_lock(&aio_job_mtx); 1666 if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) { 1667 mtx_unlock(&aio_job_mtx); 1668 aio_cancel(job); 1669 return; 1670 } 1671 job->handle_fn = func; 1672 TAILQ_INSERT_TAIL(&aio_jobs, job, list); 1673 aio_kick_nowait(job->userproc); 1674 mtx_unlock(&aio_job_mtx); 1675 } 1676 1677 static void 1678 aio_cancel_sync(struct kaiocb *job) 1679 { 1680 struct kaioinfo *ki; 1681 1682 ki = job->userproc->p_aioinfo; 1683 AIO_LOCK(ki); 1684 if (!aio_cancel_cleared(job)) 1685 TAILQ_REMOVE(&ki->kaio_syncqueue, job, list); 1686 AIO_UNLOCK(ki); 1687 aio_cancel(job); 1688 } 1689 1690 int 1691 aio_queue_file(struct file *fp, struct kaiocb *job) 1692 { 1693 struct aioliojob *lj; 1694 struct kaioinfo *ki; 1695 struct kaiocb *job2; 1696 struct vnode *vp; 1697 struct mount *mp; 1698 int error, opcode; 1699 bool safe; 1700 1701 lj = job->lio; 1702 ki = job->userproc->p_aioinfo; 1703 opcode = job->uaiocb.aio_lio_opcode; 1704 if (opcode == LIO_SYNC) 1705 goto queueit; 1706 1707 if ((error = aio_qphysio(job->userproc, job)) == 0) 1708 goto done; 1709 #if 0 1710 /* 1711 * XXX: This means qphysio() failed with EFAULT. The current 1712 * behavior is to retry the operation via fo_read/fo_write. 1713 * Wouldn't it be better to just complete the request with an 1714 * error here? 1715 */ 1716 if (error > 0) 1717 goto done; 1718 #endif 1719 queueit: 1720 safe = false; 1721 if (fp->f_type == DTYPE_VNODE) { 1722 vp = fp->f_vnode; 1723 if (vp->v_type == VREG || vp->v_type == VDIR) { 1724 mp = fp->f_vnode->v_mount; 1725 if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0) 1726 safe = true; 1727 } 1728 } 1729 if (!(safe || enable_aio_unsafe)) { 1730 counted_warning(&unsafe_warningcnt, 1731 "is attempting to use unsafe AIO requests"); 1732 return (EOPNOTSUPP); 1733 } 1734 1735 if (opcode == LIO_SYNC) { 1736 AIO_LOCK(ki); 1737 TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) { 1738 if (job2->fd_file == job->fd_file && 1739 job2->uaiocb.aio_lio_opcode != LIO_SYNC && 1740 job2->seqno < job->seqno) { 1741 job2->jobflags |= KAIOCB_CHECKSYNC; 1742 job->pending++; 1743 } 1744 } 1745 if (job->pending != 0) { 1746 if (!aio_set_cancel_function_locked(job, 1747 aio_cancel_sync)) { 1748 AIO_UNLOCK(ki); 1749 aio_cancel(job); 1750 return (0); 1751 } 1752 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list); 1753 AIO_UNLOCK(ki); 1754 return (0); 1755 } 1756 AIO_UNLOCK(ki); 1757 } 1758 1759 switch (opcode) { 1760 case LIO_READ: 1761 case LIO_WRITE: 1762 aio_schedule(job, aio_process_rw); 1763 error = 0; 1764 break; 1765 case LIO_SYNC: 1766 aio_schedule(job, aio_process_sync); 1767 error = 0; 1768 break; 1769 default: 1770 error = EINVAL; 1771 } 1772 done: 1773 return (error); 1774 } 1775 1776 static void 1777 aio_kick_nowait(struct proc *userp) 1778 { 1779 struct kaioinfo *ki = userp->p_aioinfo; 1780 struct aioproc *aiop; 1781 1782 mtx_assert(&aio_job_mtx, MA_OWNED); 1783 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1784 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1785 aiop->aioprocflags &= ~AIOP_FREE; 1786 wakeup(aiop->aioproc); 1787 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1788 ki->kaio_active_count + num_aio_resv_start < 1789 ki->kaio_maxactive_count) { 1790 taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task); 1791 } 1792 } 1793 1794 static int 1795 aio_kick(struct proc *userp) 1796 { 1797 struct kaioinfo *ki = userp->p_aioinfo; 1798 struct aioproc *aiop; 1799 int error, ret = 0; 1800 1801 mtx_assert(&aio_job_mtx, MA_OWNED); 1802 retryproc: 1803 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1804 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1805 aiop->aioprocflags &= ~AIOP_FREE; 1806 wakeup(aiop->aioproc); 1807 } else if (num_aio_resv_start + num_aio_procs < max_aio_procs && 1808 ki->kaio_active_count + num_aio_resv_start < 1809 ki->kaio_maxactive_count) { 1810 num_aio_resv_start++; 1811 mtx_unlock(&aio_job_mtx); 1812 error = aio_newproc(&num_aio_resv_start); 1813 mtx_lock(&aio_job_mtx); 1814 if (error) { 1815 num_aio_resv_start--; 1816 goto retryproc; 1817 } 1818 } else { 1819 ret = -1; 1820 } 1821 return (ret); 1822 } 1823 1824 static void 1825 aio_kick_helper(void *context, int pending) 1826 { 1827 struct proc *userp = context; 1828 1829 mtx_lock(&aio_job_mtx); 1830 while (--pending >= 0) { 1831 if (aio_kick(userp)) 1832 break; 1833 } 1834 mtx_unlock(&aio_job_mtx); 1835 } 1836 1837 /* 1838 * Support the aio_return system call, as a side-effect, kernel resources are 1839 * released. 