kern_alq.c revision 241896
1/*- 2 * Copyright (c) 2002, Jeffrey Roberson <jeff@freebsd.org> 3 * Copyright (c) 2008-2009, Lawrence Stewart <lstewart@freebsd.org> 4 * Copyright (c) 2009-2010, The FreeBSD Foundation 5 * All rights reserved. 6 * 7 * Portions of this software were developed at the Centre for Advanced 8 * Internet Architectures, Swinburne University of Technology, Melbourne, 9 * Australia by Lawrence Stewart under sponsorship from the FreeBSD Foundation. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice unmodified, this list of conditions, and the following 16 * disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 22 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 23 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 24 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 26 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 27 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 28 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 29 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 30 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 31 */ 32 33#include <sys/cdefs.h> 34__FBSDID("$FreeBSD: head/sys/kern/kern_alq.c 241896 2012-10-22 17:50:54Z kib $"); 35 36#include "opt_mac.h" 37 38#include <sys/param.h> 39#include <sys/systm.h> 40#include <sys/kernel.h> 41#include <sys/kthread.h> 42#include <sys/lock.h> 43#include <sys/mount.h> 44#include <sys/mutex.h> 45#include <sys/namei.h> 46#include <sys/proc.h> 47#include <sys/vnode.h> 48#include <sys/alq.h> 49#include <sys/malloc.h> 50#include <sys/unistd.h> 51#include <sys/fcntl.h> 52#include <sys/eventhandler.h> 53 54#include <security/mac/mac_framework.h> 55 56/* Async. Logging Queue */ 57struct alq { 58 char *aq_entbuf; /* Buffer for stored entries */ 59 int aq_entmax; /* Max entries */ 60 int aq_entlen; /* Entry length */ 61 int aq_freebytes; /* Bytes available in buffer */ 62 int aq_buflen; /* Total length of our buffer */ 63 int aq_writehead; /* Location for next write */ 64 int aq_writetail; /* Flush starts at this location */ 65 int aq_wrapearly; /* # bytes left blank at end of buf */ 66 int aq_flags; /* Queue flags */ 67 int aq_waiters; /* Num threads waiting for resources 68 * NB: Used as a wait channel so must 69 * not be first field in the alq struct 70 */ 71 struct ale aq_getpost; /* ALE for use by get/post */ 72 struct mtx aq_mtx; /* Queue lock */ 73 struct vnode *aq_vp; /* Open vnode handle */ 74 struct ucred *aq_cred; /* Credentials of the opening thread */ 75 LIST_ENTRY(alq) aq_act; /* List of active queues */ 76 LIST_ENTRY(alq) aq_link; /* List of all queues */ 77}; 78 79#define AQ_WANTED 0x0001 /* Wakeup sleeper when io is done */ 80#define AQ_ACTIVE 0x0002 /* on the active list */ 81#define AQ_FLUSHING 0x0004 /* doing IO */ 82#define AQ_SHUTDOWN 0x0008 /* Queue no longer valid */ 83#define AQ_ORDERED 0x0010 /* Queue enforces ordered writes */ 84#define AQ_LEGACY 0x0020 /* Legacy queue (fixed length writes) */ 85 86#define ALQ_LOCK(alq) mtx_lock_spin(&(alq)->aq_mtx) 87#define ALQ_UNLOCK(alq) mtx_unlock_spin(&(alq)->aq_mtx) 88 89#define HAS_PENDING_DATA(alq) ((alq)->aq_freebytes != (alq)->aq_buflen) 90 91static MALLOC_DEFINE(M_ALD, "ALD", "ALD"); 92 93/* 94 * The ald_mtx protects the ald_queues list and the ald_active list. 95 */ 96static struct mtx ald_mtx; 97static LIST_HEAD(, alq) ald_queues; 98static