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