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, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2047 2048 /* 2049 * Note: we no longer distinguish between VMIO and non-VMIO 2050 * buffers. 2051 */ 2052 KASSERT((bp->b_flags & B_DELWRI) == 0, 2053 ("delwri buffer %p found in queue %d", bp, qindex)); 2054 2055 /* 2056 * Do not try to reuse a buffer with a non-zero b_refs. 2057 * This is an unsynchronized test. A synchronized test 2058 * is also performed after we lock the buffer. 2059 */ 2060 if (bp->b_refs) 2061 continue; 2062 2063 /* 2064 * If we are defragging then we need a buffer with 2065 * b_kvasize != 0. XXX this situation should no longer 2066 * occur, if defrag is non-zero the buffer's b_kvasize 2067 * should also be non-zero at this point. XXX 2068 */ 2069 if (defrag && bp->b_kvasize == 0) { 2070 kprintf("Warning: defrag empty buffer %p\n", bp); 2071 continue; 2072 } 2073 2074 /* 2075 * Start freeing the bp. This is somewhat involved. nbp 2076 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 2077 * on the clean list must be disassociated from their 2078 * current vnode. Buffers on the empty[kva] lists have 2079 * already been disassociated. 2080 */ 2081 2082 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2083 spin_unlock(&bufqspin); 2084 tsleep(&bd_request, 0, "gnbxxx", hz / 100); 2085 goto restart; 2086 } 2087 if (bp->b_qindex != qindex) { 2088 spin_unlock(&bufqspin); 2089 kprintf("getnewbuf: warning, BUF_LOCK blocked " 2090 "unexpectedly on buf %p index %d->%d, " 2091 "race corrected\n", 2092 bp, qindex, bp->b_qindex); 2093 BUF_UNLOCK(bp); 2094 goto restart; 2095 } 2096 bremfree_locked(bp); 2097 spin_unlock(&bufqspin); 2098 2099 /* 2100 * Dependancies must be handled before we disassociate the 2101 * vnode. 2102 * 2103 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2104 * be immediately disassociated. HAMMER then becomes 2105 * responsible for releasing the buffer. 2106 * 2107 * NOTE: bufqspin is UNLOCKED now. 2108 */ 2109 if (LIST_FIRST(&bp->b_dep) != NULL) { 2110 buf_deallocate(bp); 2111 if (bp->b_flags & B_LOCKED) { 2112 bqrelse(bp); 2113 goto restart; 2114 } 2115 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2116 } 2117 2118 if (qindex == BQUEUE_CLEAN) { 2119 if (bp->b_flags & B_VMIO) 2120 vfs_vmio_release(bp); 2121 if (bp->b_vp) 2122 brelvp(bp); 2123 } 2124 2125 /* 2126 * NOTE: nbp is now entirely invalid. We can only restart 2127 * the scan from this point on. 2128 * 2129 * Get the rest of the buffer freed up. b_kva* is still 2130 * valid after this operation. 2131 */ 2132 2133 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex)); 2134 KKASSERT((bp->b_flags & B_HASHED) == 0); 2135 2136 /* 2137 * critical section protection is not required when 2138 * scrapping a buffer's contents because it is already 2139 * wired. 2140 */ 2141 if (bp->b_bufsize) 2142 allocbuf(bp, 0); 2143 2144 bp->b_flags = B_BNOCLIP; 2145 bp->b_cmd = BUF_CMD_DONE; 2146 bp->b_vp = NULL; 2147 bp->b_error = 0; 2148 bp->b_resid = 0; 2149 bp->b_bcount = 0; 2150 bp->b_xio.xio_npages = 0; 2151 bp->b_dirtyoff = bp->b_dirtyend = 0; 2152 bp->b_act_count = ACT_INIT; 2153 reinitbufbio(bp); 2154 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2155 buf_dep_init(bp); 2156 if (blkflags & GETBLK_BHEAVY) 2157 bp->b_flags |= B_HEAVY; 2158 2159 /* 2160 * If we are defragging then free the buffer. 2161 */ 2162 if (defrag) { 2163 bp->b_flags |= B_INVAL; 2164 bfreekva(bp); 2165 brelse(bp); 2166 defrag = 0; 2167 goto restart; 2168 } 2169 2170 /* 2171 * If we are overcomitted then recover the buffer and its 2172 * KVM space. This occurs in rare situations when multiple 2173 * processes are blocked in getnewbuf() or allocbuf(). 2174 */ 2175 if (bufspace >= hibufspace) 2176 flushingbufs = 1; 2177 if (flushingbufs && bp->b_kvasize != 0) { 2178 bp->b_flags |= B_INVAL; 2179 bfreekva(bp); 2180 brelse(bp); 2181 goto restart; 2182 } 2183 if (bufspace < lobufspace) 2184 flushingbufs = 0; 2185 2186 /* 2187 * The brelvp() above interlocked the buffer, test b_refs 2188 * to determine if the buffer can be reused. b_refs 2189 * interlocks lookup/blocking-lock operations and allowing 2190 * buffer reuse can create deadlocks depending on what 2191 * (vp,loffset) is assigned to the reused buffer (see getblk). 2192 */ 2193 if (bp->b_refs) { 2194 bp->b_flags |= B_INVAL; 2195 bfreekva(bp); 2196 brelse(bp); 2197 goto restart; 2198 } 2199 2200 break; 2201 /* NOT REACHED, bufqspin not held */ 2202 } 2203 2204 /* 2205 * If we exhausted our list, sleep as appropriate. We may have to 2206 * wakeup various daemons and write out some dirty buffers. 2207 * 2208 * Generally we are sleeping due to insufficient buffer space. 2209 * 2210 * NOTE: bufqspin is held if bp is NULL, else it is not held. 2211 */ 2212 if (bp == NULL) { 2213 int flags; 2214 char *waitmsg; 2215 2216 spin_unlock(&bufqspin); 2217 if (defrag) { 2218 flags = VFS_BIO_NEED_BUFSPACE; 2219 waitmsg = "nbufkv"; 2220 } else if (bufspace >= hibufspace) { 2221 waitmsg = "nbufbs"; 2222 flags = VFS_BIO_NEED_BUFSPACE; 2223 } else { 2224 waitmsg = "newbuf"; 2225 flags = VFS_BIO_NEED_ANY; 2226 } 2227 2228 bd_speedup(); /* heeeelp */ 2229 spin_lock(&bufcspin); 2230 needsbuffer |= flags; 2231 while (needsbuffer & flags) { 2232 if (ssleep(&needsbuffer, &bufcspin, 2233 slpflags, waitmsg, slptimeo)) { 2234 spin_unlock(&bufcspin); 2235 return (NULL); 2236 } 2237 } 2238 spin_unlock(&bufcspin); 2239 } else { 2240 /* 2241 * We finally have a valid bp. We aren't quite out of the 2242 * woods, we still have to reserve kva space. In order 2243 * to keep fragmentation sane we only allocate kva in 2244 * BKVASIZE chunks. 2245 * 2246 * (bufqspin is not held) 2247 */ 2248 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2249 2250 if (maxsize != bp->b_kvasize) { 2251 vm_offset_t addr = 0; 2252 int count; 2253 2254 bfreekva(bp); 2255 2256 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2257 vm_map_lock(&buffer_map); 2258 2259 if (vm_map_findspace(&buffer_map, 2260 vm_map_min(&buffer_map), maxsize, 2261 maxsize, 0, &addr)) { 2262 /* 2263 * Uh oh. Buffer map is too fragmented. We 2264 * must defragment the map. 2265 */ 2266 vm_map_unlock(&buffer_map); 2267 vm_map_entry_release(count); 2268 ++bufdefragcnt; 2269 defrag = 1; 2270 bp->b_flags |= B_INVAL; 2271 brelse(bp); 2272 goto restart; 2273 } 2274 if (addr) { 2275 vm_map_insert(&buffer_map, &count, 2276 NULL, 0, 2277 addr, addr + maxsize, 2278 VM_MAPTYPE_NORMAL, 2279 VM_PROT_ALL, VM_PROT_ALL, 2280 MAP_NOFAULT); 2281 2282 bp->b_kvabase = (caddr_t) addr; 2283 bp->b_kvasize = maxsize; 2284 bufspace += bp->b_kvasize; 2285 ++bufreusecnt; 2286 } 2287 vm_map_unlock(&buffer_map); 2288 vm_map_entry_release(count); 2289 } 2290 bp->b_data = bp->b_kvabase; 2291 } 2292 return(bp); 2293 } 2294 2295 /* 2296 * This routine is called in an emergency to recover VM pages from the 2297 * buffer cache by cashing in clean buffers. The idea is to recover 2298 * enough pages to be able to satisfy a stuck bio_page_alloc(). 2299 * 2300 * MPSAFE 2301 */ 2302 static int 2303 recoverbufpages(void) 2304 { 2305 struct buf *bp; 2306 int bytes = 0; 2307 2308 ++recoverbufcalls; 2309 2310 spin_lock(&bufqspin); 2311 while (bytes < MAXBSIZE) { 2312 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 2313 if (bp == NULL) 2314 break; 2315 2316 /* 2317 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2318 * cycles through the queue twice before being selected. 2319 */ 2320 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) { 2321 bp->b_flags |= B_AGE; 2322 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 2323 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], 2324 bp, b_freelist); 2325 continue; 2326 } 2327 2328 /* 2329 * Sanity Checks 2330 */ 2331 KKASSERT(bp->b_qindex == BQUEUE_CLEAN); 2332 KKASSERT((bp->b_flags & B_DELWRI) == 0); 2333 2334 /* 2335 * Start freeing the bp. This is somewhat involved. 2336 * 2337 * Buffers on the clean list must be disassociated from 2338 * their current vnode 2339 */ 2340 2341 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2342 kprintf("recoverbufpages: warning, locked buf %p, " 2343 "race corrected\n", 2344 bp); 2345 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100); 2346 continue; 2347 } 2348 if (bp->b_qindex != BQUEUE_CLEAN) { 2349 kprintf("recoverbufpages: warning, BUF_LOCK blocked " 2350 "unexpectedly on buf %p index %d, race " 2351 "corrected\n", 2352 bp, bp->b_qindex); 2353 BUF_UNLOCK(bp); 2354 continue; 2355 } 2356 bremfree_locked(bp); 2357 spin_unlock(&bufqspin); 2358 2359 /* 2360 * Dependancies must be handled before we disassociate the 2361 * vnode. 2362 * 2363 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2364 * be immediately disassociated. HAMMER then becomes 2365 * responsible for releasing the buffer. 2366 */ 2367 if (LIST_FIRST(&bp->b_dep) != NULL) { 2368 buf_deallocate(bp); 2369 if (bp->b_flags & B_LOCKED) { 2370 bqrelse(bp); 2371 spin_lock(&bufqspin); 2372 continue; 2373 } 2374 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2375 } 2376 2377 bytes += bp->b_bufsize; 2378 2379 if (bp->b_flags & B_VMIO) { 2380 bp->b_flags |= B_DIRECT; /* try to free pages */ 2381 vfs_vmio_release(bp); 2382 } 2383 if (bp->b_vp) 2384 brelvp(bp); 2385 2386 KKASSERT(bp->b_vp == NULL); 2387 KKASSERT((bp->b_flags & B_HASHED) == 0); 2388 2389 /* 2390 * critical section protection is not required when 2391 * scrapping a buffer's contents because it is already 2392 * wired. 2393 */ 2394 if (bp->b_bufsize) 2395 allocbuf(bp, 0); 2396 2397 bp->b_flags = B_BNOCLIP; 2398 bp->b_cmd = BUF_CMD_DONE; 2399 bp->b_vp = NULL; 2400 bp->b_error = 0; 2401 bp->b_resid = 0; 2402 bp->b_bcount = 0; 2403 bp->b_xio.xio_npages = 0; 2404 bp->b_dirtyoff = bp->b_dirtyend = 0; 2405 reinitbufbio(bp); 2406 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2407 buf_dep_init(bp); 2408 bp->b_flags |= B_INVAL; 2409 /* bfreekva(bp); */ 2410 brelse(bp); 2411 spin_lock(&bufqspin); 2412 } 2413 spin_unlock(&bufqspin); 2414 return(bytes); 2415 } 2416 2417 /* 2418 * buf_daemon: 2419 * 2420 * Buffer flushing daemon. Buffers are normally flushed by the 2421 * update daemon but if it cannot keep up this process starts to 2422 * take the load in an attempt to prevent getnewbuf() from blocking. 