1 /* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $ 15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $ 16 */ 17 18 /* 19 * this file contains a new buffer I/O scheme implementing a coherent 20 * VM object and buffer cache scheme. Pains have been taken to make 21 * sure that the performance degradation associated with schemes such 22 * as this is not realized. 23 * 24 * Author: John S. Dyson 25 * Significant help during the development and debugging phases 26 * had been provided by David Greenman, also of the FreeBSD core team. 27 * 28 * see man buf(9) for more info. 29 */ 30 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/buf.h> 34 #include <sys/conf.h> 35 #include <sys/devicestat.h> 36 #include <sys/eventhandler.h> 37 #include <sys/lock.h> 38 #include <sys/malloc.h> 39 #include <sys/mount.h> 40 #include <sys/kernel.h> 41 #include <sys/kthread.h> 42 #include <sys/proc.h> 43 #include <sys/reboot.h> 44 #include <sys/resourcevar.h> 45 #include <sys/sysctl.h> 46 #include <sys/vmmeter.h> 47 #include <sys/vnode.h> 48 #include <sys/dsched.h> 49 #include <sys/proc.h> 50 #include <vm/vm.h> 51 #include <vm/vm_param.h> 52 #include <vm/vm_kern.h> 53 #include <vm/vm_pageout.h> 54 #include <vm/vm_page.h> 55 #include <vm/vm_object.h> 56 #include <vm/vm_extern.h> 57 #include <vm/vm_map.h> 58 #include <vm/vm_pager.h> 59 #include <vm/swap_pager.h> 60 61 #include <sys/buf2.h> 62 #include <sys/thread2.h> 63 #include <sys/spinlock2.h> 64 #include <sys/mplock2.h> 65 #include <vm/vm_page2.h> 66 67 #include "opt_ddb.h" 68 #ifdef DDB 69 #include <ddb/ddb.h> 70 #endif 71 72 /* 73 * Buffer queues. 74 */ 75 enum bufq_type { 76 BQUEUE_NONE, /* not on any queue */ 77 BQUEUE_LOCKED, /* locked buffers */ 78 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 79 BQUEUE_DIRTY, /* B_DELWRI buffers */ 80 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */ 81 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 82 BQUEUE_EMPTY, /* empty buffer headers */ 83 84 BUFFER_QUEUES /* number of buffer queues */ 85 }; 86 87 typedef enum bufq_type bufq_type_t; 88 89 #define BD_WAKE_SIZE 16384 90 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1) 91 92 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 93 static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin); 94 static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin); 95 96 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 97 98 struct buf *buf; /* buffer header pool */ 99 100 static void vfs_clean_pages(struct buf *bp); 101 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m); 102 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m); 103 static void vfs_vmio_release(struct buf *bp); 104 static int flushbufqueues(bufq_type_t q); 105 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit); 106 107 static void bd_signal(int totalspace); 108 static void buf_daemon(void); 109 static void buf_daemon_hw(void); 110 111 /* 112 * bogus page -- for I/O to/from partially complete buffers 113 * this is a temporary solution to the problem, but it is not 114 * really that bad. it would be better to split the buffer 115 * for input in the case of buffers partially already in memory, 116 * but the code is intricate enough already. 117 */ 118 vm_page_t bogus_page; 119 120 /* 121 * These are all static, but make the ones we export globals so we do 122 * not need to use compiler magic. 123 */ 124 int bufspace; /* locked by buffer_map */ 125 int maxbufspace; 126 static int bufmallocspace; /* atomic ops */ 127 int maxbufmallocspace, lobufspace, hibufspace; 128 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 129 static int lorunningspace; 130 static int hirunningspace; 131 static int runningbufreq; /* locked by bufcspin */ 132 static int dirtybufspace; /* locked by bufcspin */ 133 static int dirtybufcount; /* locked by bufcspin */ 134 static int dirtybufspacehw; /* locked by bufcspin */ 135 static int dirtybufcounthw; /* locked by bufcspin */ 136 static int runningbufspace; /* locked by bufcspin */ 137 static int runningbufcount; /* locked by bufcspin */ 138 int lodirtybufspace; 139 int hidirtybufspace; 140 static int getnewbufcalls; 141 static int getnewbufrestarts; 142 static int recoverbufcalls; 143 static int needsbuffer; /* locked by bufcspin */ 144 static int bd_request; /* locked by bufcspin */ 145 static int bd_request_hw; /* locked by bufcspin */ 146 static u_int bd_wake_ary[BD_WAKE_SIZE]; 147 static u_int bd_wake_index; 148 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 149 static int debug_commit; 150 151 static struct thread *bufdaemon_td; 152 static struct thread *bufdaemonhw_td; 153 static u_int lowmempgallocs; 154 static u_int lowmempgfails; 155 156 /* 157 * Sysctls for operational control of the buffer cache. 158 */ 159 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 160 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 161 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 162 "High watermark used to trigger explicit flushing of dirty buffers"); 163 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 164 "Minimum amount of buffer space required for active I/O"); 165 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 166 "Maximum amount of buffer space to usable for active I/O"); 167 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 168 "Page allocations done during periods of very low free memory"); 169 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 170 "Page allocations which failed during periods of very low free memory"); 171 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 172 "Recycle pages to active or inactive queue transition pt 0-64"); 173 /* 174 * Sysctls determining current state of the buffer cache. 175 */ 176 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 177 "Total number of buffers in buffer cache"); 178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 179 "Pending bytes of dirty buffers (all)"); 180 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 181 "Pending bytes of dirty buffers (heavy weight)"); 182 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 183 "Pending number of dirty buffers"); 184 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 185 "Pending number of dirty buffers (heavy weight)"); 186 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 187 "I/O bytes currently in progress due to asynchronous writes"); 188 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 189 "I/O buffers currently in progress due to asynchronous writes"); 190 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 191 "Hard limit on maximum amount of memory usable for buffer space"); 192 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 193 "Soft limit on maximum amount of memory usable for buffer space"); 194 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 195 "Minimum amount of memory to reserve for system buffer space"); 196 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 197 "Amount of memory available for buffers"); 198 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 199 0, "Maximum amount of memory reserved for buffers using malloc"); 200 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 201 "Amount of memory left for buffers using malloc-scheme"); 202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 203 "New buffer header acquisition requests"); 204 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 205 0, "New buffer header acquisition restarts"); 206 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 207 "Recover VM space in an emergency"); 208 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 209 "Buffer acquisition restarts due to fragmented buffer map"); 210 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 211 "Amount of time KVA space was deallocated in an arbitrary buffer"); 212 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 213 "Amount of time buffer re-use operations were successful"); 214 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, ""); 215 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 216 "sizeof(struct buf)"); 217 218 char *buf_wmesg = BUF_WMESG; 219 220 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 221 #define VFS_BIO_NEED_UNUSED02 0x02 222 #define VFS_BIO_NEED_UNUSED04 0x04 223 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 224 225 /* 226 * bufspacewakeup: 227 * 228 * Called when buffer space is potentially available for recovery. 229 * getnewbuf() will block on this flag when it is unable to free 230 * sufficient buffer space. Buffer space becomes recoverable when 231 * bp's get placed back in the queues. 232 */ 233 static __inline void 234 bufspacewakeup(void) 235 { 236 /* 237 * If someone is waiting for BUF space, wake them up. Even 238 * though we haven't freed the kva space yet, the waiting 239 * process will be able to now. 240 */ 241 spin_lock(&bufcspin); 242 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 243 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 244 spin_unlock(&bufcspin); 245 wakeup(&needsbuffer); 246 } else { 247 spin_unlock(&bufcspin); 248 } 249 } 250 251 /* 252 * runningbufwakeup: 253 * 254 * Accounting for I/O in progress. 255 * 256 */ 257 static __inline void 258 runningbufwakeup(struct buf *bp) 259 { 260 int totalspace; 261 int limit; 262 263 if ((totalspace = bp->b_runningbufspace) != 0) { 264 spin_lock(&bufcspin); 265 runningbufspace -= totalspace; 266 --runningbufcount; 267 bp->b_runningbufspace = 0; 268 269 /* 270 * see waitrunningbufspace() for limit test. 271 */ 272 limit = hirunningspace * 4 / 6; 273 if (runningbufreq && runningbufspace <= limit) { 274 runningbufreq = 0; 275 spin_unlock(&bufcspin); 276 wakeup(&runningbufreq); 277 } else { 278 spin_unlock(&bufcspin); 279 } 280 bd_signal(totalspace); 281 } 282 } 283 284 /* 285 * bufcountwakeup: 286 * 287 * Called when a buffer has been added to one of the free queues to 288 * account for the buffer and to wakeup anyone waiting for free buffers. 289 * This typically occurs when large amounts of metadata are being handled 290 * by the buffer cache ( else buffer space runs out first, usually ). 291 * 292 * MPSAFE 293 */ 294 static __inline void 295 bufcountwakeup(void) 296 { 297 spin_lock(&bufcspin); 298 if (needsbuffer) { 299 needsbuffer &= ~VFS_BIO_NEED_ANY; 300 spin_unlock(&bufcspin); 301 wakeup(&needsbuffer); 302 } else { 303 spin_unlock(&bufcspin); 304 } 305 } 306 307 /* 308 * waitrunningbufspace() 309 * 310 * Wait for the amount of running I/O to drop to hirunningspace * 4 / 6. 311 * This is the point where write bursting stops so we don't want to wait 312 * for the running amount to drop below it (at least if we still want bioq 313 * to burst writes). 314 * 315 * The caller may be using this function to block in a tight loop, we 316 * must block while runningbufspace is greater then or equal to 317 * hirunningspace * 4 / 6. 318 * 319 * And even with that it may not be enough, due to the presence of 320 * B_LOCKED dirty buffers, so also wait for at least one running buffer 321 * to complete. 322 */ 323 void 324 waitrunningbufspace(void) 325 { 326 int limit = hirunningspace * 4 / 6; 327 int dummy; 328 329 spin_lock(&bufcspin); 330 if (runningbufspace > limit) { 331 while (runningbufspace > limit) { 332 ++runningbufreq; 333 ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0); 334 } 335 spin_unlock(&bufcspin); 336 } else if (runningbufspace > limit / 2) { 337 ++runningbufreq; 338 spin_unlock(&bufcspin); 339 tsleep(&dummy, 0, "wdrn2", 1); 340 } else { 341 spin_unlock(&bufcspin); 342 } 343 } 344 345 /* 346 * buf_dirty_count_severe: 347 * 348 * Return true if we have too many dirty buffers. 349 */ 350 int 351 buf_dirty_count_severe(void) 352 { 353 return (runningbufspace + dirtybufspace >= hidirtybufspace || 354 dirtybufcount >= nbuf / 2); 355 } 356 357 /* 358 * Return true if the amount of running I/O is severe and BIOQ should 359 * start bursting. 360 */ 361 int 362 buf_runningbufspace_severe(void) 363 { 364 return (runningbufspace >= hirunningspace * 4 / 6); 365 } 366 367 /* 368 * vfs_buf_test_cache: 369 * 370 * Called when a buffer is extended. This function clears the B_CACHE 371 * bit if the newly extended portion of the buffer does not contain 372 * valid data. 373 * 374 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 375 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 376 * them while a clean buffer was present. 377 */ 378 static __inline__ 379 void 380 vfs_buf_test_cache(struct buf *bp, 381 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 382 vm_page_t m) 383 { 384 if (bp->b_flags & B_CACHE) { 385 int base = (foff + off) & PAGE_MASK; 386 if (vm_page_is_valid(m, base, size) == 0) 387 bp->b_flags &= ~B_CACHE; 388 } 389 } 390 391 /* 392 * bd_speedup() 393 * 394 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 395 * low water mark. 396 * 397 * MPSAFE 398 */ 399 static __inline__ 400 void 401 bd_speedup(void) 402 { 403 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2) 404 return; 405 406 if (bd_request == 0 && 407 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 || 408 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 409 spin_lock(&bufcspin); 410 bd_request = 1; 411 spin_unlock(&bufcspin); 412 wakeup(&bd_request); 413 } 414 if (bd_request_hw == 0 && 415 (dirtybufspacehw > lodirtybufspace / 2 || 416 dirtybufcounthw >= nbuf / 2)) { 417 spin_lock(&bufcspin); 418 bd_request_hw = 1; 419 spin_unlock(&bufcspin); 420 wakeup(&bd_request_hw); 421 } 422 } 423 424 /* 425 * bd_heatup() 426 * 427 * Get the buf_daemon heated up when the number of running and dirty 428 * buffers exceeds the mid-point. 429 * 430 * Return the total number of dirty bytes past the second mid point 431 * as a measure of how much excess dirty data there is in the system. 432 * 433 * MPSAFE 434 */ 435 int 436 bd_heatup(void) 437 { 438 int mid1; 439 int mid2; 440 int totalspace; 441 442 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 443 444 totalspace = runningbufspace + dirtybufspace; 445 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 446 bd_speedup(); 447 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 448 if (totalspace >= mid2) 449 return(totalspace - mid2); 450 } 451 return(0); 452 } 453 454 /* 455 * bd_wait() 456 * 457 * Wait for the buffer cache to flush (totalspace) bytes worth of 458 * buffers, then return. 459 * 460 * Regardless this function blocks while the number of dirty buffers 461 * exceeds hidirtybufspace. 462 * 463 * MPSAFE 464 */ 465 void 466 bd_wait(int totalspace) 467 { 468 u_int i; 469 int count; 470 471 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 472 return; 473 474 while (totalspace > 0) { 475 bd_heatup(); 476 if (totalspace > runningbufspace + dirtybufspace) 477 totalspace = runningbufspace + dirtybufspace; 478 count = totalspace / BKVASIZE; 479 if (count >= BD_WAKE_SIZE) 480 count = BD_WAKE_SIZE - 1; 481 482 spin_lock(&bufcspin); 483 i = (bd_wake_index + count) & BD_WAKE_MASK; 484 ++bd_wake_ary[i]; 485 486 /* 487 * This is not a strict interlock, so we play a bit loose 488 * with locking access to dirtybufspace* 489 */ 490 tsleep_interlock(&bd_wake_ary[i], 0); 491 spin_unlock(&bufcspin); 492 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz); 493 494 totalspace = runningbufspace + dirtybufspace - hidirtybufspace; 495 } 496 } 497 498 /* 499 * bd_signal() 500 * 501 * This function is called whenever runningbufspace or dirtybufspace 502 * is reduced. Track threads waiting for run+dirty buffer I/O 503 * complete. 504 * 505 * MPSAFE 506 */ 507 static void 508 bd_signal(int totalspace) 509 { 510 u_int i; 511 512 if (totalspace > 0) { 513 if (totalspace > BKVASIZE * BD_WAKE_SIZE) 514 totalspace = BKVASIZE * BD_WAKE_SIZE; 515 spin_lock(&bufcspin); 516 while (totalspace > 0) { 517 i = bd_wake_index++; 518 i &= BD_WAKE_MASK; 519 if (bd_wake_ary[i]) { 520 bd_wake_ary[i] = 0; 521 spin_unlock(&bufcspin); 522 wakeup(&bd_wake_ary[i]); 523 spin_lock(&bufcspin); 524 } 525 totalspace -= BKVASIZE; 526 } 527 spin_unlock(&bufcspin); 528 } 529 } 530 531 /* 532 * BIO tracking support routines. 533 * 534 * Release a ref on a bio_track. Wakeup requests are atomically released 535 * along with the last reference so bk_active will never wind up set to 536 * only 0x80000000. 537 * 538 * MPSAFE 539 */ 540 static 541 void 542 bio_track_rel(struct bio_track *track) 543 { 544 int active; 545 int desired; 546 547 /* 548 * Shortcut 549 */ 550 active = track->bk_active; 551 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 552 return; 553 554 /* 555 * Full-on. Note that the wait flag is only atomically released on 556 * the 1->0 count transition. 557 * 558 * We check for a negative count transition using bit 30 since bit 31 559 * has a different meaning. 560 */ 561 for (;;) { 562 desired = (active & 0x7FFFFFFF) - 1; 563 if (desired) 564 desired |= active & 0x80000000; 565 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 566 if (desired & 0x40000000) 567 panic("bio_track_rel: bad count: %p\n", track); 568 if (active & 0x80000000) 569 wakeup(track); 570 break; 571 } 572 active = track->bk_active; 573 } 574 } 575 576 /* 577 * Wait for the tracking count to reach 0. 578 * 579 * Use atomic ops such that the wait flag is only set atomically when 580 * bk_active is non-zero. 581 * 582 * MPSAFE 583 */ 584 int 585 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 586 { 587 int active; 588 int desired; 589 int error; 590 591 /* 592 * Shortcut 593 */ 594 if (track->bk_active == 0) 595 return(0); 596 597 /* 598 * Full-on. Note that the wait flag may only be atomically set if 599 * the active count is non-zero. 600 * 601 * NOTE: We cannot optimize active == desired since a wakeup could 602 * clear active prior to our tsleep_interlock(). 