1840 */ 1841 static int 1842 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 1843 { 1844 struct proc *p = td->td_proc; 1845 struct kaiocb *job; 1846 struct kaioinfo *ki; 1847 long status, error; 1848 1849 ki = p->p_aioinfo; 1850 if (ki == NULL) 1851 return (EINVAL); 1852 AIO_LOCK(ki); 1853 TAILQ_FOREACH(job, &ki->kaio_done, plist) { 1854 if (job->ujob == ujob) 1855 break; 1856 } 1857 if (job != NULL) { 1858 MPASS(job->jobflags & KAIOCB_FINISHED); 1859 status = job->uaiocb._aiocb_private.status; 1860 error = job->uaiocb._aiocb_private.error; 1861 td->td_retval[0] = status; 1862 td->td_ru.ru_oublock += job->outblock; 1863 td->td_ru.ru_inblock += job->inblock; 1864 td->td_ru.ru_msgsnd += job->msgsnd; 1865 td->td_ru.ru_msgrcv += job->msgrcv; 1866 aio_free_entry(job); 1867 AIO_UNLOCK(ki); 1868 ops->store_error(ujob, error); 1869 ops->store_status(ujob, status); 1870 } else { 1871 error = EINVAL; 1872 AIO_UNLOCK(ki); 1873 } 1874 return (error); 1875 } 1876 1877 int 1878 sys_aio_return(struct thread *td, struct aio_return_args *uap) 1879 { 1880 1881 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops)); 1882 } 1883 1884 /* 1885 * Allow a process to wakeup when any of the I/O requests are completed. 1886 */ 1887 static int 1888 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist, 1889 struct timespec *ts) 1890 { 1891 struct proc *p = td->td_proc; 1892 struct timeval atv; 1893 struct kaioinfo *ki; 1894 struct kaiocb *firstjob, *job; 1895 int error, i, timo; 1896 1897 timo = 0; 1898 if (ts) { 1899 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000) 1900 return (EINVAL); 1901 1902 TIMESPEC_TO_TIMEVAL(&atv, ts); 1903 if (itimerfix(&atv)) 1904 return (EINVAL); 1905 timo = tvtohz(&atv); 1906 } 1907 1908 ki = p->p_aioinfo; 1909 if (ki == NULL) 1910 return (EAGAIN); 1911 1912 if (njoblist == 0) 1913 return (0); 1914 1915 AIO_LOCK(ki); 1916 for (;;) { 1917 firstjob = NULL; 1918 error = 0; 1919 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 1920 for (i = 0; i < njoblist; i++) { 1921 if (job->ujob == ujoblist[i]) { 1922 if (firstjob == NULL) 1923 firstjob = job; 1924 if (job->jobflags & KAIOCB_FINISHED) 1925 goto RETURN; 1926 } 1927 } 1928 } 1929 /* All tasks were finished. */ 1930 if (firstjob == NULL) 1931 break; 1932 1933 ki->kaio_flags |= KAIO_WAKEUP; 1934 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 1935 "aiospn", timo); 1936 if (error == ERESTART) 1937 error = EINTR; 1938 if (error) 1939 break; 1940 } 1941 RETURN: 1942 AIO_UNLOCK(ki); 1943 return (error); 1944 } 1945 1946 int 1947 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap) 1948 { 1949 struct timespec ts, *tsp; 1950 struct aiocb **ujoblist; 1951 int error; 1952 1953 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 1954 return (EINVAL); 1955 1956 if (uap->timeout) { 1957 /* Get timespec struct. */ 1958 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 1959 return (error); 1960 tsp = &ts; 1961 } else 1962 tsp = NULL; 1963 1964 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK); 1965 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0])); 1966 if (error == 0) 1967 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 1968 free(ujoblist, M_AIOS); 1969 return (error); 1970 } 1971 1972 /* 1973 * aio_cancel cancels any non-physio aio operations not currently in 1974 * progress. 1975 */ 1976 int 1977 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap) 1978 { 1979 struct proc *p = td->td_proc; 1980 struct kaioinfo *ki; 1981 struct kaiocb *job, *jobn; 1982 struct file *fp; 1983 cap_rights_t rights; 1984 int error; 1985 int cancelled = 0; 1986 int notcancelled = 0; 1987 struct vnode *vp; 1988 1989 /* Lookup file object. */ 1990 error = fget(td, uap->fd, cap_rights_init(&rights), &fp); 1991 if (error) 1992 return (error); 1993 1994 ki = p->p_aioinfo; 1995 if (ki == NULL) 1996 goto done; 1997 1998 if (fp->f_type == DTYPE_VNODE) { 1999 vp = fp->f_vnode; 2000 if (vn_isdisk(vp, &error)) { 2001 fdrop(fp, td); 2002 td->td_retval[0] = AIO_NOTCANCELED; 2003 return (0); 2004 } 2005 } 2006 2007 AIO_LOCK(ki); 2008 TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) { 2009 if ((uap->fd == job->uaiocb.aio_fildes) && 2010 ((uap->aiocbp == NULL) || 