2423 * 2424 * Once a flush is initiated it does not stop until the number 2425 * of buffers falls below lodirtybuffers, but we will wake up anyone 2426 * waiting at the mid-point. 2427 */ 2428 2429 static struct kproc_desc buf_kp = { 2430 "bufdaemon", 2431 buf_daemon, 2432 &bufdaemon_td 2433 }; 2434 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2435 kproc_start, &buf_kp) 2436 2437 static struct kproc_desc bufhw_kp = { 2438 "bufdaemon_hw", 2439 buf_daemon_hw, 2440 &bufdaemonhw_td 2441 }; 2442 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2443 kproc_start, &bufhw_kp) 2444 2445 /* 2446 * MPSAFE thread 2447 */ 2448 static void 2449 buf_daemon(void) 2450 { 2451 int limit; 2452 2453 /* 2454 * This process needs to be suspended prior to shutdown sync. 2455 */ 2456 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2457 bufdaemon_td, SHUTDOWN_PRI_LAST); 2458 curthread->td_flags |= TDF_SYSTHREAD; 2459 2460 /* 2461 * This process is allowed to take the buffer cache to the limit 2462 */ 2463 for (;;) { 2464 kproc_suspend_loop(); 2465 2466 /* 2467 * Do the flush as long as the number of dirty buffers 2468 * (including those running) exceeds lodirtybufspace. 2469 * 2470 * When flushing limit running I/O to hirunningspace 2471 * Do the flush. Limit the amount of in-transit I/O we 2472 * allow to build up, otherwise we would completely saturate 2473 * the I/O system. Wakeup any waiting processes before we 2474 * normally would so they can run in parallel with our drain. 2475 * 2476 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2477 * but because we split the operation into two threads we 2478 * have to cut it in half for each thread. 2479 */ 2480 waitrunningbufspace(); 2481 limit = lodirtybufspace / 2; 2482 while (runningbufspace + dirtybufspace > limit || 2483 dirtybufcount - dirtybufcounthw >= nbuf / 2) { 2484 if (flushbufqueues(BQUEUE_DIRTY) == 0) 2485 break; 2486 if (runningbufspace < hirunningspace) 2487 continue; 2488 waitrunningbufspace(); 2489 } 2490 2491 /* 2492 * We reached our low water mark, reset the 2493 * request and sleep until we are needed again. 2494 * The sleep is just so the suspend code works. 2495 */ 2496 spin_lock(&bufcspin); 2497 if (bd_request == 0) 2498 ssleep(&bd_request, &bufcspin, 0, "psleep", hz); 2499 bd_request = 0; 2500 spin_unlock(&bufcspin); 2501 } 2502 } 2503 2504 /* 2505 * MPSAFE thread 2506 */ 2507 static void 2508 buf_daemon_hw(void) 2509 { 2510 int limit; 2511 2512 /* 2513 * This process needs to be suspended prior to shutdown sync. 2514 */ 2515 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2516 bufdaemonhw_td, SHUTDOWN_PRI_LAST); 2517 curthread->td_flags |= TDF_SYSTHREAD; 2518 2519 /* 2520 * This process is allowed to take the buffer cache to the limit 2521 */ 2522 for (;;) { 2523 kproc_suspend_loop(); 2524 2525 /* 2526 * Do the flush. Limit the amount of in-transit I/O we 2527 * allow to build up, otherwise we would completely saturate 2528 * the I/O system. Wakeup any waiting processes before we 2529 * normally would so they can run in parallel with our drain. 2530 * 2531 * Once we decide to flush push the queued I/O up to 2532 * hirunningspace in order to trigger bursting by the bioq 2533 * subsystem. 2534 * 2535 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2536 * but because we split the operation into two threads we 2537 * have to cut it in half for each thread. 2538 */ 2539 waitrunningbufspace(); 2540 limit = lodirtybufspace / 2; 2541 while (runningbufspace + dirtybufspacehw > limit || 2542 dirtybufcounthw >= nbuf / 2) { 2543 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0) 2544 break; 2545 if (runningbufspace < hirunningspace) 2546 continue; 2547 waitrunningbufspace(); 2548 } 2549 2550 /* 2551 * We reached our low water mark, reset the 2552 * request and sleep until we are needed again. 2553 * The sleep is just so the suspend code works. 2554 */ 2555 spin_lock(&bufcspin); 2556 if (bd_request_hw == 0) 2557 ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz); 2558 bd_request_hw = 0; 2559 spin_unlock(&bufcspin); 2560 } 2561 } 2562 2563 /* 2564 * flushbufqueues: 2565 * 2566 * Try to flush a buffer in the dirty queue. We must be careful to 2567 * free up B_INVAL buffers instead of write them, which NFS is 2568 * particularly sensitive to. 2569 * 2570 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate 2571 * that we really want to try to get the buffer out and reuse it 2572 * due to the write load on the machine. 2573 * 2574 * We must lock the buffer in order to check its validity before we 2575 * can mess with its contents. bufqspin isn't enough. 2576 */ 2577 static int 2578 flushbufqueues(bufq_type_t q) 2579 { 2580 struct buf *bp; 2581 int r = 0; 2582 int spun; 2583 2584 spin_lock(&bufqspin); 2585 spun = 1; 2586 2587 bp = TAILQ_FIRST(&bufqueues[q]); 2588 while (bp) { 2589 if ((bp->b_flags & B_DELWRI) == 0) { 2590 kprintf("Unexpected clean buffer %p\n", bp); 2591 bp = TAILQ_NEXT(bp, b_freelist); 2592 continue; 2593 } 2594 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2595 bp = TAILQ_NEXT(bp, b_freelist); 2596 continue; 2597 } 2598 KKASSERT(bp->b_qindex == q); 2599 2600 /* 2601 * Must recheck B_DELWRI after successfully locking 2602 * the buffer. 2603 */ 2604 if ((bp->b_flags & B_DELWRI) == 0) { 2605 BUF_UNLOCK(bp); 2606 bp = TAILQ_NEXT(bp, b_freelist); 2607 continue; 2608 } 2609 2610 if (bp->b_flags & B_INVAL) { 2611 _bremfree(bp); 2612 spin_unlock(&bufqspin); 2613 spun = 0; 2614 brelse(bp); 2615 ++r; 2616 break; 2617 } 2618 2619 spin_unlock(&bufqspin); 2620 spun = 0; 2621 2622 if (LIST_FIRST(&bp->b_dep) != NULL && 2623 (bp->b_flags & B_DEFERRED) == 0 && 2624 buf_countdeps(bp, 0)) { 2625 spin_lock(&bufqspin); 2626 spun = 1; 2627 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist); 2628 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist); 2629 bp->b_flags |= B_DEFERRED; 2630 BUF_UNLOCK(bp); 2631 bp = TAILQ_FIRST(&bufqueues[q]); 2632 continue; 2633 } 2634 2635 /* 2636 * If the buffer has a dependancy, buf_checkwrite() must 2637 * also return 0 for us to be able to initate the write. 2638 * 2639 * If the buffer is flagged B_ERROR it may be requeued 2640 * over and over again, we try to avoid a live lock. 2641 * 2642 * NOTE: buf_checkwrite is MPSAFE. 2643 */ 2644 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2645 bremfree(bp); 2646 brelse(bp); 2647 } else if (bp->b_flags & B_ERROR) { 2648 tsleep(bp, 0, "bioer", 1); 2649 bp->b_flags &= ~B_AGE; 2650 vfs_bio_awrite(bp); 2651 } else { 2652 bp->b_flags |= B_AGE; 2653 vfs_bio_awrite(bp); 2654 } 2655 ++r; 2656 break; 2657 } 2658 if (spun) 2659 spin_unlock(&bufqspin); 2660 return (r); 2661 } 2662 2663 /* 2664 * inmem: 2665 * 2666 * Returns true if no I/O is needed to access the associated VM object. 2667 * This is like findblk except it also hunts around in the VM system for 2668 * the data. 2669 * 2670 * Note that we ignore vm_page_free() races from interrupts against our 2671 * lookup, since if the caller is not protected our return value will not 2672 * be any more valid then otherwise once we exit the critical section. 2673 */ 2674 int 2675 inmem(struct vnode *vp, off_t loffset) 2676 { 2677 vm_object_t obj; 2678 vm_offset_t toff, tinc, size; 2679 vm_page_t m; 2680 2681 if (findblk(vp, loffset, FINDBLK_TEST)) 2682 return 1; 2683 if (vp->v_mount == NULL) 2684 return 0; 2685 if ((obj = vp->v_object) == NULL) 2686 return 0; 2687 2688 size = PAGE_SIZE; 2689 if (size > vp->v_mount->mnt_stat.f_iosize) 2690 size = vp->v_mount->mnt_stat.f_iosize; 2691 2692 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2693 lwkt_gettoken(&vm_token); 2694 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2695 lwkt_reltoken(&vm_token); 2696 if (m == NULL) 2697 return 0; 2698 tinc = size; 2699 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2700 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2701 if (vm_page_is_valid(m, 2702 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) 2703 return 0; 2704 } 2705 return 1; 2706 } 2707 2708 /* 2709 * findblk: 2710 * 2711 * Locate and return the specified buffer. Unless flagged otherwise, 2712 * a locked buffer will be returned if it exists or NULL if it does not. 2713 * 2714 * findblk()'d buffers are still on the bufqueues and if you intend 2715 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2716 * and possibly do other stuff to it. 2717 * 2718 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2719 * for locking the buffer and ensuring that it remains 2720 * the desired buffer after locking. 2721 * 2722 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2723 * to acquire the lock we return NULL, even if the 2724 * buffer exists. 2725 * 2726 * FINDBLK_REF - Returns the buffer ref'd, which prevents reuse 2727 * by getnewbuf() but does not prevent disassociation 2728 * while we are locked. Used to avoid deadlocks 2729 * against random (vp,loffset)s due to reassignment. 2730 * 2731 * (0) - Lock the buffer blocking. 2732 * 2733 * MPSAFE 2734 */ 2735 struct buf * 2736 findblk(struct vnode *vp, off_t loffset, int flags) 2737 { 2738 struct buf *bp; 2739 int lkflags; 2740 2741 lkflags = LK_EXCLUSIVE; 2742 if (flags & FINDBLK_NBLOCK) 2743 lkflags |= LK_NOWAIT; 2744 2745 for (;;) { 2746 /* 2747 * Lookup. Ref the buf while holding v_token to prevent 2748 * reuse (but does not prevent diassociation). 2749 */ 2750 lwkt_gettoken(&vp->v_token); 2751 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2752 if (bp == NULL) { 2753 lwkt_reltoken(&vp->v_token); 2754 return(NULL); 2755 } 2756 atomic_add_int(&bp->b_refs, 1); 2757 lwkt_reltoken(&vp->v_token); 2758 2759 /* 2760 * If testing only break and return bp, do not lock. 2761 */ 2762 if (flags & FINDBLK_TEST) 2763 break; 2764 2765 /* 2766 * Lock the buffer, return an error if the lock fails. 