vfs_aio.c revision 280258
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: stable/10/sys/kern/vfs_aio.c 280258 2015-03-19 13:37:36Z rwatson $"); 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 63#include <machine/atomic.h> 64 65#include <vm/vm.h> 66#include <vm/vm_extern.h> 67#include <vm/pmap.h> 68#include <vm/vm_map.h> 69#include <vm/vm_object.h> 70#include <vm/uma.h> 71#include <sys/aio.h> 72 73#include "opt_vfs_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 */ 79static u_long jobrefid; 80 81/* 82 * Counter for aio_fsync. 83 */ 84static uint64_t jobseqno; 85 86#define JOBST_NULL 0 87#define JOBST_JOBQSOCK 1 88#define JOBST_JOBQGLOBAL 2 89#define JOBST_JOBRUNNING 3 90#define JOBST_JOBFINISHED 4 91#define JOBST_JOBQBUF 5 92#define JOBST_JOBQSYNC 6 93 94#ifndef MAX_AIO_PER_PROC 95#define MAX_AIO_PER_PROC 32 96#endif 97 98#ifndef MAX_AIO_QUEUE_PER_PROC 99#define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */ 100#endif 101 102#ifndef MAX_AIO_PROCS 103#define MAX_AIO_PROCS 32 104#endif 105 106#ifndef MAX_AIO_QUEUE 107#define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */ 108#endif 109 110#ifndef TARGET_AIO_PROCS 111#define TARGET_AIO_PROCS 4 112#endif 113 114#ifndef MAX_BUF_AIO 115#define MAX_BUF_AIO 16 116#endif 117 118#ifndef AIOD_TIMEOUT_DEFAULT 119#define AIOD_TIMEOUT_DEFAULT (10 * hz) 120#endif 121 122#ifndef AIOD_LIFETIME_DEFAULT 123#define AIOD_LIFETIME_DEFAULT (30 * hz) 124#endif 125 126FEATURE(aio, "Asynchronous I/O"); 127 128static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list"); 129 130static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management"); 131 132static int max_aio_procs = MAX_AIO_PROCS; 133SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, 134 CTLFLAG_RW, &max_aio_procs, 0, 135 "Maximum number of kernel threads to use for handling async IO "); 136 137static int num_aio_procs = 0; 138SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, 139 CTLFLAG_RD, &num_aio_procs, 0, 140 "Number of presently active kernel threads for async IO"); 141 142/* 143 * The code will adjust the actual number of AIO processes towards this 144 * number when it gets a chance. 145 */ 146static int target_aio_procs = TARGET_AIO_PROCS; 147SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs, 148 0, "Preferred number of ready kernel threads for async IO"); 149 150static int max_queue_count = MAX_AIO_QUEUE; 151SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0, 152 "Maximum number of aio requests to queue, globally"); 153 154static int num_queue_count = 0; 155SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0, 156 "Number of queued aio requests"); 157 158static int num_buf_aio = 0; 159SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0, 160 "Number of aio requests presently handled by the buf subsystem"); 161 162/* Number of async I/O thread in the process of being started */ 163/* XXX This should be local to aio_aqueue() */ 164static int num_aio_resv_start = 0; 165 166static int aiod_timeout; 167SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0, 168 "Timeout value for synchronous aio operations"); 169 170static int aiod_lifetime; 171SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0, 172 "Maximum lifetime for idle aiod"); 173 174static int unloadable = 0; 175SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0, 176 "Allow unload of aio (not recommended)"); 177 178 179static int max_aio_per_proc = MAX_AIO_PER_PROC; 180SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc, 181 0, "Maximum active aio requests per process (stored in the process)"); 182 183static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC; 184SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW, 185 &max_aio_queue_per_proc, 0, 186 "Maximum queued aio requests per process (stored in the process)"); 187 188static int max_buf_aio = MAX_BUF_AIO; 189SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0, 190 "Maximum buf aio requests per process (stored in the process)"); 191 192typedef struct oaiocb { 193 int aio_fildes; /* File descriptor */ 194 off_t aio_offset; /* File offset for I/O */ 195 volatile void *aio_buf; /* I/O buffer in process space */ 196 size_t aio_nbytes; /* Number of bytes for I/O */ 197 struct osigevent aio_sigevent; /* Signal to deliver */ 198 int aio_lio_opcode; /* LIO opcode */ 199 int aio_reqprio; /* Request priority -- ignored */ 200 struct __aiocb_private _aiocb_private; 201} oaiocb_t; 202 203/* 204 * Below is a key of locks used to protect each member of struct aiocblist 205 * aioliojob and kaioinfo and any backends. 206 * 207 * * - need not protected 208 * a - locked by kaioinfo lock 209 * b - locked by backend lock, the backend lock can be null in some cases, 210 * for example, BIO belongs to this type, in this case, proc lock is 211 * reused. 212 * c - locked by aio_job_mtx, the lock for the generic file I/O backend. 213 */ 214 215/* 216 * Current, there is only two backends: BIO and generic file I/O. 217 * socket I/O is served by generic file I/O, this is not a good idea, since 218 * disk file I/O and any other types without O_NONBLOCK flag can block daemon 219 * threads, if there is no thread to serve socket I/O, the socket I/O will be 220 * delayed too long or starved, we should create some threads dedicated to 221 * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O 222 * systems we really need non-blocking interface, fiddling O_NONBLOCK in file 223 * structure is not safe because there is race between userland and aio 224 * daemons. 225 */ 226 227struct aiocblist { 228 TAILQ_ENTRY(aiocblist) list; /* (b) internal list of for backend */ 229 TAILQ_ENTRY(aiocblist) plist; /* (a) list of jobs for each backend */ 230 TAILQ_ENTRY(aiocblist) allist; /* (a) list of all jobs in proc */ 231 int jobflags; /* (a) job flags */ 232 int jobstate; /* (b) job state */ 233 int inputcharge; /* (*) input blockes */ 234 int outputcharge; /* (*) output blockes */ 235 struct buf *bp; /* (*) private to BIO backend, 236 * buffer pointer 237 */ 238 struct proc *userproc; /* (*) user process */ 239 struct ucred *cred; /* (*) active credential when created */ 240 struct file *fd_file; /* (*) pointer to file structure */ 241 struct aioliojob *lio; /* (*) optional lio job */ 242 struct aiocb *uuaiocb; /* (*) pointer in userspace of aiocb */ 243 struct knlist klist; /* (a) list of knotes */ 244 struct aiocb uaiocb; /* (*) kernel I/O control block */ 245 ksiginfo_t ksi; /* (a) realtime signal info */ 246 struct task biotask; /* (*) private to BIO backend */ 247 uint64_t seqno; /* (*) job number */ 248 int pending; /* (a) number of pending I/O, aio_fsync only */ 249}; 250 251/* jobflags */ 252#define AIOCBLIST_DONE 0x01 253#define AIOCBLIST_BUFDONE 0x02 254#define AIOCBLIST_RUNDOWN 0x04 255#define AIOCBLIST_CHECKSYNC 0x08 256 257/* 258 * AIO process info 259 */ 260#define AIOP_FREE 0x1 /* proc on free queue */ 261 262struct aiothreadlist { 263 int aiothreadflags; /* (c) AIO proc flags */ 264 TAILQ_ENTRY(aiothreadlist) list; /* (c) list of processes */ 265 struct thread *aiothread; /* (*) the AIO thread */ 266}; 267 268/* 269 * data-structure for lio signal management 270 */ 271struct aioliojob { 272 int lioj_flags; /* (a) listio flags */ 273 int lioj_count; /* (a) listio flags */ 274 int lioj_finished_count; /* (a) listio flags */ 275 struct sigevent lioj_signal; /* (a) signal on all I/O done */ 276 TAILQ_ENTRY(aioliojob) lioj_list; /* (a) lio list */ 277 struct knlist klist; /* (a) list of knotes */ 278 ksiginfo_t lioj_ksi; /* (a) Realtime signal info */ 279}; 280 281#define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */ 282#define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */ 283#define LIOJ_KEVENT_POSTED 0x4 /* kevent triggered */ 284 285/* 286 * per process aio data structure 287 */ 288struct kaioinfo { 289 struct mtx kaio_mtx; /* the lock to protect this struct */ 290 int kaio_flags; /* (a) per process kaio flags */ 291 int kaio_maxactive_count; /* (*) maximum number of AIOs */ 292 int kaio_active_count; /* (c) number of currently used AIOs */ 293 int kaio_qallowed_count; /* (*) maxiumu size of AIO queue */ 294 int kaio_count; /* (a) size of AIO queue */ 295 int kaio_ballowed_count; /* (*) maximum number of buffers */ 296 int kaio_buffer_count; /* (a) number of physio buffers */ 297 TAILQ_HEAD(,aiocblist) kaio_all; /* (a) all AIOs in the process */ 298 TAILQ_HEAD(,aiocblist) kaio_done; /* (a) done queue for process */ 299 TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */ 300 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* (a) job queue for process */ 301 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* (a) buffer job queue for process */ 302 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* (a) queue for aios waiting on sockets, 303 * NOT USED YET. 304 */ 305 TAILQ_HEAD(,aiocblist) kaio_syncqueue; /* (a) queue for aio_fsync */ 306 struct task kaio_task; /* (*) task to kick aio threads */ 307}; 308 309#define AIO_LOCK(ki) mtx_lock(&(ki)->kaio_mtx) 310#define AIO_UNLOCK(ki) mtx_unlock(&(ki)->kaio_mtx) 311#define AIO_LOCK_ASSERT(ki, f) mtx_assert(&(ki)->kaio_mtx, (f)) 312#define AIO_MTX(ki) (&(ki)->kaio_mtx) 313 314#define KAIO_RUNDOWN 0x1 /* process is being run down */ 315#define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */ 316 317/* 318 * Operations used to interact with userland aio control blocks. 319 * Different ABIs provide their own operations. 