LIST_HEAD(, alq) ald_active; 99static int ald_shutingdown = 0; 100struct thread *ald_thread; 101static struct proc *ald_proc; 102 103#define ALD_LOCK() mtx_lock(&ald_mtx) 104#define ALD_UNLOCK() mtx_unlock(&ald_mtx) 105 106/* Daemon functions */ 107static int ald_add(struct alq *); 108static int ald_rem(struct alq *); 109static void ald_startup(void *); 110static void ald_daemon(void); 111static void ald_shutdown(void *, int); 112static void ald_activate(struct alq *); 113static void ald_deactivate(struct alq *); 114 115/* Internal queue functions */ 116static void alq_shutdown(struct alq *); 117static void alq_destroy(struct alq *); 118static int alq_doio(struct alq *); 119 120 121/* 122 * Add a new queue to the global list. Fail if we're shutting down. 123 */ 124static int 125ald_add(struct alq *alq) 126{ 127 int error; 128 129 error = 0; 130 131 ALD_LOCK(); 132 if (ald_shutingdown) { 133 error = EBUSY; 134 goto done; 135 } 136 LIST_INSERT_HEAD(&ald_queues, alq, aq_link); 137done: 138 ALD_UNLOCK(); 139 return (error); 140} 141 142/* 143 * Remove a queue from the global list unless we're shutting down. If so, 144 * the ald will take care of cleaning up it's resources. 145 */ 146static int 147ald_rem(struct alq *alq) 148{ 149 int error; 150 151 error = 0; 152 153 ALD_LOCK(); 154 if (ald_shutingdown) { 155 error = EBUSY; 156 goto done; 157 } 158 LIST_REMOVE(alq, aq_link); 159done: 160 ALD_UNLOCK(); 161 return (error); 162} 163 164/* 165 * Put a queue on the active list. This will schedule it for writing. 166 */ 167static void 168ald_activate(struct alq *alq) 169{ 170 LIST_INSERT_HEAD(&ald_active, alq, aq_act); 171 wakeup(&ald_active); 172} 173 174static void 175ald_deactivate(struct alq *alq) 176{ 177 LIST_REMOVE(alq, aq_act); 178 alq->aq_flags &= ~AQ_ACTIVE; 179} 180 181static void 182ald_startup(void *unused) 183{ 184 mtx_init(&ald_mtx, "ALDmtx", NULL, MTX_DEF|MTX_QUIET); 185 LIST_INIT(&ald_queues); 186 LIST_INIT(&ald_active); 187} 188 189static void 190ald_daemon(void) 191{ 192 int needwakeup; 193 struct alq *alq; 194 195 ald_thread = FIRST_THREAD_IN_PROC(ald_proc); 196 197 EVENTHANDLER_REGISTER(shutdown_pre_sync, ald_shutdown, NULL, 198 SHUTDOWN_PRI_FIRST); 199 200 ALD_LOCK(); 201 202 for (;;) { 203 while ((alq = LIST_FIRST(&ald_active)) == NULL && 204 !ald_shutingdown) 205 mtx_sleep(&ald_active, &ald_mtx, PWAIT, "aldslp", 0); 206 207 /* Don't shutdown until all active ALQs are flushed. */ 208 if (ald_shutingdown && alq == NULL) { 209 ALD_UNLOCK(); 210 break; 211 } 212 213 ALQ_LOCK(alq); 214 ald_deactivate(alq); 215 ALD_UNLOCK(); 216 needwakeup = alq_doio(alq); 217 ALQ_UNLOCK(alq); 218 if (needwakeup) 219 wakeup_one(alq); 220 ALD_LOCK(); 221 } 222 223 kproc_exit(0); 224} 225 226static void 227ald_shutdown(void *arg, int howto) 228{ 229 struct alq *alq; 230 231 ALD_LOCK(); 232 233 /* Ensure no new queues can be created. */ 234 ald_shutingdown = 1; 235 236 /* Shutdown all ALQs prior to terminating the ald_daemon. */ 237 while ((alq = LIST_FIRST(&ald_queues)) != NULL) { 238 LIST_REMOVE(alq, aq_link); 239 ALD_UNLOCK(); 240 alq_shutdown(alq); 241 ALD_LOCK(); 242 } 243 244 /* At this point, all ALQs are flushed and shutdown. */ 245 246 /* 247 * Wake ald_daemon so that it exits. It won't be able to do 248 * anything until we mtx_sleep because we hold the ald_mtx. 249 */ 250 wakeup(&ald_active); 251 252 /* Wait for ald_daemon to exit. */ 253 mtx_sleep(ald_proc, &ald_mtx, PWAIT, "aldslp", 0); 254 255 ALD_UNLOCK(); 256} 257 258static void 259alq_shutdown(struct alq *alq) 260{ 261 ALQ_LOCK(alq); 262 263 /* Stop any new writers. */ 264 alq->aq_flags |= AQ_SHUTDOWN; 265 266 /* 267 * If the ALQ isn't active but has unwritten data (possible if 268 * the ALQ_NOACTIVATE flag has been used), explicitly activate the 269 * ALQ here so that the pending data gets flushed by the ald_daemon. 