603 */ 604 error = 0; 605 while ((active = track->bk_active) != 0) { 606 cpu_ccfence(); 607 desired = active | 0x80000000; 608 tsleep_interlock(track, slp_flags); 609 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 610 error = tsleep(track, slp_flags | PINTERLOCKED, 611 "trwait", slp_timo); 612 if (error) 613 break; 614 } 615 } 616 return (error); 617 } 618 619 /* 620 * bufinit: 621 * 622 * Load time initialisation of the buffer cache, called from machine 623 * dependant initialization code. 624 */ 625 void 626 bufinit(void) 627 { 628 struct buf *bp; 629 vm_offset_t bogus_offset; 630 int i; 631 632 /* next, make a null set of free lists */ 633 for (i = 0; i < BUFFER_QUEUES; i++) 634 TAILQ_INIT(&bufqueues[i]); 635 636 /* finally, initialize each buffer header and stick on empty q */ 637 for (i = 0; i < nbuf; i++) { 638 bp = &buf[i]; 639 bzero(bp, sizeof *bp); 640 bp->b_flags = B_INVAL; /* we're just an empty header */ 641 bp->b_cmd = BUF_CMD_DONE; 642 bp->b_qindex = BQUEUE_EMPTY; 643 initbufbio(bp); 644 xio_init(&bp->b_xio); 645 buf_dep_init(bp); 646 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist); 647 } 648 649 /* 650 * maxbufspace is the absolute maximum amount of buffer space we are 651 * allowed to reserve in KVM and in real terms. The absolute maximum 652 * is nominally used by buf_daemon. hibufspace is the nominal maximum 653 * used by most other processes. The differential is required to 654 * ensure that buf_daemon is able to run when other processes might 655 * be blocked waiting for buffer space. 656 * 657 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 658 * this may result in KVM fragmentation which is not handled optimally 659 * by the system. 660 */ 661 maxbufspace = nbuf * BKVASIZE; 662 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 663 lobufspace = hibufspace - MAXBSIZE; 664 665 lorunningspace = 512 * 1024; 666 /* hirunningspace -- see below */ 667 668 /* 669 * Limit the amount of malloc memory since it is wired permanently 670 * into the kernel space. Even though this is accounted for in 671 * the buffer allocation, we don't want the malloced region to grow 672 * uncontrolled. The malloc scheme improves memory utilization 673 * significantly on average (small) directories. 674 */ 675 maxbufmallocspace = hibufspace / 20; 676 677 /* 678 * Reduce the chance of a deadlock occuring by limiting the number 679 * of delayed-write dirty buffers we allow to stack up. 680 * 681 * We don't want too much actually queued to the device at once 682 * (XXX this needs to be per-mount!), because the buffers will 683 * wind up locked for a very long period of time while the I/O 684 * drains. 685 */ 686 hidirtybufspace = hibufspace / 2; /* dirty + running */ 687 hirunningspace = hibufspace / 16; /* locked & queued to device */ 688 if (hirunningspace < 1024 * 1024) 689 hirunningspace = 1024 * 1024; 690 691 dirtybufspace = 0; 692 dirtybufspacehw = 0; 693 694 lodirtybufspace = hidirtybufspace / 2; 695 696 /* 697 * Maximum number of async ops initiated per buf_daemon loop. This is 698 * somewhat of a hack at the moment, we really need to limit ourselves 699 * based on the number of bytes of I/O in-transit that were initiated 700 * from buf_daemon. 701 */ 702 703 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 704 bogus_page = vm_page_alloc(&kernel_object, 705 (bogus_offset >> PAGE_SHIFT), 706 VM_ALLOC_NORMAL); 707 vmstats.v_wire_count++; 708 709 } 710 711 /* 712 * Initialize the embedded bio structures, typically used by 713 * deprecated code which tries to allocate its own struct bufs. 714 */ 715 void 716 initbufbio(struct buf *bp) 717 { 718 bp->b_bio1.bio_buf = bp; 719 bp->b_bio1.bio_prev = NULL; 720 bp->b_bio1.bio_offset = NOOFFSET; 721 bp->b_bio1.bio_next = &bp->b_bio2; 722 bp->b_bio1.bio_done = NULL; 723 bp->b_bio1.bio_flags = 0; 724 725 bp->b_bio2.bio_buf = bp; 726 bp->b_bio2.bio_prev = &bp->b_bio1; 727 bp->b_bio2.bio_offset = NOOFFSET; 728 bp->b_bio2.bio_next = NULL; 729 bp->b_bio2.bio_done = NULL; 730 bp->b_bio2.bio_flags = 0; 731 732 BUF_LOCKINIT(bp); 733 } 734 735 /* 736 * Reinitialize the embedded bio structures as well as any additional 737 * translation cache layers. 738 */ 739 void 740 reinitbufbio(struct buf *bp) 741 { 742 struct bio *bio; 743 744 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 745 bio->bio_done = NULL; 746 bio->bio_offset = NOOFFSET; 747 } 748 } 749 750 /* 751 * Undo the effects of an initbufbio(). 752 */ 753 void 754 uninitbufbio(struct buf *bp) 755 { 756 dsched_exit_buf(bp); 757 BUF_LOCKFREE(bp); 758 } 759 760 /* 761 * Push another BIO layer onto an existing BIO and return it. The new 762 * BIO layer may already exist, holding cached translation data. 763 */ 764 struct bio * 765 push_bio(struct bio *bio) 766 { 767 struct bio *nbio; 768 769 if ((nbio = bio->bio_next) == NULL) { 770 int index = bio - &bio->bio_buf->b_bio_array[0]; 771 if (index >= NBUF_BIO - 1) { 772 panic("push_bio: too many layers bp %p\n", 773 bio->bio_buf); 774 } 775 nbio = &bio->bio_buf->b_bio_array[index + 1]; 776 bio->bio_next = nbio; 777 nbio->bio_prev = bio; 778 nbio->bio_buf = bio->bio_buf; 779 nbio->bio_offset = NOOFFSET; 780 nbio->bio_done = NULL; 781 nbio->bio_next = NULL; 782 } 783 KKASSERT(nbio->bio_done == NULL); 784 return(nbio); 785 } 786 787 /* 788 * Pop a BIO translation layer, returning the previous layer. The 789 * must have been previously pushed. 790 */ 791 struct bio * 792 pop_bio(struct bio *bio) 793 { 794 return(bio->bio_prev); 795 } 796 797 void 798 clearbiocache(struct bio *bio) 799 { 800 while (bio) { 801 bio->bio_offset = NOOFFSET; 802 bio = bio->bio_next; 803 } 804 } 805 806 /* 807 * bfreekva: 808 * 809 * Free the KVA allocation for buffer 'bp'. 810 * 811 * Must be called from a critical section as this is the only locking for 812 * buffer_map. 813 * 814 * Since this call frees up buffer space, we call bufspacewakeup(). 815 * 816 * MPALMOSTSAFE 817 */ 818 static void 819 bfreekva(struct buf *bp) 820 { 821 int count; 822 823 if (bp->b_kvasize) { 824 ++buffreekvacnt; 825 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 826 vm_map_lock(&buffer_map); 827 bufspace -= bp->b_kvasize; 828 vm_map_delete(&buffer_map, 829 (vm_offset_t) bp->b_kvabase, 830 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 831 &count 832 ); 833 vm_map_unlock(&buffer_map); 834 vm_map_entry_release(count); 835 bp->b_kvasize = 0; 836 bp->b_kvabase = NULL; 837 bufspacewakeup(); 838 } 839 } 840 841 /* 842 * bremfree: 843 * 844 * Remove the buffer from the appropriate free list. 845 */ 846 static __inline void 847 _bremfree(struct buf *bp) 848 { 849 if (bp->b_qindex != BQUEUE_NONE) { 850 KASSERT(BUF_REFCNTNB(bp) == 1, 851 ("bremfree: bp %p not locked",bp)); 852 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 853 bp->b_qindex = BQUEUE_NONE; 854 } else { 855 if (BUF_REFCNTNB(bp) <= 1) 856 panic("bremfree: removing a buffer not on a queue"); 857 } 858 } 859 860 void 861 bremfree(struct buf *bp) 862 { 863 spin_lock(&bufqspin); 864 _bremfree(bp); 865 spin_unlock(&bufqspin); 866 } 867 868 static void 869 bremfree_locked(struct buf *bp) 870 { 871 _bremfree(bp); 872 } 873 874 /* 875 * bread: 876 * 877 * Get a buffer with the specified data. Look in the cache first. We 878 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 879 * is set, the buffer is valid and we do not have to do anything ( see 880 * getblk() ). 881 * 882 * MPALMOSTSAFE 883 */ 884 int 885 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp) 886 { 887 struct buf *bp; 888 889 bp = getblk(vp, loffset, size, 0, 0); 890 *bpp = bp; 891 892 /* if not found in cache, do some I/O */ 893 if ((bp->b_flags & B_CACHE) == 0) { 894 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 895 bp->b_cmd = BUF_CMD_READ; 896 bp->b_bio1.bio_done = biodone_sync; 897 bp->b_bio1.bio_flags |= BIO_SYNC; 898 vfs_busy_pages(vp, bp); 899 vn_strategy(vp, &bp->b_bio1); 900 return (biowait(&bp->b_bio1, "biord")); 901 } 902 return (0); 903 } 904 905 /* 906 * breadn: 907 * 908 * Operates like bread, but also starts asynchronous I/O on 909 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 910 * to initiating I/O . If B_CACHE is set, the buffer is valid 911 * and we do not have to do anything. 912 * 913 * MPALMOSTSAFE 914 */ 915 int 916 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset, 917 int *rabsize, int cnt, struct buf **bpp) 918 { 919 struct buf *bp, *rabp; 920 int i; 921 int rv = 0, readwait = 0; 922 923 *bpp = bp = getblk(vp, loffset, size, 0, 0); 924 925 /* if not found in cache, do some I/O */ 926 if ((bp->b_flags & B_CACHE) == 0) { 927 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 928 bp->b_cmd = BUF_CMD_READ; 929 bp->b_bio1.bio_done = biodone_sync; 930 bp->b_bio1.bio_flags |= BIO_SYNC; 931 vfs_busy_pages(vp, bp); 932 vn_strategy(vp, &bp->b_bio1); 933 ++readwait; 934 } 935 936 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 937 if (inmem(vp, *raoffset)) 938 continue; 939 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 940 941 if ((rabp->b_flags & B_CACHE) == 0) { 942 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 943 rabp->b_cmd = BUF_CMD_READ; 944 vfs_busy_pages(vp, rabp); 945 BUF_KERNPROC(rabp); 946 vn_strategy(vp, &rabp->b_bio1); 947 } else { 948 brelse(rabp); 949 } 950 } 951 if (readwait) 952 rv = biowait(&bp->b_bio1, "biord"); 953 return (rv); 954 } 955 956 /* 957 * bwrite: 958 * 959 * Synchronous write, waits for completion. 960 * 961 * Write, release buffer on completion. (Done by iodone 962 * if async). Do not bother writing anything if the buffer 963 * is invalid. 964 * 965 * Note that we set B_CACHE here, indicating that buffer is 966 * fully valid and thus cacheable. This is true even of NFS 967 * now so we set it generally. This could be set either here 968 * or in biodone() since the I/O is synchronous. We put it 969 * here. 970 */ 971 int 972 bwrite(struct buf *bp) 973 { 974 int error; 975 976 if (bp->b_flags & B_INVAL) { 977 brelse(bp); 978 return (0); 979 } 980 if (BUF_REFCNTNB(bp) == 0) 981 panic("bwrite: buffer is not busy???"); 982 983 /* Mark the buffer clean */ 984 bundirty(bp); 985 986 bp->b_flags &= ~(B_ERROR | B_EINTR); 987 bp->b_flags |= B_CACHE; 988 bp->b_cmd = BUF_CMD_WRITE; 989 bp->b_bio1.bio_done = biodone_sync; 990 bp->b_bio1.bio_flags |= BIO_SYNC; 991 vfs_busy_pages(bp->b_vp, bp); 992 993 /* 994 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 995 * valid for vnode-backed buffers. 996 */ 997 bsetrunningbufspace(bp, bp->b_bufsize); 998 vn_strategy(bp->b_vp, &bp->b_bio1); 999 error = biowait(&bp->b_bio1, "biows"); 1000 brelse(bp); 1001 1002 return (error); 1003 } 1004 1005 /* 1006 * bawrite: 1007 * 1008 * Asynchronous write. Start output on a buffer, but do not wait for 1009 * it to complete. The buffer is released when the output completes. 1010 * 1011 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1012 * B_INVAL buffers. Not us. 1013 */ 1014 void 1015 bawrite(struct buf *bp) 1016 { 1017 if (bp->b_flags & B_INVAL) { 1018 brelse(bp); 1019 return; 1020 } 1021 if (BUF_REFCNTNB(bp) == 0) 1022 panic("bwrite: buffer is not busy???"); 1023 1024 /* Mark the buffer clean */ 1025 bundirty(bp); 1026 1027 bp->b_flags &= ~(B_ERROR | B_EINTR); 1028 bp->b_flags |= B_CACHE; 1029 bp->b_cmd = BUF_CMD_WRITE; 1030 KKASSERT(bp->b_bio1.bio_done == NULL); 1031 vfs_busy_pages(bp->b_vp, bp); 1032 1033 /* 1034 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1035 * valid for vnode-backed buffers. 1036 */ 1037 bsetrunningbufspace(bp, bp->b_bufsize); 1038 BUF_KERNPROC(bp); 1039 vn_strategy(bp->b_vp, &bp->b_bio1); 1040 } 1041 1042 /* 1043 * bowrite: 1044 * 1045 * Ordered write. Start output on a buffer, and flag it so that the 1046 * device will write it in the order it was queued. The buffer is 1047 * released when the output completes. bwrite() ( or the VOP routine 1048 * anyway ) is responsible for handling B_INVAL buffers. 1049 */ 1050 int 1051 bowrite(struct buf *bp) 1052 { 1053 bp->b_flags |= B_ORDERED; 1054 bawrite(bp); 1055 return (0); 1056 } 1057 1058 /* 1059 * bdwrite: 1060 * 1061 * Delayed write. (Buffer is marked dirty). Do not bother writing 1062 * anything if the buffer is marked invalid. 1063 * 1064 * Note that since the buffer must be completely valid, we can safely 1065 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1066 * biodone() in order to prevent getblk from writing the buffer 1067 * out synchronously. 1068 */ 1069 void 1070 bdwrite(struct buf *bp) 1071 { 1072 if (BUF_REFCNTNB(bp) == 0) 1073 panic("bdwrite: buffer is not busy"); 1074 1075 if (bp->b_flags & B_INVAL) { 1076 brelse(bp); 1077 return; 1078 } 1079 bdirty(bp); 1080 1081 if (dsched_is_clear_buf_priv(bp)) 1082 dsched_new_buf(bp); 1083 1084 /* 1085 * Set B_CACHE, indicating that the buffer is fully valid. This is 1086 * true even of NFS now. 1087 */ 1088 bp->b_flags |= B_CACHE; 1089 1090 /* 1091 * This bmap keeps the system from needing to do the bmap later, 1092 * perhaps when the system is attempting to do a sync. Since it 1093 * is likely that the indirect block -- or whatever other datastructure 1094 * that the filesystem needs is still in memory now, it is a good 1095 * thing to do this. Note also, that if the pageout daemon is 1096 * requesting a sync -- there might not be enough memory to do 1097 * the bmap then... So, this is important to do. 1098 */ 1099 if (bp->b_bio2.bio_offset == NOOFFSET) { 1100 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1101 NULL, NULL, BUF_CMD_WRITE); 1102 } 1103 1104 /* 1105 * Because the underlying pages may still be mapped and 1106 * writable trying to set the dirty buffer (b_dirtyoff/end) 1107 * range here will be inaccurate. 1108 * 1109 * However, we must still clean the pages to satisfy the 1110 * vnode_pager and pageout daemon, so theythink the pages 1111 * have been "cleaned". What has really occured is that 1112 * they've been earmarked for later writing by the buffer 1113 * cache. 1114 * 1115 * So we get the b_dirtyoff/end update but will not actually 1116 * depend on it (NFS that is) until the pages are busied for 1117 * writing later on. 1118 */ 1119 vfs_clean_pages(bp); 1120 bqrelse(bp); 1121 1122 /* 1123 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1124 * due to the softdep code. 1125 */ 1126 } 1127 1128 /* 1129 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1130 * This is used by tmpfs. 1131 * 1132 * It is important for any VFS using this routine to NOT use it for 1133 * IO_SYNC or IO_ASYNC operations which occur when the system really 1134 * wants to flush VM pages to backing store. 1135 */ 1136 void 1137 buwrite(struct buf *bp) 1138 { 1139 vm_page_t m; 1140 int i; 1141 1142 /* 1143 * Only works for VMIO buffers. If the buffer is already 1144 * marked for delayed-write we can't avoid the bdwrite(). 1145 */ 1146 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1147 bdwrite(bp); 1148 return; 1149 } 1150 1151 /* 1152 * Set valid & dirty. 1153 * 1154 * WARNING! vfs_dirty_one_page() assumes vm_token is held for now. 1155 */ 1156 lwkt_gettoken(&vm_token); 1157 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1158 m = bp->b_xio.xio_pages[i]; 1159 vfs_dirty_one_page(bp, i, m); 1160 } 1161 lwkt_reltoken(&vm_token); 1162 bqrelse(bp); 1163 } 1164 1165 /* 1166 * bdirty: 1167 * 1168 * Turn buffer into delayed write request by marking it B_DELWRI. 1169 * B_RELBUF and B_NOCACHE must be cleared. 1170 * 1171 * We reassign the buffer to itself to properly update it in the 1172 * dirty/clean lists. 1173 * 1174 * Must be called from a critical section. 1175 * The buffer must be on BQUEUE_NONE. 1176 */ 1177 void 1178 bdirty(struct buf *bp) 1179 { 1180 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1181 if (bp->b_flags & B_NOCACHE) { 1182 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1183 bp->b_flags &= ~B_NOCACHE; 1184 } 1185 if (bp->b_flags & B_INVAL) { 1186 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1187 } 1188 bp->b_flags &= ~B_RELBUF; 1189 1190 if ((bp->b_flags & B_DELWRI) == 0) { 1191 lwkt_gettoken(&bp->b_vp->v_token); 1192 bp->b_flags |= B_DELWRI; 1193 reassignbuf(bp); 1194 lwkt_reltoken(&bp->b_vp->v_token); 1195 1196 spin_lock(&bufcspin); 1197 ++dirtybufcount; 1198 dirtybufspace += bp->b_bufsize; 1199 if (bp->b_flags & B_HEAVY) { 1200 ++dirtybufcounthw; 1201 dirtybufspacehw += bp->b_bufsize; 1202 } 1203 spin_unlock(&bufcspin); 1204 1205 bd_heatup(); 1206 } 1207 } 1208 1209 /* 1210 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1211 * needs to be flushed with a different buf_daemon thread to avoid 1212 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1213 */ 1214 void 1215 bheavy(struct buf *bp) 1216 { 1217 if ((bp->b_flags & B_HEAVY) == 0) { 1218 bp->b_flags |= B_HEAVY; 1219 if (bp->b_flags & B_DELWRI) { 1220 spin_lock(&bufcspin); 1221 ++dirtybufcounthw; 1222 dirtybufspacehw += bp->b_bufsize; 1223 spin_unlock(&bufcspin); 1224 } 1225 } 1226 } 1227 1228 /* 1229 * bundirty: 1230 * 1231 * Clear B_DELWRI for buffer. 1232 * 1233 * Must be called from a critical section. 1234 * 1235 * The buffer is typically on BQUEUE_NONE but there is one case in 1236 * brelse() that calls this function after placing the buffer on 1237 * a different queue. 1238 * 1239 * MPSAFE 1240 */ 1241 void 1242 bundirty(struct buf *bp) 1243 { 1244 if (bp->b_flags & B_DELWRI) { 1245 lwkt_gettoken(&bp->b_vp->v_token); 1246 bp->b_flags &= ~B_DELWRI; 1247 reassignbuf(bp); 1248 lwkt_reltoken(&bp->b_vp->v_token); 1249 1250 spin_lock(&bufcspin); 1251 --dirtybufcount; 1252 dirtybufspace -= bp->b_bufsize; 1253 if (bp->b_flags & B_HEAVY) { 1254 --dirtybufcounthw; 1255 dirtybufspacehw -= bp->b_bufsize; 1256 } 1257 spin_unlock(&bufcspin); 1258 1259 bd_signal(bp->b_bufsize); 1260 } 1261 /* 1262 * Since it is now being written, we can clear its deferred write flag. 1263 */ 1264 bp->b_flags &= ~B_DEFERRED; 1265 } 1266 1267 /* 1268 * Set the b_runningbufspace field, used to track how much I/O is 1269 * in progress at any given moment. 1270 */ 1271 void 1272 bsetrunningbufspace(struct buf *bp, int bytes) 1273 { 1274 bp->b_runningbufspace = bytes; 1275 if (bytes) { 1276 spin_lock(&bufcspin); 1277 runningbufspace += bytes; 1278 ++runningbufcount; 1279 spin_unlock(&bufcspin); 1280 } 1281 } 1282 1283 /* 1284 * brelse: 1285 * 1286 * Release a busy buffer and, if requested, free its resources. The 1287 * buffer will be stashed in the appropriate bufqueue[] allowing it 1288 * to be accessed later as a cache entity or reused for other purposes. 