320 */ 321struct aiocb_ops { 322 int (*copyin)(struct aiocb *ujob, struct aiocb *kjob); 323 long (*fetch_status)(struct aiocb *ujob); 324 long (*fetch_error)(struct aiocb *ujob); 325 int (*store_status)(struct aiocb *ujob, long status); 326 int (*store_error)(struct aiocb *ujob, long error); 327 int (*store_kernelinfo)(struct aiocb *ujob, long jobref); 328 int (*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob); 329}; 330 331static TAILQ_HEAD(,aiothreadlist) aio_freeproc; /* (c) Idle daemons */ 332static struct sema aio_newproc_sem; 333static struct mtx aio_job_mtx; 334static struct mtx aio_sock_mtx; 335static TAILQ_HEAD(,aiocblist) aio_jobs; /* (c) Async job list */ 336static struct unrhdr *aiod_unr; 337 338void aio_init_aioinfo(struct proc *p); 339static int aio_onceonly(void); 340static int aio_free_entry(struct aiocblist *aiocbe); 341static void aio_process_rw(struct aiocblist *aiocbe); 342static void aio_process_sync(struct aiocblist *aiocbe); 343static void aio_process_mlock(struct aiocblist *aiocbe); 344static int aio_newproc(int *); 345int aio_aqueue(struct thread *td, struct aiocb *job, 346 struct aioliojob *lio, int type, struct aiocb_ops *ops); 347static void aio_physwakeup(struct buf *bp); 348static void aio_proc_rundown(void *arg, struct proc *p); 349static void aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp); 350static int aio_qphysio(struct proc *p, struct aiocblist *iocb); 351static void biohelper(void *, int); 352static void aio_daemon(void *param); 353static void aio_swake_cb(struct socket *, struct sockbuf *); 354static int aio_unload(void); 355static void aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type); 356#define DONE_BUF 1 357#define DONE_QUEUE 2 358static int aio_kick(struct proc *userp); 359static void aio_kick_nowait(struct proc *userp); 360static void aio_kick_helper(void *context, int pending); 361static int filt_aioattach(struct knote *kn); 362static void filt_aiodetach(struct knote *kn); 363static int filt_aio(struct knote *kn, long hint); 364static int filt_lioattach(struct knote *kn); 365static void filt_liodetach(struct knote *kn); 366static int filt_lio(struct knote *kn, long hint); 367 368/* 369 * Zones for: 370 * kaio Per process async io info 371 * aiop async io thread data 372 * aiocb async io jobs 373 * aiol list io job pointer - internal to aio_suspend XXX 374 * aiolio list io jobs 375 */ 376static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone; 377 378/* kqueue filters for aio */ 379static struct filterops aio_filtops = { 380 .f_isfd = 0, 381 .f_attach = filt_aioattach, 382 .f_detach = filt_aiodetach, 383 .f_event = filt_aio, 384}; 385static struct filterops lio_filtops = { 386 .f_isfd = 0, 387 .f_attach = filt_lioattach, 388 .f_detach = filt_liodetach, 389 .f_event = filt_lio 390}; 391 392static eventhandler_tag exit_tag, exec_tag; 393 394TASKQUEUE_DEFINE_THREAD(aiod_bio); 395 396/* 397 * Main operations function for use as a kernel module. 398 */ 399static int 400aio_modload(struct module *module, int cmd, void *arg) 401{ 402 int error = 0; 403 404 switch (cmd) { 405 case MOD_LOAD: 406 aio_onceonly(); 407 break; 408 case MOD_UNLOAD: 409 error = aio_unload(); 410 break; 411 case MOD_SHUTDOWN: 412 break; 413 default: 414 error = EINVAL; 415 break; 416 } 417 return (error); 418} 419 420static moduledata_t aio_mod = { 421 "aio", 422 &aio_modload, 423 NULL 424}; 425 426static struct syscall_helper_data aio_syscalls[] = { 427 SYSCALL_INIT_HELPER(aio_cancel), 428 SYSCALL_INIT_HELPER(aio_error), 429 SYSCALL_INIT_HELPER(aio_fsync), 430 SYSCALL_INIT_HELPER(aio_mlock), 431 SYSCALL_INIT_HELPER(aio_read), 432 SYSCALL_INIT_HELPER(aio_return), 433 SYSCALL_INIT_HELPER(aio_suspend), 434 SYSCALL_INIT_HELPER(aio_waitcomplete), 435 SYSCALL_INIT_HELPER(aio_write), 436 SYSCALL_INIT_HELPER(lio_listio), 437 SYSCALL_INIT_HELPER(oaio_read), 438 SYSCALL_INIT_HELPER(oaio_write), 439 SYSCALL_INIT_HELPER(olio_listio), 440 SYSCALL_INIT_LAST 441}; 442 443#ifdef COMPAT_FREEBSD32 444#include <sys/mount.h> 445#include <sys/socket.h> 446#include <compat/freebsd32/freebsd32.h> 447#include <compat/freebsd32/freebsd32_proto.h> 448#include <compat/freebsd32/freebsd32_signal.h> 449#include <compat/freebsd32/freebsd32_syscall.h> 450#include <compat/freebsd32/freebsd32_util.h> 451 452static struct syscall_helper_data aio32_syscalls[] = { 453 SYSCALL32_INIT_HELPER(freebsd32_aio_return), 454 SYSCALL32_INIT_HELPER(freebsd32_aio_suspend), 455 SYSCALL32_INIT_HELPER(freebsd32_aio_cancel), 456 SYSCALL32_INIT_HELPER(freebsd32_aio_error), 457 SYSCALL32_INIT_HELPER(freebsd32_aio_fsync), 458 SYSCALL32_INIT_HELPER(freebsd32_aio_mlock), 459 SYSCALL32_INIT_HELPER(freebsd32_aio_read), 460 SYSCALL32_INIT_HELPER(freebsd32_aio_write), 461 SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete), 462 SYSCALL32_INIT_HELPER(freebsd32_lio_listio), 463 SYSCALL32_INIT_HELPER(freebsd32_oaio_read), 464 SYSCALL32_INIT_HELPER(freebsd32_oaio_write), 465 SYSCALL32_INIT_HELPER(freebsd32_olio_listio), 466 SYSCALL_INIT_LAST 467}; 468#endif 469 470DECLARE_MODULE(aio, aio_mod, 471 SI_SUB_VFS, SI_ORDER_ANY); 472MODULE_VERSION(aio, 1); 473 474/* 475 * Startup initialization 476 */ 477static int 478aio_onceonly(void) 479{ 480 int error; 481 482 /* XXX: should probably just use so->callback */ 483 aio_swake = &aio_swake_cb; 484 exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL, 485 EVENTHANDLER_PRI_ANY); 486 exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL, 487 EVENTHANDLER_PRI_ANY); 488 kqueue_add_filteropts(EVFILT_AIO, &aio_filtops); 489 kqueue_add_filteropts(EVFILT_LIO, &lio_filtops); 490 TAILQ_INIT(&aio_freeproc); 491 sema_init(&aio_newproc_sem, 0, "aio_new_proc"); 492 mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF); 493 mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF); 494 TAILQ_INIT(&aio_jobs); 495 aiod_unr = new_unrhdr(1, INT_MAX, NULL); 496 kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL, 497 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 498 aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL, 499 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 500 aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL, 501 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 502 aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL, 503 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 504 aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL, 505 NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 506 aiod_timeout = AIOD_TIMEOUT_DEFAULT; 507 aiod_lifetime = AIOD_LIFETIME_DEFAULT; 508 jobrefid = 1; 509 async_io_version = _POSIX_VERSION; 510 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX); 511 p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE); 512 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0); 513 514 error = syscall_helper_register(aio_syscalls); 515 if (error) 516 return (error); 517#ifdef COMPAT_FREEBSD32 518 error = syscall32_helper_register(aio32_syscalls); 519 if (error) 520 return (error); 521#endif 522 return (0); 523} 524 525/* 526 * Callback for unload of AIO when used as a module. 527 */ 528static int 529aio_unload(void) 530{ 531 int error; 532 533 /* 534 * XXX: no unloads by default, it's too dangerous. 535 * perhaps we could do it if locked out callers and then 536 * did an aio_proc_rundown() on each process. 537 * 538 * jhb: aio_proc_rundown() needs to run on curproc though, 539 * so I don't think that would fly. 540 */ 541 if (!unloadable) 542 return (EOPNOTSUPP); 543 544#ifdef COMPAT_FREEBSD32 545 syscall32_helper_unregister(aio32_syscalls); 546#endif 547 syscall_helper_unregister(aio_syscalls); 548 549 error = kqueue_del_filteropts(EVFILT_AIO); 550 if (error) 551 return error; 552 error = kqueue_del_filteropts(EVFILT_LIO); 553 if (error) 554 return error; 555 async_io_version = 0; 556 aio_swake = NULL; 557 taskqueue_free(taskqueue_aiod_bio); 558 delete_unrhdr(aiod_unr); 559 uma_zdestroy(kaio_zone); 560 uma_zdestroy(aiop_zone); 561 uma_zdestroy(aiocb_zone); 562 uma_zdestroy(aiol_zone); 563 uma_zdestroy(aiolio_zone); 564 EVENTHANDLER_DEREGISTER(process_exit, exit_tag); 565 EVENTHANDLER_DEREGISTER(process_exec, exec_tag); 566 mtx_destroy(&aio_job_mtx); 567 mtx_destroy(&aio_sock_mtx); 568 sema_destroy(&aio_newproc_sem); 569 p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1); 570 p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1); 571 p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1); 572 return (0); 573} 574 575/* 576 * Init the per-process aioinfo structure. The aioinfo limits are set 577 * per-process for user limit (resource) management. 578 */ 579void 580aio_init_aioinfo(struct proc *p) 581{ 582 struct kaioinfo *ki; 583 584 ki = uma_zalloc(kaio_zone, M_WAITOK); 585 mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF); 586 ki->kaio_flags = 0; 587 ki->kaio_maxactive_count = max_aio_per_proc; 588 ki->kaio_active_count = 0; 589 ki->kaio_qallowed_count = max_aio_queue_per_proc; 590 ki->kaio_count = 0; 591 ki->kaio_ballowed_count = max_buf_aio; 592 ki->kaio_buffer_count = 0; 593 TAILQ_INIT(&ki->kaio_all); 594 TAILQ_INIT(&ki->kaio_done); 595 TAILQ_INIT(&ki->kaio_jobqueue); 596 TAILQ_INIT(&ki->kaio_bufqueue); 597 TAILQ_INIT(&ki->kaio_liojoblist); 598 TAILQ_INIT(&ki->kaio_sockqueue); 599 TAILQ_INIT(&ki->kaio_syncqueue); 600 TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p); 601 PROC_LOCK(p); 602 if (p->p_aioinfo == NULL) { 603 p->p_aioinfo = ki; 604 PROC_UNLOCK(p); 605 } else { 606 PROC_UNLOCK(p); 607 mtx_destroy(&ki->kaio_mtx); 608 uma_zfree(kaio_zone, ki); 609 } 610 611 while (num_aio_procs < MIN(target_aio_procs, max_aio_procs)) 612 aio_newproc(NULL); 613} 614 615static int 616aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi) 617{ 618 struct thread *td; 619 int error; 620 621 error = sigev_findtd(p, sigev, &td); 622 if (error) 623 return (error); 624 if (!KSI_ONQ(ksi)) { 625 ksiginfo_set_sigev(ksi, sigev); 626 ksi->ksi_code = SI_ASYNCIO; 627 ksi->ksi_flags |= KSI_EXT | KSI_INS; 628 tdsendsignal(p, td, ksi->ksi_signo, ksi); 629 } 630 PROC_UNLOCK(p); 631 return (error); 632} 633 634/* 635 * Free a job entry. Wait for completion if it is currently active, but don't 636 * delay forever. If we delay, we return a flag that says that we have to 637 * restart the queue scan. 638 */ 639static int 640aio_free_entry(struct aiocblist *aiocbe) 641{ 642 struct kaioinfo *ki; 643 struct aioliojob *lj; 644 struct proc *p; 645 646 p = aiocbe->userproc; 647 MPASS(curproc == p); 648 ki = p->p_aioinfo; 649 MPASS(ki != NULL); 650 651 AIO_LOCK_ASSERT(ki, MA_OWNED); 652 MPASS(aiocbe->jobstate == JOBST_JOBFINISHED); 653 654 atomic_subtract_int(&num_queue_count, 1); 655 656 ki->kaio_count--; 657 MPASS(ki->kaio_count >= 0); 658 659 TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist); 660 TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist); 661 662 lj = aiocbe->lio; 663 if (lj) { 664 lj->lioj_count--; 665 lj->lioj_finished_count--; 666 667 if (lj->lioj_count == 0) { 668 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 669 /* lio is going away, we need to destroy any knotes */ 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 } 676 } 677 678 /* aiocbe is going away, we need to destroy any knotes */ 679 knlist_delete(&aiocbe->klist, curthread, 1); 680 PROC_LOCK(p); 681 sigqueue_take(&aiocbe->ksi); 682 PROC_UNLOCK(p); 683 684 MPASS(aiocbe->bp == NULL); 685 aiocbe->jobstate = JOBST_NULL; 686 AIO_UNLOCK(ki); 687 688 /* 689 * The thread argument here is used to find the owning process 690 * and is also passed to fo_close() which may pass it to various 691 * places such as devsw close() routines. Because of that, we 692 * need a thread pointer from the process owning the job that is 693 * persistent and won't disappear out from under us or move to 694 * another process. 