270 */ 271 if (!(alq->aq_flags & AQ_ACTIVE) && HAS_PENDING_DATA(alq)) { 272 alq->aq_flags |= AQ_ACTIVE; 273 ALQ_UNLOCK(alq); 274 ALD_LOCK(); 275 ald_activate(alq); 276 ALD_UNLOCK(); 277 ALQ_LOCK(alq); 278 } 279 280 /* Drain IO */ 281 while (alq->aq_flags & AQ_ACTIVE) { 282 alq->aq_flags |= AQ_WANTED; 283 msleep_spin(alq, &alq->aq_mtx, "aldclose", 0); 284 } 285 ALQ_UNLOCK(alq); 286 287 vn_close(alq->aq_vp, FWRITE, alq->aq_cred, 288 curthread); 289 crfree(alq->aq_cred); 290} 291 292void 293alq_destroy(struct alq *alq) 294{ 295 /* Drain all pending IO. */ 296 alq_shutdown(alq); 297 298 mtx_destroy(&alq->aq_mtx); 299 free(alq->aq_entbuf, M_ALD); 300 free(alq, M_ALD); 301} 302 303/* 304 * Flush all pending data to disk. This operation will block. 305 */ 306static int 307alq_doio(struct alq *alq) 308{ 309 struct thread *td; 310 struct mount *mp; 311 struct vnode *vp; 312 struct uio auio; 313 struct iovec aiov[2]; 314 int totlen; 315 int iov; 316 int wrapearly; 317 318 KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__)); 319 320 vp = alq->aq_vp; 321 td = curthread; 322 totlen = 0; 323 iov = 1; 324 wrapearly = alq->aq_wrapearly; 325 326 bzero(&aiov, sizeof(aiov)); 327 bzero(&auio, sizeof(auio)); 328 329 /* Start the write from the location of our buffer tail pointer. */ 330 aiov[0].iov_base = alq->aq_entbuf + alq->aq_writetail; 331 332 if (alq->aq_writetail < alq->aq_writehead) { 333 /* Buffer not wrapped. */ 334 totlen = aiov[0].iov_len = alq->aq_writehead - alq->aq_writetail; 335 } else if (alq->aq_writehead == 0) { 336 /* Buffer not wrapped (special case to avoid an empty iov). */ 337 totlen = aiov[0].iov_len = alq->aq_buflen - alq->aq_writetail - 338 wrapearly; 339 } else { 340 /* 341 * Buffer wrapped, requires 2 aiov entries: 342 * - first is from writetail to end of buffer 343 * - second is from start of buffer to writehead 344 */ 345 aiov[0].iov_len = alq->aq_buflen - alq->aq_writetail - 346 wrapearly; 347 iov++; 348 aiov[1].iov_base = alq->aq_entbuf; 349 aiov[1].iov_len = alq->aq_writehead; 350 totlen = aiov[0].iov_len + aiov[1].iov_len; 351 } 352 353 alq->aq_flags |= AQ_FLUSHING; 354 ALQ_UNLOCK(alq); 355 356 auio.uio_iov = &aiov[0]; 357 auio.uio_offset = 0; 358 auio.uio_segflg = UIO_SYSSPACE; 359 auio.uio_rw = UIO_WRITE; 360 auio.uio_iovcnt = iov; 361 auio.uio_resid = totlen; 362 auio.uio_td = td; 363 364 /* 365 * Do all of the junk required to write now. 366 */ 367 vn_start_write(vp, &mp, V_WAIT); 368 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 369 /* 370 * XXX: VOP_WRITE error checks are ignored. 371 */ 372#ifdef MAC 373 if (mac_vnode_check_write(alq->aq_cred, NOCRED, vp) == 0) 374#endif 375 VOP_WRITE(vp, &auio, IO_UNIT | IO_APPEND, alq->aq_cred); 376 VOP_UNLOCK(vp, 0); 377 vn_finished_write(mp); 378 379 ALQ_LOCK(alq); 380 alq->aq_flags &= ~AQ_FLUSHING; 381 382 /* Adjust writetail as required, taking into account wrapping. */ 383 alq->aq_writetail = (alq->aq_writetail + totlen + wrapearly) % 384 alq->aq_buflen; 385 alq->aq_freebytes += totlen + wrapearly; 386 387 /* 388 * If we just flushed part of the buffer which wrapped, reset the 389 * wrapearly indicator. 