1289 * 1290 * MPALMOSTSAFE 1291 */ 1292 void 1293 brelse(struct buf *bp) 1294 { 1295 #ifdef INVARIANTS 1296 int saved_flags = bp->b_flags; 1297 #endif 1298 1299 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1300 1301 /* 1302 * If B_NOCACHE is set we are being asked to destroy the buffer and 1303 * its backing store. Clear B_DELWRI. 1304 * 1305 * B_NOCACHE is set in two cases: (1) when the caller really wants 1306 * to destroy the buffer and backing store and (2) when the caller 1307 * wants to destroy the buffer and backing store after a write 1308 * completes. 1309 */ 1310 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1311 bundirty(bp); 1312 } 1313 1314 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1315 /* 1316 * A re-dirtied buffer is only subject to destruction 1317 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1318 */ 1319 /* leave buffer intact */ 1320 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1321 (bp->b_bufsize <= 0)) { 1322 /* 1323 * Either a failed read or we were asked to free or not 1324 * cache the buffer. This path is reached with B_DELWRI 1325 * set only if B_INVAL is already set. B_NOCACHE governs 1326 * backing store destruction. 1327 * 1328 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1329 * buffer cannot be immediately freed. 1330 */ 1331 bp->b_flags |= B_INVAL; 1332 if (LIST_FIRST(&bp->b_dep) != NULL) 1333 buf_deallocate(bp); 1334 if (bp->b_flags & B_DELWRI) { 1335 spin_lock(&bufcspin); 1336 --dirtybufcount; 1337 dirtybufspace -= bp->b_bufsize; 1338 if (bp->b_flags & B_HEAVY) { 1339 --dirtybufcounthw; 1340 dirtybufspacehw -= bp->b_bufsize; 1341 } 1342 spin_unlock(&bufcspin); 1343 1344 bd_signal(bp->b_bufsize); 1345 } 1346 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1347 } 1348 1349 /* 1350 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set. 1351 * If vfs_vmio_release() is called with either bit set, the 1352 * underlying pages may wind up getting freed causing a previous 1353 * write (bdwrite()) to get 'lost' because pages associated with 1354 * a B_DELWRI bp are marked clean. Pages associated with a 1355 * B_LOCKED buffer may be mapped by the filesystem. 1356 * 1357 * If we want to release the buffer ourselves (rather then the 1358 * originator asking us to release it), give the originator a 1359 * chance to countermand the release by setting B_LOCKED. 1360 * 1361 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1362 * if B_DELWRI is set. 1363 * 1364 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1365 * on pages to return pages to the VM page queues. 1366 */ 1367 if (bp->b_flags & (B_DELWRI | B_LOCKED)) { 1368 bp->b_flags &= ~B_RELBUF; 1369 } else if (vm_page_count_severe()) { 1370 if (LIST_FIRST(&bp->b_dep) != NULL) 1371 buf_deallocate(bp); /* can set B_LOCKED */ 1372 if (bp->b_flags & (B_DELWRI | B_LOCKED)) 1373 bp->b_flags &= ~B_RELBUF; 1374 else 1375 bp->b_flags |= B_RELBUF; 1376 } 1377 1378 /* 1379 * Make sure b_cmd is clear. It may have already been cleared by 1380 * biodone(). 1381 * 1382 * At this point destroying the buffer is governed by the B_INVAL 1383 * or B_RELBUF flags. 1384 */ 1385 bp->b_cmd = BUF_CMD_DONE; 1386 dsched_exit_buf(bp); 1387 1388 /* 1389 * VMIO buffer rundown. Make sure the VM page array is restored 1390 * after an I/O may have replaces some of the pages with bogus pages 1391 * in order to not destroy dirty pages in a fill-in read. 1392 * 1393 * Note that due to the code above, if a buffer is marked B_DELWRI 1394 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1395 * B_INVAL may still be set, however. 1396 * 1397 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1398 * but not the backing store. B_NOCACHE will destroy the backing 1399 * store. 1400 * 1401 * Note that dirty NFS buffers contain byte-granular write ranges 1402 * and should not be destroyed w/ B_INVAL even if the backing store 1403 * is left intact. 1404 */ 1405 if (bp->b_flags & B_VMIO) { 1406 /* 1407 * Rundown for VMIO buffers which are not dirty NFS buffers. 1408 */ 1409 int i, j, resid; 1410 vm_page_t m; 1411 off_t foff; 1412 vm_pindex_t poff; 1413 vm_object_t obj; 1414 struct vnode *vp; 1415 1416 vp = bp->b_vp; 1417 1418 /* 1419 * Get the base offset and length of the buffer. Note that 1420 * in the VMIO case if the buffer block size is not 1421 * page-aligned then b_data pointer may not be page-aligned. 1422 * But our b_xio.xio_pages array *IS* page aligned. 1423 * 1424 * block sizes less then DEV_BSIZE (usually 512) are not 1425 * supported due to the page granularity bits (m->valid, 1426 * m->dirty, etc...). 1427 * 1428 * See man buf(9) for more information 1429 */ 1430 1431 resid = bp->b_bufsize; 1432 foff = bp->b_loffset; 1433 1434 lwkt_gettoken(&vm_token); 1435 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1436 m = bp->b_xio.xio_pages[i]; 1437 vm_page_flag_clear(m, PG_ZERO); 1438 /* 1439 * If we hit a bogus page, fixup *all* of them 1440 * now. Note that we left these pages wired 1441 * when we removed them so they had better exist, 1442 * and they cannot be ripped out from under us so 1443 * no critical section protection is necessary. 1444 */ 1445 if (m == bogus_page) { 1446 obj = vp->v_object; 1447 poff = OFF_TO_IDX(bp->b_loffset); 1448 1449 for (j = i; j < bp->b_xio.xio_npages; j++) { 1450 vm_page_t mtmp; 1451 1452 mtmp = bp->b_xio.xio_pages[j]; 1453 if (mtmp == bogus_page) { 1454 mtmp = vm_page_lookup(obj, poff + j); 1455 if (!mtmp) { 1456 panic("brelse: page missing"); 1457 } 1458 bp->b_xio.xio_pages[j] = mtmp; 1459 } 1460 } 1461 bp->b_flags &= ~B_HASBOGUS; 1462 1463 if ((bp->b_flags & B_INVAL) == 0) { 1464 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1465 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1466 } 1467 m = bp->b_xio.xio_pages[i]; 1468 } 1469 1470 /* 1471 * Invalidate the backing store if B_NOCACHE is set 1472 * (e.g. used with vinvalbuf()). If this is NFS 1473 * we impose a requirement that the block size be 1474 * a multiple of PAGE_SIZE and create a temporary 1475 * hack to basically invalidate the whole page. The 1476 * problem is that NFS uses really odd buffer sizes 1477 * especially when tracking piecemeal writes and 1478 * it also vinvalbuf()'s a lot, which would result 1479 * in only partial page validation and invalidation 1480 * here. If the file page is mmap()'d, however, 1481 * all the valid bits get set so after we invalidate 1482 * here we would end up with weird m->valid values 1483 * like 0xfc. nfs_getpages() can't handle this so 1484 * we clear all the valid bits for the NFS case 1485 * instead of just some of them. 1486 * 1487 * The real bug is the VM system having to set m->valid 1488 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1489 * itself is an artifact of the whole 512-byte 1490 * granular mess that exists to support odd block 1491 * sizes and UFS meta-data block sizes (e.g. 6144). 1492 * A complete rewrite is required. 1493 * 1494 * XXX 1495 */ 1496 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1497 int poffset = foff & PAGE_MASK; 1498 int presid; 1499 1500 presid = PAGE_SIZE - poffset; 1501 if (bp->b_vp->v_tag == VT_NFS && 1502 bp->b_vp->v_type == VREG) { 1503 ; /* entire page */ 1504 } else if (presid > resid) { 1505 presid = resid; 1506 } 1507 KASSERT(presid >= 0, ("brelse: extra page")); 1508 vm_page_set_invalid(m, poffset, presid); 1509 1510 /* 1511 * Also make sure any swap cache is removed 1512 * as it is now stale (HAMMER in particular 1513 * uses B_NOCACHE to deal with buffer 1514 * aliasing). 1515 */ 1516 swap_pager_unswapped(m); 1517 } 1518 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1519 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1520 } 1521 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1522 vfs_vmio_release(bp); 1523 lwkt_reltoken(&vm_token); 1524 } else { 1525 /* 1526 * Rundown for non-VMIO buffers. 1527 */ 1528 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1529 if (bp->b_bufsize) 1530 allocbuf(bp, 0); 1531 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1532 if (bp->b_vp) 1533 brelvp(bp); 1534 } 1535 } 1536 1537 if (bp->b_qindex != BQUEUE_NONE) 1538 panic("brelse: free buffer onto another queue???"); 1539 if (BUF_REFCNTNB(bp) > 1) { 1540 /* Temporary panic to verify exclusive locking */ 1541 /* This panic goes away when we allow shared refs */ 1542 panic("brelse: multiple refs"); 1543 /* NOT REACHED */ 1544 return; 1545 } 1546 1547 /* 1548 * Figure out the correct queue to place the cleaned up buffer on. 1549 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1550 * disassociated from their vnode. 1551 */ 1552 spin_lock(&bufqspin); 1553 if (bp->b_flags & B_LOCKED) { 1554 /* 1555 * Buffers that are locked are placed in the locked queue 1556 * immediately, regardless of their state. 1557 */ 1558 bp->b_qindex = BQUEUE_LOCKED; 1559 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1560 } else if (bp->b_bufsize == 0) { 1561 /* 1562 * Buffers with no memory. Due to conditionals near the top 1563 * of brelse() such buffers should probably already be 1564 * marked B_INVAL and disassociated from their vnode. 1565 */ 1566 bp->b_flags |= B_INVAL; 1567 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1568 KKASSERT((bp->b_flags & B_HASHED) == 0); 1569 if (bp->b_kvasize) { 1570 bp->b_qindex = BQUEUE_EMPTYKVA; 1571 } else { 1572 bp->b_qindex = BQUEUE_EMPTY; 1573 } 1574 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1575 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1576 /* 1577 * Buffers with junk contents. Again these buffers had better 1578 * already be disassociated from their vnode. 1579 */ 1580 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1581 KKASSERT((bp->b_flags & B_HASHED) == 0); 1582 bp->b_flags |= B_INVAL; 1583 bp->b_qindex = BQUEUE_CLEAN; 1584 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1585 } else { 1586 /* 1587 * Remaining buffers. These buffers are still associated with 1588 * their vnode. 1589 */ 1590 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1591 case B_DELWRI: 1592 bp->b_qindex = BQUEUE_DIRTY; 1593 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1594 break; 1595 case B_DELWRI | B_HEAVY: 1596 bp->b_qindex = BQUEUE_DIRTY_HW; 1597 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp, 1598 b_freelist); 1599 break; 1600 default: 1601 /* 1602 * NOTE: Buffers are always placed at the end of the 1603 * queue. If B_AGE is not set the buffer will cycle 1604 * through the queue twice. 1605 */ 1606 bp->b_qindex = BQUEUE_CLEAN; 1607 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1608 break; 1609 } 1610 } 1611 spin_unlock(&bufqspin); 1612 1613 /* 1614 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1615 * on the correct queue. 1616 */ 1617 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1618 bundirty(bp); 1619 1620 /* 1621 * The bp is on an appropriate queue unless locked. If it is not 1622 * locked or dirty we can wakeup threads waiting for buffer space. 1623 * 1624 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1625 * if B_INVAL is set ). 1626 */ 1627 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1628 bufcountwakeup(); 1629 1630 /* 1631 * Something we can maybe free or reuse 1632 */ 1633 if (bp->b_bufsize || bp->b_kvasize) 1634 bufspacewakeup(); 1635 1636 /* 1637 * Clean up temporary flags and unlock the buffer. 1638 */ 1639 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT); 1640 BUF_UNLOCK(bp); 1641 } 1642 1643 /* 1644 * bqrelse: 1645 * 1646 * Release a buffer back to the appropriate queue but do not try to free 1647 * it. The buffer is expected to be used again soon. 1648 * 1649 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1650 * biodone() to requeue an async I/O on completion. It is also used when 1651 * known good buffers need to be requeued but we think we may need the data 1652 * again soon. 1653 * 1654 * XXX we should be able to leave the B_RELBUF hint set on completion. 1655 * 1656 * MPSAFE 1657 */ 1658 void 1659 bqrelse(struct buf *bp) 1660 { 1661 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1662 1663 if (bp->b_qindex != BQUEUE_NONE) 1664 panic("bqrelse: free buffer onto another queue???"); 1665 if (BUF_REFCNTNB(bp) > 1) { 1666 /* do not release to free list */ 1667 panic("bqrelse: multiple refs"); 1668 return; 1669 } 1670 1671 buf_act_advance(bp); 1672 1673 spin_lock(&bufqspin); 1674 if (bp->b_flags & B_LOCKED) { 1675 /* 1676 * Locked buffers are released to the locked queue. However, 1677 * if the buffer is dirty it will first go into the dirty 1678 * queue and later on after the I/O completes successfully it 1679 * will be released to the locked queue. 1680 */ 1681 bp->b_qindex = BQUEUE_LOCKED; 1682 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1683 } else if (bp->b_flags & B_DELWRI) { 1684 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1685 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1686 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1687 } else if (vm_page_count_severe()) { 1688 /* 1689 * We are too low on memory, we have to try to free the 1690 * buffer (most importantly: the wired pages making up its 1691 * backing store) *now*. 1692 */ 1693 spin_unlock(&bufqspin); 1694 brelse(bp); 1695 return; 1696 } else { 1697 bp->b_qindex = BQUEUE_CLEAN; 1698 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1699 } 1700 spin_unlock(&bufqspin); 1701 1702 if ((bp->b_flags & B_LOCKED) == 0 && 1703 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1704 bufcountwakeup(); 1705 } 1706 1707 /* 1708 * Something we can maybe free or reuse. 1709 */ 1710 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1711 bufspacewakeup(); 1712 1713 /* 1714 * Final cleanup and unlock. Clear bits that are only used while a 1715 * buffer is actively locked. 1716 */ 1717 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF); 1718 dsched_exit_buf(bp); 1719 BUF_UNLOCK(bp); 1720 } 1721 1722 /* 1723 * vfs_vmio_release: 1724 * 1725 * Return backing pages held by the buffer 'bp' back to the VM system 1726 * if possible. The pages are freed if they are no longer valid or 1727 * attempt to free if it was used for direct I/O otherwise they are 1728 * sent to the page cache. 1729 * 1730 * Pages that were marked busy are left alone and skipped. 1731 * 1732 * The KVA mapping (b_data) for the underlying pages is removed by 1733 * this function. 1734 */ 1735 static void 1736 vfs_vmio_release(struct buf *bp) 1737 { 1738 int i; 1739 vm_page_t m; 1740 1741 lwkt_gettoken(&vm_token); 1742 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1743 m = bp->b_xio.xio_pages[i]; 1744 bp->b_xio.xio_pages[i] = NULL; 1745 1746 /* 1747 * The VFS is telling us this is not a meta-data buffer 1748 * even if it is backed by a block device. 1749 */ 1750 if (bp->b_flags & B_NOTMETA) 1751 vm_page_flag_set(m, PG_NOTMETA); 1752 1753 /* 1754 * This is a very important bit of code. We try to track 1755 * VM page use whether the pages are wired into the buffer 1756 * cache or not. While wired into the buffer cache the 1757 * bp tracks the act_count. 1758 * 1759 * We can choose to place unwired pages on the inactive 1760 * queue (0) or active queue (1). If we place too many 1761 * on the active queue the queue will cycle the act_count 1762 * on pages we'd like to keep, just from single-use pages 1763 * (such as when doing a tar-up or file scan). 1764 */ 1765 if (bp->b_act_count < vm_cycle_point) 1766 vm_page_unwire(m, 0); 1767 else 1768 vm_page_unwire(m, 1); 1769 1770 /* 1771 * We don't mess with busy pages, it is the responsibility 1772 * of the process that busied the pages to deal with them. 1773 * 1774 * However, the caller may have marked the page invalid and 1775 * we must still make sure the page is no longer mapped. 1776 */ 1777 if ((m->flags & PG_BUSY) || (m->busy != 0)) { 1778 vm_page_protect(m, VM_PROT_NONE); 1779 continue; 1780 } 1781 1782 if (m->wire_count == 0) { 1783 vm_page_flag_clear(m, PG_ZERO); 1784 /* 1785 * Might as well free the page if we can and it has 1786 * no valid data. We also free the page if the 1787 * buffer was used for direct I/O. 1788 */ 1789 #if 0 1790 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1791 m->hold_count == 0) { 1792 vm_page_busy(m); 1793 vm_page_protect(m, VM_PROT_NONE); 1794 vm_page_free(m); 1795 } else 1796 #endif 1797 if (bp->b_flags & B_DIRECT) { 1798 vm_page_try_to_free(m); 1799 } else if (vm_page_count_severe()) { 1800 m->act_count = bp->b_act_count; 1801 vm_page_try_to_cache(m); 1802 } else { 1803 m->act_count = bp->b_act_count; 1804 } 1805 } 1806 } 1807 lwkt_reltoken(&vm_token); 1808 1809 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1810 bp->b_xio.xio_npages); 1811 if (bp->b_bufsize) { 1812 bufspacewakeup(); 1813 bp->b_bufsize = 0; 1814 } 1815 bp->b_xio.xio_npages = 0; 1816 bp->b_flags &= ~B_VMIO; 1817 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1818 if (bp->b_vp) 1819 brelvp(bp); 1820 } 1821 1822 /* 1823 * vfs_bio_awrite: 1824 * 1825 * Implement clustered async writes for clearing out B_DELWRI buffers. 1826 * This is much better then the old way of writing only one buffer at 1827 * a time. Note that we may not be presented with the buffers in the 1828 * correct order, so we search for the cluster in both directions. 1829 * 1830 * The buffer is locked on call. 1831 */ 1832 int 1833 vfs_bio_awrite(struct buf *bp) 1834 { 1835 int i; 1836 int j; 1837 off_t loffset = bp->b_loffset; 1838 struct vnode *vp = bp->b_vp; 1839 int nbytes; 1840 struct buf *bpa; 1841 int nwritten; 1842 int size; 1843 1844 /* 1845 * right now we support clustered writing only to regular files. If 1846 * we find a clusterable block we could be in the middle of a cluster 1847 * rather then at the beginning. 1848 * 1849 * NOTE: b_bio1 contains the logical loffset and is aliased 1850 * to b_loffset. b_bio2 contains the translated block number. 