695 * 696 * Currently, all the callers of this function call it to remove 697 * an aiocblist from the current process' job list either via a 698 * syscall or due to the current process calling exit() or 699 * execve(). Thus, we know that p == curproc. We also know that 700 * curthread can't exit since we are curthread. 701 * 702 * Therefore, we use curthread as the thread to pass to 703 * knlist_delete(). This does mean that it is possible for the 704 * thread pointer at close time to differ from the thread pointer 705 * at open time, but this is already true of file descriptors in 706 * a multithreaded process. 707 */ 708 if (aiocbe->fd_file) 709 fdrop(aiocbe->fd_file, curthread); 710 crfree(aiocbe->cred); 711 uma_zfree(aiocb_zone, aiocbe); 712 AIO_LOCK(ki); 713 714 return (0); 715} 716 717static void 718aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused) 719{ 720 aio_proc_rundown(arg, p); 721} 722 723/* 724 * Rundown the jobs for a given process. 725 */ 726static void 727aio_proc_rundown(void *arg, struct proc *p) 728{ 729 struct kaioinfo *ki; 730 struct aioliojob *lj; 731 struct aiocblist *cbe, *cbn; 732 struct file *fp; 733 struct socket *so; 734 int remove; 735 736 KASSERT(curthread->td_proc == p, 737 ("%s: called on non-curproc", __func__)); 738 ki = p->p_aioinfo; 739 if (ki == NULL) 740 return; 741 742 AIO_LOCK(ki); 743 ki->kaio_flags |= KAIO_RUNDOWN; 744 745restart: 746 747 /* 748 * Try to cancel all pending requests. This code simulates 749 * aio_cancel on all pending I/O requests. 750 */ 751 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) { 752 remove = 0; 753 mtx_lock(&aio_job_mtx); 754 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 755 TAILQ_REMOVE(&aio_jobs, cbe, list); 756 remove = 1; 757 } else if (cbe->jobstate == JOBST_JOBQSOCK) { 758 fp = cbe->fd_file; 759 MPASS(fp->f_type == DTYPE_SOCKET); 760 so = fp->f_data; 761 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 762 remove = 1; 763 } else if (cbe->jobstate == JOBST_JOBQSYNC) { 764 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list); 765 remove = 1; 766 } 767 mtx_unlock(&aio_job_mtx); 768 769 if (remove) { 770 cbe->jobstate = JOBST_JOBFINISHED; 771 cbe->uaiocb._aiocb_private.status = -1; 772 cbe->uaiocb._aiocb_private.error = ECANCELED; 773 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 774 aio_bio_done_notify(p, cbe, DONE_QUEUE); 775 } 776 } 777 778 /* Wait for all running I/O to be finished */ 779 if (TAILQ_FIRST(&ki->kaio_bufqueue) || 780 TAILQ_FIRST(&ki->kaio_jobqueue)) { 781 ki->kaio_flags |= KAIO_WAKEUP; 782 msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz); 783 goto restart; 784 } 785 786 /* Free all completed I/O requests. */ 787 while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL) 788 aio_free_entry(cbe); 789 790 while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) { 791 if (lj->lioj_count == 0) { 792 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 793 knlist_delete(&lj->klist, curthread, 1); 794 PROC_LOCK(p); 795 sigqueue_take(&lj->lioj_ksi); 796 PROC_UNLOCK(p); 797 uma_zfree(aiolio_zone, lj); 798 } else { 799 panic("LIO job not cleaned up: C:%d, FC:%d\n", 800 lj->lioj_count, lj->lioj_finished_count); 801 } 802 } 803 AIO_UNLOCK(ki); 804 taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task); 805 mtx_destroy(&ki->kaio_mtx); 806 uma_zfree(kaio_zone, ki); 807 p->p_aioinfo = NULL; 808} 809 810/* 811 * Select a job to run (called by an AIO daemon). 812 */ 813static struct aiocblist * 814aio_selectjob(struct aiothreadlist *aiop) 815{ 816 struct aiocblist *aiocbe; 817 struct kaioinfo *ki; 818 struct proc *userp; 819 820 mtx_assert(&aio_job_mtx, MA_OWNED); 821 TAILQ_FOREACH(aiocbe, &aio_jobs, list) { 822 userp = aiocbe->userproc; 823 ki = userp->p_aioinfo; 824 825 if (ki->kaio_active_count < ki->kaio_maxactive_count) { 826 TAILQ_REMOVE(&aio_jobs, aiocbe, list); 827 /* Account for currently active jobs. */ 828 ki->kaio_active_count++; 829 aiocbe->jobstate = JOBST_JOBRUNNING; 830 break; 831 } 832 } 833 return (aiocbe); 834} 835 836/* 837 * Move all data to a permanent storage device, this code 838 * simulates fsync syscall. 839 */ 840static int 841aio_fsync_vnode(struct thread *td, struct vnode *vp) 842{ 843 struct mount *mp; 844 int error; 845 846 if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 847 goto drop; 848 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 849 if (vp->v_object != NULL) { 850 VM_OBJECT_WLOCK(vp->v_object); 851 vm_object_page_clean(vp->v_object, 0, 0, 0); 852 VM_OBJECT_WUNLOCK(vp->v_object); 853 } 854 error = VOP_FSYNC(vp, MNT_WAIT, td); 855 856 VOP_UNLOCK(vp, 0); 857 vn_finished_write(mp); 858drop: 859 return (error); 860} 861 862/* 863 * The AIO processing activity for LIO_READ/LIO_WRITE. This is the code that 864 * does the I/O request for the non-physio version of the operations. The 865 * normal vn operations are used, and this code should work in all instances 866 * for every type of file, including pipes, sockets, fifos, and regular files. 867 * 868 * XXX I don't think it works well for socket, pipe, and fifo. 869 */ 870static void 871aio_process_rw(struct aiocblist *aiocbe) 872{ 873 struct ucred *td_savedcred; 874 struct thread *td; 875 struct aiocb *cb; 876 struct file *fp; 877 struct socket *so; 878 struct uio auio; 879 struct iovec aiov; 880 int cnt; 881 int error; 882 int oublock_st, oublock_end; 883 int inblock_st, inblock_end; 884 885 KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_READ || 886 aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE, 887 ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode)); 888 889 td = curthread; 890 td_savedcred = td->td_ucred; 891 td->td_ucred = aiocbe->cred; 892 cb = &aiocbe->uaiocb; 893 fp = aiocbe->fd_file; 894 895 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf; 896 aiov.iov_len = cb->aio_nbytes; 897 898 auio.uio_iov = &aiov; 899 auio.uio_iovcnt = 1; 900 auio.uio_offset = cb->aio_offset; 901 auio.uio_resid = cb->aio_nbytes; 902 cnt = cb->aio_nbytes; 903 auio.uio_segflg = UIO_USERSPACE; 904 auio.uio_td = td; 905 906 inblock_st = td->td_ru.ru_inblock; 907 oublock_st = td->td_ru.ru_oublock; 908 /* 909 * aio_aqueue() acquires a reference to the file that is 910 * released in aio_free_entry(). 911 */ 912 if (cb->aio_lio_opcode == LIO_READ) { 913 auio.uio_rw = UIO_READ; 914 if (auio.uio_resid == 0) 915 error = 0; 916 else 917 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td); 918 } else { 919 if (fp->f_type == DTYPE_VNODE) 920 bwillwrite(); 921 auio.uio_rw = UIO_WRITE; 922 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td); 923 } 924 inblock_end = td->td_ru.ru_inblock; 925 oublock_end = td->td_ru.ru_oublock; 926 927 aiocbe->inputcharge = inblock_end - inblock_st; 928 aiocbe->outputcharge = oublock_end - oublock_st; 929 930 if ((error) && (auio.uio_resid != cnt)) { 931 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK) 932 error = 0; 933 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) { 934 int sigpipe = 1; 935 if (fp->f_type == DTYPE_SOCKET) { 936 so = fp->f_data; 937 if (so->so_options & SO_NOSIGPIPE) 938 sigpipe = 0; 939 } 940 if (sigpipe) { 941 PROC_LOCK(aiocbe->userproc); 942 kern_psignal(aiocbe->userproc, SIGPIPE); 943 PROC_UNLOCK(aiocbe->userproc); 944 } 945 } 946 } 947 948 cnt -= auio.uio_resid; 949 cb->_aiocb_private.error = error; 950 cb->_aiocb_private.status = cnt; 951 td->td_ucred = td_savedcred; 952} 953 954static void 955aio_process_sync(struct aiocblist *aiocbe) 956{ 957 struct thread *td = curthread; 958 struct ucred *td_savedcred = td->td_ucred; 959 struct aiocb *cb = &aiocbe->uaiocb; 960 struct file *fp = aiocbe->fd_file; 961 int error = 0; 962 963 KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_SYNC, 964 ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode)); 965 966 td->td_ucred = aiocbe->cred; 967 if (fp->f_vnode != NULL) 968 error = aio_fsync_vnode(td, fp->f_vnode); 969 cb->_aiocb_private.error = error; 970 cb->_aiocb_private.status = 0; 971 td->td_ucred = td_savedcred; 972} 973 974static void 975aio_process_mlock(struct aiocblist *aiocbe) 976{ 977 struct aiocb *cb = &aiocbe->uaiocb; 978 int error; 979 980 KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_MLOCK, 981 ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode)); 982 983 error = vm_mlock(aiocbe->userproc, aiocbe->cred, 984 __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes); 985 cb->_aiocb_private.error = error; 986 cb->_aiocb_private.status = 0; 987} 988 989static void 990aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type) 991{ 992 struct aioliojob *lj; 993 struct kaioinfo *ki; 994 struct aiocblist *scb, *scbn; 995 int lj_done; 996 997 ki = userp->p_aioinfo; 998 AIO_LOCK_ASSERT(ki, MA_OWNED); 999 lj = aiocbe->lio; 1000 lj_done = 0; 1001 if (lj) { 1002 lj->lioj_finished_count++; 1003 if (lj->lioj_count == lj->lioj_finished_count) 1004 lj_done = 1; 1005 } 1006 if (type == DONE_QUEUE) { 1007 aiocbe->jobflags |= AIOCBLIST_DONE; 1008 } else { 1009 aiocbe->jobflags |= AIOCBLIST_BUFDONE; 1010 } 1011 TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist); 1012 aiocbe->jobstate = JOBST_JOBFINISHED; 1013 1014 if (ki->kaio_flags & KAIO_RUNDOWN) 1015 goto notification_done; 1016 1017 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1018 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) 1019 aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi); 1020 1021 KNOTE_LOCKED(&aiocbe->klist, 1); 1022 1023 if (lj_done) { 1024 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 1025 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 1026 KNOTE_LOCKED(&lj->klist, 1); 1027 } 1028 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 1029 == LIOJ_SIGNAL 1030 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 1031 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 1032 aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi); 1033 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 1034 } 1035 } 1036 1037notification_done: 1038 if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) { 1039 TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) { 1040 if (aiocbe->fd_file == scb->fd_file && 1041 aiocbe->seqno < scb->seqno) { 1042 if (--scb->pending == 0) { 1043 mtx_lock(&aio_job_mtx); 1044 scb->jobstate = JOBST_JOBQGLOBAL; 1045 TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list); 1046 TAILQ_INSERT_TAIL(&aio_jobs, scb, list); 1047 aio_kick_nowait(userp); 1048 mtx_unlock(&aio_job_mtx); 1049 } 1050 } 1051 } 1052 } 1053 if (ki->kaio_flags & KAIO_WAKEUP) { 1054 ki->kaio_flags &= ~KAIO_WAKEUP; 1055 wakeup(&userp->p_aioinfo); 1056 } 1057} 1058 1059/* 1060 * The AIO daemon, most of the actual work is done in aio_process_*, 1061 * but the setup (and address space mgmt) is done in this routine. 