390 */ 391 if (wrapearly) 392 alq->aq_wrapearly = 0; 393 394 /* 395 * If we just flushed the buffer completely, reset indexes to 0 to 396 * minimise buffer wraps. 397 * This is also required to ensure alq_getn() can't wedge itself. 398 */ 399 if (!HAS_PENDING_DATA(alq)) 400 alq->aq_writehead = alq->aq_writetail = 0; 401 402 KASSERT((alq->aq_writetail >= 0 && alq->aq_writetail < alq->aq_buflen), 403 ("%s: aq_writetail < 0 || aq_writetail >= aq_buflen", __func__)); 404 405 if (alq->aq_flags & AQ_WANTED) { 406 alq->aq_flags &= ~AQ_WANTED; 407 return (1); 408 } 409 410 return(0); 411} 412 413static struct kproc_desc ald_kp = { 414 "ALQ Daemon", 415 ald_daemon, 416 &ald_proc 417}; 418 419SYSINIT(aldthread, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, kproc_start, &ald_kp); 420SYSINIT(ald, SI_SUB_LOCK, SI_ORDER_ANY, ald_startup, NULL); 421 422 423/* User visible queue functions */ 424 425/* 426 * Create the queue data structure, allocate the buffer, and open the file. 427 */ 428 429int 430alq_open_flags(struct alq **alqp, const char *file, struct ucred *cred, int cmode, 431 int size, int flags) 432{ 433 struct thread *td; 434 struct nameidata nd; 435 struct alq *alq; 436 int oflags; 437 int error; 438 439 KASSERT((size > 0), ("%s: size <= 0", __func__)); 440 441 *alqp = NULL; 442 td = curthread; 443 444 NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, file, td); 445 oflags = FWRITE | O_NOFOLLOW | O_CREAT; 446 447 error = vn_open_cred(&nd, &oflags, cmode, 0, cred, NULL); 448 if (error) 449 return (error); 450 451 NDFREE(&nd, NDF_ONLY_PNBUF); 452 /* We just unlock so we hold a reference */ 453 VOP_UNLOCK(nd.ni_vp, 0); 454 455 alq = malloc(sizeof(*alq), M_ALD, M_WAITOK|M_ZERO); 456 alq->aq_vp = nd.ni_vp; 457 alq->aq_cred = crhold(cred); 458 459 mtx_init(&alq->aq_mtx, "ALD Queue", NULL, MTX_SPIN|MTX_QUIET); 460 461 alq->aq_buflen = size; 462 alq->aq_entmax = 0; 463 alq->aq_entlen = 0; 464 465 alq->aq_freebytes = alq->aq_buflen; 466 alq->aq_entbuf = malloc(alq->aq_buflen, M_ALD, M_WAITOK|M_ZERO); 467 alq->aq_writehead = alq->aq_writetail = 0; 468 if (flags & ALQ_ORDERED) 469 alq->aq_flags |= AQ_ORDERED; 470 471 if ((error = ald_add(alq)) != 0) { 472 alq_destroy(alq); 473 return (error); 474 } 475 476 *alqp = alq; 477 478 return (0); 479} 480 481int 482alq_open(struct alq **alqp, const char *file, struct ucred *cred, int cmode, 483 int size, int count) 484{ 485 int ret; 486 487 KASSERT((count >= 0), ("%s: count < 0", __func__)); 488 489 if (count > 0) { 490 ret = alq_open_flags(alqp, file, cred, cmode, size*count, 0); 491 (*alqp)->aq_flags |= AQ_LEGACY; 492 (*alqp)->aq_entmax = count; 493 (*alqp)->aq_entlen = size; 494 } else 495 ret = alq_open_flags(alqp, file, cred, cmode, size, 0); 496 497 return (ret); 498} 499 500 501/* 502 * Copy a new entry into the queue. If the operation would block either 503 * wait or return an error depending on the value of waitok. 