2767 * (only FINDBLK_NBLOCK can cause the lock to fail). 2768 */ 2769 if (BUF_LOCK(bp, lkflags)) { 2770 atomic_subtract_int(&bp->b_refs, 1); 2771 /* bp = NULL; not needed */ 2772 return(NULL); 2773 } 2774 2775 /* 2776 * Revalidate the locked buf before allowing it to be 2777 * returned. 2778 */ 2779 if (bp->b_vp == vp && bp->b_loffset == loffset) 2780 break; 2781 atomic_subtract_int(&bp->b_refs, 1); 2782 BUF_UNLOCK(bp); 2783 } 2784 2785 /* 2786 * Success 2787 */ 2788 if ((flags & FINDBLK_REF) == 0) 2789 atomic_subtract_int(&bp->b_refs, 1); 2790 return(bp); 2791 } 2792 2793 void 2794 unrefblk(struct buf *bp) 2795 { 2796 atomic_subtract_int(&bp->b_refs, 1); 2797 } 2798 2799 /* 2800 * getcacheblk: 2801 * 2802 * Similar to getblk() except only returns the buffer if it is 2803 * B_CACHE and requires no other manipulation. Otherwise NULL 2804 * is returned. 2805 * 2806 * If B_RAM is set the buffer might be just fine, but we return 2807 * NULL anyway because we want the code to fall through to the 2808 * cluster read. Otherwise read-ahead breaks. 2809 */ 2810 struct buf * 2811 getcacheblk(struct vnode *vp, off_t loffset) 2812 { 2813 struct buf *bp; 2814 2815 bp = findblk(vp, loffset, 0); 2816 if (bp) { 2817 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) { 2818 bp->b_flags &= ~B_AGE; 2819 bremfree(bp); 2820 } else { 2821 BUF_UNLOCK(bp); 2822 bp = NULL; 2823 } 2824 } 2825 return (bp); 2826 } 2827 2828 /* 2829 * getblk: 2830 * 2831 * Get a block given a specified block and offset into a file/device. 2832 * B_INVAL may or may not be set on return. The caller should clear 2833 * B_INVAL prior to initiating a READ. 2834 * 2835 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2836 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2837 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2838 * without doing any of those things the system will likely believe 2839 * the buffer to be valid (especially if it is not B_VMIO), and the 2840 * next getblk() will return the buffer with B_CACHE set. 2841 * 2842 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2843 * an existing buffer. 2844 * 2845 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2846 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2847 * and then cleared based on the backing VM. If the previous buffer is 2848 * non-0-sized but invalid, B_CACHE will be cleared. 2849 * 2850 * If getblk() must create a new buffer, the new buffer is returned with 2851 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2852 * case it is returned with B_INVAL clear and B_CACHE set based on the 2853 * backing VM. 2854 * 2855 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2856 * B_CACHE bit is clear. 2857 * 2858 * What this means, basically, is that the caller should use B_CACHE to 2859 * determine whether the buffer is fully valid or not and should clear 2860 * B_INVAL prior to issuing a read. If the caller intends to validate 2861 * the buffer by loading its data area with something, the caller needs 2862 * to clear B_INVAL. If the caller does this without issuing an I/O, 2863 * the caller should set B_CACHE ( as an optimization ), else the caller 2864 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2865 * a write attempt or if it was a successfull read. If the caller 2866 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2867 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2868 * 2869 * getblk flags: 2870 * 2871 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2872 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2873 * 2874 * MPALMOSTSAFE 2875 */ 2876 struct buf * 2877 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2878 { 2879 struct buf *bp; 2880 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2881 int error; 2882 int lkflags; 2883 2884 if (size > MAXBSIZE) 2885 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2886 if (vp->v_object == NULL) 2887 panic("getblk: vnode %p has no object!", vp); 2888 2889 loop: 2890 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2891 /* 2892 * The buffer was found in the cache, but we need to lock it. 2893 * We must acquire a ref on the bp to prevent reuse, but 2894 * this will not prevent disassociation (brelvp()) so we 2895 * must recheck (vp,loffset) after acquiring the lock. 2896 * 2897 * Without the ref the buffer could potentially be reused 2898 * before we acquire the lock and create a deadlock 2899 * situation between the thread trying to reuse the buffer 2900 * and us due to the fact that we would wind up blocking 2901 * on a random (vp,loffset). 2902 */ 2903 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2904 if (blkflags & GETBLK_NOWAIT) { 2905 unrefblk(bp); 2906 return(NULL); 2907 } 2908 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2909 if (blkflags & GETBLK_PCATCH) 2910 lkflags |= LK_PCATCH; 2911 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2912 if (error) { 2913 unrefblk(bp); 2914 if (error == ENOLCK) 2915 goto loop; 2916 return (NULL); 2917 } 2918 /* buffer may have changed on us */ 2919 } 2920 unrefblk(bp); 2921 2922 /* 2923 * Once the buffer has been locked, make sure we didn't race 2924 * a buffer recyclement. Buffers that are no longer hashed 2925 * will have b_vp == NULL, so this takes care of that check 2926 * as well. 2927 */ 2928 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2929 kprintf("Warning buffer %p (vp %p loffset %lld) " 2930 "was recycled\n", 2931 bp, vp, (long long)loffset); 2932 BUF_UNLOCK(bp); 2933 goto loop; 2934 } 2935 2936 /* 2937 * If SZMATCH any pre-existing buffer must be of the requested 2938 * size or NULL is returned. The caller absolutely does not 2939 * want getblk() to bwrite() the buffer on a size mismatch. 2940 */ 2941 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2942 BUF_UNLOCK(bp); 2943 return(NULL); 2944 } 2945 2946 /* 2947 * All vnode-based buffers must be backed by a VM object. 2948 */ 2949 KKASSERT(bp->b_flags & B_VMIO); 2950 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2951 bp->b_flags &= ~B_AGE; 2952 2953 /* 2954 * Make sure that B_INVAL buffers do not have a cached 2955 * block number translation. 2956 */ 2957 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2958 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2959 " did not have cleared bio_offset cache\n", 2960 bp, vp, (long long)loffset); 2961 clearbiocache(&bp->b_bio2); 2962 } 2963 2964 /* 2965 * The buffer is locked. B_CACHE is cleared if the buffer is 2966 * invalid. 2967 */ 2968 if (bp->b_flags & B_INVAL) 2969 bp->b_flags &= ~B_CACHE; 2970 bremfree(bp); 2971 2972 /* 2973 * Any size inconsistancy with a dirty buffer or a buffer 2974 * with a softupdates dependancy must be resolved. Resizing 2975 * the buffer in such circumstances can lead to problems. 2976 * 2977 * Dirty or dependant buffers are written synchronously. 2978 * Other types of buffers are simply released and 2979 * reconstituted as they may be backed by valid, dirty VM 2980 * pages (but not marked B_DELWRI). 2981 * 2982 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2983 * and may be left over from a prior truncation (and thus 2984 * no longer represent the actual EOF point), so we 2985 * definitely do not want to B_NOCACHE the backing store. 2986 */ 2987 if (size != bp->b_bcount) { 2988 if (bp->b_flags & B_DELWRI) { 2989 bp->b_flags |= B_RELBUF; 2990 bwrite(bp); 2991 } else if (LIST_FIRST(&bp->b_dep)) { 2992 bp->b_flags |= B_RELBUF; 2993 bwrite(bp); 2994 } else { 2995 bp->b_flags |= B_RELBUF; 2996 brelse(bp); 2997 } 2998 goto loop; 2999 } 3000 KKASSERT(size <= bp->b_kvasize); 3001 KASSERT(bp->b_loffset != NOOFFSET, 3002 ("getblk: no buffer offset")); 3003 3004 /* 3005 * A buffer with B_DELWRI set and B_CACHE clear must 3006 * be committed before we can return the buffer in 3007 * order to prevent the caller from issuing a read 3008 * ( due to B_CACHE not being set ) and overwriting 3009 * it. 3010 * 3011 * Most callers, including NFS and FFS, need this to 3012 * operate properly either because they assume they 3013 * can issue a read if B_CACHE is not set, or because 3014 * ( for example ) an uncached B_DELWRI might loop due 3015 * to softupdates re-dirtying the buffer. In the latter 3016 * case, B_CACHE is set after the first write completes, 3017 * preventing further loops. 3018 * 3019 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3020 * above while extending the buffer, we cannot allow the 3021 * buffer to remain with B_CACHE set after the write 3022 * completes or it will represent a corrupt state. To 3023 * deal with this we set B_NOCACHE to scrap the buffer 3024 * after the write. 3025 * 3026 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3027 * I'm not even sure this state is still possible 3028 * now that getblk() writes out any dirty buffers 3029 * on size changes. 3030 * 3031 * We might be able to do something fancy, like setting 3032 * B_CACHE in bwrite() except if B_DELWRI is already set, 3033 * so the below call doesn't set B_CACHE, but that gets real 3034 * confusing. This is much easier. 3035 */ 3036 3037 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3038 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3039 "and CACHE clear, b_flags %08x\n", 3040 bp, (intmax_t)bp->b_loffset, bp->b_flags); 3041 bp->b_flags |= B_NOCACHE; 3042 bwrite(bp); 3043 goto loop; 3044 } 3045 } else { 3046 /* 3047 * Buffer is not in-core, create new buffer. The buffer 3048 * returned by getnewbuf() is locked. Note that the returned 3049 * buffer is also considered valid (not marked B_INVAL). 