2011 (uap->aiocbp == job->ujob))) { 2012 if (aio_cancel_job(p, ki, job)) { 2013 cancelled++; 2014 } else { 2015 notcancelled++; 2016 } 2017 if (uap->aiocbp != NULL) 2018 break; 2019 } 2020 } 2021 AIO_UNLOCK(ki); 2022 2023 done: 2024 fdrop(fp, td); 2025 2026 if (uap->aiocbp != NULL) { 2027 if (cancelled) { 2028 td->td_retval[0] = AIO_CANCELED; 2029 return (0); 2030 } 2031 } 2032 2033 if (notcancelled) { 2034 td->td_retval[0] = AIO_NOTCANCELED; 2035 return (0); 2036 } 2037 2038 if (cancelled) { 2039 td->td_retval[0] = AIO_CANCELED; 2040 return (0); 2041 } 2042 2043 td->td_retval[0] = AIO_ALLDONE; 2044 2045 return (0); 2046 } 2047 2048 /* 2049 * aio_error is implemented in the kernel level for compatibility purposes 2050 * only. For a user mode async implementation, it would be best to do it in 2051 * a userland subroutine. 2052 */ 2053 static int 2054 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops) 2055 { 2056 struct proc *p = td->td_proc; 2057 struct kaiocb *job; 2058 struct kaioinfo *ki; 2059 int status; 2060 2061 ki = p->p_aioinfo; 2062 if (ki == NULL) { 2063 td->td_retval[0] = EINVAL; 2064 return (0); 2065 } 2066 2067 AIO_LOCK(ki); 2068 TAILQ_FOREACH(job, &ki->kaio_all, allist) { 2069 if (job->ujob == ujob) { 2070 if (job->jobflags & KAIOCB_FINISHED) 2071 td->td_retval[0] = 2072 job->uaiocb._aiocb_private.error; 2073 else 2074 td->td_retval[0] = EINPROGRESS; 2075 AIO_UNLOCK(ki); 2076 return (0); 2077 } 2078 } 2079 AIO_UNLOCK(ki); 2080 2081 /* 2082 * Hack for failure of aio_aqueue. 2083 */ 2084 status = ops->fetch_status(ujob); 2085 if (status == -1) { 2086 td->td_retval[0] = ops->fetch_error(ujob); 2087 return (0); 2088 } 2089 2090 td->td_retval[0] = EINVAL; 2091 return (0); 2092 } 2093 2094 int 2095 sys_aio_error(struct thread *td, struct aio_error_args *uap) 2096 { 2097 2098 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops)); 2099 } 2100 2101 /* syscall - asynchronous read from a file (REALTIME) */ 2102 #ifdef COMPAT_FREEBSD6 2103 int 2104 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap) 2105 { 2106 2107 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2108 &aiocb_ops_osigevent)); 2109 } 2110 #endif 2111 2112 int 2113 sys_aio_read(struct thread *td, struct aio_read_args *uap) 2114 { 2115 2116 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops)); 2117 } 2118 2119 /* syscall - asynchronous write to a file (REALTIME) */ 2120 #ifdef COMPAT_FREEBSD6 2121 int 2122 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap) 2123 { 2124 2125 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2126 &aiocb_ops_osigevent)); 2127 } 2128 #endif 2129 2130 int 2131 sys_aio_write(struct thread *td, struct aio_write_args *uap) 2132 { 2133 2134 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops)); 2135 } 2136 2137 int 2138 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap) 2139 { 2140 2141 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops)); 2142 } 2143 2144 static int 2145 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list, 2146 struct aiocb **acb_list, int nent, struct sigevent *sig, 2147 struct aiocb_ops *ops) 2148 { 2149 struct proc *p = td->td_proc; 2150 struct aiocb *job; 2151 struct kaioinfo *ki; 2152 struct aioliojob *lj; 2153 struct kevent kev; 2154 int error; 2155 int nerror; 2156 int i; 2157 2158 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT)) 2159 return (EINVAL); 2160 2161 if (nent < 0 || nent > max_aio_queue_per_proc) 2162 return (EINVAL); 2163 2164 if (p->p_aioinfo == NULL) 2165 aio_init_aioinfo(p); 2166 2167 ki = p->p_aioinfo; 2168 2169 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2170 lj->lioj_flags = 0; 2171 lj->lioj_count = 0; 2172 lj->lioj_finished_count = 0; 2173 knlist_init_mtx(&lj->klist, AIO_MTX(ki)); 2174 ksiginfo_init(&lj->lioj_ksi); 2175 2176 /* 2177 * Setup signal. 