1062 */ 1063static void 1064aio_daemon(void *_id) 1065{ 1066 struct aiocblist *aiocbe; 1067 struct aiothreadlist *aiop; 1068 struct kaioinfo *ki; 1069 struct proc *curcp, *mycp, *userp; 1070 struct vmspace *myvm, *tmpvm; 1071 struct thread *td = curthread; 1072 int id = (intptr_t)_id; 1073 1074 /* 1075 * Local copies of curproc (cp) and vmspace (myvm) 1076 */ 1077 mycp = td->td_proc; 1078 myvm = mycp->p_vmspace; 1079 1080 KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp")); 1081 1082 /* 1083 * Allocate and ready the aio control info. There is one aiop structure 1084 * per daemon. 1085 */ 1086 aiop = uma_zalloc(aiop_zone, M_WAITOK); 1087 aiop->aiothread = td; 1088 aiop->aiothreadflags = 0; 1089 1090 /* The daemon resides in its own pgrp. */ 1091 sys_setsid(td, NULL); 1092 1093 /* 1094 * Wakeup parent process. (Parent sleeps to keep from blasting away 1095 * and creating too many daemons.) 1096 */ 1097 sema_post(&aio_newproc_sem); 1098 1099 mtx_lock(&aio_job_mtx); 1100 for (;;) { 1101 /* 1102 * curcp is the current daemon process context. 1103 * userp is the current user process context. 1104 */ 1105 curcp = mycp; 1106 1107 /* 1108 * Take daemon off of free queue 1109 */ 1110 if (aiop->aiothreadflags & AIOP_FREE) { 1111 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1112 aiop->aiothreadflags &= ~AIOP_FREE; 1113 } 1114 1115 /* 1116 * Check for jobs. 1117 */ 1118 while ((aiocbe = aio_selectjob(aiop)) != NULL) { 1119 mtx_unlock(&aio_job_mtx); 1120 userp = aiocbe->userproc; 1121 1122 /* 1123 * Connect to process address space for user program. 1124 */ 1125 if (userp != curcp) { 1126 /* 1127 * Save the current address space that we are 1128 * connected to. 1129 */ 1130 tmpvm = mycp->p_vmspace; 1131 1132 /* 1133 * Point to the new user address space, and 1134 * refer to it. 1135 */ 1136 mycp->p_vmspace = userp->p_vmspace; 1137 atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1); 1138 1139 /* Activate the new mapping. */ 1140 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 1141 1142 /* 1143 * If the old address space wasn't the daemons 1144 * own address space, then we need to remove the 1145 * daemon's reference from the other process 1146 * that it was acting on behalf of. 1147 */ 1148 if (tmpvm != myvm) { 1149 vmspace_free(tmpvm); 1150 } 1151 curcp = userp; 1152 } 1153 1154 ki = userp->p_aioinfo; 1155 1156 /* Do the I/O function. */ 1157 switch(aiocbe->uaiocb.aio_lio_opcode) { 1158 case LIO_READ: 1159 case LIO_WRITE: 1160 aio_process_rw(aiocbe); 1161 break; 1162 case LIO_SYNC: 1163 aio_process_sync(aiocbe); 1164 break; 1165 case LIO_MLOCK: 1166 aio_process_mlock(aiocbe); 1167 break; 1168 } 1169 1170 mtx_lock(&aio_job_mtx); 1171 /* Decrement the active job count. */ 1172 ki->kaio_active_count--; 1173 mtx_unlock(&aio_job_mtx); 1174 1175 AIO_LOCK(ki); 1176 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist); 1177 aio_bio_done_notify(userp, aiocbe, DONE_QUEUE); 1178 AIO_UNLOCK(ki); 1179 1180 mtx_lock(&aio_job_mtx); 1181 } 1182 1183 /* 1184 * Disconnect from user address space. 1185 */ 1186 if (curcp != mycp) { 1187 1188 mtx_unlock(&aio_job_mtx); 1189 1190 /* Get the user address space to disconnect from. */ 1191 tmpvm = mycp->p_vmspace; 1192 1193 /* Get original address space for daemon. */ 1194 mycp->p_vmspace = myvm; 1195 1196 /* Activate the daemon's address space. */ 1197 pmap_activate(FIRST_THREAD_IN_PROC(mycp)); 1198#ifdef DIAGNOSTIC 1199 if (tmpvm == myvm) { 1200 printf("AIOD: vmspace problem -- %d\n", 1201 mycp->p_pid); 1202 } 1203#endif 1204 /* Remove our vmspace reference. */ 1205 vmspace_free(tmpvm); 1206 1207 curcp = mycp; 1208 1209 mtx_lock(&aio_job_mtx); 1210 /* 1211 * We have to restart to avoid race, we only sleep if 1212 * no job can be selected, that should be 1213 * curcp == mycp. 1214 */ 1215 continue; 1216 } 1217 1218 mtx_assert(&aio_job_mtx, MA_OWNED); 1219 1220 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list); 1221 aiop->aiothreadflags |= AIOP_FREE; 1222 1223 /* 1224 * If daemon is inactive for a long time, allow it to exit, 1225 * thereby freeing resources. 1226 */ 1227 if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy", 1228 aiod_lifetime)) { 1229 if (TAILQ_EMPTY(&aio_jobs)) { 1230 if ((aiop->aiothreadflags & AIOP_FREE) && 1231 (num_aio_procs > target_aio_procs)) { 1232 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1233 num_aio_procs--; 1234 mtx_unlock(&aio_job_mtx); 1235 uma_zfree(aiop_zone, aiop); 1236 free_unr(aiod_unr, id); 1237#ifdef DIAGNOSTIC 1238 if (mycp->p_vmspace->vm_refcnt <= 1) { 1239 printf("AIOD: bad vm refcnt for" 1240 " exiting daemon: %d\n", 1241 mycp->p_vmspace->vm_refcnt); 1242 } 1243#endif 1244 kproc_exit(0); 1245 } 1246 } 1247 } 1248 } 1249 mtx_unlock(&aio_job_mtx); 1250 panic("shouldn't be here\n"); 1251} 1252 1253/* 1254 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The 1255 * AIO daemon modifies its environment itself. 1256 */ 1257static int 1258aio_newproc(int *start) 1259{ 1260 int error; 1261 struct proc *p; 1262 int id; 1263 1264 id = alloc_unr(aiod_unr); 1265 error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p, 1266 RFNOWAIT, 0, "aiod%d", id); 1267 if (error == 0) { 1268 /* 1269 * Wait until daemon is started. 1270 */ 1271 sema_wait(&aio_newproc_sem); 1272 mtx_lock(&aio_job_mtx); 1273 num_aio_procs++; 1274 if (start != NULL) 1275 (*start)--; 1276 mtx_unlock(&aio_job_mtx); 1277 } else { 1278 free_unr(aiod_unr, id); 1279 } 1280 return (error); 1281} 1282 1283/* 1284 * Try the high-performance, low-overhead physio method for eligible 1285 * VCHR devices. This method doesn't use an aio helper thread, and 1286 * thus has very low overhead. 1287 * 1288 * Assumes that the caller, aio_aqueue(), has incremented the file 1289 * structure's reference count, preventing its deallocation for the 1290 * duration of this call. 1291 */ 1292static int 1293aio_qphysio(struct proc *p, struct aiocblist *aiocbe) 1294{ 1295 struct aiocb *cb; 1296 struct file *fp; 1297 struct buf *bp; 1298 struct vnode *vp; 1299 struct cdevsw *csw; 1300 struct cdev *dev; 1301 struct kaioinfo *ki; 1302 struct aioliojob *lj; 1303 int error, ref; 1304 1305 cb = &aiocbe->uaiocb; 1306 fp = aiocbe->fd_file; 1307 1308 if (fp == NULL || fp->f_type != DTYPE_VNODE) 1309 return (-1); 1310 1311 vp = fp->f_vnode; 1312 1313 /* 1314 * If its not a disk, we don't want to return a positive error. 1315 * It causes the aio code to not fall through to try the thread 1316 * way when you're talking to a regular file. 1317 */ 1318 if (!vn_isdisk(vp, &error)) { 1319 if (error == ENOTBLK) 1320 return (-1); 1321 else 1322 return (error); 1323 } 1324 1325 if (vp->v_bufobj.bo_bsize == 0) 1326 return (-1); 1327 1328 if (cb->aio_nbytes % vp->v_bufobj.bo_bsize) 1329 return (-1); 1330 1331 if (cb->aio_nbytes > 1332 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK)) 1333 return (-1); 1334 1335 ki = p->p_aioinfo; 1336 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) 1337 return (-1); 1338 1339 ref = 0; 1340 csw = devvn_refthread(vp, &dev, &ref); 1341 if (csw == NULL) 1342 return (ENXIO); 1343 if (cb->aio_nbytes > dev->si_iosize_max) { 1344 error = -1; 1345 goto unref; 1346 } 1347 1348 /* Create and build a buffer header for a transfer. */ 1349 bp = (struct buf *)getpbuf(NULL); 1350 BUF_KERNPROC(bp); 1351 1352 AIO_LOCK(ki); 1353 ki->kaio_count++; 1354 ki->kaio_buffer_count++; 1355 lj = aiocbe->lio; 1356 if (lj) 1357 lj->lioj_count++; 1358 AIO_UNLOCK(ki); 1359 1360 /* 1361 * Get a copy of the kva from the physical buffer. 1362 */ 1363 error = 0; 1364 1365 bp->b_bcount = cb->aio_nbytes; 1366 bp->b_bufsize = cb->aio_nbytes; 1367 bp->b_iodone = aio_physwakeup; 1368 bp->b_saveaddr = bp->b_data; 1369 bp->b_data = (void *)(uintptr_t)cb->aio_buf; 1370 bp->b_offset = cb->aio_offset; 1371 bp->b_iooffset = cb->aio_offset; 1372 bp->b_blkno = btodb(cb->aio_offset); 1373 bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ; 1374 1375 /* 1376 * Bring buffer into kernel space. 1377 */ 1378 if (vmapbuf(bp, (dev->si_flags & SI_UNMAPPED) == 0) < 0) { 1379 error = EFAULT; 1380 goto doerror; 1381 } 1382 1383 AIO_LOCK(ki); 1384 aiocbe->bp = bp; 1385 bp->b_caller1 = (void *)aiocbe; 1386 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist); 1387 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1388 aiocbe->jobstate = JOBST_JOBQBUF; 1389 cb->_aiocb_private.status = cb->aio_nbytes; 1390 AIO_UNLOCK(ki); 1391 1392 atomic_add_int(&num_queue_count, 1); 1393 atomic_add_int(&num_buf_aio, 1); 1394 1395 bp->b_error = 0; 1396 1397 TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe); 1398 1399 /* Perform transfer. */ 1400 dev_strategy_csw(dev, csw, bp); 1401 dev_relthread(dev, ref); 1402 return (0); 1403 1404doerror: 1405 AIO_LOCK(ki); 1406 ki->kaio_count--; 1407 ki->kaio_buffer_count--; 1408 if (lj) 1409 lj->lioj_count--; 1410 aiocbe->bp = NULL; 1411 AIO_UNLOCK(ki); 1412 relpbuf(bp, NULL); 1413unref: 1414 dev_relthread(dev, ref); 1415 return (error); 1416} 1417 1418/* 1419 * Wake up aio requests that may be serviceable now. 1420 */ 1421static void 1422aio_swake_cb(struct socket *so, struct sockbuf *sb) 1423{ 1424 struct aiocblist *cb, *cbn; 1425 int opcode; 1426 1427 SOCKBUF_LOCK_ASSERT(sb); 1428 if (sb == &so->so_snd) 1429 opcode = LIO_WRITE; 1430 else 1431 opcode = LIO_READ; 1432 1433 sb->sb_flags &= ~SB_AIO; 1434 mtx_lock(&aio_job_mtx); 1435 TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) { 1436 if (opcode == cb->uaiocb.aio_lio_opcode) { 1437 if (cb->jobstate != JOBST_JOBQSOCK) 1438 panic("invalid queue value"); 1439 /* XXX 1440 * We don't have actual sockets backend yet, 1441 * so we simply move the requests to the generic 1442 * file I/O backend. 1443 */ 1444 TAILQ_REMOVE(&so->so_aiojobq, cb, list); 1445 TAILQ_INSERT_TAIL(&aio_jobs, cb, list); 1446 aio_kick_nowait(cb->userproc); 1447 } 1448 } 1449 mtx_unlock(&aio_job_mtx); 1450} 1451 1452static int 1453convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig) 1454{ 1455 1456 /* 1457 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 1458 * supported by AIO with the old sigevent structure. 1459 */ 1460 nsig->sigev_notify = osig->sigev_notify; 1461 switch (nsig->sigev_notify) { 1462 case SIGEV_NONE: 1463 break; 1464 case SIGEV_SIGNAL: 1465 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 1466 break; 1467 case SIGEV_KEVENT: 1468 nsig->sigev_notify_kqueue = 1469 osig->__sigev_u.