3050 * 3051 * Calculating the offset for the I/O requires figuring out 3052 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3053 * the mount's f_iosize otherwise. If the vnode does not 3054 * have an associated mount we assume that the passed size is 3055 * the block size. 3056 * 3057 * Note that vn_isdisk() cannot be used here since it may 3058 * return a failure for numerous reasons. Note that the 3059 * buffer size may be larger then the block size (the caller 3060 * will use block numbers with the proper multiple). Beware 3061 * of using any v_* fields which are part of unions. In 3062 * particular, in DragonFly the mount point overloading 3063 * mechanism uses the namecache only and the underlying 3064 * directory vnode is not a special case. 3065 */ 3066 int bsize, maxsize; 3067 3068 if (vp->v_type == VBLK || vp->v_type == VCHR) 3069 bsize = DEV_BSIZE; 3070 else if (vp->v_mount) 3071 bsize = vp->v_mount->mnt_stat.f_iosize; 3072 else 3073 bsize = size; 3074 3075 maxsize = size + (loffset & PAGE_MASK); 3076 maxsize = imax(maxsize, bsize); 3077 3078 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3079 if (bp == NULL) { 3080 if (slpflags || slptimeo) 3081 return NULL; 3082 goto loop; 3083 } 3084 3085 /* 3086 * Atomically insert the buffer into the hash, so that it can 3087 * be found by findblk(). 3088 * 3089 * If bgetvp() returns non-zero a collision occured, and the 3090 * bp will not be associated with the vnode. 3091 * 3092 * Make sure the translation layer has been cleared. 3093 */ 3094 bp->b_loffset = loffset; 3095 bp->b_bio2.bio_offset = NOOFFSET; 3096 /* bp->b_bio2.bio_next = NULL; */ 3097 3098 if (bgetvp(vp, bp, size)) { 3099 bp->b_flags |= B_INVAL; 3100 brelse(bp); 3101 goto loop; 3102 } 3103 3104 /* 3105 * All vnode-based buffers must be backed by a VM object. 3106 */ 3107 KKASSERT(vp->v_object != NULL); 3108 bp->b_flags |= B_VMIO; 3109 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3110 3111 allocbuf(bp, size); 3112 } 3113 KKASSERT(dsched_is_clear_buf_priv(bp)); 3114 return (bp); 3115 } 3116 3117 /* 3118 * regetblk(bp) 3119 * 3120 * Reacquire a buffer that was previously released to the locked queue, 3121 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3122 * set B_LOCKED (which handles the acquisition race). 3123 * 3124 * To this end, either B_LOCKED must be set or the dependancy list must be 3125 * non-empty. 3126 * 3127 * MPSAFE 3128 */ 3129 void 3130 regetblk(struct buf *bp) 3131 { 3132 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3133 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3134 bremfree(bp); 3135 } 3136 3137 /* 3138 * geteblk: 3139 * 3140 * Get an empty, disassociated buffer of given size. The buffer is 3141 * initially set to B_INVAL. 3142 * 3143 * critical section protection is not required for the allocbuf() 3144 * call because races are impossible here. 3145 * 3146 * MPALMOSTSAFE 3147 */ 3148 struct buf * 3149 geteblk(int size) 3150 { 3151 struct buf *bp; 3152 int maxsize; 3153 3154 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3155 3156 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 3157 ; 3158 allocbuf(bp, size); 3159 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3160 KKASSERT(dsched_is_clear_buf_priv(bp)); 3161 return (bp); 3162 } 3163 3164 3165 /* 3166 * allocbuf: 3167 * 3168 * This code constitutes the buffer memory from either anonymous system 3169 * memory (in the case of non-VMIO operations) or from an associated 3170 * VM object (in the case of VMIO operations). This code is able to 3171 * resize a buffer up or down. 3172 * 3173 * Note that this code is tricky, and has many complications to resolve 3174 * deadlock or inconsistant data situations. Tread lightly!!! 3175 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3176 * the caller. Calling this code willy nilly can result in the loss of 3177 * data. 3178 * 3179 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3180 * B_CACHE for the non-VMIO case. 3181 * 3182 * This routine does not need to be called from a critical section but you 3183 * must own the buffer. 3184 * 3185 * MPSAFE 3186 */ 3187 int 3188 allocbuf(struct buf *bp, int size) 3189 { 3190 int newbsize, mbsize; 3191 int i; 3192 3193 if (BUF_REFCNT(bp) == 0) 3194 panic("allocbuf: buffer not busy"); 3195 3196 if (bp->b_kvasize < size) 3197 panic("allocbuf: buffer too small"); 3198 3199 if ((bp->b_flags & B_VMIO) == 0) { 3200 caddr_t origbuf; 3201 int origbufsize; 3202 /* 3203 * Just get anonymous memory from the kernel. Don't 3204 * mess with B_CACHE. 3205 */ 3206 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3207 if (bp->b_flags & B_MALLOC) 3208 newbsize = mbsize; 3209 else 3210 newbsize = round_page(size); 3211 3212 if (newbsize < bp->b_bufsize) { 3213 /* 3214 * Malloced buffers are not shrunk 3215 */ 3216 if (bp->b_flags & B_MALLOC) { 3217 if (newbsize) { 3218 bp->b_bcount = size; 3219 } else { 3220 kfree(bp->b_data, M_BIOBUF); 3221 if (bp->b_bufsize) { 3222 atomic_subtract_int(&bufmallocspace, bp->b_bufsize); 3223 bufspacewakeup(); 3224 bp->b_bufsize = 0; 3225 } 3226 bp->b_data = bp->b_kvabase; 3227 bp->b_bcount = 0; 3228 bp->b_flags &= ~B_MALLOC; 3229 } 3230 return 1; 3231 } 3232 vm_hold_free_pages( 3233 bp, 3234 (vm_offset_t) bp->b_data + newbsize, 3235 (vm_offset_t) bp->b_data + bp->b_bufsize); 3236 } else if (newbsize > bp->b_bufsize) { 3237 /* 3238 * We only use malloced memory on the first allocation. 3239 * and revert to page-allocated memory when the buffer 3240 * grows. 3241 */ 3242 if ((bufmallocspace < maxbufmallocspace) && 3243 (bp->b_bufsize == 0) && 3244 (mbsize <= PAGE_SIZE/2)) { 3245 3246 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3247 bp->b_bufsize = mbsize; 3248 bp->b_bcount = size; 3249 bp->b_flags |= B_MALLOC; 3250 atomic_add_int(&bufmallocspace, mbsize); 3251 return 1; 3252 } 3253 origbuf = NULL; 3254 origbufsize = 0; 3255 /* 3256 * If the buffer is growing on its other-than-first 3257 * allocation, then we revert to the page-allocation 3258 * scheme. 3259 */ 3260 if (bp->b_flags & B_MALLOC) { 3261 origbuf = bp->b_data; 3262 origbufsize = bp->b_bufsize; 3263 bp->b_data = bp->b_kvabase; 3264 if (bp->b_bufsize) { 3265 atomic_subtract_int(&bufmallocspace, 3266 bp->b_bufsize); 3267 bufspacewakeup(); 3268 bp->b_bufsize = 0; 3269 } 3270 bp->b_flags &= ~B_MALLOC; 3271 newbsize = round_page(newbsize); 3272 } 3273 vm_hold_load_pages( 3274 bp, 3275 (vm_offset_t) bp->b_data + bp->b_bufsize, 3276 (vm_offset_t) bp->b_data + newbsize); 3277 if (origbuf) { 3278 bcopy(origbuf, bp->b_data, origbufsize); 3279 kfree(origbuf, M_BIOBUF); 3280 } 3281 } 3282 } else { 3283 vm_page_t m; 3284 int desiredpages; 3285 3286 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3287 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3288 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3289 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3290 3291 if (bp->b_flags & B_MALLOC) 3292 panic("allocbuf: VMIO buffer can't be malloced"); 3293 /* 3294 * Set B_CACHE initially if buffer is 0 length or will become 3295 * 0-length. 3296 */ 3297 if (size == 0 || bp->b_bufsize == 0) 3298 bp->b_flags |= B_CACHE; 3299 3300 if (newbsize < bp->b_bufsize) { 3301 /* 3302 * DEV_BSIZE aligned new buffer size is less then the 3303 * DEV_BSIZE aligned existing buffer size. Figure out 3304 * if we have to remove any pages. 3305 */ 3306 if (desiredpages < bp->b_xio.xio_npages) { 3307 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3308 /* 3309 * the page is not freed here -- it 3310 * is the responsibility of 3311 * vnode_pager_setsize 3312 */ 3313 m = bp->b_xio.xio_pages[i]; 3314 KASSERT(m != bogus_page, 3315 ("allocbuf: bogus page found")); 3316 while (vm_page_sleep_busy(m, TRUE, "biodep")) 3317 ; 3318 3319 bp->b_xio.xio_pages[i] = NULL; 3320 vm_page_unwire(m, 0); 3321 } 3322 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3323 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3324 bp->b_xio.xio_npages = desiredpages; 3325 } 3326 } else if (size > bp->b_bcount) { 3327 /* 3328 * We are growing the buffer, possibly in a 3329 * byte-granular fashion. 3330 */ 3331 struct vnode *vp; 3332 vm_object_t obj; 3333 vm_offset_t toff; 3334 vm_offset_t tinc; 3335 3336 /* 3337 * Step 1, bring in the VM pages from the object, 3338 * allocating them if necessary. We must clear 3339 * B_CACHE if these pages are not valid for the 3340 * range covered by the buffer. 3341 * 3342 * critical section protection is required to protect 3343 * against interrupts unbusying and freeing pages 3344 * between our vm_page_lookup() and our 3345 * busycheck/wiring call. 3346 */ 3347 vp = bp->b_vp; 3348 obj = vp->v_object; 3349 3350 lwkt_gettoken(&vm_token); 3351 while (bp->b_xio.xio_npages < desiredpages) { 3352 vm_page_t m; 3353 vm_pindex_t pi; 3354 3355 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages; 3356 if ((m = vm_page_lookup(obj, pi)) == NULL) { 3357 /* 3358 * note: must allocate system pages 3359 * since blocking here could intefere 3360 * with paging I/O, no matter which 3361 * process we are. 3362 */ 3363 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages); 3364 if (m) { 3365 vm_page_wire(m); 3366 vm_page_wakeup(m); 3367 vm_page_flag_clear(m, PG_ZERO); 3368 bp->b_flags &= ~B_CACHE; 3369 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3370 ++bp->b_xio.xio_npages; 3371 } 3372 continue; 3373 } 3374 3375 /* 3376 * We found a page. If we have to sleep on it, 3377 * retry because it might have gotten freed out 3378 * from under us. 3379 * 3380 * We can only test PG_BUSY here. Blocking on 3381 * m->busy might lead to a deadlock: 3382 * 3383 * vm_fault->getpages->cluster_read->allocbuf 3384 * 3385 */ 3386 3387 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 3388 continue; 3389 vm_page_flag_clear(m, PG_ZERO); 3390 vm_page_wire(m); 3391 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3392 ++bp->b_xio.xio_npages; 3393 if (bp->b_act_count < m->act_count) 3394 bp->b_act_count = m->act_count; 3395 } 3396 lwkt_reltoken(&vm_token); 3397 3398 /* 3399 * Step 2. We've loaded the pages into the buffer, 3400 * we have to figure out if we can still have B_CACHE 3401 * set. Note that B_CACHE is set according to the 3402 * byte-granular range ( bcount and size ), not the 3403 * aligned range ( newbsize ). 