2178 */ 2179 if (sig && (mode == LIO_NOWAIT)) { 2180 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal)); 2181 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2182 /* Assume only new style KEVENT */ 2183 kev.filter = EVFILT_LIO; 2184 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2185 kev.ident = (uintptr_t)uacb_list; /* something unique */ 2186 kev.data = (intptr_t)lj; 2187 /* pass user defined sigval data */ 2188 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2189 error = kqfd_register( 2190 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1); 2191 if (error) { 2192 uma_zfree(aiolio_zone, lj); 2193 return (error); 2194 } 2195 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2196 ; 2197 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2198 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2199 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2200 uma_zfree(aiolio_zone, lj); 2201 return EINVAL; 2202 } 2203 lj->lioj_flags |= LIOJ_SIGNAL; 2204 } else { 2205 uma_zfree(aiolio_zone, lj); 2206 return EINVAL; 2207 } 2208 } 2209 2210 AIO_LOCK(ki); 2211 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2212 /* 2213 * Add extra aiocb count to avoid the lio to be freed 2214 * by other threads doing aio_waitcomplete or aio_return, 2215 * and prevent event from being sent until we have queued 2216 * all tasks. 2217 */ 2218 lj->lioj_count = 1; 2219 AIO_UNLOCK(ki); 2220 2221 /* 2222 * Get pointers to the list of I/O requests. 2223 */ 2224 nerror = 0; 2225 for (i = 0; i < nent; i++) { 2226 job = acb_list[i]; 2227 if (job != NULL) { 2228 error = aio_aqueue(td, job, lj, LIO_NOP, ops); 2229 if (error != 0) 2230 nerror++; 2231 } 2232 } 2233 2234 error = 0; 2235 AIO_LOCK(ki); 2236 if (mode == LIO_WAIT) { 2237 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2238 ki->kaio_flags |= KAIO_WAKEUP; 2239 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2240 PRIBIO | PCATCH, "aiospn", 0); 2241 if (error == ERESTART) 2242 error = EINTR; 2243 if (error) 2244 break; 2245 } 2246 } else { 2247 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2248 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2249 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2250 KNOTE_LOCKED(&lj->klist, 1); 2251 } 2252 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 2253 == LIOJ_SIGNAL 2254 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2255 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2256 aio_sendsig(p, &lj->lioj_signal, 2257 &lj->lioj_ksi); 2258 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2259 } 2260 } 2261 } 2262 lj->lioj_count--; 2263 if (lj->lioj_count == 0) { 2264 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2265 knlist_delete(&lj->klist, curthread, 1); 2266 PROC_LOCK(p); 2267 sigqueue_take(&lj->lioj_ksi); 2268 PROC_UNLOCK(p); 2269 AIO_UNLOCK(ki); 2270 uma_zfree(aiolio_zone, lj); 2271 } else 2272 AIO_UNLOCK(ki); 2273 2274 if (nerror) 2275 return (EIO); 2276 return (error); 2277 } 2278 2279 /* syscall - list directed I/O (REALTIME) */ 2280 #ifdef COMPAT_FREEBSD6 2281 int 2282 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap) 2283 { 2284 struct aiocb **acb_list; 2285 struct sigevent *sigp, sig; 2286 struct osigevent osig; 2287 int error, nent; 2288 2289 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2290 return (EINVAL); 2291 2292 nent = uap->nent; 2293 if (nent < 0 || nent > max_aio_queue_per_proc) 2294 return (EINVAL); 2295 2296 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2297 error = copyin(uap->sig, &osig, sizeof(osig)); 2298 if (error) 2299 return (error); 2300 error = convert_old_sigevent(&osig, &sig); 2301 if (error) 2302 return (error); 2303 sigp = &sig; 2304 } else 2305 sigp = NULL; 2306 2307 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2308 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2309 if (error == 0) 2310 error = kern_lio_listio(td, uap->mode, 2311 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2312 &aiocb_ops_osigevent); 2313 free(acb_list, M_LIO); 2314 return (error); 2315 } 2316 #endif 2317 2318 /* syscall - list directed I/O (REALTIME) */ 2319 int 2320 sys_lio_listio(struct thread *td, struct lio_listio_args *uap) 2321 { 2322 struct aiocb **acb_list; 2323 struct sigevent *sigp, sig; 2324 int error, nent; 2325 2326 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2327 return (EINVAL); 2328 2329 nent = uap->nent; 2330 if (nent < 0 || nent > max_aio_queue_per_proc) 2331 return (EINVAL); 2332 2333 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2334 error = copyin(uap->sig, &sig, sizeof(sig)); 2335 if (error) 2336 return (error); 2337 sigp = &sig; 2338 } else 2339 sigp = NULL; 2340 2341 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2342 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2343 if (error == 0) 2344 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list, 2345 