1851 */ 1852 if ((vp->v_type == VREG) && 1853 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1854 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1855 1856 size = vp->v_mount->mnt_stat.f_iosize; 1857 1858 for (i = size; i < MAXPHYS; i += size) { 1859 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) && 1860 BUF_REFCNT(bpa) == 0 && 1861 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1862 (B_DELWRI | B_CLUSTEROK)) && 1863 (bpa->b_bufsize == size)) { 1864 if ((bpa->b_bio2.bio_offset == NOOFFSET) || 1865 (bpa->b_bio2.bio_offset != 1866 bp->b_bio2.bio_offset + i)) 1867 break; 1868 } else { 1869 break; 1870 } 1871 } 1872 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) { 1873 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) && 1874 BUF_REFCNT(bpa) == 0 && 1875 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1876 (B_DELWRI | B_CLUSTEROK)) && 1877 (bpa->b_bufsize == size)) { 1878 if ((bpa->b_bio2.bio_offset == NOOFFSET) || 1879 (bpa->b_bio2.bio_offset != 1880 bp->b_bio2.bio_offset - j)) 1881 break; 1882 } else { 1883 break; 1884 } 1885 } 1886 j -= size; 1887 nbytes = (i + j); 1888 1889 /* 1890 * this is a possible cluster write 1891 */ 1892 if (nbytes != size) { 1893 BUF_UNLOCK(bp); 1894 nwritten = cluster_wbuild(vp, size, 1895 loffset - j, nbytes); 1896 return nwritten; 1897 } 1898 } 1899 1900 /* 1901 * default (old) behavior, writing out only one block 1902 * 1903 * XXX returns b_bufsize instead of b_bcount for nwritten? 1904 */ 1905 nwritten = bp->b_bufsize; 1906 bremfree(bp); 1907 bawrite(bp); 1908 1909 return nwritten; 1910 } 1911 1912 /* 1913 * getnewbuf: 1914 * 1915 * Find and initialize a new buffer header, freeing up existing buffers 1916 * in the bufqueues as necessary. The new buffer is returned locked. 1917 * 1918 * Important: B_INVAL is not set. If the caller wishes to throw the 1919 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1920 * 1921 * We block if: 1922 * We have insufficient buffer headers 1923 * We have insufficient buffer space 1924 * buffer_map is too fragmented ( space reservation fails ) 1925 * If we have to flush dirty buffers ( but we try to avoid this ) 1926 * 1927 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1928 * Instead we ask the buf daemon to do it for us. We attempt to 1929 * avoid piecemeal wakeups of the pageout daemon. 1930 * 1931 * MPALMOSTSAFE 1932 */ 1933 static struct buf * 1934 getnewbuf(int blkflags, int slptimeo, int size, int maxsize) 1935 { 1936 struct buf *bp; 1937 struct buf *nbp; 1938 int defrag = 0; 1939 int nqindex; 1940 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1941 static int flushingbufs; 1942 1943 /* 1944 * We can't afford to block since we might be holding a vnode lock, 1945 * which may prevent system daemons from running. We deal with 1946 * low-memory situations by proactively returning memory and running 1947 * async I/O rather then sync I/O. 1948 */ 1949 1950 ++getnewbufcalls; 1951 --getnewbufrestarts; 1952 restart: 1953 ++getnewbufrestarts; 1954 1955 /* 1956 * Setup for scan. If we do not have enough free buffers, 1957 * we setup a degenerate case that immediately fails. Note 1958 * that if we are specially marked process, we are allowed to 1959 * dip into our reserves. 1960 * 1961 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1962 * 1963 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1964 * However, there are a number of cases (defragging, reusing, ...) 1965 * where we cannot backup. 1966 */ 1967 nqindex = BQUEUE_EMPTYKVA; 1968 spin_lock(&bufqspin); 1969 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]); 1970 1971 if (nbp == NULL) { 1972 /* 1973 * If no EMPTYKVA buffers and we are either 1974 * defragging or reusing, locate a CLEAN buffer 1975 * to free or reuse. If bufspace useage is low 1976 * skip this step so we can allocate a new buffer. 1977 */ 1978 if (defrag || bufspace >= lobufspace) { 1979 nqindex = BQUEUE_CLEAN; 1980 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 1981 } 1982 1983 /* 1984 * If we could not find or were not allowed to reuse a 1985 * CLEAN buffer, check to see if it is ok to use an EMPTY 1986 * buffer. We can only use an EMPTY buffer if allocating 1987 * its KVA would not otherwise run us out of buffer space. 1988 */ 1989 if (nbp == NULL && defrag == 0 && 1990 bufspace + maxsize < hibufspace) { 1991 nqindex = BQUEUE_EMPTY; 1992 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]); 1993 } 1994 } 1995 1996 /* 1997 * Run scan, possibly freeing data and/or kva mappings on the fly 1998 * depending. 1999 * 2000 * WARNING! bufqspin is held! 2001 */ 2002 while ((bp = nbp) != NULL) { 2003 int qindex = nqindex; 2004 2005 nbp = TAILQ_NEXT(bp, b_freelist); 2006 2007 /* 2008 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2009 * cycles through the queue twice before being selected. 2010 */ 2011 if (qindex == BQUEUE_CLEAN && 2012 (bp->b_flags & B_AGE) == 0 && nbp) { 2013 bp->b_flags |= B_AGE; 2014 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist); 2015 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist); 2016 continue; 2017 } 2018 2019 /* 2020 * Calculate next bp ( we can only use it if we do not block 2021 * or do other fancy things ). 2022 */ 2023 if (nbp == NULL) { 2024 switch(qindex) { 2025 case BQUEUE_EMPTY: 2026 nqindex = BQUEUE_EMPTYKVA; 2027 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]))) 2028 break; 2029 /* fall through */ 2030 case BQUEUE_EMPTYKVA: 2031 nqindex = BQUEUE_CLEAN; 2032 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]))) 2033 break; 2034 /* fall through */ 2035 case BQUEUE_CLEAN: 2036 /* 2037 * nbp is NULL. 2038 */ 2039 break; 2040 } 2041 } 2042 2043 /* 2044 * Sanity Checks 2045 */ 2046 KASSERT(bp->b_qindex == qindex, 2047 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2048 2049 /* 2050 * Note: we no longer distinguish between VMIO and non-VMIO 2051 * buffers. 2052 */ 2053 KASSERT((bp->b_flags & B_DELWRI) == 0, 2054 ("delwri buffer %p found in queue %d", bp, qindex)); 2055 2056 /* 2057 * Do not try to reuse a buffer with a non-zero b_refs. 2058 * This is an unsynchronized test. A synchronized test 2059 * is also performed after we lock the buffer. 2060 */ 2061 if (bp->b_refs) 2062 continue; 2063 2064 /* 2065 * If we are defragging then we need a buffer with 2066 * b_kvasize != 0. XXX this situation should no longer 2067 * occur, if defrag is non-zero the buffer's b_kvasize 2068 * should also be non-zero at this point. XXX 2069 */ 2070 if (defrag && bp->b_kvasize == 0) { 2071 kprintf("Warning: defrag empty buffer %p\n", bp); 2072 continue; 2073 } 2074 2075 /* 2076 * Start freeing the bp. This is somewhat involved. nbp 2077 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 2078 * on the clean list must be disassociated from their 2079 * current vnode. Buffers on the empty[kva] lists have 2080 * already been disassociated. 2081 */ 2082 2083 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2084 spin_unlock(&bufqspin); 2085 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2086 goto restart; 2087 } 2088 if (bp->b_qindex != qindex) { 2089 spin_unlock(&bufqspin); 2090 kprintf("getnewbuf: warning, BUF_LOCK blocked " 2091 "unexpectedly on buf %p index %d->%d, " 2092 "race corrected\n", 2093 bp, qindex, bp->b_qindex); 2094 BUF_UNLOCK(bp); 2095 goto restart; 2096 } 2097 bremfree_locked(bp); 2098 spin_unlock(&bufqspin); 2099 2100 /* 2101 * Dependancies must be handled before we disassociate the 2102 * vnode. 2103 * 2104 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2105 * be immediately disassociated. HAMMER then becomes 2106 * responsible for releasing the buffer. 2107 * 2108 * NOTE: bufqspin is UNLOCKED now. 2109 */ 2110 if (LIST_FIRST(&bp->b_dep) != NULL) { 2111 buf_deallocate(bp); 2112 if (bp->b_flags & B_LOCKED) { 2113 bqrelse(bp); 2114 goto restart; 2115 } 2116 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2117 } 2118 2119 if (qindex == BQUEUE_CLEAN) { 2120 if (bp->b_flags & B_VMIO) 2121 vfs_vmio_release(bp); 2122 if (bp->b_vp) 2123 brelvp(bp); 2124 } 2125 2126 /* 2127 * NOTE: nbp is now entirely invalid. We can only restart 2128 * the scan from this point on. 2129 * 2130 * Get the rest of the buffer freed up. b_kva* is still 2131 * valid after this operation. 2132 */ 2133 2134 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex)); 2135 KKASSERT((bp->b_flags & B_HASHED) == 0); 2136 2137 /* 2138 * critical section protection is not required when 2139 * scrapping a buffer's contents because it is already 2140 * wired. 2141 */ 2142 if (bp->b_bufsize) 2143 allocbuf(bp, 0); 2144 2145 bp->b_flags = B_BNOCLIP; 2146 bp->b_cmd = BUF_CMD_DONE; 2147 bp->b_vp = NULL; 2148 bp->b_error = 0; 2149 bp->b_resid = 0; 2150 bp->b_bcount = 0; 2151 bp->b_xio.xio_npages = 0; 2152 bp->b_dirtyoff = bp->b_dirtyend = 0; 2153 bp->b_act_count = ACT_INIT; 2154 reinitbufbio(bp); 2155 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2156 buf_dep_init(bp); 2157 if (blkflags & GETBLK_BHEAVY) 2158 bp->b_flags |= B_HEAVY; 2159 2160 /* 2161 * If we are defragging then free the buffer. 2162 */ 2163 if (defrag) { 2164 bp->b_flags |= B_INVAL; 2165 bfreekva(bp); 2166 brelse(bp); 2167 defrag = 0; 2168 goto restart; 2169 } 2170 2171 /* 2172 * If we are overcomitted then recover the buffer and its 2173 * KVM space. This occurs in rare situations when multiple 2174 * processes are blocked in getnewbuf() or allocbuf(). 2175 */ 2176 if (bufspace >= hibufspace) 2177 flushingbufs = 1; 2178 if (flushingbufs && bp->b_kvasize != 0) { 2179 bp->b_flags |= B_INVAL; 2180 bfreekva(bp); 2181 brelse(bp); 2182 goto restart; 2183 } 2184 if (bufspace < lobufspace) 2185 flushingbufs = 0; 2186 2187 /* 2188 * The brelvp() above interlocked the buffer, test b_refs 2189 * to determine if the buffer can be reused. b_refs 2190 * interlocks lookup/blocking-lock operations and allowing 2191 * buffer reuse can create deadlocks depending on what 2192 * (vp,loffset) is assigned to the reused buffer (see getblk). 2193 */ 2194 if (bp->b_refs) { 2195 bp->b_flags |= B_INVAL; 2196 bfreekva(bp); 2197 brelse(bp); 2198 goto restart; 2199 } 2200 2201 break; 2202 /* NOT REACHED, bufqspin not held */ 2203 } 2204 2205 /* 2206 * If we exhausted our list, sleep as appropriate. We may have to 2207 * wakeup various daemons and write out some dirty buffers. 2208 * 2209 * Generally we are sleeping due to insufficient buffer space. 2210 * 2211 * NOTE: bufqspin is held if bp is NULL, else it is not held. 2212 */ 2213 if (bp == NULL) { 2214 int flags; 2215 char *waitmsg; 2216 2217 spin_unlock(&bufqspin); 2218 if (defrag) { 2219 flags = VFS_BIO_NEED_BUFSPACE; 2220 waitmsg = "nbufkv"; 2221 } else if (bufspace >= hibufspace) { 2222 waitmsg = "nbufbs"; 2223 flags = VFS_BIO_NEED_BUFSPACE; 2224 } else { 2225 waitmsg = "newbuf"; 2226 flags = VFS_BIO_NEED_ANY; 2227 } 2228 2229 bd_speedup(); /* heeeelp */ 2230 spin_lock(&bufcspin); 2231 needsbuffer |= flags; 2232 while (needsbuffer & flags) { 2233 if (ssleep(&needsbuffer, &bufcspin, 2234 slpflags, waitmsg, slptimeo)) { 2235 spin_unlock(&bufcspin); 2236 return (NULL); 2237 } 2238 } 2239 spin_unlock(&bufcspin); 2240 } else { 2241 /* 2242 * We finally have a valid bp. We aren't quite out of the 2243 * woods, we still have to reserve kva space. In order 2244 * to keep fragmentation sane we only allocate kva in 2245 * BKVASIZE chunks. 2246 * 2247 * (bufqspin is not held) 2248 */ 2249 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2250 2251 if (maxsize != bp->b_kvasize) { 2252 vm_offset_t addr = 0; 2253 int count; 2254 2255 bfreekva(bp); 2256 2257 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2258 vm_map_lock(&buffer_map); 2259 2260 if (vm_map_findspace(&buffer_map, 2261 vm_map_min(&buffer_map), maxsize, 2262 maxsize, 0, &addr)) { 2263 /* 2264 * Uh oh. Buffer map is too fragmented. We 2265 * must defragment the map. 2266 */ 2267 vm_map_unlock(&buffer_map); 2268 vm_map_entry_release(count); 2269 ++bufdefragcnt; 2270 defrag = 1; 2271 bp->b_flags |= B_INVAL; 2272 brelse(bp); 2273 goto restart; 2274 } 2275 if (addr) { 2276 vm_map_insert(&buffer_map, &count, 2277 NULL, 0, 2278 addr, addr + maxsize, 2279 VM_MAPTYPE_NORMAL, 2280 VM_PROT_ALL, VM_PROT_ALL, 2281 MAP_NOFAULT); 2282 2283 bp->b_kvabase = (caddr_t) addr; 2284 bp->b_kvasize = maxsize; 2285 bufspace += bp->b_kvasize; 2286 ++bufreusecnt; 2287 } 2288 vm_map_unlock(&buffer_map); 2289 vm_map_entry_release(count); 2290 } 2291 bp->b_data = bp->b_kvabase; 2292 } 2293 return(bp); 2294 } 2295 2296 /* 2297 * This routine is called in an emergency to recover VM pages from the 2298 * buffer cache by cashing in clean buffers. The idea is to recover 2299 * enough pages to be able to satisfy a stuck bio_page_alloc(). 2300 * 2301 * MPSAFE 2302 */ 2303 static int 2304 recoverbufpages(void) 2305 { 2306 struct buf *bp; 2307 int bytes = 0; 2308 2309 ++recoverbufcalls; 2310 2311 spin_lock(&bufqspin); 2312 while (bytes < MAXBSIZE) { 2313 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 2314 if (bp == NULL) 2315 break; 2316 2317 /* 2318 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2319 * cycles through the queue twice before being selected. 2320 */ 2321 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) { 2322 bp->b_flags |= B_AGE; 2323 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 2324 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], 2325 bp, b_freelist); 2326 continue; 2327 } 2328 2329 /* 2330 * Sanity Checks 2331 */ 2332 KKASSERT(bp->b_qindex == BQUEUE_CLEAN); 2333 KKASSERT((bp->b_flags & B_DELWRI) == 0); 2334 2335 /* 2336 * Start freeing the bp. This is somewhat involved. 2337 * 2338 * Buffers on the clean list must be disassociated from 2339 * their current vnode 2340 */ 2341 2342 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2343 kprintf("recoverbufpages: warning, locked buf %p, " 2344 "race corrected\n", 2345 bp); 2346 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100); 2347 continue; 2348 } 2349 if (bp->b_qindex != BQUEUE_CLEAN) { 2350 kprintf("recoverbufpages: warning, BUF_LOCK blocked " 2351 "unexpectedly on buf %p index %d, race " 2352 "corrected\n", 2353 bp, bp->b_qindex); 2354 BUF_UNLOCK(bp); 2355 continue; 2356 } 2357 bremfree_locked(bp); 2358 spin_unlock(&bufqspin); 2359 2360 /* 2361 * Dependancies must be handled before we disassociate the 2362 * vnode. 2363 * 2364 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2365 * be immediately disassociated. HAMMER then becomes 2366 * responsible for releasing the buffer. 2367 */ 2368 if (LIST_FIRST(&bp->b_dep) != NULL) { 2369 buf_deallocate(bp); 2370 if (bp->b_flags & B_LOCKED) { 2371 bqrelse(bp); 2372 spin_lock(&bufqspin); 2373 continue; 2374 } 2375 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2376 } 2377 2378 bytes += bp->b_bufsize; 2379 2380 if (bp->b_flags & B_VMIO) { 2381 bp->b_flags |= B_DIRECT; /* try to free pages */ 2382 vfs_vmio_release(bp); 2383 } 2384 if (bp->b_vp) 2385 brelvp(bp); 2386 2387 KKASSERT(bp->b_vp == NULL); 2388 KKASSERT((bp->b_flags & B_HASHED) == 0); 2389 2390 /* 2391 * critical section protection is not required when 2392 * scrapping a buffer's contents because it is already 2393 * wired. 2394 */ 2395 if (bp->b_bufsize) 2396 allocbuf(bp, 0); 2397 2398 bp->b_flags = B_BNOCLIP; 2399 bp->b_cmd = BUF_CMD_DONE; 2400 bp->b_vp = NULL; 2401 bp->b_error = 0; 2402 bp->b_resid = 0; 2403 bp->b_bcount = 0; 2404 bp->b_xio.xio_npages = 0; 2405 bp->b_dirtyoff = bp->b_dirtyend = 0; 2406 reinitbufbio(bp); 2407 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2408 buf_dep_init(bp); 2409 bp->b_flags |= B_INVAL; 2410 /* bfreekva(bp); */ 2411 brelse(bp); 2412 spin_lock(&bufqspin); 2413 } 2414 spin_unlock(&bufqspin); 2415 return(bytes); 2416 } 2417 2418 /* 2419 * buf_daemon: 2420 * 2421 * Buffer flushing daemon. Buffers are normally flushed by the 2422 * update daemon but if it cannot keep up this process starts to 2423 * take the load in an attempt to prevent getnewbuf() from blocking. 2424 * 2425 * Once a flush is initiated it does not stop until the number 2426 * of buffers falls below lodirtybuffers, but we will wake up anyone 2427 * waiting at the mid-point. 2428 */ 2429 2430 static struct kproc_desc buf_kp = { 2431 "bufdaemon", 2432 buf_daemon, 2433 &bufdaemon_td 2434 }; 2435 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2436 kproc_start, &buf_kp) 2437 2438 static struct kproc_desc bufhw_kp = { 2439 "bufdaemon_hw", 2440 buf_daemon_hw, 2441 &bufdaemonhw_td 2442 }; 2443 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2444 kproc_start, &bufhw_kp) 2445 2446 /* 2447 * MPSAFE thread 2448 */ 2449 static void 2450 buf_daemon(void) 2451 { 2452 int limit; 2453 2454 /* 2455 * This process needs to be suspended prior to shutdown sync. 2456 */ 2457 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2458 bufdaemon_td, SHUTDOWN_PRI_LAST); 2459 curthread->td_flags |= TDF_SYSTHREAD; 2460 2461 /* 2462 * This process is allowed to take the buffer cache to the limit 2463 */ 2464 for (;;) { 2465 kproc_suspend_loop(); 2466 2467 /* 2468 * Do the flush as long as the number of dirty buffers 2469 * (including those running) exceeds lodirtybufspace. 2470 * 2471 * When flushing limit running I/O to hirunningspace 2472 * Do the flush. Limit the amount of in-transit I/O we 2473 * allow to build up, otherwise we would completely saturate 2474 * the I/O system. Wakeup any waiting processes before we 2475 * normally would so they can run in parallel with our drain. 2476 * 2477 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2478 * but because we split the operation into two threads we 2479 * have to cut it in half for each thread. 2480 */ 2481 waitrunningbufspace(); 2482 limit = lodirtybufspace / 2; 2483 while (runningbufspace + dirtybufspace > limit || 2484 dirtybufcount - dirtybufcounthw >= nbuf / 2) { 2485 if (flushbufqueues(BQUEUE_DIRTY) == 0) 2486 break; 2487 if (runningbufspace < hirunningspace) 2488 continue; 2489 waitrunningbufspace(); 2490 } 2491 2492 /* 2493 * We reached our low water mark, reset the 2494 * request and sleep until we are needed again. 2495 * The sleep is just so the suspend code works. 