3404 * 3405 * The VM test is against m->valid, which is DEV_BSIZE 3406 * aligned. Needless to say, the validity of the data 3407 * needs to also be DEV_BSIZE aligned. Note that this 3408 * fails with NFS if the server or some other client 3409 * extends the file's EOF. If our buffer is resized, 3410 * B_CACHE may remain set! XXX 3411 */ 3412 3413 toff = bp->b_bcount; 3414 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3415 3416 while ((bp->b_flags & B_CACHE) && toff < size) { 3417 vm_pindex_t pi; 3418 3419 if (tinc > (size - toff)) 3420 tinc = size - toff; 3421 3422 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3423 PAGE_SHIFT; 3424 3425 vfs_buf_test_cache( 3426 bp, 3427 bp->b_loffset, 3428 toff, 3429 tinc, 3430 bp->b_xio.xio_pages[pi] 3431 ); 3432 toff += tinc; 3433 tinc = PAGE_SIZE; 3434 } 3435 3436 /* 3437 * Step 3, fixup the KVM pmap. Remember that 3438 * bp->b_data is relative to bp->b_loffset, but 3439 * bp->b_loffset may be offset into the first page. 3440 */ 3441 3442 bp->b_data = (caddr_t) 3443 trunc_page((vm_offset_t)bp->b_data); 3444 pmap_qenter( 3445 (vm_offset_t)bp->b_data, 3446 bp->b_xio.xio_pages, 3447 bp->b_xio.xio_npages 3448 ); 3449 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3450 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3451 } 3452 } 3453 3454 /* adjust space use on already-dirty buffer */ 3455 if (bp->b_flags & B_DELWRI) { 3456 spin_lock(&bufcspin); 3457 dirtybufspace += newbsize - bp->b_bufsize; 3458 if (bp->b_flags & B_HEAVY) 3459 dirtybufspacehw += newbsize - bp->b_bufsize; 3460 spin_unlock(&bufcspin); 3461 } 3462 if (newbsize < bp->b_bufsize) 3463 bufspacewakeup(); 3464 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3465 bp->b_bcount = size; /* requested buffer size */ 3466 return 1; 3467 } 3468 3469 /* 3470 * biowait: 3471 * 3472 * Wait for buffer I/O completion, returning error status. B_EINTR 3473 * is converted into an EINTR error but not cleared (since a chain 3474 * of biowait() calls may occur). 3475 * 3476 * On return bpdone() will have been called but the buffer will remain 3477 * locked and will not have been brelse()'d. 3478 * 3479 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3480 * likely still in progress on return. 3481 * 3482 * NOTE! This operation is on a BIO, not a BUF. 3483 * 3484 * NOTE! BIO_DONE is cleared by vn_strategy() 3485 * 3486 * MPSAFE 3487 */ 3488 static __inline int 3489 _biowait(struct bio *bio, const char *wmesg, int to) 3490 { 3491 struct buf *bp = bio->bio_buf; 3492 u_int32_t flags; 3493 u_int32_t nflags; 3494 int error; 3495 3496 KKASSERT(bio == &bp->b_bio1); 3497 for (;;) { 3498 flags = bio->bio_flags; 3499 if (flags & BIO_DONE) 3500 break; 3501 tsleep_interlock(bio, 0); 3502 nflags = flags | BIO_WANT; 3503 tsleep_interlock(bio, 0); 3504 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3505 if (wmesg) 3506 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3507 else if (bp->b_cmd == BUF_CMD_READ) 3508 error = tsleep(bio, PINTERLOCKED, "biord", to); 3509 else 3510 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3511 if (error) { 3512 kprintf("tsleep error biowait %d\n", error); 3513 return (error); 3514 } 3515 } 3516 } 3517 3518 /* 3519 * Finish up. 3520 */ 3521 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3522 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3523 if (bp->b_flags & B_EINTR) 3524 return (EINTR); 3525 if (bp->b_flags & B_ERROR) 3526 return (bp->b_error ? bp->b_error : EIO); 3527 return (0); 3528 } 3529 3530 int 3531 biowait(struct bio *bio, const char *wmesg) 3532 { 3533 return(_biowait(bio, wmesg, 0)); 3534 } 3535 3536 int 3537 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3538 { 3539 return(_biowait(bio, wmesg, to)); 3540 } 3541 3542 /* 3543 * This associates a tracking count with an I/O. vn_strategy() and 3544 * dev_dstrategy() do this automatically but there are a few cases 3545 * where a vnode or device layer is bypassed when a block translation 3546 * is cached. In such cases bio_start_transaction() may be called on 3547 * the bypassed layers so the system gets an I/O in progress indication 3548 * for those higher layers. 3549 */ 3550 void 3551 bio_start_transaction(struct bio *bio, struct bio_track *track) 3552 { 3553 bio->bio_track = track; 3554 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3555 dsched_new_buf(bio->bio_buf); 3556 bio_track_ref(track); 3557 } 3558 3559 /* 3560 * Initiate I/O on a vnode. 3561 * 3562 * SWAPCACHE OPERATION: 3563 * 3564 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3565 * devfs also uses b_vp for fake buffers so we also have to check 3566 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3567 * underlying block device. The swap assignments are related to the 3568 * buffer cache buffer's b_vp, not the passed vp. 3569 * 3570 * The passed vp == bp->b_vp only in the case where the strategy call 3571 * is made on the vp itself for its own buffers (a regular file or 3572 * block device vp). The filesystem usually then re-calls vn_strategy() 3573 * after translating the request to an underlying device. 3574 * 3575 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3576 * underlying buffer cache buffers. 3577 * 3578 * We can only deal with page-aligned buffers at the moment, because 3579 * we can't tell what the real dirty state for pages straddling a buffer 3580 * are. 3581 * 3582 * In order to call swap_pager_strategy() we must provide the VM object 3583 * and base offset for the underlying buffer cache pages so it can find 3584 * the swap blocks. 3585 */ 3586 void 3587 vn_strategy(struct vnode *vp, struct bio *bio) 3588 { 3589 struct bio_track *track; 3590 struct buf *bp = bio->bio_buf; 3591 3592 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3593 3594 /* 3595 * Set when an I/O is issued on the bp. Cleared by consumers 3596 * (aka HAMMER), allowing the consumer to determine if I/O had 3597 * actually occurred. 3598 */ 3599 bp->b_flags |= B_IODEBUG; 3600 3601 /* 3602 * Handle the swap cache intercept. 3603 */ 3604 if (vn_cache_strategy(vp, bio)) 3605 return; 3606 3607 /* 3608 * Otherwise do the operation through the filesystem 3609 */ 3610 if (bp->b_cmd == BUF_CMD_READ) 3611 track = &vp->v_track_read; 3612 else 3613 track = &vp->v_track_write; 3614 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3615 bio->bio_track = track; 3616 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3617 dsched_new_buf(bio->bio_buf); 3618 bio_track_ref(track); 3619 vop_strategy(*vp->v_ops, vp, bio); 3620 } 3621 3622 static void vn_cache_strategy_callback(struct bio *bio); 3623 3624 int 3625 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3626 { 3627 struct buf *bp = bio->bio_buf; 3628 struct bio *nbio; 3629 vm_object_t object; 3630 vm_page_t m; 3631 int i; 3632 3633 /* 3634 * Is this buffer cache buffer suitable for reading from 3635 * the swap cache? 3636 */ 3637 if (vm_swapcache_read_enable == 0 || 3638 bp->b_cmd != BUF_CMD_READ || 3639 ((bp->b_flags & B_CLUSTER) == 0 && 3640 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3641 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3642 (bp->b_bcount & PAGE_MASK) != 0) { 3643 return(0); 3644 } 3645 3646 /* 3647 * Figure out the original VM object (it will match the underlying 3648 * VM pages). Note that swap cached data uses page indices relative 3649 * to that object, not relative to bio->bio_offset. 3650 */ 3651 if (bp->b_flags & B_CLUSTER) 3652 object = vp->v_object; 3653 else 3654 object = bp->b_vp->v_object; 3655 3656 /* 3657 * In order to be able to use the swap cache all underlying VM 3658 * pages must be marked as such, and we can't have any bogus pages. 3659 */ 3660 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3661 m = bp->b_xio.xio_pages[i]; 3662 if ((m->flags & PG_SWAPPED) == 0) 3663 break; 3664 if (m == bogus_page) 3665 break; 3666 } 3667 3668 /* 3669 * If we are good then issue the I/O using swap_pager_strategy(). 3670 */ 3671 if (i == bp->b_xio.xio_npages) { 3672 m = bp->b_xio.xio_pages[0]; 3673 nbio = push_bio(bio); 3674 nbio->bio_done = vn_cache_strategy_callback; 3675 nbio->bio_offset = ptoa(m->pindex); 3676 KKASSERT(m->object == object); 3677 swap_pager_strategy(object, nbio); 3678 return(1); 3679 } 3680 return(0); 3681 } 3682 3683 /* 3684 * This is a bit of a hack but since the vn_cache_strategy() function can 3685 * override a VFS's strategy function we must make sure that the bio, which 3686 * is probably bio2, doesn't leak an unexpected offset value back to the 3687 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3688 * bio went through its own file strategy function and the the bio2 offset 3689 * is a cached disk offset when, in fact, it isn't. 3690 */ 3691 static void 3692 vn_cache_strategy_callback(struct bio *bio) 3693 { 3694 bio->bio_offset = NOOFFSET; 3695 biodone(pop_bio(bio)); 3696 } 3697 3698 /* 3699 * bpdone: 3700 * 3701 * Finish I/O on a buffer after all BIOs have been processed. 3702 * Called when the bio chain is exhausted or by biowait. If called 3703 * by biowait, elseit is typically 0. 3704 * 3705 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3706 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3707 * assuming B_INVAL is clear. 3708 * 3709 * For the VMIO case, we set B_CACHE if the op was a read and no 3710 * read error occured, or if the op was a write. B_CACHE is never 3711 * set if the buffer is invalid or otherwise uncacheable. 3712 * 3713 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3714 * initiator to leave B_INVAL set to brelse the buffer out of existance 3715 * in the biodone routine. 3716 */ 3717 void 3718 bpdone(struct buf *bp, int elseit) 3719 { 3720 buf_cmd_t cmd; 3721 3722 KASSERT(BUF_REFCNTNB(bp) > 0, 3723 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3724 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3725 ("biodone: bp %p already done!", bp)); 3726 3727 /* 3728 * No more BIOs are left. All completion functions have been dealt 3729 * with, now we clean up the buffer. 