504 */ 505int 506alq_writen(struct alq *alq, void *data, int len, int flags) 507{ 508 int activate, copy, ret; 509 void *waitchan; 510 511 KASSERT((len > 0 && len <= alq->aq_buflen), 512 ("%s: len <= 0 || len > aq_buflen", __func__)); 513 514 activate = ret = 0; 515 copy = len; 516 waitchan = NULL; 517 518 ALQ_LOCK(alq); 519 520 /* 521 * Fail to perform the write and return EWOULDBLOCK if: 522 * - The message is larger than our underlying buffer. 523 * - The ALQ is being shutdown. 524 * - There is insufficient free space in our underlying buffer 525 * to accept the message and the user can't wait for space. 526 * - There is insufficient free space in our underlying buffer 527 * to accept the message and the alq is inactive due to prior 528 * use of the ALQ_NOACTIVATE flag (which would lead to deadlock). 529 */ 530 if (len > alq->aq_buflen || 531 alq->aq_flags & AQ_SHUTDOWN || 532 (((flags & ALQ_NOWAIT) || (!(alq->aq_flags & AQ_ACTIVE) && 533 HAS_PENDING_DATA(alq))) && alq->aq_freebytes < len)) { 534 ALQ_UNLOCK(alq); 535 return (EWOULDBLOCK); 536 } 537 538 /* 539 * If we want ordered writes and there is already at least one thread 540 * waiting for resources to become available, sleep until we're woken. 541 */ 542 if (alq->aq_flags & AQ_ORDERED && alq->aq_waiters > 0) { 543 KASSERT(!(flags & ALQ_NOWAIT), 544 ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__)); 545 alq->aq_waiters++; 546 msleep_spin(&alq->aq_waiters, &alq->aq_mtx, "alqwnord", 0); 547 alq->aq_waiters--; 548 } 549 550 /* 551 * (ALQ_WAITOK && aq_freebytes < len) or aq_freebytes >= len, either 552 * enter while loop and sleep until we have enough free bytes (former) 553 * or skip (latter). If AQ_ORDERED is set, only 1 thread at a time will 554 * be in this loop. Otherwise, multiple threads may be sleeping here 555 * competing for ALQ resources. 556 */ 557 while (alq->aq_freebytes < len && !(alq->aq_flags & AQ_SHUTDOWN)) { 558 KASSERT(!(flags & ALQ_NOWAIT), 559 ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__)); 560 alq->aq_flags |= AQ_WANTED; 561 alq->aq_waiters++; 562 if (waitchan) 563 wakeup(waitchan); 564 msleep_spin(alq, &alq->aq_mtx, "alqwnres", 0); 565 alq->aq_waiters--; 566 567 /* 568 * If we're the first thread to wake after an AQ_WANTED wakeup 569 * but there isn't enough free space for us, we're going to loop 570 * and sleep again. If there are other threads waiting in this 571 * loop, schedule a wakeup so that they can see if the space 572 * they require is available. 573 */ 574 if (alq->aq_waiters > 0 && !(alq->aq_flags & AQ_ORDERED) && 575 alq->aq_freebytes < len && !(alq->aq_flags & AQ_WANTED)) 576 waitchan = alq; 577 else 578 waitchan = NULL; 579 } 580 581 /* 582 * If there are waiters, we need to signal the waiting threads after we 583 * complete our work. The alq ptr is used as a wait channel for threads 584 * requiring resources to be freed up. In the AQ_ORDERED case, threads 585 * are not allowed to concurrently compete for resources in the above 586 * while loop, so we use a different wait channel in this case. 587 */ 588 if (alq->aq_waiters > 0) { 589 if (alq->aq_flags & AQ_ORDERED) 590 waitchan = &alq->aq_waiters; 591 else 592 waitchan = alq; 593 } else 594 waitchan = NULL; 595 596 /* Bail if we're shutting down. */ 597 if (alq->aq_flags & AQ_SHUTDOWN) { 598 ret = EWOULDBLOCK; 599 goto unlock; 600 } 601 602 /* 603 * If we need to wrap the buffer to accommodate the write, 604 * we'll need 2 calls to bcopy. 