2496 */ 2497 spin_lock(&bufcspin); 2498 if (bd_request == 0) 2499 ssleep(&bd_request, &bufcspin, 0, "psleep", hz); 2500 bd_request = 0; 2501 spin_unlock(&bufcspin); 2502 } 2503 } 2504 2505 /* 2506 * MPSAFE thread 2507 */ 2508 static void 2509 buf_daemon_hw(void) 2510 { 2511 int limit; 2512 2513 /* 2514 * This process needs to be suspended prior to shutdown sync. 2515 */ 2516 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2517 bufdaemonhw_td, SHUTDOWN_PRI_LAST); 2518 curthread->td_flags |= TDF_SYSTHREAD; 2519 2520 /* 2521 * This process is allowed to take the buffer cache to the limit 2522 */ 2523 for (;;) { 2524 kproc_suspend_loop(); 2525 2526 /* 2527 * Do the flush. Limit the amount of in-transit I/O we 2528 * allow to build up, otherwise we would completely saturate 2529 * the I/O system. Wakeup any waiting processes before we 2530 * normally would so they can run in parallel with our drain. 2531 * 2532 * Once we decide to flush push the queued I/O up to 2533 * hirunningspace in order to trigger bursting by the bioq 2534 * subsystem. 2535 * 2536 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2537 * but because we split the operation into two threads we 2538 * have to cut it in half for each thread. 2539 */ 2540 waitrunningbufspace(); 2541 limit = lodirtybufspace / 2; 2542 while (runningbufspace + dirtybufspacehw > limit || 2543 dirtybufcounthw >= nbuf / 2) { 2544 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0) 2545 break; 2546 if (runningbufspace < hirunningspace) 2547 continue; 2548 waitrunningbufspace(); 2549 } 2550 2551 /* 2552 * We reached our low water mark, reset the 2553 * request and sleep until we are needed again. 2554 * The sleep is just so the suspend code works. 2555 */ 2556 spin_lock(&bufcspin); 2557 if (bd_request_hw == 0) 2558 ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz); 2559 bd_request_hw = 0; 2560 spin_unlock(&bufcspin); 2561 } 2562 } 2563 2564 /* 2565 * flushbufqueues: 2566 * 2567 * Try to flush a buffer in the dirty queue. We must be careful to 2568 * free up B_INVAL buffers instead of write them, which NFS is 2569 * particularly sensitive to. 2570 * 2571 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate 2572 * that we really want to try to get the buffer out and reuse it 2573 * due to the write load on the machine. 2574 * 2575 * We must lock the buffer in order to check its validity before we 2576 * can mess with its contents. bufqspin isn't enough. 2577 */ 2578 static int 2579 flushbufqueues(bufq_type_t q) 2580 { 2581 struct buf *bp; 2582 int r = 0; 2583 int spun; 2584 2585 spin_lock(&bufqspin); 2586 spun = 1; 2587 2588 bp = TAILQ_FIRST(&bufqueues[q]); 2589 while (bp) { 2590 if ((bp->b_flags & B_DELWRI) == 0) { 2591 kprintf("Unexpected clean buffer %p\n", bp); 2592 bp = TAILQ_NEXT(bp, b_freelist); 2593 continue; 2594 } 2595 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2596 bp = TAILQ_NEXT(bp, b_freelist); 2597 continue; 2598 } 2599 KKASSERT(bp->b_qindex == q); 2600 2601 /* 2602 * Must recheck B_DELWRI after successfully locking 2603 * the buffer. 2604 */ 2605 if ((bp->b_flags & B_DELWRI) == 0) { 2606 BUF_UNLOCK(bp); 2607 bp = TAILQ_NEXT(bp, b_freelist); 2608 continue; 2609 } 2610 2611 if (bp->b_flags & B_INVAL) { 2612 _bremfree(bp); 2613 spin_unlock(&bufqspin); 2614 spun = 0; 2615 brelse(bp); 2616 ++r; 2617 break; 2618 } 2619 2620 spin_unlock(&bufqspin); 2621 spun = 0; 2622 2623 if (LIST_FIRST(&bp->b_dep) != NULL && 2624 (bp->b_flags & B_DEFERRED) == 0 && 2625 buf_countdeps(bp, 0)) { 2626 spin_lock(&bufqspin); 2627 spun = 1; 2628 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist); 2629 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist); 2630 bp->b_flags |= B_DEFERRED; 2631 BUF_UNLOCK(bp); 2632 bp = TAILQ_FIRST(&bufqueues[q]); 2633 continue; 2634 } 2635 2636 /* 2637 * If the buffer has a dependancy, buf_checkwrite() must 2638 * also return 0 for us to be able to initate the write. 2639 * 2640 * If the buffer is flagged B_ERROR it may be requeued 2641 * over and over again, we try to avoid a live lock. 2642 * 2643 * NOTE: buf_checkwrite is MPSAFE. 2644 */ 2645 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2646 bremfree(bp); 2647 brelse(bp); 2648 } else if (bp->b_flags & B_ERROR) { 2649 tsleep(bp, 0, "bioer", 1); 2650 bp->b_flags &= ~B_AGE; 2651 vfs_bio_awrite(bp); 2652 } else { 2653 bp->b_flags |= B_AGE; 2654 vfs_bio_awrite(bp); 2655 } 2656 ++r; 2657 break; 2658 } 2659 if (spun) 2660 spin_unlock(&bufqspin); 2661 return (r); 2662 } 2663 2664 /* 2665 * inmem: 2666 * 2667 * Returns true if no I/O is needed to access the associated VM object. 2668 * This is like findblk except it also hunts around in the VM system for 2669 * the data. 2670 * 2671 * Note that we ignore vm_page_free() races from interrupts against our 2672 * lookup, since if the caller is not protected our return value will not 2673 * be any more valid then otherwise once we exit the critical section. 2674 */ 2675 int 2676 inmem(struct vnode *vp, off_t loffset) 2677 { 2678 vm_object_t obj; 2679 vm_offset_t toff, tinc, size; 2680 vm_page_t m; 2681 2682 if (findblk(vp, loffset, FINDBLK_TEST)) 2683 return 1; 2684 if (vp->v_mount == NULL) 2685 return 0; 2686 if ((obj = vp->v_object) == NULL) 2687 return 0; 2688 2689 size = PAGE_SIZE; 2690 if (size > vp->v_mount->mnt_stat.f_iosize) 2691 size = vp->v_mount->mnt_stat.f_iosize; 2692 2693 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2694 lwkt_gettoken(&vm_token); 2695 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2696 lwkt_reltoken(&vm_token); 2697 if (m == NULL) 2698 return 0; 2699 tinc = size; 2700 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2701 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2702 if (vm_page_is_valid(m, 2703 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) 2704 return 0; 2705 } 2706 return 1; 2707 } 2708 2709 /* 2710 * findblk: 2711 * 2712 * Locate and return the specified buffer. Unless flagged otherwise, 2713 * a locked buffer will be returned if it exists or NULL if it does not. 2714 * 2715 * findblk()'d buffers are still on the bufqueues and if you intend 2716 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2717 * and possibly do other stuff to it. 2718 * 2719 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2720 * for locking the buffer and ensuring that it remains 2721 * the desired buffer after locking. 2722 * 2723 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2724 * to acquire the lock we return NULL, even if the 2725 * buffer exists. 2726 * 2727 * FINDBLK_REF - Returns the buffer ref'd, which prevents reuse 2728 * by getnewbuf() but does not prevent disassociation 2729 * while we are locked. Used to avoid deadlocks 2730 * against random (vp,loffset)s due to reassignment. 2731 * 2732 * (0) - Lock the buffer blocking. 2733 * 2734 * MPSAFE 2735 */ 2736 struct buf * 2737 findblk(struct vnode *vp, off_t loffset, int flags) 2738 { 2739 struct buf *bp; 2740 int lkflags; 2741 2742 lkflags = LK_EXCLUSIVE; 2743 if (flags & FINDBLK_NBLOCK) 2744 lkflags |= LK_NOWAIT; 2745 2746 for (;;) { 2747 /* 2748 * Lookup. Ref the buf while holding v_token to prevent 2749 * reuse (but does not prevent diassociation). 2750 */ 2751 lwkt_gettoken(&vp->v_token); 2752 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2753 if (bp == NULL) { 2754 lwkt_reltoken(&vp->v_token); 2755 return(NULL); 2756 } 2757 atomic_add_int(&bp->b_refs, 1); 2758 lwkt_reltoken(&vp->v_token); 2759 2760 /* 2761 * If testing only break and return bp, do not lock. 2762 */ 2763 if (flags & FINDBLK_TEST) 2764 break; 2765 2766 /* 2767 * Lock the buffer, return an error if the lock fails. 2768 * (only FINDBLK_NBLOCK can cause the lock to fail). 2769 */ 2770 if (BUF_LOCK(bp, lkflags)) { 2771 atomic_subtract_int(&bp->b_refs, 1); 2772 /* bp = NULL; not needed */ 2773 return(NULL); 2774 } 2775 2776 /* 2777 * Revalidate the locked buf before allowing it to be 2778 * returned. 2779 */ 2780 if (bp->b_vp == vp && bp->b_loffset == loffset) 2781 break; 2782 atomic_subtract_int(&bp->b_refs, 1); 2783 BUF_UNLOCK(bp); 2784 } 2785 2786 /* 2787 * Success 2788 */ 2789 if ((flags & FINDBLK_REF) == 0) 2790 atomic_subtract_int(&bp->b_refs, 1); 2791 return(bp); 2792 } 2793 2794 void 2795 unrefblk(struct buf *bp) 2796 { 2797 atomic_subtract_int(&bp->b_refs, 1); 2798 } 2799 2800 /* 2801 * getcacheblk: 2802 * 2803 * Similar to getblk() except only returns the buffer if it is 2804 * B_CACHE and requires no other manipulation. Otherwise NULL 2805 * is returned. 2806 * 2807 * If B_RAM is set the buffer might be just fine, but we return 2808 * NULL anyway because we want the code to fall through to the 2809 * cluster read. Otherwise read-ahead breaks. 2810 * 2811 * If blksize is 0 the buffer cache buffer must already be fully 2812 * cached. 2813 * 2814 * If blksize is non-zero getblk() will be used, allowing a buffer 2815 * to be reinstantiated from its VM backing store. The buffer must 2816 * still be fully cached after reinstantiation to be returned. 2817 */ 2818 struct buf * 2819 getcacheblk(struct vnode *vp, off_t loffset, int blksize) 2820 { 2821 struct buf *bp; 2822 2823 if (blksize) { 2824 bp = getblk(vp, loffset, blksize, 0, 0); 2825 if (bp) { 2826 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2827 B_CACHE) { 2828 bp->b_flags &= ~B_AGE; 2829 } else { 2830 brelse(bp); 2831 bp = NULL; 2832 } 2833 } 2834 } else { 2835 bp = findblk(vp, loffset, 0); 2836 if (bp) { 2837 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2838 B_CACHE) { 2839 bp->b_flags &= ~B_AGE; 2840 bremfree(bp); 2841 } else { 2842 BUF_UNLOCK(bp); 2843 bp = NULL; 2844 } 2845 } 2846 } 2847 return (bp); 2848 } 2849 2850 /* 2851 * getblk: 2852 * 2853 * Get a block given a specified block and offset into a file/device. 2854 * B_INVAL may or may not be set on return. The caller should clear 2855 * B_INVAL prior to initiating a READ. 2856 * 2857 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2858 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2859 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2860 * without doing any of those things the system will likely believe 2861 * the buffer to be valid (especially if it is not B_VMIO), and the 2862 * next getblk() will return the buffer with B_CACHE set. 2863 * 2864 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2865 * an existing buffer. 2866 * 2867 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2868 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2869 * and then cleared based on the backing VM. If the previous buffer is 2870 * non-0-sized but invalid, B_CACHE will be cleared. 2871 * 2872 * If getblk() must create a new buffer, the new buffer is returned with 2873 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2874 * case it is returned with B_INVAL clear and B_CACHE set based on the 2875 * backing VM. 2876 * 2877 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2878 * B_CACHE bit is clear. 2879 * 2880 * What this means, basically, is that the caller should use B_CACHE to 2881 * determine whether the buffer is fully valid or not and should clear 2882 * B_INVAL prior to issuing a read. If the caller intends to validate 2883 * the buffer by loading its data area with something, the caller needs 2884 * to clear B_INVAL. If the caller does this without issuing an I/O, 2885 * the caller should set B_CACHE ( as an optimization ), else the caller 2886 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2887 * a write attempt or if it was a successfull read. If the caller 2888 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2889 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2890 * 2891 * getblk flags: 2892 * 2893 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2894 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2895 * 2896 * MPALMOSTSAFE 2897 */ 2898 struct buf * 2899 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2900 { 2901 struct buf *bp; 2902 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2903 int error; 2904 int lkflags; 2905 2906 if (size > MAXBSIZE) 2907 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2908 if (vp->v_object == NULL) 2909 panic("getblk: vnode %p has no object!", vp); 2910 2911 loop: 2912 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2913 /* 2914 * The buffer was found in the cache, but we need to lock it. 2915 * We must acquire a ref on the bp to prevent reuse, but 2916 * this will not prevent disassociation (brelvp()) so we 2917 * must recheck (vp,loffset) after acquiring the lock. 2918 * 2919 * Without the ref the buffer could potentially be reused 2920 * before we acquire the lock and create a deadlock 2921 * situation between the thread trying to reuse the buffer 2922 * and us due to the fact that we would wind up blocking 2923 * on a random (vp,loffset). 2924 */ 2925 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2926 if (blkflags & GETBLK_NOWAIT) { 2927 unrefblk(bp); 2928 return(NULL); 2929 } 2930 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2931 if (blkflags & GETBLK_PCATCH) 2932 lkflags |= LK_PCATCH; 2933 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2934 if (error) { 2935 unrefblk(bp); 2936 if (error == ENOLCK) 2937 goto loop; 2938 return (NULL); 2939 } 2940 /* buffer may have changed on us */ 2941 } 2942 unrefblk(bp); 2943 2944 /* 2945 * Once the buffer has been locked, make sure we didn't race 2946 * a buffer recyclement. Buffers that are no longer hashed 2947 * will have b_vp == NULL, so this takes care of that check 2948 * as well. 2949 */ 2950 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2951 kprintf("Warning buffer %p (vp %p loffset %lld) " 2952 "was recycled\n", 2953 bp, vp, (long long)loffset); 2954 BUF_UNLOCK(bp); 2955 goto loop; 2956 } 2957 2958 /* 2959 * If SZMATCH any pre-existing buffer must be of the requested 2960 * size or NULL is returned. The caller absolutely does not 2961 * want getblk() to bwrite() the buffer on a size mismatch. 2962 */ 2963 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2964 BUF_UNLOCK(bp); 2965 return(NULL); 2966 } 2967 2968 /* 2969 * All vnode-based buffers must be backed by a VM object. 2970 */ 2971 KKASSERT(bp->b_flags & B_VMIO); 2972 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2973 bp->b_flags &= ~B_AGE; 2974 2975 /* 2976 * Make sure that B_INVAL buffers do not have a cached 2977 * block number translation. 2978 */ 2979 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2980 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2981 " did not have cleared bio_offset cache\n", 2982 bp, vp, (long long)loffset); 2983 clearbiocache(&bp->b_bio2); 2984 } 2985 2986 /* 2987 * The buffer is locked. B_CACHE is cleared if the buffer is 2988 * invalid. 2989 */ 2990 if (bp->b_flags & B_INVAL) 2991 bp->b_flags &= ~B_CACHE; 2992 bremfree(bp); 2993 2994 /* 2995 * Any size inconsistancy with a dirty buffer or a buffer 2996 * with a softupdates dependancy must be resolved. Resizing 2997 * the buffer in such circumstances can lead to problems. 2998 * 2999 * Dirty or dependant buffers are written synchronously. 3000 * Other types of buffers are simply released and 3001 * reconstituted as they may be backed by valid, dirty VM 3002 * pages (but not marked B_DELWRI). 3003 * 3004 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 3005 * and may be left over from a prior truncation (and thus 3006 * no longer represent the actual EOF point), so we 3007 * definitely do not want to B_NOCACHE the backing store. 3008 */ 3009 if (size != bp->b_bcount) { 3010 if (bp->b_flags & B_DELWRI) { 3011 bp->b_flags |= B_RELBUF; 3012 bwrite(bp); 3013 } else if (LIST_FIRST(&bp->b_dep)) { 3014 bp->b_flags |= B_RELBUF; 3015 bwrite(bp); 3016 } else { 3017 bp->b_flags |= B_RELBUF; 3018 brelse(bp); 3019 } 3020 goto loop; 3021 } 3022 KKASSERT(size <= bp->b_kvasize); 3023 KASSERT(bp->b_loffset != NOOFFSET, 3024 ("getblk: no buffer offset")); 3025 3026 /* 3027 * A buffer with B_DELWRI set and B_CACHE clear must 3028 * be committed before we can return the buffer in 3029 * order to prevent the caller from issuing a read 3030 * ( due to B_CACHE not being set ) and overwriting 3031 * it. 3032 * 3033 * Most callers, including NFS and FFS, need this to 3034 * operate properly either because they assume they 3035 * can issue a read if B_CACHE is not set, or because 3036 * ( for example ) an uncached B_DELWRI might loop due 3037 * to softupdates re-dirtying the buffer. In the latter 3038 * case, B_CACHE is set after the first write completes, 3039 * preventing further loops. 3040 * 3041 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3042 * above while extending the buffer, we cannot allow the 3043 * buffer to remain with B_CACHE set after the write 3044 * completes or it will represent a corrupt state. To 3045 * deal with this we set B_NOCACHE to scrap the buffer 3046 * after the write. 3047 * 3048 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3049 * I'm not even sure this state is still possible 3050 * now that getblk() writes out any dirty buffers 3051 * on size changes. 3052 * 3053 * We might be able to do something fancy, like setting 3054 * B_CACHE in bwrite() except if B_DELWRI is already set, 3055 * so the below call doesn't set B_CACHE, but that gets real 3056 * confusing. This is much easier. 3057 */ 3058 3059 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3060 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3061 "and CACHE clear, b_flags %08x\n", 3062 bp, (intmax_t)bp->b_loffset, bp->b_flags); 3063 bp->b_flags |= B_NOCACHE; 3064 bwrite(bp); 3065 goto loop; 3066 } 3067 } else { 3068 /* 3069 * Buffer is not in-core, create new buffer. The buffer 3070 * returned by getnewbuf() is locked. Note that the returned 3071 * buffer is also considered valid (not marked B_INVAL). 