3730 */ 3731 cmd = bp->b_cmd; 3732 bp->b_cmd = BUF_CMD_DONE; 3733 3734 /* 3735 * Only reads and writes are processed past this point. 3736 */ 3737 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3738 if (cmd == BUF_CMD_FREEBLKS) 3739 bp->b_flags |= B_NOCACHE; 3740 if (elseit) 3741 brelse(bp); 3742 return; 3743 } 3744 3745 /* 3746 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3747 * a lot worse. XXX - move this above the clearing of b_cmd 3748 */ 3749 if (LIST_FIRST(&bp->b_dep) != NULL) 3750 buf_complete(bp); /* MPSAFE */ 3751 3752 /* 3753 * A failed write must re-dirty the buffer unless B_INVAL 3754 * was set. Only applicable to normal buffers (with VPs). 3755 * vinum buffers may not have a vp. 3756 */ 3757 if (cmd == BUF_CMD_WRITE && 3758 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3759 bp->b_flags &= ~B_NOCACHE; 3760 if (bp->b_vp) 3761 bdirty(bp); 3762 } 3763 3764 if (bp->b_flags & B_VMIO) { 3765 int i; 3766 vm_ooffset_t foff; 3767 vm_page_t m; 3768 vm_object_t obj; 3769 int iosize; 3770 struct vnode *vp = bp->b_vp; 3771 3772 obj = vp->v_object; 3773 3774 #if defined(VFS_BIO_DEBUG) 3775 if (vp->v_auxrefs == 0) 3776 panic("biodone: zero vnode hold count"); 3777 if ((vp->v_flag & VOBJBUF) == 0) 3778 panic("biodone: vnode is not setup for merged cache"); 3779 #endif 3780 3781 foff = bp->b_loffset; 3782 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 3783 KASSERT(obj != NULL, ("biodone: missing VM object")); 3784 3785 #if defined(VFS_BIO_DEBUG) 3786 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3787 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n", 3788 obj->paging_in_progress, bp->b_xio.xio_npages); 3789 } 3790 #endif 3791 3792 /* 3793 * Set B_CACHE if the op was a normal read and no error 3794 * occured. B_CACHE is set for writes in the b*write() 3795 * routines. 3796 */ 3797 iosize = bp->b_bcount - bp->b_resid; 3798 if (cmd == BUF_CMD_READ && 3799 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3800 bp->b_flags |= B_CACHE; 3801 } 3802 3803 lwkt_gettoken(&vm_token); 3804 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3805 int bogusflag = 0; 3806 int resid; 3807 3808 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3809 if (resid > iosize) 3810 resid = iosize; 3811 3812 /* 3813 * cleanup bogus pages, restoring the originals. Since 3814 * the originals should still be wired, we don't have 3815 * to worry about interrupt/freeing races destroying 3816 * the VM object association. 3817 */ 3818 m = bp->b_xio.xio_pages[i]; 3819 if (m == bogus_page) { 3820 bogusflag = 1; 3821 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3822 if (m == NULL) 3823 panic("biodone: page disappeared"); 3824 bp->b_xio.xio_pages[i] = m; 3825 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3826 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3827 } 3828 #if defined(VFS_BIO_DEBUG) 3829 if (OFF_TO_IDX(foff) != m->pindex) { 3830 kprintf("biodone: foff(%lu)/m->pindex(%ld) " 3831 "mismatch\n", 3832 (unsigned long)foff, (long)m->pindex); 3833 } 3834 #endif 3835 3836 /* 3837 * In the write case, the valid and clean bits are 3838 * already changed correctly (see bdwrite()), so we 3839 * only need to do this here in the read case. 3840 */ 3841 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3842 vfs_clean_one_page(bp, i, m); 3843 } 3844 vm_page_flag_clear(m, PG_ZERO); 3845 3846 /* 3847 * when debugging new filesystems or buffer I/O 3848 * methods, this is the most common error that pops 3849 * up. if you see this, you have not set the page 3850 * busy flag correctly!!! 3851 */ 3852 if (m->busy == 0) { 3853 kprintf("biodone: page busy < 0, " 3854 "pindex: %d, foff: 0x(%x,%x), " 3855 "resid: %d, index: %d\n", 3856 (int) m->pindex, (int)(foff >> 32), 3857 (int) foff & 0xffffffff, resid, i); 3858 if (!vn_isdisk(vp, NULL)) 3859 kprintf(" iosize: %ld, loffset: %lld, " 3860 "flags: 0x%08x, npages: %d\n", 3861 bp->b_vp->v_mount->mnt_stat.f_iosize, 3862 (long long)bp->b_loffset, 3863 bp->b_flags, bp->b_xio.xio_npages); 3864 else 3865 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3866 (long long)bp->b_loffset, 3867 bp->b_flags, bp->b_xio.xio_npages); 3868 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3869 m->valid, m->dirty, m->wire_count); 3870 panic("biodone: page busy < 0"); 3871 } 3872 vm_page_io_finish(m); 3873 vm_object_pip_subtract(obj, 1); 3874 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3875 iosize -= resid; 3876 } 3877 bp->b_flags &= ~B_HASBOGUS; 3878 if (obj) 3879 vm_object_pip_wakeupn(obj, 0); 3880 lwkt_reltoken(&vm_token); 3881 } 3882 3883 /* 3884 * Finish up by releasing the buffer. There are no more synchronous 3885 * or asynchronous completions, those were handled by bio_done 3886 * callbacks. 3887 */ 3888 if (elseit) { 3889 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3890 brelse(bp); 3891 else 3892 bqrelse(bp); 3893 } 3894 } 3895 3896 /* 3897 * Normal biodone. 3898 */ 3899 void 3900 biodone(struct bio *bio) 3901 { 3902 struct buf *bp = bio->bio_buf; 3903 3904 runningbufwakeup(bp); 3905 3906 /* 3907 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3908 */ 3909 while (bio) { 3910 biodone_t *done_func; 3911 struct bio_track *track; 3912 3913 /* 3914 * BIO tracking. Most but not all BIOs are tracked. 3915 */ 3916 if ((track = bio->bio_track) != NULL) { 3917 bio_track_rel(track); 3918 bio->bio_track = NULL; 3919 } 3920 3921 /* 3922 * A bio_done function terminates the loop. The function 3923 * will be responsible for any further chaining and/or 3924 * buffer management. 3925 * 3926 * WARNING! The done function can deallocate the buffer! 3927 */ 3928 if ((done_func = bio->bio_done) != NULL) { 3929 bio->bio_done = NULL; 3930 done_func(bio); 3931 return; 3932 } 3933 bio = bio->bio_prev; 3934 } 3935 3936 /* 3937 * If we've run out of bio's do normal [a]synchronous completion. 3938 */ 3939 bpdone(bp, 1); 3940 } 3941 3942 /* 3943 * Synchronous biodone - this terminates a synchronous BIO. 3944 * 3945 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3946 * but still locked. The caller must brelse() the buffer after waiting 3947 * for completion. 3948 */ 3949 void 3950 biodone_sync(struct bio *bio) 3951 { 3952 struct buf *bp = bio->bio_buf; 3953 int flags; 3954 int nflags; 3955 3956 KKASSERT(bio == &bp->b_bio1); 3957 bpdone(bp, 0); 3958 3959 for (;;) { 3960 flags = bio->bio_flags; 3961 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3962 3963 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3964 if (flags & BIO_WANT) 3965 wakeup(bio); 3966 break; 3967 } 3968 } 3969 } 3970 3971 /* 3972 * vfs_unbusy_pages: 3973 * 3974 * This routine is called in lieu of iodone in the case of 3975 * incomplete I/O. This keeps the busy status for pages 3976 * consistant. 3977 */ 3978 void 3979 vfs_unbusy_pages(struct buf *bp) 3980 { 3981 int i; 3982 3983 runningbufwakeup(bp); 3984 3985 lwkt_gettoken(&vm_token); 3986 if (bp->b_flags & B_VMIO) { 3987 struct vnode *vp = bp->b_vp; 3988 vm_object_t obj; 3989 3990 obj = vp->v_object; 3991 3992 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3993 vm_page_t m = bp->b_xio.xio_pages[i]; 3994 3995 /* 3996 * When restoring bogus changes the original pages 3997 * should still be wired, so we are in no danger of 3998 * losing the object association and do not need 3999 * critical section protection particularly. 4000 */ 4001 if (m == bogus_page) { 4002 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 4003 if (!m) { 4004 panic("vfs_unbusy_pages: page missing"); 4005 } 4006 bp->b_xio.xio_pages[i] = m; 4007 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4008 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4009 } 4010 vm_object_pip_subtract(obj, 1); 4011 vm_page_flag_clear(m, PG_ZERO); 4012 vm_page_io_finish(m); 4013 } 4014 bp->b_flags &= ~B_HASBOGUS; 4015 vm_object_pip_wakeupn(obj, 0); 4016 } 4017 lwkt_reltoken(&vm_token); 4018 } 4019 4020 /* 4021 * vfs_busy_pages: 4022 * 4023 * This routine is called before a device strategy routine. 4024 * It is used to tell the VM system that paging I/O is in 4025 * progress, and treat the pages associated with the buffer 4026 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4027 * flag is handled to make sure that the object doesn't become 4028 * inconsistant. 4029 * 4030 * Since I/O has not been initiated yet, certain buffer flags 4031 * such as B_ERROR or B_INVAL may be in an inconsistant state 4032 * and should be ignored. 4033 * 4034 * MPSAFE 4035 */ 4036 void 4037 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4038 { 4039 int i, bogus; 4040 struct lwp *lp = curthread->td_lwp; 4041 4042 /* 4043 * The buffer's I/O command must already be set. If reading, 4044 * B_CACHE must be 0 (double check against callers only doing 4045 * I/O when B_CACHE is 0). 4046 */ 4047 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4048 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4049 4050 if (bp->b_flags & B_VMIO) { 4051 vm_object_t obj; 4052 4053 lwkt_gettoken(&vm_token); 4054 4055 obj = vp->v_object; 4056 KASSERT(bp->b_loffset != NOOFFSET, 4057 ("vfs_busy_pages: no buffer offset")); 4058 4059 /* 4060 * Loop until none of the pages are busy. 4061 */ 4062 retry: 4063 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4064 vm_page_t m = bp->b_xio.xio_pages[i]; 4065 4066 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 4067 goto retry; 4068 } 4069 4070 /* 4071 * Setup for I/O, soft-busy the page right now because 4072 * the next loop may block. 4073 */ 4074 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4075 vm_page_t m = bp->b_xio.xio_pages[i]; 4076 4077 vm_page_flag_clear(m, PG_ZERO); 4078 if ((bp->b_flags & B_CLUSTER) == 0) { 4079 vm_object_pip_add(obj, 1); 4080 vm_page_io_start(m); 4081 } 4082 } 4083 4084 /* 4085 * Adjust protections for I/O and do bogus-page mapping. 4086 * Assume that vm_page_protect() can block (it can block 4087 * if VM_PROT_NONE, don't take any chances regardless). 