605 */ 606 if ((alq->aq_buflen - alq->aq_writehead) < len) 607 copy = alq->aq_buflen - alq->aq_writehead; 608 609 /* Copy message (or part thereof if wrap required) to the buffer. */ 610 bcopy(data, alq->aq_entbuf + alq->aq_writehead, copy); 611 alq->aq_writehead += copy; 612 613 if (alq->aq_writehead >= alq->aq_buflen) { 614 KASSERT((alq->aq_writehead == alq->aq_buflen), 615 ("%s: alq->aq_writehead (%d) > alq->aq_buflen (%d)", 616 __func__, 617 alq->aq_writehead, 618 alq->aq_buflen)); 619 alq->aq_writehead = 0; 620 } 621 622 if (copy != len) { 623 /* 624 * Wrap the buffer by copying the remainder of our message 625 * to the start of the buffer and resetting aq_writehead. 626 */ 627 bcopy(((uint8_t *)data)+copy, alq->aq_entbuf, len - copy); 628 alq->aq_writehead = len - copy; 629 } 630 631 KASSERT((alq->aq_writehead >= 0 && alq->aq_writehead < alq->aq_buflen), 632 ("%s: aq_writehead < 0 || aq_writehead >= aq_buflen", __func__)); 633 634 alq->aq_freebytes -= len; 635 636 if (!(alq->aq_flags & AQ_ACTIVE) && !(flags & ALQ_NOACTIVATE)) { 637 alq->aq_flags |= AQ_ACTIVE; 638 activate = 1; 639 } 640 641 KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__)); 642 643unlock: 644 ALQ_UNLOCK(alq); 645 646 if (activate) { 647 ALD_LOCK(); 648 ald_activate(alq); 649 ALD_UNLOCK(); 650 } 651 652 /* NB: We rely on wakeup_one waking threads in a FIFO manner. */ 653 if (waitchan != NULL) 654 wakeup_one(waitchan); 655 656 return (ret); 657} 658 659int 660alq_write(struct alq *alq, void *data, int flags) 661{ 662 /* Should only be called in fixed length message (legacy) mode. */ 663 KASSERT((alq->aq_flags & AQ_LEGACY), 664 ("%s: fixed length write on variable length queue", __func__)); 665 return (alq_writen(alq, data, alq->aq_entlen, flags)); 666} 667 668/* 669 * Retrieve a pointer for the ALQ to write directly into, avoiding bcopy. 670 */ 671struct ale * 672alq_getn(struct alq *alq, int len, int flags) 673{ 674 int contigbytes; 675 void *waitchan; 676 677 KASSERT((len > 0 && len <= alq->aq_buflen), 678 ("%s: len <= 0 || len > alq->aq_buflen", __func__)); 679 680 waitchan = NULL; 681 682 ALQ_LOCK(alq); 683 684 /* 685 * Determine the number of free contiguous bytes. 686 * We ensure elsewhere that if aq_writehead == aq_writetail because 687 * the buffer is empty, they will both be set to 0 and therefore 688 * aq_freebytes == aq_buflen and is fully contiguous. 689 * If they are equal and the buffer is not empty, aq_freebytes will 690 * be 0 indicating the buffer is full. 691 */ 692 if (alq->aq_writehead <= alq->aq_writetail) 693 contigbytes = alq->aq_freebytes; 694 else { 695 contigbytes = alq->aq_buflen - alq->aq_writehead; 696 697 if (contigbytes < len) { 698 /* 699 * Insufficient space at end of buffer to handle a 700 * contiguous write. Wrap early if there's space at 701 * the beginning. This will leave a hole at the end 702 * of the buffer which we will have to skip over when 703 * flushing the buffer to disk. 704 */ 705 if (alq->aq_writetail >= len || flags & ALQ_WAITOK) { 706 /* Keep track of # bytes left blank. */ 707 alq->aq_wrapearly = contigbytes; 708 /* Do the wrap and adjust counters. */ 709 contigbytes = alq->aq_freebytes = 710 alq->aq_writetail; 711 alq->aq_writehead = 0; 712 } 713 } 714 } 715 716 /* 717 * Return a NULL ALE if: 718 * - The message is larger than our underlying buffer. 719 * - The ALQ is being shutdown. 720 * - There is insufficient free space in our underlying buffer 721 * to accept the message and the user can't wait for space. 