3072 * 3073 * Calculating the offset for the I/O requires figuring out 3074 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3075 * the mount's f_iosize otherwise. If the vnode does not 3076 * have an associated mount we assume that the passed size is 3077 * the block size. 3078 * 3079 * Note that vn_isdisk() cannot be used here since it may 3080 * return a failure for numerous reasons. Note that the 3081 * buffer size may be larger then the block size (the caller 3082 * will use block numbers with the proper multiple). Beware 3083 * of using any v_* fields which are part of unions. In 3084 * particular, in DragonFly the mount point overloading 3085 * mechanism uses the namecache only and the underlying 3086 * directory vnode is not a special case. 3087 */ 3088 int bsize, maxsize; 3089 3090 if (vp->v_type == VBLK || vp->v_type == VCHR) 3091 bsize = DEV_BSIZE; 3092 else if (vp->v_mount) 3093 bsize = vp->v_mount->mnt_stat.f_iosize; 3094 else 3095 bsize = size; 3096 3097 maxsize = size + (loffset & PAGE_MASK); 3098 maxsize = imax(maxsize, bsize); 3099 3100 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3101 if (bp == NULL) { 3102 if (slpflags || slptimeo) 3103 return NULL; 3104 goto loop; 3105 } 3106 3107 /* 3108 * Atomically insert the buffer into the hash, so that it can 3109 * be found by findblk(). 3110 * 3111 * If bgetvp() returns non-zero a collision occured, and the 3112 * bp will not be associated with the vnode. 3113 * 3114 * Make sure the translation layer has been cleared. 3115 */ 3116 bp->b_loffset = loffset; 3117 bp->b_bio2.bio_offset = NOOFFSET; 3118 /* bp->b_bio2.bio_next = NULL; */ 3119 3120 if (bgetvp(vp, bp, size)) { 3121 bp->b_flags |= B_INVAL; 3122 brelse(bp); 3123 goto loop; 3124 } 3125 3126 /* 3127 * All vnode-based buffers must be backed by a VM object. 3128 */ 3129 KKASSERT(vp->v_object != NULL); 3130 bp->b_flags |= B_VMIO; 3131 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3132 3133 allocbuf(bp, size); 3134 } 3135 KKASSERT(dsched_is_clear_buf_priv(bp)); 3136 return (bp); 3137 } 3138 3139 /* 3140 * regetblk(bp) 3141 * 3142 * Reacquire a buffer that was previously released to the locked queue, 3143 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3144 * set B_LOCKED (which handles the acquisition race). 3145 * 3146 * To this end, either B_LOCKED must be set or the dependancy list must be 3147 * non-empty. 3148 * 3149 * MPSAFE 3150 */ 3151 void 3152 regetblk(struct buf *bp) 3153 { 3154 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3155 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3156 bremfree(bp); 3157 } 3158 3159 /* 3160 * geteblk: 3161 * 3162 * Get an empty, disassociated buffer of given size. The buffer is 3163 * initially set to B_INVAL. 3164 * 3165 * critical section protection is not required for the allocbuf() 3166 * call because races are impossible here. 3167 * 3168 * MPALMOSTSAFE 3169 */ 3170 struct buf * 3171 geteblk(int size) 3172 { 3173 struct buf *bp; 3174 int maxsize; 3175 3176 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3177 3178 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 3179 ; 3180 allocbuf(bp, size); 3181 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3182 KKASSERT(dsched_is_clear_buf_priv(bp)); 3183 return (bp); 3184 } 3185 3186 3187 /* 3188 * allocbuf: 3189 * 3190 * This code constitutes the buffer memory from either anonymous system 3191 * memory (in the case of non-VMIO operations) or from an associated 3192 * VM object (in the case of VMIO operations). This code is able to 3193 * resize a buffer up or down. 3194 * 3195 * Note that this code is tricky, and has many complications to resolve 3196 * deadlock or inconsistant data situations. Tread lightly!!! 3197 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3198 * the caller. Calling this code willy nilly can result in the loss of 3199 * data. 3200 * 3201 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3202 * B_CACHE for the non-VMIO case. 3203 * 3204 * This routine does not need to be called from a critical section but you 3205 * must own the buffer. 3206 * 3207 * MPSAFE 3208 */ 3209 int 3210 allocbuf(struct buf *bp, int size) 3211 { 3212 int newbsize, mbsize; 3213 int i; 3214 3215 if (BUF_REFCNT(bp) == 0) 3216 panic("allocbuf: buffer not busy"); 3217 3218 if (bp->b_kvasize < size) 3219 panic("allocbuf: buffer too small"); 3220 3221 if ((bp->b_flags & B_VMIO) == 0) { 3222 caddr_t origbuf; 3223 int origbufsize; 3224 /* 3225 * Just get anonymous memory from the kernel. Don't 3226 * mess with B_CACHE. 3227 */ 3228 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3229 if (bp->b_flags & B_MALLOC) 3230 newbsize = mbsize; 3231 else 3232 newbsize = round_page(size); 3233 3234 if (newbsize < bp->b_bufsize) { 3235 /* 3236 * Malloced buffers are not shrunk 3237 */ 3238 if (bp->b_flags & B_MALLOC) { 3239 if (newbsize) { 3240 bp->b_bcount = size; 3241 } else { 3242 kfree(bp->b_data, M_BIOBUF); 3243 if (bp->b_bufsize) { 3244 atomic_subtract_int(&bufmallocspace, bp->b_bufsize); 3245 bufspacewakeup(); 3246 bp->b_bufsize = 0; 3247 } 3248 bp->b_data = bp->b_kvabase; 3249 bp->b_bcount = 0; 3250 bp->b_flags &= ~B_MALLOC; 3251 } 3252 return 1; 3253 } 3254 vm_hold_free_pages( 3255 bp, 3256 (vm_offset_t) bp->b_data + newbsize, 3257 (vm_offset_t) bp->b_data + bp->b_bufsize); 3258 } else if (newbsize > bp->b_bufsize) { 3259 /* 3260 * We only use malloced memory on the first allocation. 3261 * and revert to page-allocated memory when the buffer 3262 * grows. 3263 */ 3264 if ((bufmallocspace < maxbufmallocspace) && 3265 (bp->b_bufsize == 0) && 3266 (mbsize <= PAGE_SIZE/2)) { 3267 3268 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3269 bp->b_bufsize = mbsize; 3270 bp->b_bcount = size; 3271 bp->b_flags |= B_MALLOC; 3272 atomic_add_int(&bufmallocspace, mbsize); 3273 return 1; 3274 } 3275 origbuf = NULL; 3276 origbufsize = 0; 3277 /* 3278 * If the buffer is growing on its other-than-first 3279 * allocation, then we revert to the page-allocation 3280 * scheme. 3281 */ 3282 if (bp->b_flags & B_MALLOC) { 3283 origbuf = bp->b_data; 3284 origbufsize = bp->b_bufsize; 3285 bp->b_data = bp->b_kvabase; 3286 if (bp->b_bufsize) { 3287 atomic_subtract_int(&bufmallocspace, 3288 bp->b_bufsize); 3289 bufspacewakeup(); 3290 bp->b_bufsize = 0; 3291 } 3292 bp->b_flags &= ~B_MALLOC; 3293 newbsize = round_page(newbsize); 3294 } 3295 vm_hold_load_pages( 3296 bp, 3297 (vm_offset_t) bp->b_data + bp->b_bufsize, 3298 (vm_offset_t) bp->b_data + newbsize); 3299 if (origbuf) { 3300 bcopy(origbuf, bp->b_data, origbufsize); 3301 kfree(origbuf, M_BIOBUF); 3302 } 3303 } 3304 } else { 3305 vm_page_t m; 3306 int desiredpages; 3307 3308 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3309 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3310 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3311 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3312 3313 if (bp->b_flags & B_MALLOC) 3314 panic("allocbuf: VMIO buffer can't be malloced"); 3315 /* 3316 * Set B_CACHE initially if buffer is 0 length or will become 3317 * 0-length. 3318 */ 3319 if (size == 0 || bp->b_bufsize == 0) 3320 bp->b_flags |= B_CACHE; 3321 3322 if (newbsize < bp->b_bufsize) { 3323 /* 3324 * DEV_BSIZE aligned new buffer size is less then the 3325 * DEV_BSIZE aligned existing buffer size. Figure out 3326 * if we have to remove any pages. 3327 */ 3328 if (desiredpages < bp->b_xio.xio_npages) { 3329 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3330 /* 3331 * the page is not freed here -- it 3332 * is the responsibility of 3333 * vnode_pager_setsize 3334 */ 3335 m = bp->b_xio.xio_pages[i]; 3336 KASSERT(m != bogus_page, 3337 ("allocbuf: bogus page found")); 3338 while (vm_page_sleep_busy(m, TRUE, "biodep")) 3339 ; 3340 3341 bp->b_xio.xio_pages[i] = NULL; 3342 vm_page_unwire(m, 0); 3343 } 3344 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3345 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3346 bp->b_xio.xio_npages = desiredpages; 3347 } 3348 } else if (size > bp->b_bcount) { 3349 /* 3350 * We are growing the buffer, possibly in a 3351 * byte-granular fashion. 3352 */ 3353 struct vnode *vp; 3354 vm_object_t obj; 3355 vm_offset_t toff; 3356 vm_offset_t tinc; 3357 3358 /* 3359 * Step 1, bring in the VM pages from the object, 3360 * allocating them if necessary. We must clear 3361 * B_CACHE if these pages are not valid for the 3362 * range covered by the buffer. 3363 * 3364 * critical section protection is required to protect 3365 * against interrupts unbusying and freeing pages 3366 * between our vm_page_lookup() and our 3367 * busycheck/wiring call. 3368 */ 3369 vp = bp->b_vp; 3370 obj = vp->v_object; 3371 3372 lwkt_gettoken(&vm_token); 3373 while (bp->b_xio.xio_npages < desiredpages) { 3374 vm_page_t m; 3375 vm_pindex_t pi; 3376 3377 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages; 3378 if ((m = vm_page_lookup(obj, pi)) == NULL) { 3379 /* 3380 * note: must allocate system pages 3381 * since blocking here could intefere 3382 * with paging I/O, no matter which 3383 * process we are. 3384 */ 3385 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages); 3386 if (m) { 3387 vm_page_wire(m); 3388 vm_page_flag_clear(m, PG_ZERO); 3389 vm_page_wakeup(m); 3390 bp->b_flags &= ~B_CACHE; 3391 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3392 ++bp->b_xio.xio_npages; 3393 } 3394 continue; 3395 } 3396 3397 /* 3398 * We found a page. If we have to sleep on it, 3399 * retry because it might have gotten freed out 3400 * from under us. 3401 * 3402 * We can only test PG_BUSY here. Blocking on 3403 * m->busy might lead to a deadlock: 3404 * 3405 * vm_fault->getpages->cluster_read->allocbuf 3406 * 3407 */ 3408 3409 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 3410 continue; 3411 vm_page_flag_clear(m, PG_ZERO); 3412 vm_page_wire(m); 3413 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3414 ++bp->b_xio.xio_npages; 3415 if (bp->b_act_count < m->act_count) 3416 bp->b_act_count = m->act_count; 3417 } 3418 lwkt_reltoken(&vm_token); 3419 3420 /* 3421 * Step 2. We've loaded the pages into the buffer, 3422 * we have to figure out if we can still have B_CACHE 3423 * set. Note that B_CACHE is set according to the 3424 * byte-granular range ( bcount and size ), not the 3425 * aligned range ( newbsize ). 3426 * 3427 * The VM test is against m->valid, which is DEV_BSIZE 3428 * aligned. Needless to say, the validity of the data 3429 * needs to also be DEV_BSIZE aligned. Note that this 3430 * fails with NFS if the server or some other client 3431 * extends the file's EOF. If our buffer is resized, 3432 * B_CACHE may remain set! XXX 3433 */ 3434 3435 toff = bp->b_bcount; 3436 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3437 3438 while ((bp->b_flags & B_CACHE) && toff < size) { 3439 vm_pindex_t pi; 3440 3441 if (tinc > (size - toff)) 3442 tinc = size - toff; 3443 3444 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3445 PAGE_SHIFT; 3446 3447 vfs_buf_test_cache( 3448 bp, 3449 bp->b_loffset, 3450 toff, 3451 tinc, 3452 bp->b_xio.xio_pages[pi] 3453 ); 3454 toff += tinc; 3455 tinc = PAGE_SIZE; 3456 } 3457 3458 /* 3459 * Step 3, fixup the KVM pmap. Remember that 3460 * bp->b_data is relative to bp->b_loffset, but 3461 * bp->b_loffset may be offset into the first page. 3462 */ 3463 3464 bp->b_data = (caddr_t) 3465 trunc_page((vm_offset_t)bp->b_data); 3466 pmap_qenter( 3467 (vm_offset_t)bp->b_data, 3468 bp->b_xio.xio_pages, 3469 bp->b_xio.xio_npages 3470 ); 3471 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3472 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3473 } 3474 } 3475 3476 /* adjust space use on already-dirty buffer */ 3477 if (bp->b_flags & B_DELWRI) { 3478 spin_lock(&bufcspin); 3479 dirtybufspace += newbsize - bp->b_bufsize; 3480 if (bp->b_flags & B_HEAVY) 3481 dirtybufspacehw += newbsize - bp->b_bufsize; 3482 spin_unlock(&bufcspin); 3483 } 3484 if (newbsize < bp->b_bufsize) 3485 bufspacewakeup(); 3486 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3487 bp->b_bcount = size; /* requested buffer size */ 3488 return 1; 3489 } 3490 3491 /* 3492 * biowait: 3493 * 3494 * Wait for buffer I/O completion, returning error status. B_EINTR 3495 * is converted into an EINTR error but not cleared (since a chain 3496 * of biowait() calls may occur). 3497 * 3498 * On return bpdone() will have been called but the buffer will remain 3499 * locked and will not have been brelse()'d. 3500 * 3501 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3502 * likely still in progress on return. 3503 * 3504 * NOTE! This operation is on a BIO, not a BUF. 3505 * 3506 * NOTE! BIO_DONE is cleared by vn_strategy() 3507 * 3508 * MPSAFE 3509 */ 3510 static __inline int 3511 _biowait(struct bio *bio, const char *wmesg, int to) 3512 { 3513 struct buf *bp = bio->bio_buf; 3514 u_int32_t flags; 3515 u_int32_t nflags; 3516 int error; 3517 3518 KKASSERT(bio == &bp->b_bio1); 3519 for (;;) { 3520 flags = bio->bio_flags; 3521 if (flags & BIO_DONE) 3522 break; 3523 tsleep_interlock(bio, 0); 3524 nflags = flags | BIO_WANT; 3525 tsleep_interlock(bio, 0); 3526 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3527 if (wmesg) 3528 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3529 else if (bp->b_cmd == BUF_CMD_READ) 3530 error = tsleep(bio, PINTERLOCKED, "biord", to); 3531 else 3532 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3533 if (error) { 3534 kprintf("tsleep error biowait %d\n", error); 3535 return (error); 3536 } 3537 } 3538 } 3539 3540 /* 3541 * Finish up. 3542 */ 3543 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3544 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3545 if (bp->b_flags & B_EINTR) 3546 return (EINTR); 3547 if (bp->b_flags & B_ERROR) 3548 return (bp->b_error ? bp->b_error : EIO); 3549 return (0); 3550 } 3551 3552 int 3553 biowait(struct bio *bio, const char *wmesg) 3554 { 3555 return(_biowait(bio, wmesg, 0)); 3556 } 3557 3558 int 3559 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3560 { 3561 return(_biowait(bio, wmesg, to)); 3562 } 3563 3564 /* 3565 * This associates a tracking count with an I/O. vn_strategy() and 3566 * dev_dstrategy() do this automatically but there are a few cases 3567 * where a vnode or device layer is bypassed when a block translation 3568 * is cached. In such cases bio_start_transaction() may be called on 3569 * the bypassed layers so the system gets an I/O in progress indication 3570 * for those higher layers. 3571 */ 3572 void 3573 bio_start_transaction(struct bio *bio, struct bio_track *track) 3574 { 3575 bio->bio_track = track; 3576 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3577 dsched_new_buf(bio->bio_buf); 3578 bio_track_ref(track); 3579 } 3580 3581 /* 3582 * Initiate I/O on a vnode. 3583 * 3584 * SWAPCACHE OPERATION: 3585 * 3586 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3587 * devfs also uses b_vp for fake buffers so we also have to check 3588 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3589 * underlying block device. The swap assignments are related to the 3590 * buffer cache buffer's b_vp, not the passed vp. 3591 * 3592 * The passed vp == bp->b_vp only in the case where the strategy call 3593 * is made on the vp itself for its own buffers (a regular file or 3594 * block device vp). The filesystem usually then re-calls vn_strategy() 3595 * after translating the request to an underlying device. 3596 * 3597 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3598 * underlying buffer cache buffers. 3599 * 3600 * We can only deal with page-aligned buffers at the moment, because 3601 * we can't tell what the real dirty state for pages straddling a buffer 3602 * are. 3603 * 3604 * In order to call swap_pager_strategy() we must provide the VM object 3605 * and base offset for the underlying buffer cache pages so it can find 3606 * the swap blocks. 3607 */ 3608 void 3609 vn_strategy(struct vnode *vp, struct bio *bio) 3610 { 3611 struct bio_track *track; 3612 struct buf *bp = bio->bio_buf; 3613 3614 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3615 3616 /* 3617 * Set when an I/O is issued on the bp. Cleared by consumers 3618 * (aka HAMMER), allowing the consumer to determine if I/O had 3619 * actually occurred. 3620 */ 3621 bp->b_flags |= B_IODEBUG; 3622 3623 /* 3624 * Handle the swap cache intercept. 3625 */ 3626 if (vn_cache_strategy(vp, bio)) 3627 return; 3628 3629 /* 3630 * Otherwise do the operation through the filesystem 3631 */ 3632 if (bp->b_cmd == BUF_CMD_READ) 3633 track = &vp->v_track_read; 3634 else 3635 track = &vp->v_track_write; 3636 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3637 bio->bio_track = track; 3638 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3639 dsched_new_buf(bio->bio_buf); 3640 bio_track_ref(track); 3641 vop_strategy(*vp->v_ops, vp, bio); 3642 } 3643 3644 static void vn_cache_strategy_callback(struct bio *bio); 3645 3646 int 3647 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3648 { 3649 struct buf *bp = bio->bio_buf; 3650 struct bio *nbio; 3651 vm_object_t object; 3652 vm_page_t m; 3653 int i; 3654 3655 /* 3656 * Is this buffer cache buffer suitable for reading from 3657 * the swap cache? 3658 */ 3659 if (vm_swapcache_read_enable == 0 || 3660 bp->b_cmd != BUF_CMD_READ || 3661 ((bp->b_flags & B_CLUSTER) == 0 && 3662 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3663 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3664 (bp->b_bcount & PAGE_MASK) != 0) { 3665 return(0); 3666 } 3667 3668 /* 3669 * Figure out the original VM object (it will match the underlying 3670 * VM pages). Note that swap cached data uses page indices relative 3671 * to that object, not relative to bio->bio_offset. 3672 */ 3673 if (bp->b_flags & B_CLUSTER) 3674 object = vp->v_object; 3675 else 3676 object = bp->b_vp->v_object; 3677 3678 /* 3679 * In order to be able to use the swap cache all underlying VM 3680 * pages must be marked as such, and we can't have any bogus pages. 3681 */ 3682 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3683 m = bp->b_xio.xio_pages[i]; 3684 if ((m->flags & PG_SWAPPED) == 0) 3685 break; 3686 if (m == bogus_page) 3687 break; 3688 } 3689 3690 /* 3691 * If we are good then issue the I/O using swap_pager_strategy(). 