__sigev_notify_kqueue; 1470 nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr; 1471 break; 1472 default: 1473 return (EINVAL); 1474 } 1475 return (0); 1476} 1477 1478static int 1479aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob) 1480{ 1481 struct oaiocb *ojob; 1482 int error; 1483 1484 bzero(kjob, sizeof(struct aiocb)); 1485 error = copyin(ujob, kjob, sizeof(struct oaiocb)); 1486 if (error) 1487 return (error); 1488 ojob = (struct oaiocb *)kjob; 1489 return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent)); 1490} 1491 1492static int 1493aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob) 1494{ 1495 1496 return (copyin(ujob, kjob, sizeof(struct aiocb))); 1497} 1498 1499static long 1500aiocb_fetch_status(struct aiocb *ujob) 1501{ 1502 1503 return (fuword(&ujob->_aiocb_private.status)); 1504} 1505 1506static long 1507aiocb_fetch_error(struct aiocb *ujob) 1508{ 1509 1510 return (fuword(&ujob->_aiocb_private.error)); 1511} 1512 1513static int 1514aiocb_store_status(struct aiocb *ujob, long status) 1515{ 1516 1517 return (suword(&ujob->_aiocb_private.status, status)); 1518} 1519 1520static int 1521aiocb_store_error(struct aiocb *ujob, long error) 1522{ 1523 1524 return (suword(&ujob->_aiocb_private.error, error)); 1525} 1526 1527static int 1528aiocb_store_kernelinfo(struct aiocb *ujob, long jobref) 1529{ 1530 1531 return (suword(&ujob->_aiocb_private.kernelinfo, jobref)); 1532} 1533 1534static int 1535aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 1536{ 1537 1538 return (suword(ujobp, (long)ujob)); 1539} 1540 1541static struct aiocb_ops aiocb_ops = { 1542 .copyin = aiocb_copyin, 1543 .fetch_status = aiocb_fetch_status, 1544 .fetch_error = aiocb_fetch_error, 1545 .store_status = aiocb_store_status, 1546 .store_error = aiocb_store_error, 1547 .store_kernelinfo = aiocb_store_kernelinfo, 1548 .store_aiocb = aiocb_store_aiocb, 1549}; 1550 1551static struct aiocb_ops aiocb_ops_osigevent = { 1552 .copyin = aiocb_copyin_old_sigevent, 1553 .fetch_status = aiocb_fetch_status, 1554 .fetch_error = aiocb_fetch_error, 1555 .store_status = aiocb_store_status, 1556 .store_error = aiocb_store_error, 1557 .store_kernelinfo = aiocb_store_kernelinfo, 1558 .store_aiocb = aiocb_store_aiocb, 1559}; 1560 1561/* 1562 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR 1563 * technique is done in this code. 1564 */ 1565int 1566aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj, 1567 int type, struct aiocb_ops *ops) 1568{ 1569 struct proc *p = td->td_proc; 1570 cap_rights_t rights; 1571 struct file *fp; 1572 struct socket *so; 1573 struct aiocblist *aiocbe, *cb; 1574 struct kaioinfo *ki; 1575 struct kevent kev; 1576 struct sockbuf *sb; 1577 int opcode; 1578 int error; 1579 int fd, kqfd; 1580 int jid; 1581 u_short evflags; 1582 1583 if (p->p_aioinfo == NULL) 1584 aio_init_aioinfo(p); 1585 1586 ki = p->p_aioinfo; 1587 1588 ops->store_status(job, -1); 1589 ops->store_error(job, 0); 1590 ops->store_kernelinfo(job, -1); 1591 1592 if (num_queue_count >= max_queue_count || 1593 ki->kaio_count >= ki->kaio_qallowed_count) { 1594 ops->store_error(job, EAGAIN); 1595 return (EAGAIN); 1596 } 1597 1598 aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO); 1599 knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki)); 1600 1601 error = ops->copyin(job, &aiocbe->uaiocb); 1602 if (error) { 1603 ops->store_error(job, error); 1604 uma_zfree(aiocb_zone, aiocbe); 1605 return (error); 1606 } 1607 1608 /* XXX: aio_nbytes is later casted to signed types. */ 1609 if (aiocbe->uaiocb.aio_nbytes > INT_MAX) { 1610 uma_zfree(aiocb_zone, aiocbe); 1611 return (EINVAL); 1612 } 1613 1614 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT && 1615 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL && 1616 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID && 1617 aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) { 1618 ops->store_error(job, EINVAL); 1619 uma_zfree(aiocb_zone, aiocbe); 1620 return (EINVAL); 1621 } 1622 1623 if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL || 1624 aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) && 1625 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) { 1626 uma_zfree(aiocb_zone, aiocbe); 1627 return (EINVAL); 1628 } 1629 1630 ksiginfo_init(&aiocbe->ksi); 1631 1632 /* Save userspace address of the job info. */ 1633 aiocbe->uuaiocb = job; 1634 1635 /* Get the opcode. */ 1636 if (type != LIO_NOP) 1637 aiocbe->uaiocb.aio_lio_opcode = type; 1638 opcode = aiocbe->uaiocb.aio_lio_opcode; 1639 1640 /* 1641 * Validate the opcode and fetch the file object for the specified 1642 * file descriptor. 1643 * 1644 * XXXRW: Moved the opcode validation up here so that we don't 1645 * retrieve a file descriptor without knowing what the capabiltity 1646 * should be. 1647 */ 1648 fd = aiocbe->uaiocb.aio_fildes; 1649 switch (opcode) { 1650 case LIO_WRITE: 1651 error = fget_write(td, fd, 1652 cap_rights_init(&rights, CAP_PWRITE), &fp); 1653 break; 1654 case LIO_READ: 1655 error = fget_read(td, fd, 1656 cap_rights_init(&rights, CAP_PREAD), &fp); 1657 break; 1658 case LIO_SYNC: 1659 error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp); 1660 break; 1661 case LIO_MLOCK: 1662 fp = NULL; 1663 break; 1664 case LIO_NOP: 1665 error = fget(td, fd, cap_rights_init(&rights), &fp); 1666 break; 1667 default: 1668 error = EINVAL; 1669 } 1670 if (error) { 1671 uma_zfree(aiocb_zone, aiocbe); 1672 ops->store_error(job, error); 1673 return (error); 1674 } 1675 1676 if (opcode == LIO_SYNC && fp->f_vnode == NULL) { 1677 error = EINVAL; 1678 goto aqueue_fail; 1679 } 1680 1681 if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) { 1682 error = EINVAL; 1683 goto aqueue_fail; 1684 } 1685 1686 aiocbe->fd_file = fp; 1687 1688 mtx_lock(&aio_job_mtx); 1689 jid = jobrefid++; 1690 aiocbe->seqno = jobseqno++; 1691 mtx_unlock(&aio_job_mtx); 1692 error = ops->store_kernelinfo(job, jid); 1693 if (error) { 1694 error = EINVAL; 1695 goto aqueue_fail; 1696 } 1697 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid; 1698 1699 if (opcode == LIO_NOP) { 1700 fdrop(fp, td); 1701 uma_zfree(aiocb_zone, aiocbe); 1702 return (0); 1703 } 1704 1705 if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT) 1706 goto no_kqueue; 1707 evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags; 1708 if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) { 1709 error = EINVAL; 1710 goto aqueue_fail; 1711 } 1712 kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue; 1713 kev.ident = (uintptr_t)aiocbe->uuaiocb; 1714 kev.filter = EVFILT_AIO; 1715 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags; 1716 kev.data = (intptr_t)aiocbe; 1717 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr; 1718 error = kqfd_register(kqfd, &kev, td, 1); 1719aqueue_fail: 1720 if (error) { 1721 if (fp) 1722 fdrop(fp, td); 1723 uma_zfree(aiocb_zone, aiocbe); 1724 ops->store_error(job, error); 1725 goto done; 1726 } 1727no_kqueue: 1728 1729 ops->store_error(job, EINPROGRESS); 1730 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS; 1731 aiocbe->userproc = p; 1732 aiocbe->cred = crhold(td->td_ucred); 1733 aiocbe->jobflags = 0; 1734 aiocbe->lio = lj; 1735 1736 if (opcode == LIO_SYNC) 1737 goto queueit; 1738 1739 if (fp && fp->f_type == DTYPE_SOCKET) { 1740 /* 1741 * Alternate queueing for socket ops: Reach down into the 1742 * descriptor to get the socket data. Then check to see if the 1743 * socket is ready to be read or written (based on the requested 1744 * operation). 1745 * 1746 * If it is not ready for io, then queue the aiocbe on the 1747 * socket, and set the flags so we get a call when sbnotify() 1748 * happens. 1749 * 1750 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock 1751 * and unlock the snd sockbuf for no reason. 1752 */ 1753 so = fp->f_data; 1754 sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd; 1755 SOCKBUF_LOCK(sb); 1756 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode == 1757 LIO_WRITE) && (!sowriteable(so)))) { 1758 sb->sb_flags |= SB_AIO; 1759 1760 mtx_lock(&aio_job_mtx); 1761 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list); 1762 mtx_unlock(&aio_job_mtx); 1763 1764 AIO_LOCK(ki); 1765 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1766 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1767 aiocbe->jobstate = JOBST_JOBQSOCK; 1768 ki->kaio_count++; 1769 if (lj) 1770 lj->lioj_count++; 1771 AIO_UNLOCK(ki); 1772 SOCKBUF_UNLOCK(sb); 1773 atomic_add_int(&num_queue_count, 1); 1774 error = 0; 1775 goto done; 1776 } 1777 SOCKBUF_UNLOCK(sb); 1778 } 1779 1780 if ((error = aio_qphysio(p, aiocbe)) == 0) 1781 goto done; 1782#if 0 1783 if (error > 0) { 1784 aiocbe->uaiocb._aiocb_private.error = error; 1785 ops->store_error(job, error); 1786 goto done; 1787 } 1788#endif 1789queueit: 1790 /* No buffer for daemon I/O. */ 1791 aiocbe->bp = NULL; 1792 atomic_add_int(&num_queue_count, 1); 1793 1794 AIO_LOCK(ki); 1795 ki->kaio_count++; 1796 if (lj) 1797 lj->lioj_count++; 1798 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist); 1799 TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist); 1800 if (opcode == LIO_SYNC) { 1801 TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) { 1802 if (cb->fd_file == aiocbe->fd_file && 1803 cb->uaiocb.aio_lio_opcode != LIO_SYNC && 1804 cb->seqno < aiocbe->seqno) { 1805 cb->jobflags |= AIOCBLIST_CHECKSYNC; 1806 aiocbe->pending++; 1807 } 1808 } 1809 TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) { 1810 if (cb->fd_file == aiocbe->fd_file && 1811 cb->uaiocb.aio_lio_opcode != LIO_SYNC && 1812 cb->seqno < aiocbe->seqno) { 1813 cb->jobflags |= AIOCBLIST_CHECKSYNC; 1814 aiocbe->pending++; 1815 } 1816 } 1817 if (aiocbe->pending != 0) { 1818 TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list); 1819 aiocbe->jobstate = JOBST_JOBQSYNC; 1820 AIO_UNLOCK(ki); 1821 goto done; 1822 } 1823 } 1824 mtx_lock(&aio_job_mtx); 1825 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list); 1826 aiocbe->jobstate = JOBST_JOBQGLOBAL; 1827 aio_kick_nowait(p); 1828 mtx_unlock(&aio_job_mtx); 1829 AIO_UNLOCK(ki); 1830 error = 0; 1831done: 1832 return (error); 1833} 1834 1835static void 1836aio_kick_nowait(struct proc *userp) 1837{ 1838 struct kaioinfo *ki = userp->p_aioinfo; 1839 struct aiothreadlist *aiop; 1840 1841 mtx_assert(&aio_job_mtx, MA_OWNED); 1842 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1843 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1844 aiop->aiothreadflags &= ~AIOP_FREE; 1845 wakeup(aiop->aiothread); 1846 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1847 ((ki->kaio_active_count + num_aio_resv_start) < 1848 ki->kaio_maxactive_count)) { 1849 taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task); 1850 } 1851} 1852 1853static int 1854aio_kick(struct proc *userp) 1855{ 1856 struct kaioinfo *ki = userp->p_aioinfo; 1857 struct aiothreadlist *aiop; 1858 int error, ret = 0; 1859 1860 mtx_assert(&aio_job_mtx, MA_OWNED); 1861retryproc: 1862 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) { 1863 TAILQ_REMOVE(&aio_freeproc, aiop, list); 1864 aiop->aiothreadflags &= ~AIOP_FREE; 1865 wakeup(aiop->aiothread); 1866 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) && 1867 ((ki->kaio_active_count + num_aio_resv_start) < 1868 ki->kaio_maxactive_count)) { 1869 num_aio_resv_start++; 1870 mtx_unlock(&aio_job_mtx); 1871 error = aio_newproc(&num_aio_resv_start); 1872 mtx_lock(&aio_job_mtx); 1873 if (error) { 1874 num_aio_resv_start--; 1875 goto retryproc; 1876 } 1877 } else { 1878 ret = -1; 1879 } 1880 return (ret); 1881} 1882 1883static void 1884aio_kick_helper(void *context, int pending) 1885{ 1886 struct proc *userp = context; 1887 1888 mtx_lock(&aio_job_mtx); 1889 while (--pending >= 0) { 1890 if (aio_kick(userp)) 1891 break; 1892 } 1893 mtx_unlock(&aio_job_mtx); 1894} 1895 1896/* 1897 * Support the aio_return system call, as a side-effect, kernel resources are 1898 * released. 