nent, sigp, &aiocb_ops); 2346 free(acb_list, M_LIO); 2347 return (error); 2348 } 2349 2350 static void 2351 aio_physwakeup(struct bio *bp) 2352 { 2353 struct kaiocb *job = (struct kaiocb *)bp->bio_caller1; 2354 struct proc *userp; 2355 struct kaioinfo *ki; 2356 size_t nbytes; 2357 int error, nblks; 2358 2359 /* Release mapping into kernel space. */ 2360 userp = job->userproc; 2361 ki = userp->p_aioinfo; 2362 if (job->pbuf) { 2363 pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages); 2364 relpbuf(job->pbuf, NULL); 2365 job->pbuf = NULL; 2366 atomic_subtract_int(&num_buf_aio, 1); 2367 AIO_LOCK(ki); 2368 ki->kaio_buffer_count--; 2369 AIO_UNLOCK(ki); 2370 } 2371 vm_page_unhold_pages(job->pages, job->npages); 2372 2373 bp = job->bp; 2374 job->bp = NULL; 2375 nbytes = job->uaiocb.aio_nbytes - bp->bio_resid; 2376 error = 0; 2377 if (bp->bio_flags & BIO_ERROR) 2378 error = bp->bio_error; 2379 nblks = btodb(nbytes); 2380 if (job->uaiocb.aio_lio_opcode == LIO_WRITE) 2381 job->outblock += nblks; 2382 else 2383 job->inblock += nblks; 2384 2385 if (error) 2386 aio_complete(job, -1, error); 2387 else 2388 aio_complete(job, nbytes, 0); 2389 2390 g_destroy_bio(bp); 2391 } 2392 2393 /* syscall - wait for the next completion of an aio request */ 2394 static int 2395 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp, 2396 struct timespec *ts, struct aiocb_ops *ops) 2397 { 2398 struct proc *p = td->td_proc; 2399 struct timeval atv; 2400 struct kaioinfo *ki; 2401 struct kaiocb *job; 2402 struct aiocb *ujob; 2403 long error, status; 2404 int timo; 2405 2406 ops->store_aiocb(ujobp, NULL); 2407 2408 if (ts == NULL) { 2409 timo = 0; 2410 } else if (ts->tv_sec == 0 && ts->tv_nsec == 0) { 2411 timo = -1; 2412 } else { 2413 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000)) 2414 return (EINVAL); 2415 2416 TIMESPEC_TO_TIMEVAL(&atv, ts); 2417 if (itimerfix(&atv)) 2418 return (EINVAL); 2419 timo = tvtohz(&atv); 2420 } 2421 2422 if (p->p_aioinfo == NULL) 2423 aio_init_aioinfo(p); 2424 ki = p->p_aioinfo; 2425 2426 error = 0; 2427 job = NULL; 2428 AIO_LOCK(ki); 2429 while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2430 if (timo == -1) { 2431 error = EWOULDBLOCK; 2432 break; 2433 } 2434 ki->kaio_flags |= KAIO_WAKEUP; 2435 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2436 "aiowc", timo); 2437 if (timo && error == ERESTART) 2438 error = EINTR; 2439 if (error) 2440 break; 2441 } 2442 2443 if (job != NULL) { 2444 MPASS(job->jobflags & KAIOCB_FINISHED); 2445 ujob = job->ujob; 2446 status = job->uaiocb._aiocb_private.status; 2447 error = job->uaiocb._aiocb_private.error; 2448 td->td_retval[0] = status; 2449 td->td_ru.ru_oublock += job->outblock; 2450 td->td_ru.ru_inblock += job->inblock; 2451 td->td_ru.ru_msgsnd += job->msgsnd; 2452 td->td_ru.ru_msgrcv += job->msgrcv; 2453 aio_free_entry(job); 2454 AIO_UNLOCK(ki); 2455 ops->store_aiocb(ujobp, ujob); 2456 ops->store_error(ujob, error); 2457 ops->store_status(ujob, status); 2458 } else 2459 AIO_UNLOCK(ki); 2460 2461 return (error); 2462 } 2463 2464 int 2465 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2466 { 2467 struct timespec ts, *tsp; 2468 int error; 2469 2470 if (uap->timeout) { 2471 /* Get timespec struct. */ 2472 error = copyin(uap->timeout, &ts, sizeof(ts)); 2473 if (error) 2474 return (error); 2475 tsp = &ts; 2476 } else 2477 tsp = NULL; 2478 2479 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops)); 2480 } 2481 2482 static int 2483 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob, 2484 struct aiocb_ops *ops) 2485 { 2486 2487 if (op != O_SYNC) /* XXX lack of O_DSYNC */ 2488 return (EINVAL); 2489 return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops)); 2490 } 2491 2492 int 2493 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2494 { 2495 2496 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops)); 2497 } 2498 2499 /* kqueue attach function */ 2500 static int 2501 filt_aioattach(struct knote *kn) 2502 { 2503 struct kaiocb *job; 2504 2505 job = (struct kaiocb *)(uintptr_t)kn->kn_sdata; 2506 2507 /* 2508 * The job pointer must be validated before using it, so 2509 * registration is restricted to the kernel; the user cannot 2510 * set EV_FLAG1. 