3692 */ 3693 if (i == bp->b_xio.xio_npages) { 3694 m = bp->b_xio.xio_pages[0]; 3695 nbio = push_bio(bio); 3696 nbio->bio_done = vn_cache_strategy_callback; 3697 nbio->bio_offset = ptoa(m->pindex); 3698 KKASSERT(m->object == object); 3699 swap_pager_strategy(object, nbio); 3700 return(1); 3701 } 3702 return(0); 3703 } 3704 3705 /* 3706 * This is a bit of a hack but since the vn_cache_strategy() function can 3707 * override a VFS's strategy function we must make sure that the bio, which 3708 * is probably bio2, doesn't leak an unexpected offset value back to the 3709 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3710 * bio went through its own file strategy function and the the bio2 offset 3711 * is a cached disk offset when, in fact, it isn't. 3712 */ 3713 static void 3714 vn_cache_strategy_callback(struct bio *bio) 3715 { 3716 bio->bio_offset = NOOFFSET; 3717 biodone(pop_bio(bio)); 3718 } 3719 3720 /* 3721 * bpdone: 3722 * 3723 * Finish I/O on a buffer after all BIOs have been processed. 3724 * Called when the bio chain is exhausted or by biowait. If called 3725 * by biowait, elseit is typically 0. 3726 * 3727 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3728 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3729 * assuming B_INVAL is clear. 3730 * 3731 * For the VMIO case, we set B_CACHE if the op was a read and no 3732 * read error occured, or if the op was a write. B_CACHE is never 3733 * set if the buffer is invalid or otherwise uncacheable. 3734 * 3735 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3736 * initiator to leave B_INVAL set to brelse the buffer out of existance 3737 * in the biodone routine. 3738 */ 3739 void 3740 bpdone(struct buf *bp, int elseit) 3741 { 3742 buf_cmd_t cmd; 3743 3744 KASSERT(BUF_REFCNTNB(bp) > 0, 3745 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3746 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3747 ("biodone: bp %p already done!", bp)); 3748 3749 /* 3750 * No more BIOs are left. All completion functions have been dealt 3751 * with, now we clean up the buffer. 3752 */ 3753 cmd = bp->b_cmd; 3754 bp->b_cmd = BUF_CMD_DONE; 3755 3756 /* 3757 * Only reads and writes are processed past this point. 3758 */ 3759 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3760 if (cmd == BUF_CMD_FREEBLKS) 3761 bp->b_flags |= B_NOCACHE; 3762 if (elseit) 3763 brelse(bp); 3764 return; 3765 } 3766 3767 /* 3768 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3769 * a lot worse. XXX - move this above the clearing of b_cmd 3770 */ 3771 if (LIST_FIRST(&bp->b_dep) != NULL) 3772 buf_complete(bp); /* MPSAFE */ 3773 3774 /* 3775 * A failed write must re-dirty the buffer unless B_INVAL 3776 * was set. Only applicable to normal buffers (with VPs). 3777 * vinum buffers may not have a vp. 3778 */ 3779 if (cmd == BUF_CMD_WRITE && 3780 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3781 bp->b_flags &= ~B_NOCACHE; 3782 if (bp->b_vp) 3783 bdirty(bp); 3784 } 3785 3786 if (bp->b_flags & B_VMIO) { 3787 int i; 3788 vm_ooffset_t foff; 3789 vm_page_t m; 3790 vm_object_t obj; 3791 int iosize; 3792 struct vnode *vp = bp->b_vp; 3793 3794 obj = vp->v_object; 3795 3796 #if defined(VFS_BIO_DEBUG) 3797 if (vp->v_auxrefs == 0) 3798 panic("biodone: zero vnode hold count"); 3799 if ((vp->v_flag & VOBJBUF) == 0) 3800 panic("biodone: vnode is not setup for merged cache"); 3801 #endif 3802 3803 foff = bp->b_loffset; 3804 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 3805 KASSERT(obj != NULL, ("biodone: missing VM object")); 3806 3807 #if defined(VFS_BIO_DEBUG) 3808 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3809 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n", 3810 obj->paging_in_progress, bp->b_xio.xio_npages); 3811 } 3812 #endif 3813 3814 /* 3815 * Set B_CACHE if the op was a normal read and no error 3816 * occured. B_CACHE is set for writes in the b*write() 3817 * routines. 3818 */ 3819 iosize = bp->b_bcount - bp->b_resid; 3820 if (cmd == BUF_CMD_READ && 3821 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3822 bp->b_flags |= B_CACHE; 3823 } 3824 3825 lwkt_gettoken(&vm_token); 3826 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3827 int bogusflag = 0; 3828 int resid; 3829 3830 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3831 if (resid > iosize) 3832 resid = iosize; 3833 3834 /* 3835 * cleanup bogus pages, restoring the originals. Since 3836 * the originals should still be wired, we don't have 3837 * to worry about interrupt/freeing races destroying 3838 * the VM object association. 3839 */ 3840 m = bp->b_xio.xio_pages[i]; 3841 if (m == bogus_page) { 3842 bogusflag = 1; 3843 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3844 if (m == NULL) 3845 panic("biodone: page disappeared"); 3846 bp->b_xio.xio_pages[i] = m; 3847 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3848 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3849 } 3850 #if defined(VFS_BIO_DEBUG) 3851 if (OFF_TO_IDX(foff) != m->pindex) { 3852 kprintf("biodone: foff(%lu)/m->pindex(%ld) " 3853 "mismatch\n", 3854 (unsigned long)foff, (long)m->pindex); 3855 } 3856 #endif 3857 3858 /* 3859 * In the write case, the valid and clean bits are 3860 * already changed correctly (see bdwrite()), so we 3861 * only need to do this here in the read case. 3862 */ 3863 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3864 vfs_clean_one_page(bp, i, m); 3865 } 3866 vm_page_flag_clear(m, PG_ZERO); 3867 3868 /* 3869 * when debugging new filesystems or buffer I/O 3870 * methods, this is the most common error that pops 3871 * up. if you see this, you have not set the page 3872 * busy flag correctly!!! 3873 */ 3874 if (m->busy == 0) { 3875 kprintf("biodone: page busy < 0, " 3876 "pindex: %d, foff: 0x(%x,%x), " 3877 "resid: %d, index: %d\n", 3878 (int) m->pindex, (int)(foff >> 32), 3879 (int) foff & 0xffffffff, resid, i); 3880 if (!vn_isdisk(vp, NULL)) 3881 kprintf(" iosize: %ld, loffset: %lld, " 3882 "flags: 0x%08x, npages: %d\n", 3883 bp->b_vp->v_mount->mnt_stat.f_iosize, 3884 (long long)bp->b_loffset, 3885 bp->b_flags, bp->b_xio.xio_npages); 3886 else 3887 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3888 (long long)bp->b_loffset, 3889 bp->b_flags, bp->b_xio.xio_npages); 3890 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3891 m->valid, m->dirty, m->wire_count); 3892 panic("biodone: page busy < 0"); 3893 } 3894 vm_page_io_finish(m); 3895 vm_object_pip_subtract(obj, 1); 3896 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3897 iosize -= resid; 3898 } 3899 bp->b_flags &= ~B_HASBOGUS; 3900 if (obj) 3901 vm_object_pip_wakeupn(obj, 0); 3902 lwkt_reltoken(&vm_token); 3903 } 3904 3905 /* 3906 * Finish up by releasing the buffer. There are no more synchronous 3907 * or asynchronous completions, those were handled by bio_done 3908 * callbacks. 3909 */ 3910 if (elseit) { 3911 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3912 brelse(bp); 3913 else 3914 bqrelse(bp); 3915 } 3916 } 3917 3918 /* 3919 * Normal biodone. 3920 */ 3921 void 3922 biodone(struct bio *bio) 3923 { 3924 struct buf *bp = bio->bio_buf; 3925 3926 runningbufwakeup(bp); 3927 3928 /* 3929 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3930 */ 3931 while (bio) { 3932 biodone_t *done_func; 3933 struct bio_track *track; 3934 3935 /* 3936 * BIO tracking. Most but not all BIOs are tracked. 3937 */ 3938 if ((track = bio->bio_track) != NULL) { 3939 bio_track_rel(track); 3940 bio->bio_track = NULL; 3941 } 3942 3943 /* 3944 * A bio_done function terminates the loop. The function 3945 * will be responsible for any further chaining and/or 3946 * buffer management. 3947 * 3948 * WARNING! The done function can deallocate the buffer! 3949 */ 3950 if ((done_func = bio->bio_done) != NULL) { 3951 bio->bio_done = NULL; 3952 done_func(bio); 3953 return; 3954 } 3955 bio = bio->bio_prev; 3956 } 3957 3958 /* 3959 * If we've run out of bio's do normal [a]synchronous completion. 3960 */ 3961 bpdone(bp, 1); 3962 } 3963 3964 /* 3965 * Synchronous biodone - this terminates a synchronous BIO. 3966 * 3967 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3968 * but still locked. The caller must brelse() the buffer after waiting 3969 * for completion. 3970 */ 3971 void 3972 biodone_sync(struct bio *bio) 3973 { 3974 struct buf *bp = bio->bio_buf; 3975 int flags; 3976 int nflags; 3977 3978 KKASSERT(bio == &bp->b_bio1); 3979 bpdone(bp, 0); 3980 3981 for (;;) { 3982 flags = bio->bio_flags; 3983 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3984 3985 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3986 if (flags & BIO_WANT) 3987 wakeup(bio); 3988 break; 3989 } 3990 } 3991 } 3992 3993 /* 3994 * vfs_unbusy_pages: 3995 * 3996 * This routine is called in lieu of iodone in the case of 3997 * incomplete I/O. This keeps the busy status for pages 3998 * consistant. 3999 */ 4000 void 4001 vfs_unbusy_pages(struct buf *bp) 4002 { 4003 int i; 4004 4005 runningbufwakeup(bp); 4006 4007 lwkt_gettoken(&vm_token); 4008 if (bp->b_flags & B_VMIO) { 4009 struct vnode *vp = bp->b_vp; 4010 vm_object_t obj; 4011 4012 obj = vp->v_object; 4013 4014 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4015 vm_page_t m = bp->b_xio.xio_pages[i]; 4016 4017 /* 4018 * When restoring bogus changes the original pages 4019 * should still be wired, so we are in no danger of 4020 * losing the object association and do not need 4021 * critical section protection particularly. 4022 */ 4023 if (m == bogus_page) { 4024 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 4025 if (!m) { 4026 panic("vfs_unbusy_pages: page missing"); 4027 } 4028 bp->b_xio.xio_pages[i] = m; 4029 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4030 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4031 } 4032 vm_object_pip_subtract(obj, 1); 4033 vm_page_flag_clear(m, PG_ZERO); 4034 vm_page_io_finish(m); 4035 } 4036 bp->b_flags &= ~B_HASBOGUS; 4037 vm_object_pip_wakeupn(obj, 0); 4038 } 4039 lwkt_reltoken(&vm_token); 4040 } 4041 4042 /* 4043 * vfs_busy_pages: 4044 * 4045 * This routine is called before a device strategy routine. 4046 * It is used to tell the VM system that paging I/O is in 4047 * progress, and treat the pages associated with the buffer 4048 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4049 * flag is handled to make sure that the object doesn't become 4050 * inconsistant. 4051 * 4052 * Since I/O has not been initiated yet, certain buffer flags 4053 * such as B_ERROR or B_INVAL may be in an inconsistant state 4054 * and should be ignored. 4055 * 4056 * MPSAFE 4057 */ 4058 void 4059 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4060 { 4061 int i, bogus; 4062 struct lwp *lp = curthread->td_lwp; 4063 4064 /* 4065 * The buffer's I/O command must already be set. If reading, 4066 * B_CACHE must be 0 (double check against callers only doing 4067 * I/O when B_CACHE is 0). 4068 */ 4069 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4070 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4071 4072 if (bp->b_flags & B_VMIO) { 4073 vm_object_t obj; 4074 4075 lwkt_gettoken(&vm_token); 4076 4077 obj = vp->v_object; 4078 KASSERT(bp->b_loffset != NOOFFSET, 4079 ("vfs_busy_pages: no buffer offset")); 4080 4081 /* 4082 * Loop until none of the pages are busy. 4083 */ 4084 retry: 4085 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4086 vm_page_t m = bp->b_xio.xio_pages[i]; 4087 4088 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 4089 goto retry; 4090 } 4091 4092 /* 4093 * Setup for I/O, soft-busy the page right now because 4094 * the next loop may block. 4095 */ 4096 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4097 vm_page_t m = bp->b_xio.xio_pages[i]; 4098 4099 vm_page_flag_clear(m, PG_ZERO); 4100 if ((bp->b_flags & B_CLUSTER) == 0) { 4101 vm_object_pip_add(obj, 1); 4102 vm_page_io_start(m); 4103 } 4104 } 4105 4106 /* 4107 * Adjust protections for I/O and do bogus-page mapping. 4108 * Assume that vm_page_protect() can block (it can block 4109 * if VM_PROT_NONE, don't take any chances regardless). 4110 * 4111 * In particular note that for writes we must incorporate 4112 * page dirtyness from the VM system into the buffer's 4113 * dirty range. 4114 * 4115 * For reads we theoretically must incorporate page dirtyness 4116 * from the VM system to determine if the page needs bogus 4117 * replacement, but we shortcut the test by simply checking 4118 * that all m->valid bits are set, indicating that the page 4119 * is fully valid and does not need to be re-read. For any 4120 * VM system dirtyness the page will also be fully valid 4121 * since it was mapped at one point. 4122 */ 4123 bogus = 0; 4124 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4125 vm_page_t m = bp->b_xio.xio_pages[i]; 4126 4127 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4128 if (bp->b_cmd == BUF_CMD_WRITE) { 4129 /* 4130 * When readying a vnode-backed buffer for 4131 * a write we must zero-fill any invalid 4132 * portions of the backing VM pages, mark 4133 * it valid and clear related dirty bits. 4134 * 4135 * vfs_clean_one_page() incorporates any 4136 * VM dirtyness and updates the b_dirtyoff 4137 * range (after we've made the page RO). 4138 * 4139 * It is also expected that the pmap modified 4140 * bit has already been cleared by the 4141 * vm_page_protect(). We may not be able 4142 * to clear all dirty bits for a page if it 4143 * was also memory mapped (NFS). 4144 * 4145 * Finally be sure to unassign any swap-cache 4146 * backing store as it is now stale. 4147 */ 4148 vm_page_protect(m, VM_PROT_READ); 4149 vfs_clean_one_page(bp, i, m); 4150 swap_pager_unswapped(m); 4151 } else if (m->valid == VM_PAGE_BITS_ALL) { 4152 /* 4153 * When readying a vnode-backed buffer for 4154 * read we must replace any dirty pages with 4155 * a bogus page so dirty data is not destroyed 4156 * when filling gaps. 4157 * 4158 * To avoid testing whether the page is 4159 * dirty we instead test that the page was 4160 * at some point mapped (m->valid fully 4161 * valid) with the understanding that 4162 * this also covers the dirty case. 4163 */ 4164 bp->b_xio.xio_pages[i] = bogus_page; 4165 bp->b_flags |= B_HASBOGUS; 4166 bogus++; 4167 } else if (m->valid & m->dirty) { 4168 /* 4169 * This case should not occur as partial 4170 * dirtyment can only happen if the buffer 4171 * is B_CACHE, and this code is not entered 4172 * if the buffer is B_CACHE. 4173 */ 4174 kprintf("Warning: vfs_busy_pages - page not " 4175 "fully valid! loff=%jx bpf=%08x " 4176 "idx=%d val=%02x dir=%02x\n", 4177 (intmax_t)bp->b_loffset, bp->b_flags, 4178 i, m->valid, m->dirty); 4179 vm_page_protect(m, VM_PROT_NONE); 4180 } else { 4181 /* 4182 * The page is not valid and can be made 4183 * part of the read. 4184 */ 4185 vm_page_protect(m, VM_PROT_NONE); 4186 } 4187 } 4188 if (bogus) { 4189 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4190 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4191 } 4192 lwkt_reltoken(&vm_token); 4193 } 4194 4195 /* 4196 * This is the easiest place to put the process accounting for the I/O 4197 * for now. 4198 */ 4199 if (lp != NULL) { 4200 if (bp->b_cmd == BUF_CMD_READ) 4201 lp->lwp_ru.ru_inblock++; 4202 else 4203 lp->lwp_ru.ru_oublock++; 4204 } 4205 } 4206 4207 /* 4208 * Tell the VM system that the pages associated with this buffer 4209 * are clean. This is used for delayed writes where the data is 4210 * going to go to disk eventually without additional VM intevention. 4211 * 4212 * NOTE: While we only really need to clean through to b_bcount, we 4213 * just go ahead and clean through to b_bufsize. 4214 */ 4215 static void 4216 vfs_clean_pages(struct buf *bp) 4217 { 4218 vm_page_t m; 4219 int i; 4220 4221 if ((bp->b_flags & B_VMIO) == 0) 4222 return; 4223 4224 KASSERT(bp->b_loffset != NOOFFSET, 4225 ("vfs_clean_pages: no buffer offset")); 4226 4227 /* 4228 * vm_token must be held for vfs_clean_one_page() calls. 4229 */ 4230 lwkt_gettoken(&vm_token); 4231 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4232 m = bp->b_xio.xio_pages[i]; 4233 vfs_clean_one_page(bp, i, m); 4234 } 4235 lwkt_reltoken(&vm_token); 4236 } 4237 4238 /* 4239 * vfs_clean_one_page: 4240 * 4241 * Set the valid bits and clear the dirty bits in a page within a 4242 * buffer. The range is restricted to the buffer's size and the 4243 * buffer's logical offset might index into the first page. 4244 * 4245 * The caller has busied or soft-busied the page and it is not mapped, 4246 * test and incorporate the dirty bits into b_dirtyoff/end before 4247 * clearing them. Note that we need to clear the pmap modified bits 4248 * after determining the the page was dirty, vm_page_set_validclean() 4249 * does not do it for us. 4250 * 4251 * This routine is typically called after a read completes (dirty should 4252 * be zero in that case as we are not called on bogus-replace pages), 4253 * or before a write is initiated. 4254 * 4255 * NOTE: vm_token must be held by the caller, and vm_page_set_validclean() 4256 * currently assumes the vm_token is held. 4257 */ 4258 static void 4259 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4260 { 4261 int bcount; 4262 int xoff; 4263 int soff; 4264 int eoff; 4265 4266 /* 4267 * Calculate offset range within the page but relative to buffer's 4268 * loffset. loffset might be offset into the first page. 4269 */ 4270 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4271 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4272 4273 if (pageno == 0) { 4274 soff = xoff; 4275 eoff = PAGE_SIZE; 4276 } else { 4277 soff = (pageno << PAGE_SHIFT); 4278 eoff = soff + PAGE_SIZE; 4279 } 4280 if (eoff > bcount) 4281 eoff = bcount; 4282 if (soff >= eoff) 4283 return; 4284 4285 /* 4286 * Test dirty bits and adjust b_dirtyoff/end. 4287 * 4288 * If dirty pages are incorporated into the bp any prior 4289 * B_NEEDCOMMIT state (NFS) must be cleared because the 4290 * caller has not taken into account the new dirty data. 4291 * 4292 * If the page was memory mapped the dirty bits might go beyond the 4293 * end of the buffer, but we can't really make the assumption that 4294 * a file EOF straddles the buffer (even though this is the case for 4295 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4296 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4297 * This also saves some console spam. 4298 * 4299 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4300 * NFS can handle huge commits but not huge writes. 