1899 */ 1900static int 1901kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops) 1902{ 1903 struct proc *p = td->td_proc; 1904 struct aiocblist *cb; 1905 struct kaioinfo *ki; 1906 int status, error; 1907 1908 ki = p->p_aioinfo; 1909 if (ki == NULL) 1910 return (EINVAL); 1911 AIO_LOCK(ki); 1912 TAILQ_FOREACH(cb, &ki->kaio_done, plist) { 1913 if (cb->uuaiocb == uaiocb) 1914 break; 1915 } 1916 if (cb != NULL) { 1917 MPASS(cb->jobstate == JOBST_JOBFINISHED); 1918 status = cb->uaiocb._aiocb_private.status; 1919 error = cb->uaiocb._aiocb_private.error; 1920 td->td_retval[0] = status; 1921 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 1922 td->td_ru.ru_oublock += cb->outputcharge; 1923 cb->outputcharge = 0; 1924 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 1925 td->td_ru.ru_inblock += cb->inputcharge; 1926 cb->inputcharge = 0; 1927 } 1928 aio_free_entry(cb); 1929 AIO_UNLOCK(ki); 1930 ops->store_error(uaiocb, error); 1931 ops->store_status(uaiocb, status); 1932 } else { 1933 error = EINVAL; 1934 AIO_UNLOCK(ki); 1935 } 1936 return (error); 1937} 1938 1939int 1940sys_aio_return(struct thread *td, struct aio_return_args *uap) 1941{ 1942 1943 return (kern_aio_return(td, uap->aiocbp, &aiocb_ops)); 1944} 1945 1946/* 1947 * Allow a process to wakeup when any of the I/O requests are completed. 1948 */ 1949static int 1950kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist, 1951 struct timespec *ts) 1952{ 1953 struct proc *p = td->td_proc; 1954 struct timeval atv; 1955 struct kaioinfo *ki; 1956 struct aiocblist *cb, *cbfirst; 1957 int error, i, timo; 1958 1959 timo = 0; 1960 if (ts) { 1961 if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000) 1962 return (EINVAL); 1963 1964 TIMESPEC_TO_TIMEVAL(&atv, ts); 1965 if (itimerfix(&atv)) 1966 return (EINVAL); 1967 timo = tvtohz(&atv); 1968 } 1969 1970 ki = p->p_aioinfo; 1971 if (ki == NULL) 1972 return (EAGAIN); 1973 1974 if (njoblist == 0) 1975 return (0); 1976 1977 AIO_LOCK(ki); 1978 for (;;) { 1979 cbfirst = NULL; 1980 error = 0; 1981 TAILQ_FOREACH(cb, &ki->kaio_all, allist) { 1982 for (i = 0; i < njoblist; i++) { 1983 if (cb->uuaiocb == ujoblist[i]) { 1984 if (cbfirst == NULL) 1985 cbfirst = cb; 1986 if (cb->jobstate == JOBST_JOBFINISHED) 1987 goto RETURN; 1988 } 1989 } 1990 } 1991 /* All tasks were finished. */ 1992 if (cbfirst == NULL) 1993 break; 1994 1995 ki->kaio_flags |= KAIO_WAKEUP; 1996 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 1997 "aiospn", timo); 1998 if (error == ERESTART) 1999 error = EINTR; 2000 if (error) 2001 break; 2002 } 2003RETURN: 2004 AIO_UNLOCK(ki); 2005 return (error); 2006} 2007 2008int 2009sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap) 2010{ 2011 struct timespec ts, *tsp; 2012 struct aiocb **ujoblist; 2013 int error; 2014 2015 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX) 2016 return (EINVAL); 2017 2018 if (uap->timeout) { 2019 /* Get timespec struct. */ 2020 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0) 2021 return (error); 2022 tsp = &ts; 2023 } else 2024 tsp = NULL; 2025 2026 ujoblist = uma_zalloc(aiol_zone, M_WAITOK); 2027 error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0])); 2028 if (error == 0) 2029 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2030 uma_zfree(aiol_zone, ujoblist); 2031 return (error); 2032} 2033 2034/* 2035 * aio_cancel cancels any non-physio aio operations not currently in 2036 * progress. 2037 */ 2038int 2039sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap) 2040{ 2041 struct proc *p = td->td_proc; 2042 struct kaioinfo *ki; 2043 struct aiocblist *cbe, *cbn; 2044 struct file *fp; 2045 struct socket *so; 2046 int error; 2047 int remove; 2048 int cancelled = 0; 2049 int notcancelled = 0; 2050 struct vnode *vp; 2051 2052 /* Lookup file object. */ 2053 error = fget(td, uap->fd, NULL, &fp); 2054 if (error) 2055 return (error); 2056 2057 ki = p->p_aioinfo; 2058 if (ki == NULL) 2059 goto done; 2060 2061 if (fp->f_type == DTYPE_VNODE) { 2062 vp = fp->f_vnode; 2063 if (vn_isdisk(vp, &error)) { 2064 fdrop(fp, td); 2065 td->td_retval[0] = AIO_NOTCANCELED; 2066 return (0); 2067 } 2068 } 2069 2070 AIO_LOCK(ki); 2071 TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) { 2072 if ((uap->fd == cbe->uaiocb.aio_fildes) && 2073 ((uap->aiocbp == NULL) || 2074 (uap->aiocbp == cbe->uuaiocb))) { 2075 remove = 0; 2076 2077 mtx_lock(&aio_job_mtx); 2078 if (cbe->jobstate == JOBST_JOBQGLOBAL) { 2079 TAILQ_REMOVE(&aio_jobs, cbe, list); 2080 remove = 1; 2081 } else if (cbe->jobstate == JOBST_JOBQSOCK) { 2082 MPASS(fp->f_type == DTYPE_SOCKET); 2083 so = fp->f_data; 2084 TAILQ_REMOVE(&so->so_aiojobq, cbe, list); 2085 remove = 1; 2086 } else if (cbe->jobstate == JOBST_JOBQSYNC) { 2087 TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list); 2088 remove = 1; 2089 } 2090 mtx_unlock(&aio_job_mtx); 2091 2092 if (remove) { 2093 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist); 2094 cbe->uaiocb._aiocb_private.status = -1; 2095 cbe->uaiocb._aiocb_private.error = ECANCELED; 2096 aio_bio_done_notify(p, cbe, DONE_QUEUE); 2097 cancelled++; 2098 } else { 2099 notcancelled++; 2100 } 2101 if (uap->aiocbp != NULL) 2102 break; 2103 } 2104 } 2105 AIO_UNLOCK(ki); 2106 2107done: 2108 fdrop(fp, td); 2109 2110 if (uap->aiocbp != NULL) { 2111 if (cancelled) { 2112 td->td_retval[0] = AIO_CANCELED; 2113 return (0); 2114 } 2115 } 2116 2117 if (notcancelled) { 2118 td->td_retval[0] = AIO_NOTCANCELED; 2119 return (0); 2120 } 2121 2122 if (cancelled) { 2123 td->td_retval[0] = AIO_CANCELED; 2124 return (0); 2125 } 2126 2127 td->td_retval[0] = AIO_ALLDONE; 2128 2129 return (0); 2130} 2131 2132/* 2133 * aio_error is implemented in the kernel level for compatibility purposes 2134 * only. For a user mode async implementation, it would be best to do it in 2135 * a userland subroutine. 2136 */ 2137static int 2138kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops) 2139{ 2140 struct proc *p = td->td_proc; 2141 struct aiocblist *cb; 2142 struct kaioinfo *ki; 2143 int status; 2144 2145 ki = p->p_aioinfo; 2146 if (ki == NULL) { 2147 td->td_retval[0] = EINVAL; 2148 return (0); 2149 } 2150 2151 AIO_LOCK(ki); 2152 TAILQ_FOREACH(cb, &ki->kaio_all, allist) { 2153 if (cb->uuaiocb == aiocbp) { 2154 if (cb->jobstate == JOBST_JOBFINISHED) 2155 td->td_retval[0] = 2156 cb->uaiocb._aiocb_private.error; 2157 else 2158 td->td_retval[0] = EINPROGRESS; 2159 AIO_UNLOCK(ki); 2160 return (0); 2161 } 2162 } 2163 AIO_UNLOCK(ki); 2164 2165 /* 2166 * Hack for failure of aio_aqueue. 2167 */ 2168 status = ops->fetch_status(aiocbp); 2169 if (status == -1) { 2170 td->td_retval[0] = ops->fetch_error(aiocbp); 2171 return (0); 2172 } 2173 2174 td->td_retval[0] = EINVAL; 2175 return (0); 2176} 2177 2178int 2179sys_aio_error(struct thread *td, struct aio_error_args *uap) 2180{ 2181 2182 return (kern_aio_error(td, uap->aiocbp, &aiocb_ops)); 2183} 2184 2185/* syscall - asynchronous read from a file (REALTIME) */ 2186int 2187sys_oaio_read(struct thread *td, struct oaio_read_args *uap) 2188{ 2189 2190 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2191 &aiocb_ops_osigevent)); 2192} 2193 2194int 2195sys_aio_read(struct thread *td, struct aio_read_args *uap) 2196{ 2197 2198 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops)); 2199} 2200 2201/* syscall - asynchronous write to a file (REALTIME) */ 2202int 2203sys_oaio_write(struct thread *td, struct oaio_write_args *uap) 2204{ 2205 2206 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2207 &aiocb_ops_osigevent)); 2208} 2209 2210int 2211sys_aio_write(struct thread *td, struct aio_write_args *uap) 2212{ 2213 2214 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops)); 2215} 2216 2217int 2218sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap) 2219{ 2220 2221 return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops)); 2222} 2223 2224static int 2225kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list, 2226 struct aiocb **acb_list, int nent, struct sigevent *sig, 2227 struct aiocb_ops *ops) 2228{ 2229 struct proc *p = td->td_proc; 2230 struct aiocb *iocb; 2231 struct kaioinfo *ki; 2232 struct aioliojob *lj; 2233 struct kevent kev; 2234 int error; 2235 int nerror; 2236 int i; 2237 2238 if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT)) 2239 return (EINVAL); 2240 2241 if (nent < 0 || nent > AIO_LISTIO_MAX) 2242 return (EINVAL); 2243 2244 if (p->p_aioinfo == NULL) 2245 aio_init_aioinfo(p); 2246 2247 ki = p->p_aioinfo; 2248 2249 lj = uma_zalloc(aiolio_zone, M_WAITOK); 2250 lj->lioj_flags = 0; 2251 lj->lioj_count = 0; 2252 lj->lioj_finished_count = 0; 2253 knlist_init_mtx(&lj->klist, AIO_MTX(ki)); 2254 ksiginfo_init(&lj->lioj_ksi); 2255 2256 /* 2257 * Setup signal. 2258 */ 2259 if (sig && (mode == LIO_NOWAIT)) { 2260 bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal)); 2261 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2262 /* Assume only new style KEVENT */ 2263 kev.filter = EVFILT_LIO; 2264 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1; 2265 kev.ident = (uintptr_t)uacb_list; /* something unique */ 2266 kev.data = (intptr_t)lj; 2267 /* pass user defined sigval data */ 2268 kev.udata = lj->lioj_signal.sigev_value.sival_ptr; 2269 error = kqfd_register( 2270 lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1); 2271 if (error) { 2272 uma_zfree(aiolio_zone, lj); 2273 return (error); 2274 } 2275 } else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) { 2276 ; 2277 } else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2278 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) { 2279 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) { 2280 uma_zfree(aiolio_zone, lj); 2281 return EINVAL; 2282 } 2283 lj->lioj_flags |= LIOJ_SIGNAL; 2284 } else { 2285 uma_zfree(aiolio_zone, lj); 2286 return EINVAL; 2287 } 2288 } 2289 2290 AIO_LOCK(ki); 2291 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list); 2292 /* 2293 * Add extra aiocb count to avoid the lio to be freed 2294 * by other threads doing aio_waitcomplete or aio_return, 2295 * and prevent event from being sent until we have queued 2296 * all tasks. 