4088 * 4089 * In particular note that for writes we must incorporate 4090 * page dirtyness from the VM system into the buffer's 4091 * dirty range. 4092 * 4093 * For reads we theoretically must incorporate page dirtyness 4094 * from the VM system to determine if the page needs bogus 4095 * replacement, but we shortcut the test by simply checking 4096 * that all m->valid bits are set, indicating that the page 4097 * is fully valid and does not need to be re-read. For any 4098 * VM system dirtyness the page will also be fully valid 4099 * since it was mapped at one point. 4100 */ 4101 bogus = 0; 4102 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4103 vm_page_t m = bp->b_xio.xio_pages[i]; 4104 4105 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4106 if (bp->b_cmd == BUF_CMD_WRITE) { 4107 /* 4108 * When readying a vnode-backed buffer for 4109 * a write we must zero-fill any invalid 4110 * portions of the backing VM pages, mark 4111 * it valid and clear related dirty bits. 4112 * 4113 * vfs_clean_one_page() incorporates any 4114 * VM dirtyness and updates the b_dirtyoff 4115 * range (after we've made the page RO). 4116 * 4117 * It is also expected that the pmap modified 4118 * bit has already been cleared by the 4119 * vm_page_protect(). We may not be able 4120 * to clear all dirty bits for a page if it 4121 * was also memory mapped (NFS). 4122 * 4123 * Finally be sure to unassign any swap-cache 4124 * backing store as it is now stale. 4125 */ 4126 vm_page_protect(m, VM_PROT_READ); 4127 vfs_clean_one_page(bp, i, m); 4128 swap_pager_unswapped(m); 4129 } else if (m->valid == VM_PAGE_BITS_ALL) { 4130 /* 4131 * When readying a vnode-backed buffer for 4132 * read we must replace any dirty pages with 4133 * a bogus page so dirty data is not destroyed 4134 * when filling gaps. 4135 * 4136 * To avoid testing whether the page is 4137 * dirty we instead test that the page was 4138 * at some point mapped (m->valid fully 4139 * valid) with the understanding that 4140 * this also covers the dirty case. 4141 */ 4142 bp->b_xio.xio_pages[i] = bogus_page; 4143 bp->b_flags |= B_HASBOGUS; 4144 bogus++; 4145 } else if (m->valid & m->dirty) { 4146 /* 4147 * This case should not occur as partial 4148 * dirtyment can only happen if the buffer 4149 * is B_CACHE, and this code is not entered 4150 * if the buffer is B_CACHE. 4151 */ 4152 kprintf("Warning: vfs_busy_pages - page not " 4153 "fully valid! loff=%jx bpf=%08x " 4154 "idx=%d val=%02x dir=%02x\n", 4155 (intmax_t)bp->b_loffset, bp->b_flags, 4156 i, m->valid, m->dirty); 4157 vm_page_protect(m, VM_PROT_NONE); 4158 } else { 4159 /* 4160 * The page is not valid and can be made 4161 * part of the read. 4162 */ 4163 vm_page_protect(m, VM_PROT_NONE); 4164 } 4165 } 4166 if (bogus) { 4167 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4168 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4169 } 4170 lwkt_reltoken(&vm_token); 4171 } 4172 4173 /* 4174 * This is the easiest place to put the process accounting for the I/O 4175 * for now. 4176 */ 4177 if (lp != NULL) { 4178 if (bp->b_cmd == BUF_CMD_READ) 4179 lp->lwp_ru.ru_inblock++; 4180 else 4181 lp->lwp_ru.ru_oublock++; 4182 } 4183 } 4184 4185 /* 4186 * Tell the VM system that the pages associated with this buffer 4187 * are clean. This is used for delayed writes where the data is 4188 * going to go to disk eventually without additional VM intevention. 4189 * 4190 * NOTE: While we only really need to clean through to b_bcount, we 4191 * just go ahead and clean through to b_bufsize. 4192 */ 4193 static void 4194 vfs_clean_pages(struct buf *bp) 4195 { 4196 vm_page_t m; 4197 int i; 4198 4199 if ((bp->b_flags & B_VMIO) == 0) 4200 return; 4201 4202 KASSERT(bp->b_loffset != NOOFFSET, 4203 ("vfs_clean_pages: no buffer offset")); 4204 4205 /* 4206 * vm_token must be held for vfs_clean_one_page() calls. 4207 */ 4208 lwkt_gettoken(&vm_token); 4209 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4210 m = bp->b_xio.xio_pages[i]; 4211 vfs_clean_one_page(bp, i, m); 4212 } 4213 lwkt_reltoken(&vm_token); 4214 } 4215 4216 /* 4217 * vfs_clean_one_page: 4218 * 4219 * Set the valid bits and clear the dirty bits in a page within a 4220 * buffer. The range is restricted to the buffer's size and the 4221 * buffer's logical offset might index into the first page. 4222 * 4223 * The caller has busied or soft-busied the page and it is not mapped, 4224 * test and incorporate the dirty bits into b_dirtyoff/end before 4225 * clearing them. Note that we need to clear the pmap modified bits 4226 * after determining the the page was dirty, vm_page_set_validclean() 4227 * does not do it for us. 4228 * 4229 * This routine is typically called after a read completes (dirty should 4230 * be zero in that case as we are not called on bogus-replace pages), 4231 * or before a write is initiated. 4232 * 4233 * NOTE: vm_token must be held by the caller, and vm_page_set_validclean() 4234 * currently assumes the vm_token is held. 4235 */ 4236 static void 4237 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4238 { 4239 int bcount; 4240 int xoff; 4241 int soff; 4242 int eoff; 4243 4244 /* 4245 * Calculate offset range within the page but relative to buffer's 4246 * loffset. loffset might be offset into the first page. 4247 */ 4248 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4249 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4250 4251 if (pageno == 0) { 4252 soff = xoff; 4253 eoff = PAGE_SIZE; 4254 } else { 4255 soff = (pageno << PAGE_SHIFT); 4256 eoff = soff + PAGE_SIZE; 4257 } 4258 if (eoff > bcount) 4259 eoff = bcount; 4260 if (soff >= eoff) 4261 return; 4262 4263 /* 4264 * Test dirty bits and adjust b_dirtyoff/end. 4265 * 4266 * If dirty pages are incorporated into the bp any prior 4267 * B_NEEDCOMMIT state (NFS) must be cleared because the 4268 * caller has not taken into account the new dirty data. 4269 * 4270 * If the page was memory mapped the dirty bits might go beyond the 4271 * end of the buffer, but we can't really make the assumption that 4272 * a file EOF straddles the buffer (even though this is the case for 4273 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4274 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4275 * This also saves some console spam. 4276 * 4277 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4278 * NFS can handle huge commits but not huge writes. 4279 */ 4280 vm_page_test_dirty(m); 4281 if (m->dirty) { 4282 if ((bp->b_flags & B_NEEDCOMMIT) && 4283 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4284 if (debug_commit) 4285 kprintf("Warning: vfs_clean_one_page: bp %p " 4286 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4287 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4288 "doff/end %d %d\n", 4289 bp, (intmax_t)bp->b_loffset, bp->b_bcount, 4290 bp->b_flags, bp->b_cmd, 4291 m->valid, m->dirty, xoff, soff, eoff, 4292 bp->b_dirtyoff, bp->b_dirtyend); 4293 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4294 if (debug_commit) 4295 print_backtrace(-1); 4296 } 4297 /* 4298 * Only clear the pmap modified bits if ALL the dirty bits 4299 * are set, otherwise the system might mis-clear portions 4300 * of a page. 4301 */ 4302 if (m->dirty == VM_PAGE_BITS_ALL && 4303 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4304 pmap_clear_modify(m); 4305 } 4306 if (bp->b_dirtyoff > soff - xoff) 4307 bp->b_dirtyoff = soff - xoff; 4308 if (bp->b_dirtyend < eoff - xoff) 4309 bp->b_dirtyend = eoff - xoff; 4310 } 4311 4312 /* 4313 * Set related valid bits, clear related dirty bits. 4314 * Does not mess with the pmap modified bit. 4315 * 4316 * WARNING! We cannot just clear all of m->dirty here as the 4317 * buffer cache buffers may use a DEV_BSIZE'd aligned 4318 * block size, or have an odd size (e.g. NFS at file EOF). 4319 * The putpages code can clear m->dirty to 0. 4320 * 4321 * If a VOP_WRITE generates a buffer cache buffer which 4322 * covers the same space as mapped writable pages the 4323 * buffer flush might not be able to clear all the dirty 4324 * bits and still require a putpages from the VM system 4325 * to finish it off. 4326 * 4327 * WARNING! vm_page_set_validclean() currently assumes vm_token 4328 * is held. The page might not be busied (bdwrite() case). 4329 */ 4330 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4331 } 4332 4333 /* 4334 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4335 * The page data is assumed to be valid (there is no zeroing here). 4336 */ 4337 static void 4338 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4339 { 4340 int bcount; 4341 int xoff; 4342 int soff; 4343 int eoff; 4344 4345 /* 4346 * Calculate offset range within the page but relative to buffer's 4347 * loffset. loffset might be offset into the first page. 4348 */ 4349 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4350 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4351 4352 if (pageno == 0) { 4353 soff = xoff; 4354 eoff = PAGE_SIZE; 4355 } else { 4356 soff = (pageno << PAGE_SHIFT); 4357 eoff = soff + PAGE_SIZE; 4358 } 4359 if (eoff > bcount) 4360 eoff = bcount; 4361 if (soff >= eoff) 4362 return; 4363 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4364 } 4365 4366 /* 4367 * vfs_bio_clrbuf: 4368 * 4369 * Clear a buffer. This routine essentially fakes an I/O, so we need 4370 * to clear B_ERROR and B_INVAL. 4371 * 4372 * Note that while we only theoretically need to clear through b_bcount, 4373 * we go ahead and clear through b_bufsize. 