4301 */ 4302 vm_page_test_dirty(m); 4303 if (m->dirty) { 4304 if ((bp->b_flags & B_NEEDCOMMIT) && 4305 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4306 if (debug_commit) 4307 kprintf("Warning: vfs_clean_one_page: bp %p " 4308 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4309 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4310 "doff/end %d %d\n", 4311 bp, (intmax_t)bp->b_loffset, bp->b_bcount, 4312 bp->b_flags, bp->b_cmd, 4313 m->valid, m->dirty, xoff, soff, eoff, 4314 bp->b_dirtyoff, bp->b_dirtyend); 4315 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4316 if (debug_commit) 4317 print_backtrace(-1); 4318 } 4319 /* 4320 * Only clear the pmap modified bits if ALL the dirty bits 4321 * are set, otherwise the system might mis-clear portions 4322 * of a page. 4323 */ 4324 if (m->dirty == VM_PAGE_BITS_ALL && 4325 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4326 pmap_clear_modify(m); 4327 } 4328 if (bp->b_dirtyoff > soff - xoff) 4329 bp->b_dirtyoff = soff - xoff; 4330 if (bp->b_dirtyend < eoff - xoff) 4331 bp->b_dirtyend = eoff - xoff; 4332 } 4333 4334 /* 4335 * Set related valid bits, clear related dirty bits. 4336 * Does not mess with the pmap modified bit. 4337 * 4338 * WARNING! We cannot just clear all of m->dirty here as the 4339 * buffer cache buffers may use a DEV_BSIZE'd aligned 4340 * block size, or have an odd size (e.g. NFS at file EOF). 4341 * The putpages code can clear m->dirty to 0. 4342 * 4343 * If a VOP_WRITE generates a buffer cache buffer which 4344 * covers the same space as mapped writable pages the 4345 * buffer flush might not be able to clear all the dirty 4346 * bits and still require a putpages from the VM system 4347 * to finish it off. 4348 * 4349 * WARNING! vm_page_set_validclean() currently assumes vm_token 4350 * is held. The page might not be busied (bdwrite() case). 4351 */ 4352 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4353 } 4354 4355 /* 4356 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4357 * The page data is assumed to be valid (there is no zeroing here). 4358 */ 4359 static void 4360 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4361 { 4362 int bcount; 4363 int xoff; 4364 int soff; 4365 int eoff; 4366 4367 /* 4368 * Calculate offset range within the page but relative to buffer's 4369 * loffset. loffset might be offset into the first page. 4370 */ 4371 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4372 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4373 4374 if (pageno == 0) { 4375 soff = xoff; 4376 eoff = PAGE_SIZE; 4377 } else { 4378 soff = (pageno << PAGE_SHIFT); 4379 eoff = soff + PAGE_SIZE; 4380 } 4381 if (eoff > bcount) 4382 eoff = bcount; 4383 if (soff >= eoff) 4384 return; 4385 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4386 } 4387 4388 /* 4389 * vfs_bio_clrbuf: 4390 * 4391 * Clear a buffer. This routine essentially fakes an I/O, so we need 4392 * to clear B_ERROR and B_INVAL. 4393 * 4394 * Note that while we only theoretically need to clear through b_bcount, 4395 * we go ahead and clear through b_bufsize. 4396 */ 4397 4398 void 4399 vfs_bio_clrbuf(struct buf *bp) 4400 { 4401 int i, mask = 0; 4402 caddr_t sa, ea; 4403 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4404 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4405 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4406 (bp->b_loffset & PAGE_MASK) == 0) { 4407 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4408 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4409 bp->b_resid = 0; 4410 return; 4411 } 4412 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4413 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4414 bzero(bp->b_data, bp->b_bufsize); 4415 bp->b_xio.xio_pages[0]->valid |= mask; 4416 bp->b_resid = 0; 4417 return; 4418 } 4419 } 4420 sa = bp->b_data; 4421 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4422 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4423 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4424 ea = (caddr_t)(vm_offset_t)ulmin( 4425 (u_long)(vm_offset_t)ea, 4426 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4427 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4428 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4429 continue; 4430 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4431 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4432 bzero(sa, ea - sa); 4433 } 4434 } else { 4435 for (; sa < ea; sa += DEV_BSIZE, j++) { 4436 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4437 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4438 bzero(sa, DEV_BSIZE); 4439 } 4440 } 4441 bp->b_xio.xio_pages[i]->valid |= mask; 4442 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4443 } 4444 bp->b_resid = 0; 4445 } else { 4446 clrbuf(bp); 4447 } 4448 } 4449 4450 /* 4451 * vm_hold_load_pages: 4452 * 4453 * Load pages into the buffer's address space. The pages are 4454 * allocated from the kernel object in order to reduce interference 4455 * with the any VM paging I/O activity. The range of loaded 4456 * pages will be wired. 4457 * 4458 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4459 * retrieve the full range (to - from) of pages. 4460 * 4461 * MPSAFE 4462 */ 4463 void 4464 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4465 { 4466 vm_offset_t pg; 4467 vm_page_t p; 4468 int index; 4469 4470 to = round_page(to); 4471 from = round_page(from); 4472 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4473 4474 pg = from; 4475 while (pg < to) { 4476 /* 4477 * Note: must allocate system pages since blocking here 4478 * could intefere with paging I/O, no matter which 4479 * process we are. 4480 */ 4481 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT, 4482 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4483 if (p) { 4484 vm_page_wire(p); 4485 p->valid = VM_PAGE_BITS_ALL; 4486 vm_page_flag_clear(p, PG_ZERO); 4487 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 4488 bp->b_xio.xio_pages[index] = p; 4489 vm_page_wakeup(p); 4490 4491 pg += PAGE_SIZE; 4492 ++index; 4493 } 4494 } 4495 bp->b_xio.xio_npages = index; 4496 } 4497 4498 /* 4499 * Allocate pages for a buffer cache buffer. 4500 * 4501 * Under extremely severe memory conditions even allocating out of the 4502 * system reserve can fail. If this occurs we must allocate out of the 4503 * interrupt reserve to avoid a deadlock with the pageout daemon. 4504 * 4505 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf). 4506 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock 4507 * against the pageout daemon if pages are not freed from other sources. 4508 * 4509 * MPSAFE 4510 */ 4511 static 4512 vm_page_t 4513 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit) 4514 { 4515 vm_page_t p; 4516 4517 /* 4518 * Try a normal allocation, allow use of system reserve. 4519 */ 4520 lwkt_gettoken(&vm_token); 4521 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 4522 if (p) { 4523 lwkt_reltoken(&vm_token); 4524 return(p); 4525 } 4526 4527 /* 4528 * The normal allocation failed and we clearly have a page 4529 * deficit. Try to reclaim some clean VM pages directly 4530 * from the buffer cache. 4531 */ 4532 vm_pageout_deficit += deficit; 4533 recoverbufpages(); 4534 4535 /* 4536 * We may have blocked, the caller will know what to do if the 4537 * page now exists. 4538 */ 4539 if (vm_page_lookup(obj, pg)) { 4540 lwkt_reltoken(&vm_token); 4541 return(NULL); 4542 } 4543 4544 /* 4545 * Allocate and allow use of the interrupt reserve. 4546 * 4547 * If after all that we still can't allocate a VM page we are 4548 * in real trouble, but we slog on anyway hoping that the system 4549 * won't deadlock. 4550 */ 4551 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 4552 VM_ALLOC_INTERRUPT); 4553 if (p) { 4554 if (vm_page_count_severe()) { 4555 ++lowmempgallocs; 4556 vm_wait(hz / 20 + 1); 4557 } 4558 } else { 4559 kprintf("bio_page_alloc: Memory exhausted during bufcache " 4560 "page allocation\n"); 4561 ++lowmempgfails; 4562 vm_wait(hz); 4563 } 4564 lwkt_reltoken(&vm_token); 4565 return(p); 4566 } 4567 4568 /* 4569 * vm_hold_free_pages: 4570 * 4571 * Return pages associated with the buffer back to the VM system. 4572 * 4573 * The range of pages underlying the buffer's address space will 4574 * be unmapped and un-wired. 4575 * 4576 * MPSAFE 4577 */ 4578 void 4579 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4580 { 4581 vm_offset_t pg; 4582 vm_page_t p; 4583 int index, newnpages; 4584 4585 from = round_page(from); 4586 to = round_page(to); 4587 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4588 newnpages = index; 4589 4590 lwkt_gettoken(&vm_token); 4591 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4592 p = bp->b_xio.xio_pages[index]; 4593 if (p && (index < bp->b_xio.xio_npages)) { 4594 if (p->busy) { 4595 kprintf("vm_hold_free_pages: doffset: %lld, " 4596 "loffset: %lld\n", 4597 (long long)bp->b_bio2.bio_offset, 4598 (long long)bp->b_loffset); 4599 } 4600 bp->b_xio.xio_pages[index] = NULL; 4601 pmap_kremove(pg); 4602 vm_page_busy(p); 4603 vm_page_unwire(p, 0); 4604 vm_page_free(p); 4605 } 4606 } 4607 bp->b_xio.xio_npages = newnpages; 4608 lwkt_reltoken(&vm_token); 4609 } 4610 4611 /* 4612 * vmapbuf: 4613 * 4614 * Map a user buffer into KVM via a pbuf. On return the buffer's 4615 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4616 * initialized. 4617 */ 4618 int 4619 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4620 { 4621 caddr_t addr; 4622 vm_offset_t va; 4623 vm_page_t m; 4624 int vmprot; 4625 int error; 4626 int pidx; 4627 int i; 4628 4629 /* 4630 * bp had better have a command and it better be a pbuf. 4631 */ 4632 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4633 KKASSERT(bp->b_flags & B_PAGING); 4634 KKASSERT(bp->b_kvabase); 4635 4636 if (bytes < 0) 4637 return (-1); 4638 4639 /* 4640 * Map the user data into KVM. Mappings have to be page-aligned. 4641 */ 4642 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4643 pidx = 0; 4644 4645 vmprot = VM_PROT_READ; 4646 if (bp->b_cmd == BUF_CMD_READ) 4647 vmprot |= VM_PROT_WRITE; 4648 4649 while (addr < udata + bytes) { 4650 /* 4651 * Do the vm_fault if needed; do the copy-on-write thing 4652 * when reading stuff off device into memory. 4653 * 4654 * vm_fault_page*() returns a held VM page. 4655 */ 4656 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4657 va = trunc_page(va); 4658 4659 m = vm_fault_page_quick(va, vmprot, &error); 4660 if (m == NULL) { 4661 for (i = 0; i < pidx; ++i) { 4662 vm_page_unhold(bp->b_xio.xio_pages[i]); 4663 bp->b_xio.xio_pages[i] = NULL; 4664 } 4665 return(-1); 4666 } 4667 bp->b_xio.xio_pages[pidx] = m; 4668 addr += PAGE_SIZE; 4669 ++pidx; 4670 } 4671 4672 /* 4673 * Map the page array and set the buffer fields to point to 4674 * the mapped data buffer. 4675 */ 4676 if (pidx > btoc(MAXPHYS)) 4677 panic("vmapbuf: mapped more than MAXPHYS"); 4678 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4679 4680 bp->b_xio.xio_npages = pidx; 4681 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4682 bp->b_bcount = bytes; 4683 bp->b_bufsize = bytes; 4684 return(0); 4685 } 4686 4687 /* 4688 * vunmapbuf: 4689 * 4690 * Free the io map PTEs associated with this IO operation. 4691 * We also invalidate the TLB entries and restore the original b_addr. 4692 */ 4693 void 4694 vunmapbuf(struct buf *bp) 4695 { 4696 int pidx; 4697 int npages; 4698 4699 KKASSERT(bp->b_flags & B_PAGING); 4700 4701 npages = bp->b_xio.xio_npages; 4702 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4703 for (pidx = 0; pidx < npages; ++pidx) { 4704 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4705 bp->b_xio.xio_pages[pidx] = NULL; 4706 } 4707 bp->b_xio.xio_npages = 0; 4708 bp->b_data = bp->b_kvabase; 4709 } 4710 4711 /* 4712 * Scan all buffers in the system and issue the callback. 4713 */ 4714 int 4715 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4716 { 4717 int count = 0; 4718 int error; 4719 int n; 4720 4721 for (n = 0; n < nbuf; ++n) { 4722 if ((error = callback(&buf[n], info)) < 0) { 4723 count = error; 4724 break; 4725 } 4726 count += error; 4727 } 4728 return (count); 4729 } 4730 4731 /* 4732 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4733 * completion to the master buffer. 4734 */ 4735 static void 4736 nestiobuf_iodone(struct bio *bio) 4737 { 4738 struct bio *mbio; 4739 struct buf *mbp, *bp; 4740 struct devstat *stats; 4741 int error; 4742 int donebytes; 4743 4744 bp = bio->bio_buf; 4745 mbio = bio->bio_caller_info1.ptr; 4746 stats = bio->bio_caller_info2.ptr; 4747 mbp = mbio->bio_buf; 4748 4749 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4750 KKASSERT(mbp != bp); 4751 4752 error = bp->b_error; 4753 if (bp->b_error == 0 && 4754 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4755 /* 4756 * Not all got transfered, raise an error. We have no way to 4757 * propagate these conditions to mbp. 4758 */ 4759 error = EIO; 4760 } 4761 4762 donebytes = bp->b_bufsize; 4763 4764 relpbuf(bp, NULL); 4765 4766 nestiobuf_done(mbio, donebytes, error, stats); 4767 } 4768 4769 void 4770 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4771 { 4772 struct buf *mbp; 4773 4774 mbp = mbio->bio_buf; 4775 4776 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4777 4778 /* 4779 * If an error occured, propagate it to the master buffer. 4780 * 4781 * Several biodone()s may wind up running concurrently so 4782 * use an atomic op to adjust b_flags. 4783 */ 4784 if (error) { 4785 mbp->b_error = error; 4786 atomic_set_int(&mbp->b_flags, B_ERROR); 4787 } 4788 4789 /* 4790 * Decrement the operations in progress counter and terminate the 4791 * I/O if this was the last bit. 4792 */ 4793 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4794 mbp->b_resid = 0; 4795 if (stats) 4796 devstat_end_transaction_buf(stats, mbp); 4797 biodone(mbio); 4798 } 4799 } 4800 4801 /* 4802 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4803 * the mbio from being biodone()'d while we are still adding sub-bios to 4804 * it. 4805 */ 4806 void 4807 nestiobuf_init(struct bio *bio) 4808 { 4809 bio->bio_driver_info = (void *)1; 4810 } 4811 4812 /* 4813 * The BIOs added to the nestedio have already been started, remove the 4814 * count that placeheld our mbio and biodone() it if the count would 4815 * transition to 0. 4816 */ 4817 void 4818 nestiobuf_start(struct bio *mbio) 4819 { 4820 struct buf *mbp = mbio->bio_buf; 4821 4822 /* 4823 * Decrement the operations in progress counter and terminate the 4824 * I/O if this was the last bit. 4825 */ 4826 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4827 if (mbp->b_flags & B_ERROR) 4828 mbp->b_resid = mbp->b_bcount; 4829 else 4830 mbp->b_resid = 0; 4831 biodone(mbio); 4832 } 4833 } 4834 4835 /* 4836 * Set an intermediate error prior to calling nestiobuf_start() 4837 */ 4838 void 4839 nestiobuf_error(struct bio *mbio, int error) 4840 { 4841 struct buf *mbp = mbio->bio_buf; 4842 4843 if (error) { 4844 mbp->b_error = error; 4845 atomic_set_int(&mbp->b_flags, B_ERROR); 4846 } 4847 } 4848 4849 /* 4850 * nestiobuf_add: setup a "nested" buffer. 4851 * 4852 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4853 * => 'bp' should be a buffer allocated by getiobuf. 4854 * => 'offset' is a byte offset in the master buffer. 4855 * => 'size' is a size in bytes of this nested buffer. 4856 */ 4857 void 4858 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4859 { 4860 struct buf *mbp = mbio->bio_buf; 4861 struct vnode *vp = mbp->b_vp; 4862 4863 KKASSERT(mbp->b_bcount >= offset + size); 4864 4865 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4866 4867 /* kernel needs to own the lock for it to be released in biodone */ 4868 BUF_KERNPROC(bp); 4869 bp->b_vp = vp; 4870 bp->b_cmd = mbp->b_cmd; 4871 bp->b_bio1.bio_done = nestiobuf_iodone; 4872 bp->b_data = (char *)mbp->b_data + offset; 4873 bp->b_resid = bp->b_bcount = size; 4874 bp->b_bufsize = bp->b_bcount; 4875 4876 bp->b_bio1.bio_track = NULL; 4877 bp->b_bio1.bio_caller_info1.ptr = mbio; 4878 bp->b_bio1.bio_caller_info2.ptr = stats; 4879 } 4880 4881 /* 4882 * print out statistics from the current status of the buffer pool 4883 * this can be toggeled by the system control option debug.syncprt 4884 */ 4885 #ifdef DEBUG 4886 void 4887 vfs_bufstats(void) 4888 { 4889 int i, j, count; 4890 struct buf *bp; 4891 struct bqueues *dp; 4892 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 4893 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 4894 4895 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 4896 count = 0; 4897 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4898 counts[j] = 0; 4899 4900 spin_lock(&bufqspin); 4901 TAILQ_FOREACH(bp, dp, b_freelist) { 4902 counts[bp->b_bufsize/PAGE_SIZE]++; 4903 count++; 4904 } 4905 spin_unlock(&bufqspin); 4906 4907 kprintf("%s: total-%d", bname[i], count); 4908 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4909 if (counts[j] != 0) 4910 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]); 4911 kprintf("\n"); 4912 } 4913 } 4914 #endif 4915 4916 #ifdef DDB 4917 4918 DB_SHOW_COMMAND(buffer, db_show_buffer) 4919 { 4920 /* get args */ 4921 struct buf *bp = (struct buf *)addr; 4922 4923 if (!have_addr) { 4924 db_printf("usage: show buffer <addr>\n"); 4925 return; 4926 } 4927 4928 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4929 db_printf("b_cmd = %d\n", bp->b_cmd); 4930 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4931 "b_resid = %d\n, b_data = %p, " 4932 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4933 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4934 bp->b_data, 4935 (long long)bp->b_bio2.bio_offset, 4936 (long long)(bp->b_bio2.bio_next ? 4937 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4938 if (bp->b_xio.xio_npages) { 4939 int i; 4940 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4941 bp->b_xio.xio_npages); 4942 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4943 vm_page_t m; 4944 m = bp->b_xio.xio_pages[i]; 4945 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4946 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4947 if ((i + 1) < bp->b_xio.xio_npages) 4948 db_printf(","); 4949 } 4950 db_printf("\n"); 4951 } 4952 } 4953 #endif /* DDB */ 4954