2297 */ 2298 lj->lioj_count = 1; 2299 AIO_UNLOCK(ki); 2300 2301 /* 2302 * Get pointers to the list of I/O requests. 2303 */ 2304 nerror = 0; 2305 for (i = 0; i < nent; i++) { 2306 iocb = acb_list[i]; 2307 if (iocb != NULL) { 2308 error = aio_aqueue(td, iocb, lj, LIO_NOP, ops); 2309 if (error != 0) 2310 nerror++; 2311 } 2312 } 2313 2314 error = 0; 2315 AIO_LOCK(ki); 2316 if (mode == LIO_WAIT) { 2317 while (lj->lioj_count - 1 != lj->lioj_finished_count) { 2318 ki->kaio_flags |= KAIO_WAKEUP; 2319 error = msleep(&p->p_aioinfo, AIO_MTX(ki), 2320 PRIBIO | PCATCH, "aiospn", 0); 2321 if (error == ERESTART) 2322 error = EINTR; 2323 if (error) 2324 break; 2325 } 2326 } else { 2327 if (lj->lioj_count - 1 == lj->lioj_finished_count) { 2328 if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) { 2329 lj->lioj_flags |= LIOJ_KEVENT_POSTED; 2330 KNOTE_LOCKED(&lj->klist, 1); 2331 } 2332 if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) 2333 == LIOJ_SIGNAL 2334 && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL || 2335 lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) { 2336 aio_sendsig(p, &lj->lioj_signal, 2337 &lj->lioj_ksi); 2338 lj->lioj_flags |= LIOJ_SIGNAL_POSTED; 2339 } 2340 } 2341 } 2342 lj->lioj_count--; 2343 if (lj->lioj_count == 0) { 2344 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list); 2345 knlist_delete(&lj->klist, curthread, 1); 2346 PROC_LOCK(p); 2347 sigqueue_take(&lj->lioj_ksi); 2348 PROC_UNLOCK(p); 2349 AIO_UNLOCK(ki); 2350 uma_zfree(aiolio_zone, lj); 2351 } else 2352 AIO_UNLOCK(ki); 2353 2354 if (nerror) 2355 return (EIO); 2356 return (error); 2357} 2358 2359/* syscall - list directed I/O (REALTIME) */ 2360int 2361sys_olio_listio(struct thread *td, struct olio_listio_args *uap) 2362{ 2363 struct aiocb **acb_list; 2364 struct sigevent *sigp, sig; 2365 struct osigevent osig; 2366 int error, nent; 2367 2368 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2369 return (EINVAL); 2370 2371 nent = uap->nent; 2372 if (nent < 0 || nent > AIO_LISTIO_MAX) 2373 return (EINVAL); 2374 2375 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2376 error = copyin(uap->sig, &osig, sizeof(osig)); 2377 if (error) 2378 return (error); 2379 error = convert_old_sigevent(&osig, &sig); 2380 if (error) 2381 return (error); 2382 sigp = &sig; 2383 } else 2384 sigp = NULL; 2385 2386 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2387 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2388 if (error == 0) 2389 error = kern_lio_listio(td, uap->mode, 2390 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 2391 &aiocb_ops_osigevent); 2392 free(acb_list, M_LIO); 2393 return (error); 2394} 2395 2396/* syscall - list directed I/O (REALTIME) */ 2397int 2398sys_lio_listio(struct thread *td, struct lio_listio_args *uap) 2399{ 2400 struct aiocb **acb_list; 2401 struct sigevent *sigp, sig; 2402 int error, nent; 2403 2404 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2405 return (EINVAL); 2406 2407 nent = uap->nent; 2408 if (nent < 0 || nent > AIO_LISTIO_MAX) 2409 return (EINVAL); 2410 2411 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2412 error = copyin(uap->sig, &sig, sizeof(sig)); 2413 if (error) 2414 return (error); 2415 sigp = &sig; 2416 } else 2417 sigp = NULL; 2418 2419 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 2420 error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0])); 2421 if (error == 0) 2422 error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list, 2423 nent, sigp, &aiocb_ops); 2424 free(acb_list, M_LIO); 2425 return (error); 2426} 2427 2428/* 2429 * Called from interrupt thread for physio, we should return as fast 2430 * as possible, so we schedule a biohelper task. 2431 */ 2432static void 2433aio_physwakeup(struct buf *bp) 2434{ 2435 struct aiocblist *aiocbe; 2436 2437 aiocbe = (struct aiocblist *)bp->b_caller1; 2438 taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask); 2439} 2440 2441/* 2442 * Task routine to perform heavy tasks, process wakeup, and signals. 2443 */ 2444static void 2445biohelper(void *context, int pending) 2446{ 2447 struct aiocblist *aiocbe = context; 2448 struct buf *bp; 2449 struct proc *userp; 2450 struct kaioinfo *ki; 2451 int nblks; 2452 2453 bp = aiocbe->bp; 2454 userp = aiocbe->userproc; 2455 ki = userp->p_aioinfo; 2456 AIO_LOCK(ki); 2457 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid; 2458 aiocbe->uaiocb._aiocb_private.error = 0; 2459 if (bp->b_ioflags & BIO_ERROR) 2460 aiocbe->uaiocb._aiocb_private.error = bp->b_error; 2461 nblks = btodb(aiocbe->uaiocb.aio_nbytes); 2462 if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE) 2463 aiocbe->outputcharge += nblks; 2464 else 2465 aiocbe->inputcharge += nblks; 2466 aiocbe->bp = NULL; 2467 TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist); 2468 ki->kaio_buffer_count--; 2469 aio_bio_done_notify(userp, aiocbe, DONE_BUF); 2470 AIO_UNLOCK(ki); 2471 2472 /* Release mapping into kernel space. */ 2473 vunmapbuf(bp); 2474 relpbuf(bp, NULL); 2475 atomic_subtract_int(&num_buf_aio, 1); 2476} 2477 2478/* syscall - wait for the next completion of an aio request */ 2479static int 2480kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp, 2481 struct timespec *ts, struct aiocb_ops *ops) 2482{ 2483 struct proc *p = td->td_proc; 2484 struct timeval atv; 2485 struct kaioinfo *ki; 2486 struct aiocblist *cb; 2487 struct aiocb *uuaiocb; 2488 int error, status, timo; 2489 2490 ops->store_aiocb(aiocbp, NULL); 2491 2492 timo = 0; 2493 if (ts) { 2494 if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000)) 2495 return (EINVAL); 2496 2497 TIMESPEC_TO_TIMEVAL(&atv, ts); 2498 if (itimerfix(&atv)) 2499 return (EINVAL); 2500 timo = tvtohz(&atv); 2501 } 2502 2503 if (p->p_aioinfo == NULL) 2504 aio_init_aioinfo(p); 2505 ki = p->p_aioinfo; 2506 2507 error = 0; 2508 cb = NULL; 2509 AIO_LOCK(ki); 2510 while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) { 2511 ki->kaio_flags |= KAIO_WAKEUP; 2512 error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH, 2513 "aiowc", timo); 2514 if (timo && error == ERESTART) 2515 error = EINTR; 2516 if (error) 2517 break; 2518 } 2519 2520 if (cb != NULL) { 2521 MPASS(cb->jobstate == JOBST_JOBFINISHED); 2522 uuaiocb = cb->uuaiocb; 2523 status = cb->uaiocb._aiocb_private.status; 2524 error = cb->uaiocb._aiocb_private.error; 2525 td->td_retval[0] = status; 2526 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) { 2527 td->td_ru.ru_oublock += cb->outputcharge; 2528 cb->outputcharge = 0; 2529 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) { 2530 td->td_ru.ru_inblock += cb->inputcharge; 2531 cb->inputcharge = 0; 2532 } 2533 aio_free_entry(cb); 2534 AIO_UNLOCK(ki); 2535 ops->store_aiocb(aiocbp, uuaiocb); 2536 ops->store_error(uuaiocb, error); 2537 ops->store_status(uuaiocb, status); 2538 } else 2539 AIO_UNLOCK(ki); 2540 2541 return (error); 2542} 2543 2544int 2545sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap) 2546{ 2547 struct timespec ts, *tsp; 2548 int error; 2549 2550 if (uap->timeout) { 2551 /* Get timespec struct. */ 2552 error = copyin(uap->timeout, &ts, sizeof(ts)); 2553 if (error) 2554 return (error); 2555 tsp = &ts; 2556 } else 2557 tsp = NULL; 2558 2559 return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops)); 2560} 2561 2562static int 2563kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp, 2564 struct aiocb_ops *ops) 2565{ 2566 struct proc *p = td->td_proc; 2567 struct kaioinfo *ki; 2568 2569 if (op != O_SYNC) /* XXX lack of O_DSYNC */ 2570 return (EINVAL); 2571 ki = p->p_aioinfo; 2572 if (ki == NULL) 2573 aio_init_aioinfo(p); 2574 return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops)); 2575} 2576 2577int 2578sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap) 2579{ 2580 2581 return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops)); 2582} 2583 2584/* kqueue attach function */ 2585static int 2586filt_aioattach(struct knote *kn) 2587{ 2588 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata; 2589 2590 /* 2591 * The aiocbe pointer must be validated before using it, so 2592 * registration is restricted to the kernel; the user cannot 2593 * set EV_FLAG1. 2594 */ 2595 if ((kn->kn_flags & EV_FLAG1) == 0) 2596 return (EPERM); 2597 kn->kn_ptr.p_aio = aiocbe; 2598 kn->kn_flags &= ~EV_FLAG1; 2599 2600 knlist_add(&aiocbe->klist, kn, 0); 2601 2602 return (0); 2603} 2604 2605/* kqueue detach function */ 2606static void 2607filt_aiodetach(struct knote *kn) 2608{ 2609 struct knlist *knl; 2610 2611 knl = &kn->kn_ptr.p_aio->klist; 2612 knl->kl_lock(knl->kl_lockarg); 2613 if (!knlist_empty(knl)) 2614 knlist_remove(knl, kn, 1); 2615 knl->kl_unlock(knl->kl_lockarg); 2616} 2617 2618/* kqueue filter function */ 2619/*ARGSUSED*/ 2620static int 2621filt_aio(struct knote *kn, long hint) 2622{ 2623 struct aiocblist *aiocbe = kn->kn_ptr.p_aio; 2624 2625 kn->kn_data = aiocbe->uaiocb._aiocb_private.error; 2626 if (aiocbe->jobstate != JOBST_JOBFINISHED) 2627 return (0); 2628 kn->kn_flags |= EV_EOF; 2629 return (1); 2630} 2631 2632/* kqueue attach function */ 2633static int 2634filt_lioattach(struct knote *kn) 2635{ 2636 struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata; 2637 2638 /* 2639 * The aioliojob pointer must be validated before using it, so 2640 * registration is restricted to the kernel; the user cannot 2641 * set EV_FLAG1. 