722 * - There is insufficient free space in our underlying buffer 723 * to accept the message and the alq is inactive due to prior 724 * use of the ALQ_NOACTIVATE flag (which would lead to deadlock). 725 */ 726 if (len > alq->aq_buflen || 727 alq->aq_flags & AQ_SHUTDOWN || 728 (((flags & ALQ_NOWAIT) || (!(alq->aq_flags & AQ_ACTIVE) && 729 HAS_PENDING_DATA(alq))) && contigbytes < len)) { 730 ALQ_UNLOCK(alq); 731 return (NULL); 732 } 733 734 /* 735 * If we want ordered writes and there is already at least one thread 736 * waiting for resources to become available, sleep until we're woken. 737 */ 738 if (alq->aq_flags & AQ_ORDERED && alq->aq_waiters > 0) { 739 KASSERT(!(flags & ALQ_NOWAIT), 740 ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__)); 741 alq->aq_waiters++; 742 msleep_spin(&alq->aq_waiters, &alq->aq_mtx, "alqgnord", 0); 743 alq->aq_waiters--; 744 } 745 746 /* 747 * (ALQ_WAITOK && contigbytes < len) or contigbytes >= len, either enter 748 * while loop and sleep until we have enough contiguous free bytes 749 * (former) or skip (latter). If AQ_ORDERED is set, only 1 thread at a 750 * time will be in this loop. Otherwise, multiple threads may be 751 * sleeping here competing for ALQ resources. 752 */ 753 while (contigbytes < len && !(alq->aq_flags & AQ_SHUTDOWN)) { 754 KASSERT(!(flags & ALQ_NOWAIT), 755 ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__)); 756 alq->aq_flags |= AQ_WANTED; 757 alq->aq_waiters++; 758 if (waitchan) 759 wakeup(waitchan); 760 msleep_spin(alq, &alq->aq_mtx, "alqgnres", 0); 761 alq->aq_waiters--; 762 763 if (alq->aq_writehead <= alq->aq_writetail) 764 contigbytes = alq->aq_freebytes; 765 else 766 contigbytes = alq->aq_buflen - alq->aq_writehead; 767 768 /* 769 * If we're the first thread to wake after an AQ_WANTED wakeup 770 * but there isn't enough free space for us, we're going to loop 771 * and sleep again. If there are other threads waiting in this 772 * loop, schedule a wakeup so that they can see if the space 773 * they require is available. 774 */ 775 if (alq->aq_waiters > 0 && !(alq->aq_flags & AQ_ORDERED) && 776 contigbytes < len && !(alq->aq_flags & AQ_WANTED)) 777 waitchan = alq; 778 else 779 waitchan = NULL; 780 } 781 782 /* 783 * If there are waiters, we need to signal the waiting threads after we 784 * complete our work. The alq ptr is used as a wait channel for threads 785 * requiring resources to be freed up. In the AQ_ORDERED case, threads 786 * are not allowed to concurrently compete for resources in the above 787 * while loop, so we use a different wait channel in this case. 788 */ 789 if (alq->aq_waiters > 0) { 790 if (alq->aq_flags & AQ_ORDERED) 791 waitchan = &alq->aq_waiters; 792 else 793 waitchan = alq; 794 } else 795 waitchan = NULL; 796 797 /* Bail if we're shutting down. */ 798 if (alq->aq_flags & AQ_SHUTDOWN) { 799 ALQ_UNLOCK(alq); 800 if (waitchan != NULL) 801 wakeup_one(waitchan); 802 return (NULL); 803 } 804 805 /* 806 * If we are here, we have a contiguous number of bytes >= len 807 * available in our buffer starting at aq_writehead. 