2511 */ 2512 if ((kn->kn_flags & EV_FLAG1) == 0) 2513 return (EPERM); 2514 kn->kn_ptr.p_aio = job; 2515 kn->kn_flags &= ~EV_FLAG1; 2516 2517 knlist_add(&job->klist, kn, 0); 2518 2519 return (0); 2520 } 2521 2522 /* kqueue detach function */ 2523 static void 2524 filt_aiodetach(struct knote *kn) 2525 { 2526 struct knlist *knl; 2527 2528 knl = &kn->kn_ptr.p_aio->klist; 2529 knl->kl_lock(knl->kl_lockarg); 2530 if (!knlist_empty(knl)) 2531 knlist_remove(knl, kn, 1); 2532 knl->kl_unlock(knl->kl_lockarg); 2533 } 2534 2535 /* kqueue filter function */ 2536 /*ARGSUSED*/ 2537 static int 2538 filt_aio(struct knote *kn, long hint) 2539 { 2540 struct kaiocb *job = kn->kn_ptr.p_aio; 2541 2542 kn->kn_data = job->uaiocb._aiocb_private.error; 2543 if (!(job->jobflags & KAIOCB_FINISHED)) 2544 return (0); 2545 kn->kn_flags |= EV_EOF; 2546 return (1); 2547 } 2548 2549 /* kqueue attach function */ 2550 static int 2551 filt_lioattach(struct knote *kn) 2552 { 2553 struct aioliojob *lj; 2554 2555 lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata; 2556 2557 /* 2558 * The aioliojob pointer must be validated before using it, so 2559 * registration is restricted to the kernel; the user cannot 2560 * set EV_FLAG1. 2561 */ 2562 if ((kn->kn_flags & EV_FLAG1) == 0) 2563 return (EPERM); 2564 kn->kn_ptr.p_lio = lj; 2565 kn->kn_flags &= ~EV_FLAG1; 2566 2567 knlist_add(&lj->klist, kn, 0); 2568 2569 return (0); 2570 } 2571 2572 /* kqueue detach function */ 2573 static void 2574 filt_liodetach(struct knote *kn) 2575 { 2576 struct knlist *knl; 2577 2578 knl = &kn->kn_ptr.p_lio->klist; 2579 knl->kl_lock(knl->kl_lockarg); 2580 if (!knlist_empty(knl)) 2581 knlist_remove(knl, kn, 1); 2582 knl->kl_unlock(knl->kl_lockarg); 2583 } 2584 2585 /* kqueue filter function */ 2586 /*ARGSUSED*/ 2587 static int 2588 filt_lio(struct knote *kn, long hint) 2589 { 2590 struct aioliojob * lj = kn->kn_ptr.p_lio; 2591 2592 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2593 } 2594 2595 #ifdef COMPAT_FREEBSD32 2596 #include <sys/mount.h> 2597 #include <sys/socket.h> 2598 #include <compat/freebsd32/freebsd32.h> 2599 #include <compat/freebsd32/freebsd32_proto.h> 2600 #include <compat/freebsd32/freebsd32_signal.h> 2601 #include <compat/freebsd32/freebsd32_syscall.h> 2602 #include <compat/freebsd32/freebsd32_util.h> 2603 2604 struct __aiocb_private32 { 2605 int32_t status; 2606 int32_t error; 2607 uint32_t kernelinfo; 2608 }; 2609 2610 #ifdef COMPAT_FREEBSD6 2611 typedef struct oaiocb32 { 2612 int aio_fildes; /* File descriptor */ 2613 uint64_t aio_offset __packed; /* File offset for I/O */ 2614 uint32_t aio_buf; /* I/O buffer in process space */ 2615 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2616 struct osigevent32 aio_sigevent; /* Signal to deliver */ 2617 int aio_lio_opcode; /* LIO opcode */ 2618 int aio_reqprio; /* Request priority -- ignored */ 2619 struct __aiocb_private32 _aiocb_private; 2620 } oaiocb32_t; 2621 #endif 2622 2623 typedef struct aiocb32 { 2624 int32_t aio_fildes; /* File descriptor */ 2625 uint64_t aio_offset __packed; /* File offset for I/O */ 2626 uint32_t aio_buf; /* I/O buffer in process space */ 2627 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2628 int __spare__[2]; 2629 uint32_t __spare2__; 2630 int aio_lio_opcode; /* LIO opcode */ 2631 int aio_reqprio; /* Request priority -- ignored */ 2632 struct __aiocb_private32 _aiocb_private; 2633 struct sigevent32 aio_sigevent; /* Signal to deliver */ 2634 } aiocb32_t; 2635 2636 #ifdef COMPAT_FREEBSD6 2637 static int 2638 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig) 2639 { 2640 2641 /* 2642 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 2643 * supported by AIO with the old sigevent structure. 2644 */ 2645 CP(*osig, *nsig, sigev_notify); 2646 switch (nsig->sigev_notify) { 2647 case SIGEV_NONE: 2648 break; 2649 case SIGEV_SIGNAL: 2650 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 2651 break; 2652 case SIGEV_KEVENT: 2653 nsig->sigev_notify_kqueue = 2654 osig->__sigev_u.