2642 */ 2643 if ((kn->kn_flags & EV_FLAG1) == 0) 2644 return (EPERM); 2645 kn->kn_ptr.p_lio = lj; 2646 kn->kn_flags &= ~EV_FLAG1; 2647 2648 knlist_add(&lj->klist, kn, 0); 2649 2650 return (0); 2651} 2652 2653/* kqueue detach function */ 2654static void 2655filt_liodetach(struct knote *kn) 2656{ 2657 struct knlist *knl; 2658 2659 knl = &kn->kn_ptr.p_lio->klist; 2660 knl->kl_lock(knl->kl_lockarg); 2661 if (!knlist_empty(knl)) 2662 knlist_remove(knl, kn, 1); 2663 knl->kl_unlock(knl->kl_lockarg); 2664} 2665 2666/* kqueue filter function */ 2667/*ARGSUSED*/ 2668static int 2669filt_lio(struct knote *kn, long hint) 2670{ 2671 struct aioliojob * lj = kn->kn_ptr.p_lio; 2672 2673 return (lj->lioj_flags & LIOJ_KEVENT_POSTED); 2674} 2675 2676#ifdef COMPAT_FREEBSD32 2677 2678struct __aiocb_private32 { 2679 int32_t status; 2680 int32_t error; 2681 uint32_t kernelinfo; 2682}; 2683 2684typedef struct oaiocb32 { 2685 int aio_fildes; /* File descriptor */ 2686 uint64_t aio_offset __packed; /* File offset for I/O */ 2687 uint32_t aio_buf; /* I/O buffer in process space */ 2688 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2689 struct osigevent32 aio_sigevent; /* Signal to deliver */ 2690 int aio_lio_opcode; /* LIO opcode */ 2691 int aio_reqprio; /* Request priority -- ignored */ 2692 struct __aiocb_private32 _aiocb_private; 2693} oaiocb32_t; 2694 2695typedef struct aiocb32 { 2696 int32_t aio_fildes; /* File descriptor */ 2697 uint64_t aio_offset __packed; /* File offset for I/O */ 2698 uint32_t aio_buf; /* I/O buffer in process space */ 2699 uint32_t aio_nbytes; /* Number of bytes for I/O */ 2700 int __spare__[2]; 2701 uint32_t __spare2__; 2702 int aio_lio_opcode; /* LIO opcode */ 2703 int aio_reqprio; /* Request priority -- ignored */ 2704 struct __aiocb_private32 _aiocb_private; 2705 struct sigevent32 aio_sigevent; /* Signal to deliver */ 2706} aiocb32_t; 2707 2708static int 2709convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig) 2710{ 2711 2712 /* 2713 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are 2714 * supported by AIO with the old sigevent structure. 2715 */ 2716 CP(*osig, *nsig, sigev_notify); 2717 switch (nsig->sigev_notify) { 2718 case SIGEV_NONE: 2719 break; 2720 case SIGEV_SIGNAL: 2721 nsig->sigev_signo = osig->__sigev_u.__sigev_signo; 2722 break; 2723 case SIGEV_KEVENT: 2724 nsig->sigev_notify_kqueue = 2725 osig->__sigev_u.__sigev_notify_kqueue; 2726 PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr); 2727 break; 2728 default: 2729 return (EINVAL); 2730 } 2731 return (0); 2732} 2733 2734static int 2735aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob) 2736{ 2737 struct oaiocb32 job32; 2738 int error; 2739 2740 bzero(kjob, sizeof(struct aiocb)); 2741 error = copyin(ujob, &job32, sizeof(job32)); 2742 if (error) 2743 return (error); 2744 2745 CP(job32, *kjob, aio_fildes); 2746 CP(job32, *kjob, aio_offset); 2747 PTRIN_CP(job32, *kjob, aio_buf); 2748 CP(job32, *kjob, aio_nbytes); 2749 CP(job32, *kjob, aio_lio_opcode); 2750 CP(job32, *kjob, aio_reqprio); 2751 CP(job32, *kjob, _aiocb_private.status); 2752 CP(job32, *kjob, _aiocb_private.error); 2753 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo); 2754 return (convert_old_sigevent32(&job32.aio_sigevent, 2755 &kjob->aio_sigevent)); 2756} 2757 2758static int 2759aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob) 2760{ 2761 struct aiocb32 job32; 2762 int error; 2763 2764 error = copyin(ujob, &job32, sizeof(job32)); 2765 if (error) 2766 return (error); 2767 CP(job32, *kjob, aio_fildes); 2768 CP(job32, *kjob, aio_offset); 2769 PTRIN_CP(job32, *kjob, aio_buf); 2770 CP(job32, *kjob, aio_nbytes); 2771 CP(job32, *kjob, aio_lio_opcode); 2772 CP(job32, *kjob, aio_reqprio); 2773 CP(job32, *kjob, _aiocb_private.status); 2774 CP(job32, *kjob, _aiocb_private.error); 2775 PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo); 2776 return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent)); 2777} 2778 2779static long 2780aiocb32_fetch_status(struct aiocb *ujob) 2781{ 2782 struct aiocb32 *ujob32; 2783 2784 ujob32 = (struct aiocb32 *)ujob; 2785 return (fuword32(&ujob32->_aiocb_private.status)); 2786} 2787 2788static long 2789aiocb32_fetch_error(struct aiocb *ujob) 2790{ 2791 struct aiocb32 *ujob32; 2792 2793 ujob32 = (struct aiocb32 *)ujob; 2794 return (fuword32(&ujob32->_aiocb_private.error)); 2795} 2796 2797static int 2798aiocb32_store_status(struct aiocb *ujob, long status) 2799{ 2800 struct aiocb32 *ujob32; 2801 2802 ujob32 = (struct aiocb32 *)ujob; 2803 return (suword32(&ujob32->_aiocb_private.status, status)); 2804} 2805 2806static int 2807aiocb32_store_error(struct aiocb *ujob, long error) 2808{ 2809 struct aiocb32 *ujob32; 2810 2811 ujob32 = (struct aiocb32 *)ujob; 2812 return (suword32(&ujob32->_aiocb_private.error, error)); 2813} 2814 2815static int 2816aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref) 2817{ 2818 struct aiocb32 *ujob32; 2819 2820 ujob32 = (struct aiocb32 *)ujob; 2821 return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref)); 2822} 2823 2824static int 2825aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob) 2826{ 2827 2828 return (suword32(ujobp, (long)ujob)); 2829} 2830 2831static struct aiocb_ops aiocb32_ops = { 2832 .copyin = aiocb32_copyin, 2833 .fetch_status = aiocb32_fetch_status, 2834 .fetch_error = aiocb32_fetch_error, 2835 .store_status = aiocb32_store_status, 2836 .store_error = aiocb32_store_error, 2837 .store_kernelinfo = aiocb32_store_kernelinfo, 2838 .store_aiocb = aiocb32_store_aiocb, 2839}; 2840 2841static struct aiocb_ops aiocb32_ops_osigevent = { 2842 .copyin = aiocb32_copyin_old_sigevent, 2843 .fetch_status = aiocb32_fetch_status, 2844 .fetch_error = aiocb32_fetch_error, 2845 .store_status = aiocb32_store_status, 2846 .store_error = aiocb32_store_error, 2847 .store_kernelinfo = aiocb32_store_kernelinfo, 2848 .store_aiocb = aiocb32_store_aiocb, 2849}; 2850 2851int 2852freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap) 2853{ 2854 2855 return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2856} 2857 2858int 2859freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap) 2860{ 2861 struct timespec32 ts32; 2862 struct timespec ts, *tsp; 2863 struct aiocb **ujoblist; 2864 uint32_t *ujoblist32; 2865 int error, i; 2866 2867 if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX) 2868 return (EINVAL); 2869 2870 if (uap->timeout) { 2871 /* Get timespec struct. */ 2872 if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0) 2873 return (error); 2874 CP(ts32, ts, tv_sec); 2875 CP(ts32, ts, tv_nsec); 2876 tsp = &ts; 2877 } else 2878 tsp = NULL; 2879 2880 ujoblist = uma_zalloc(aiol_zone, M_WAITOK); 2881 ujoblist32 = (uint32_t *)ujoblist; 2882 error = copyin(uap->aiocbp, ujoblist32, uap->nent * 2883 sizeof(ujoblist32[0])); 2884 if (error == 0) { 2885 for (i = uap->nent; i > 0; i--) 2886 ujoblist[i] = PTRIN(ujoblist32[i]); 2887 2888 error = kern_aio_suspend(td, uap->nent, ujoblist, tsp); 2889 } 2890 uma_zfree(aiol_zone, ujoblist); 2891 return (error); 2892} 2893 2894int 2895freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap) 2896{ 2897 2898 return (sys_aio_cancel(td, (struct aio_cancel_args *)uap)); 2899} 2900 2901int 2902freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap) 2903{ 2904 2905 return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops)); 2906} 2907 2908int 2909freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap) 2910{ 2911 2912 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2913 &aiocb32_ops_osigevent)); 2914} 2915 2916int 2917freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap) 2918{ 2919 2920 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 2921 &aiocb32_ops)); 2922} 2923 2924int 2925freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap) 2926{ 2927 2928 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2929 &aiocb32_ops_osigevent)); 2930} 2931 2932int 2933freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap) 2934{ 2935 2936 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 2937 &aiocb32_ops)); 2938} 2939 2940int 2941freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap) 2942{ 2943 2944 return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK, 2945 &aiocb32_ops)); 2946} 2947 2948int 2949freebsd32_aio_waitcomplete(struct thread *td, 2950 struct freebsd32_aio_waitcomplete_args *uap) 2951{ 2952 struct timespec32 ts32; 2953 struct timespec ts, *tsp; 2954 int error; 2955 2956 if (uap->timeout) { 2957 /* Get timespec struct. */ 2958 error = copyin(uap->timeout, &ts32, sizeof(ts32)); 2959 if (error) 2960 return (error); 2961 CP(ts32, ts, tv_sec); 2962 CP(ts32, ts, tv_nsec); 2963 tsp = &ts; 2964 } else 2965 tsp = NULL; 2966 2967 return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp, 2968 &aiocb32_ops)); 2969} 2970 2971int 2972freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap) 2973{ 2974 2975 return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp, 2976 &aiocb32_ops)); 2977} 2978 2979int 2980freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap) 2981{ 2982 struct aiocb **acb_list; 2983 struct sigevent *sigp, sig; 2984 struct osigevent32 osig; 2985 uint32_t *acb_list32; 2986 int error, i, nent; 2987 2988 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 2989 return (EINVAL); 2990 2991 nent = uap->nent; 2992 if (nent < 0 || nent > AIO_LISTIO_MAX) 2993 return (EINVAL); 2994 2995 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 2996 error = copyin(uap->sig, &osig, sizeof(osig)); 2997 if (error) 2998 return (error); 2999 error = convert_old_sigevent32(&osig, &sig); 3000 if (error) 3001 return (error); 3002 sigp = &sig; 3003 } else 3004 sigp = NULL; 3005 3006 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3007 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3008 if (error) { 3009 free(acb_list32, M_LIO); 3010 return (error); 3011 } 3012 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3013 for (i = 0; i < nent; i++) 3014 acb_list[i] = PTRIN(acb_list32[i]); 3015 free(acb_list32, M_LIO); 3016 3017 error = kern_lio_listio(td, uap->mode, 3018 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3019 &aiocb32_ops_osigevent); 3020 free(acb_list, M_LIO); 3021 return (error); 3022} 3023 3024int 3025freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap) 3026{ 3027 struct aiocb **acb_list; 3028 struct sigevent *sigp, sig; 3029 struct sigevent32 sig32; 3030 uint32_t *acb_list32; 3031 int error, i, nent; 3032 3033 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) 3034 return (EINVAL); 3035 3036 nent = uap->nent; 3037 if (nent < 0 || nent > AIO_LISTIO_MAX) 3038 return (EINVAL); 3039 3040 if (uap->sig && (uap->mode == LIO_NOWAIT)) { 3041 error = copyin(uap->sig, &sig32, sizeof(sig32)); 3042 if (error) 3043 return (error); 3044 error = convert_sigevent32(&sig32, &sig); 3045 if (error) 3046 return (error); 3047 sigp = &sig; 3048 } else 3049 sigp = NULL; 3050 3051 acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK); 3052 error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t)); 3053 if (error) { 3054 free(acb_list32, M_LIO); 3055 return (error); 3056 } 3057 acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK); 3058 for (i = 0; i < nent; i++) 3059 acb_list[i] = PTRIN(acb_list32[i]); 3060 free(acb_list32, M_LIO); 3061 3062 error = kern_lio_listio(td, uap->mode, 3063 (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp, 3064 &aiocb32_ops); 3065 free(acb_list, M_LIO); 3066 return (error); 3067} 3068 3069#endif 3070