808 */ 809 alq->aq_getpost.ae_data = alq->aq_entbuf + alq->aq_writehead; 810 alq->aq_getpost.ae_bytesused = len; 811 812 return (&alq->aq_getpost); 813} 814 815struct ale * 816alq_get(struct alq *alq, int flags) 817{ 818 /* Should only be called in fixed length message (legacy) mode. */ 819 KASSERT((alq->aq_flags & AQ_LEGACY), 820 ("%s: fixed length get on variable length queue", __func__)); 821 return (alq_getn(alq, alq->aq_entlen, flags)); 822} 823 824void 825alq_post_flags(struct alq *alq, struct ale *ale, int flags) 826{ 827 int activate; 828 void *waitchan; 829 830 activate = 0; 831 832 if (ale->ae_bytesused > 0) { 833 if (!(alq->aq_flags & AQ_ACTIVE) && 834 !(flags & ALQ_NOACTIVATE)) { 835 alq->aq_flags |= AQ_ACTIVE; 836 activate = 1; 837 } 838 839 alq->aq_writehead += ale->ae_bytesused; 840 alq->aq_freebytes -= ale->ae_bytesused; 841 842 /* Wrap aq_writehead if we filled to the end of the buffer. */ 843 if (alq->aq_writehead == alq->aq_buflen) 844 alq->aq_writehead = 0; 845 846 KASSERT((alq->aq_writehead >= 0 && 847 alq->aq_writehead < alq->aq_buflen), 848 ("%s: aq_writehead < 0 || aq_writehead >= aq_buflen", 849 __func__)); 850 851 KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__)); 852 } 853 854 /* 855 * If there are waiters, we need to signal the waiting threads after we 856 * complete our work. The alq ptr is used as a wait channel for threads 857 * requiring resources to be freed up. In the AQ_ORDERED case, threads 858 * are not allowed to concurrently compete for resources in the 859 * alq_getn() while loop, so we use a different wait channel in this case. 860 */ 861 if (alq->aq_waiters > 0) { 862 if (alq->aq_flags & AQ_ORDERED) 863 waitchan = &alq->aq_waiters; 864 else 865 waitchan = alq; 866 } else 867 waitchan = NULL; 868 869 ALQ_UNLOCK(alq); 870 871 if (activate) { 872 ALD_LOCK(); 873 ald_activate(alq); 874 ALD_UNLOCK(); 875 } 876 877 /* NB: We rely on wakeup_one waking threads in a FIFO manner. */ 878 if (waitchan != NULL) 879 wakeup_one(waitchan); 880} 881 882void 883alq_flush(struct alq *alq) 884{ 885 int needwakeup = 0; 886 887 ALD_LOCK(); 888 ALQ_LOCK(alq); 889 890 /* 891 * Pull the lever iff there is data to flush and we're 892 * not already in the middle of a flush operation. 893 */ 894 if (HAS_PENDING_DATA(alq) && !(alq->aq_flags & AQ_FLUSHING)) { 895 if (alq->aq_flags & AQ_ACTIVE) 896 ald_deactivate(alq); 897 898 ALD_UNLOCK(); 899 needwakeup = alq_doio(alq); 900 } else 901 ALD_UNLOCK(); 902 903 ALQ_UNLOCK(alq); 904 905 if (needwakeup) 906 wakeup_one(alq); 907} 908 909/* 910 * Flush remaining data, close the file and free all resources. 911 */ 912void 913alq_close(struct alq *alq) 914{ 915 /* Only flush and destroy alq if not already shutting down. */ 916 if (ald_rem(alq) == 0) 917 alq_destroy(alq); 918} 919 920static int 921alq_load_handler(module_t mod, int what, void *arg) 922{ 923 int ret; 924 925 ret = 0; 926 927 switch (what) { 928 case MOD_LOAD: 929 case MOD_SHUTDOWN: 930 break; 931 932 case MOD_QUIESCE: 933 ALD_LOCK(); 934 /* Only allow unload if there are no open queues. */ 935 if (LIST_FIRST(&ald_queues) == NULL) { 936 ald_shutingdown = 1; 937 ALD_UNLOCK(); 938 ald_shutdown(NULL, 0); 939 mtx_destroy(&ald_mtx); 940 } else { 941 ALD_UNLOCK(); 942 ret = EBUSY; 943 } 944 break; 945 946 case MOD_UNLOAD: 947 /* If MOD_QUIESCE failed we must fail here too. */ 948 if (ald_shutingdown == 0) 949 ret = EBUSY; 950 break; 951 952 default: 953 ret = EINVAL; 954 break; 955 } 956 957 return (ret); 958} 959 960static moduledata_t alq_mod = 961{ 962 "alq", 963 alq_load_handler, 964 NULL 965}; 966 967DECLARE_MODULE(alq, alq_mod, SI_SUB_SMP, SI_ORDER_ANY); 968MODULE_VERSION(alq, 1); 969