__sigev_notify_kqueue; 2655 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr); 2656 break; 2657 default: 2658 return (EINVAL); 2659 } 2660 return (0); 2661 } 2662 2663 static int 2664 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob) 2665 { 2666 struct oaiocb32 job32; 2667 int error; 2668 2669 bzero(kjob, sizeof(struct aiocb)); 2670 error = copyin(ujob, &job32, sizeof(job32)); 2671 if (error) 2672 return (error); 2673 2674 CP(job32, *kjob, aio_fildes); 2675 CP(job32, *kjob, aio_offset); 2676 PTRIN_CP(job32, *kjob, aio_buf); 2677 CP(job32, *kjob, aio_nbytes); 2678 CP(job32, *kjob, aio_lio_opcode); 2679 CP(job32, *kjob, aio_reqprio); 2680 CP(job32, *kjob, _aiocb_private.status); 2681 CP(job32, *kjob, _aiocb_private.error); 2682 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo); 2683 return (convert_old_sigevent32(&job32.aio_sigevent, 2684 &kjob->aio_sigevent)); 2685 } 2686 #endif 2687 2688 static int 2689 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob) 2690 { 2691 struct aiocb32 job32; 2692 int error; 2693 2694 error = copyin(ujob, &job32, sizeof(job32)); 2695 if (error) 2696 return (error); 2697 CP(job32, *kjob, aio_fildes); 2698 CP(job32, *kjob, aio_offset); 2699 PTRIN_CP(job32, *kjob, aio_buf); 2700 CP(job32, *kjob, aio_nbytes); 2701 CP(job32, *kjob, aio_lio_opcode); 2702 CP(job32, *kjob, aio_reqprio); 2703 CP(job32, *kjob, _aiocb_private.status); 2704 CP(job32, *kjob, _aiocb_private.error); 2705 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo); 2706 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent)); 2707 } 2708 2709 static long 2710 aiocb32_fetch_status(struct aiocb *ujob) 2711 { 2712 struct aiocb32 *ujob32; 2713 2714 ujob32 = (struct aiocb32 *)ujob; 2715 return (fuword32(&ujob32->_aiocb_private.status)); 2716 } 2717 2718 static long 2719 aiocb32_fetch_error(struct aiocb *ujob) 2720 { 2721 struct aiocb32 *ujob32; 2722 2723 ujob32 = (struct aiocb32 *)ujob; 2724 return (fuword32(&ujob32->_aiocb_private.error)); 2725 } 2726 2727 static int 2728 aiocb32_store_status(struct aiocb *ujob, long status) 2729 { 2730 struct aiocb32 *ujob32; 2731 2732 ujob32 = (struct aiocb32 *)ujob; 2733 return (suword32(&ujob32->_aiocb_private.status, status)); 2734 } 2735 2736 static int 2737 aiocb32_store_error(struct aiocb *ujob, long error) 2738 { 2739 struct aiocb32 *ujob32; 2740 2741 ujob32 = (struct aiocb32 *)ujob; 2742 return (suword32(&ujob32->_aiocb_private.error, error)); 2743 } 2744 2745 static int 2746 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref) 2747 { 2748 struct aiocb32 *ujob32; 2749 2750 ujob32 = (struct aiocb32 *)ujob; 2751 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref)); 2752 } 2753 2754 static int 2755 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 2756 { 2757 2758 return (suword32(ujobp, (long)ujob)); 2759 } 2760 2761 static struct aiocb_ops aiocb32_ops = { 2762 .copyin = aiocb32_copyin, 2763 .fetch_status = aiocb32_fetch_status, 2764 .fetch_error = aiocb32_fetch_error, 2765 .store_status = aiocb32_store_status, 2766 .store_error = aiocb32_store_error, 2767 .store_kernelinfo = aiocb32_store_kernelinfo, 2768 .store_aiocb = aiocb32_store_aiocb, 2769 }; 2770 2771 #ifdef COMPAT_FREEBSD6 2772 static struct aiocb_ops aiocb32_ops_osigevent = { 2773 .copyin = aiocb32_copyin_old_sigevent, 2774 .fetch_status = aiocb32_fetch_status, 2775 .fetch_error = aiocb32_fetch_error, 2776 .store_status = aiocb32_store_status, 2777 .store_error = aiocb32_store_error, 2778 .store_kernelinfo = aiocb32_store_kernelinfo, 2779 .store_aiocb = aiocb32_store_aiocb, 2780 }; 2781 #endif 2782 2783 int 2784 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap) 2785 { 2786 2787 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2788 } 2789 2790 int 2791 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap) 2792 { 2793 struct timespec32 ts32; 2794 struct timespec ts, *tsp; 2795 struct aiocb **ujoblist; 2796 uint32_t *ujoblist32; 2797 int error, i; 2798 2799 if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc) 2800 return (EINVAL); 2801 2802 if (uap->timeout) { 2803 /* Get timespec struct. */ 2804 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0) 2805 return (error); 2806 CP(ts32, ts, tv_sec); 2807 CP(ts32, ts, tv_nsec); 2808 tsp = &ts; 2809 } else 2810 tsp = NULL; 2811 2812 ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK); 2813 ujoblist32 = (uint32_t *)ujoblist; 2814 error = copyin(uap->aiocbp, ujoblist32, uap->nent * 2815 sizeof(ujoblist32[0])); 2816 if (error == 0) { 2817 for (i = uap->nent - 1; i >= 0; i--) 2818 ujoblist[i] = PTRIN(ujoblist32[i]); 2819 2820 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2821 } 2822 free(ujoblist, M_AIOS); 2823 return (error); 2824 } 2825 2826 int 2827 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap) 2828 { 2829 2830 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2831 } 2832 2833 #ifdef COMPAT_FREEBSD6 2834 int 2835 freebsd6_freebsd32_aio_read(struct thread *td, 2836 struct freebsd6_freebsd32_aio_read_args *uap) 2837 { 2838 2839 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2840 &aiocb32_ops_osigevent)); 2841 } 2842 #endif 2843 2844 int 2845 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap) 2846 { 2847 2848 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2849 &aiocb32_ops)); 2850 } 2851 2852 #ifdef COMPAT_FREEBSD6 2853 int 2854 freebsd6_freebsd32_aio_write(struct thread *td, 2855 struct freebsd6_freebsd32_aio_write_args *uap) 2856 { 2857 2858 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2859 &aiocb32_ops_osigevent)); 2860 } 2861 #endif 2862 2863 int 2864 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap) 2865 { 2866 2867 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2868 &aiocb32_ops)); 2869 } 2870 2871 int 2872 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap) 2873 { 2874 2875 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK, 2876 &aiocb32_ops)); 2877 } 2878 2879 int 2880 freebsd32_aio_waitcomplete(struct thread *td, 2881 struct freebsd32_aio_waitcomplete_args *uap) 2882 { 2883 struct timespec32 ts32; 2884 struct timespec ts, *tsp; 2885 int error; 2886 2887 if (uap->timeout) { 2888 /* Get timespec struct. */ 2889 error = copyin(uap->timeout, &ts32, sizeof(ts32)); 2890 if (error) 2891 return (error); 2892 CP(ts32, ts, tv_sec); 2893 CP(ts32, ts, tv_nsec); 2894 tsp = &ts; 2895 } else 2896 tsp = NULL; 2897 2898 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp, 2899 &aiocb32_ops)); 2900 } 2901 2902 int 2903 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap) 2904 { 2905 2906 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp, 2907 &aiocb32_ops)); 2908 } 2909 2910 #ifdef COMPAT_FREEBSD6 2911 int 2912 freebsd6_freebsd32_lio_listio(struct thread *td, 2913 struct freebsd6_freebsd32_lio_listio_args *uap) 2914 { 2915 struct aiocb **acb_list; 2916 struct sigevent *sigp, sig; 2917 struct osigevent32 osig; 2918 uint32_t *acb_list32; 2919 int error, i, nent; 2920 2921 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2922 return (EINVAL); 2923 2924 nent = uap->nent; 2925 if (nent < 0 || nent > max_aio_queue_per_proc) 2926 return (EINVAL); 2927 2928 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2929 error = copyin(uap->sig, &osig, sizeof(osig)); 2930 if (error) 2931 return (error); 2932 error = convert_old_sigevent32(&osig, &sig); 2933 if (error) 2934 return (error); 2935 sigp = &sig; 2936 } else 2937 sigp = NULL; 2938 2939 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 2940 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 2941 if (error) { 2942 free(acb_list32, M_LIO); 2943 return (error); 2944 } 2945 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2946 for (i = 0; i < nent; i++) 2947 acb_list[i] = PTRIN(acb_list32[i]); 2948 free(acb_list32, M_LIO); 2949 2950 error = kern_lio_listio(td, uap->mode, 2951 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2952 &aiocb32_ops_osigevent); 2953 free(acb_list, M_LIO); 2954 return (error); 2955 } 2956 #endif 2957 2958 int 2959 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap) 2960 { 2961 struct aiocb **acb_list; 2962 struct sigevent *sigp, sig; 2963 struct sigevent32 sig32; 2964 uint32_t *acb_list32; 2965 int error, i, nent; 2966 2967 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2968 return (EINVAL); 2969 2970 nent = uap->nent; 2971 if (nent < 0 || nent > max_aio_queue_per_proc) 2972 return (EINVAL); 2973 2974 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2975 error = copyin(uap->sig, &sig32, sizeof(sig32)); 2976 if (error) 2977 return (error); 2978 error = convert_sigevent32(&sig32, &sig); 2979 if (error) 2980 return (error); 2981 sigp = &sig; 2982 } else 2983 sigp = NULL; 2984 2985 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 2986 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 2987 if (error) { 2988 free(acb_list32, M_LIO); 2989 return (error); 2990 } 2991 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2992 for (i = 0; i < nent; i++) 2993 acb_list[i] = PTRIN(acb_list32[i]); 2994 free(acb_list32, M_LIO); 2995 2996 error = kern_lio_listio(td, uap->mode, 2997 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2998 &aiocb32_ops); 2999 free(acb_list, M_LIO); 3000 return (error); 3001 } 3002 3003 #endif 3004