4374 */ 4375 4376 void 4377 vfs_bio_clrbuf(struct buf *bp) 4378 { 4379 int i, mask = 0; 4380 caddr_t sa, ea; 4381 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4382 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4383 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4384 (bp->b_loffset & PAGE_MASK) == 0) { 4385 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4386 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4387 bp->b_resid = 0; 4388 return; 4389 } 4390 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4391 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4392 bzero(bp->b_data, bp->b_bufsize); 4393 bp->b_xio.xio_pages[0]->valid |= mask; 4394 bp->b_resid = 0; 4395 return; 4396 } 4397 } 4398 sa = bp->b_data; 4399 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4400 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4401 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4402 ea = (caddr_t)(vm_offset_t)ulmin( 4403 (u_long)(vm_offset_t)ea, 4404 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4405 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4406 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4407 continue; 4408 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4409 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4410 bzero(sa, ea - sa); 4411 } 4412 } else { 4413 for (; sa < ea; sa += DEV_BSIZE, j++) { 4414 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4415 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4416 bzero(sa, DEV_BSIZE); 4417 } 4418 } 4419 bp->b_xio.xio_pages[i]->valid |= mask; 4420 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4421 } 4422 bp->b_resid = 0; 4423 } else { 4424 clrbuf(bp); 4425 } 4426 } 4427 4428 /* 4429 * vm_hold_load_pages: 4430 * 4431 * Load pages into the buffer's address space. The pages are 4432 * allocated from the kernel object in order to reduce interference 4433 * with the any VM paging I/O activity. The range of loaded 4434 * pages will be wired. 4435 * 4436 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4437 * retrieve the full range (to - from) of pages. 4438 * 4439 * MPSAFE 4440 */ 4441 void 4442 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4443 { 4444 vm_offset_t pg; 4445 vm_page_t p; 4446 int index; 4447 4448 to = round_page(to); 4449 from = round_page(from); 4450 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4451 4452 pg = from; 4453 while (pg < to) { 4454 /* 4455 * Note: must allocate system pages since blocking here 4456 * could intefere with paging I/O, no matter which 4457 * process we are. 4458 */ 4459 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT, 4460 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4461 if (p) { 4462 vm_page_wire(p); 4463 p->valid = VM_PAGE_BITS_ALL; 4464 vm_page_flag_clear(p, PG_ZERO); 4465 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 4466 bp->b_xio.xio_pages[index] = p; 4467 vm_page_wakeup(p); 4468 4469 pg += PAGE_SIZE; 4470 ++index; 4471 } 4472 } 4473 bp->b_xio.xio_npages = index; 4474 } 4475 4476 /* 4477 * Allocate pages for a buffer cache buffer. 4478 * 4479 * Under extremely severe memory conditions even allocating out of the 4480 * system reserve can fail. If this occurs we must allocate out of the 4481 * interrupt reserve to avoid a deadlock with the pageout daemon. 4482 * 4483 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf). 4484 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock 4485 * against the pageout daemon if pages are not freed from other sources. 4486 * 4487 * MPSAFE 4488 */ 4489 static 4490 vm_page_t 4491 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit) 4492 { 4493 vm_page_t p; 4494 4495 /* 4496 * Try a normal allocation, allow use of system reserve. 4497 */ 4498 lwkt_gettoken(&vm_token); 4499 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 4500 if (p) { 4501 lwkt_reltoken(&vm_token); 4502 return(p); 4503 } 4504 4505 /* 4506 * The normal allocation failed and we clearly have a page 4507 * deficit. Try to reclaim some clean VM pages directly 4508 * from the buffer cache. 4509 */ 4510 vm_pageout_deficit += deficit; 4511 recoverbufpages(); 4512 4513 /* 4514 * We may have blocked, the caller will know what to do if the 4515 * page now exists. 4516 */ 4517 if (vm_page_lookup(obj, pg)) { 4518 lwkt_reltoken(&vm_token); 4519 return(NULL); 4520 } 4521 4522 /* 4523 * Allocate and allow use of the interrupt reserve. 4524 * 4525 * If after all that we still can't allocate a VM page we are 4526 * in real trouble, but we slog on anyway hoping that the system 4527 * won't deadlock. 4528 */ 4529 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 4530 VM_ALLOC_INTERRUPT); 4531 if (p) { 4532 if (vm_page_count_severe()) { 4533 ++lowmempgallocs; 4534 vm_wait(hz / 20 + 1); 4535 } 4536 } else { 4537 kprintf("bio_page_alloc: Memory exhausted during bufcache " 4538 "page allocation\n"); 4539 ++lowmempgfails; 4540 vm_wait(hz); 4541 } 4542 lwkt_reltoken(&vm_token); 4543 return(p); 4544 } 4545 4546 /* 4547 * vm_hold_free_pages: 4548 * 4549 * Return pages associated with the buffer back to the VM system. 4550 * 4551 * The range of pages underlying the buffer's address space will 4552 * be unmapped and un-wired. 4553 * 4554 * MPSAFE 4555 */ 4556 void 4557 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4558 { 4559 vm_offset_t pg; 4560 vm_page_t p; 4561 int index, newnpages; 4562 4563 from = round_page(from); 4564 to = round_page(to); 4565 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4566 newnpages = index; 4567 4568 lwkt_gettoken(&vm_token); 4569 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4570 p = bp->b_xio.xio_pages[index]; 4571 if (p && (index < bp->b_xio.xio_npages)) { 4572 if (p->busy) { 4573 kprintf("vm_hold_free_pages: doffset: %lld, " 4574 "loffset: %lld\n", 4575 (long long)bp->b_bio2.bio_offset, 4576 (long long)bp->b_loffset); 4577 } 4578 bp->b_xio.xio_pages[index] = NULL; 4579 pmap_kremove(pg); 4580 vm_page_busy(p); 4581 vm_page_unwire(p, 0); 4582 vm_page_free(p); 4583 } 4584 } 4585 bp->b_xio.xio_npages = newnpages; 4586 lwkt_reltoken(&vm_token); 4587 } 4588 4589 /* 4590 * vmapbuf: 4591 * 4592 * Map a user buffer into KVM via a pbuf. On return the buffer's 4593 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4594 * initialized. 4595 */ 4596 int 4597 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4598 { 4599 caddr_t addr; 4600 vm_offset_t va; 4601 vm_page_t m; 4602 int vmprot; 4603 int error; 4604 int pidx; 4605 int i; 4606 4607 /* 4608 * bp had better have a command and it better be a pbuf. 4609 */ 4610 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4611 KKASSERT(bp->b_flags & B_PAGING); 4612 KKASSERT(bp->b_kvabase); 4613 4614 if (bytes < 0) 4615 return (-1); 4616 4617 /* 4618 * Map the user data into KVM. Mappings have to be page-aligned. 4619 */ 4620 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4621 pidx = 0; 4622 4623 vmprot = VM_PROT_READ; 4624 if (bp->b_cmd == BUF_CMD_READ) 4625 vmprot |= VM_PROT_WRITE; 4626 4627 while (addr < udata + bytes) { 4628 /* 4629 * Do the vm_fault if needed; do the copy-on-write thing 4630 * when reading stuff off device into memory. 4631 * 4632 * vm_fault_page*() returns a held VM page. 4633 */ 4634 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4635 va = trunc_page(va); 4636 4637 m = vm_fault_page_quick(va, vmprot, &error); 4638 if (m == NULL) { 4639 for (i = 0; i < pidx; ++i) { 4640 vm_page_unhold(bp->b_xio.xio_pages[i]); 4641 bp->b_xio.xio_pages[i] = NULL; 4642 } 4643 return(-1); 4644 } 4645 bp->b_xio.xio_pages[pidx] = m; 4646 addr += PAGE_SIZE; 4647 ++pidx; 4648 } 4649 4650 /* 4651 * Map the page array and set the buffer fields to point to 4652 * the mapped data buffer. 4653 */ 4654 if (pidx > btoc(MAXPHYS)) 4655 panic("vmapbuf: mapped more than MAXPHYS"); 4656 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4657 4658 bp->b_xio.xio_npages = pidx; 4659 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4660 bp->b_bcount = bytes; 4661 bp->b_bufsize = bytes; 4662 return(0); 4663 } 4664 4665 /* 4666 * vunmapbuf: 4667 * 4668 * Free the io map PTEs associated with this IO operation. 4669 * We also invalidate the TLB entries and restore the original b_addr. 4670 */ 4671 void 4672 vunmapbuf(struct buf *bp) 4673 { 4674 int pidx; 4675 int npages; 4676 4677 KKASSERT(bp->b_flags & B_PAGING); 4678 4679 npages = bp->b_xio.xio_npages; 4680 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4681 for (pidx = 0; pidx < npages; ++pidx) { 4682 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4683 bp->b_xio.xio_pages[pidx] = NULL; 4684 } 4685 bp->b_xio.xio_npages = 0; 4686 bp->b_data = bp->b_kvabase; 4687 } 4688 4689 /* 4690 * Scan all buffers in the system and issue the callback. 4691 */ 4692 int 4693 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4694 { 4695 int count = 0; 4696 int error; 4697 int n; 4698 4699 for (n = 0; n < nbuf; ++n) { 4700 if ((error = callback(&buf[n], info)) < 0) { 4701 count = error; 4702 break; 4703 } 4704 count += error; 4705 } 4706 return (count); 4707 } 4708 4709 /* 4710 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4711 * completion to the master buffer. 4712 */ 4713 static void 4714 nestiobuf_iodone(struct bio *bio) 4715 { 4716 struct bio *mbio; 4717 struct buf *mbp, *bp; 4718 struct devstat *stats; 4719 int error; 4720 int donebytes; 4721 4722 bp = bio->bio_buf; 4723 mbio = bio->bio_caller_info1.ptr; 4724 stats = bio->bio_caller_info2.ptr; 4725 mbp = mbio->bio_buf; 4726 4727 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4728 KKASSERT(mbp != bp); 4729 4730 error = bp->b_error; 4731 if (bp->b_error == 0 && 4732 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4733 /* 4734 * Not all got transfered, raise an error. We have no way to 4735 * propagate these conditions to mbp. 4736 */ 4737 error = EIO; 4738 } 4739 4740 donebytes = bp->b_bufsize; 4741 4742 relpbuf(bp, NULL); 4743 4744 nestiobuf_done(mbio, donebytes, error, stats); 4745 } 4746 4747 void 4748 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4749 { 4750 struct buf *mbp; 4751 4752 mbp = mbio->bio_buf; 4753 4754 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4755 4756 /* 4757 * If an error occured, propagate it to the master buffer. 4758 * 4759 * Several biodone()s may wind up running concurrently so 4760 * use an atomic op to adjust b_flags. 4761 */ 4762 if (error) { 4763 mbp->b_error = error; 4764 atomic_set_int(&mbp->b_flags, B_ERROR); 4765 } 4766 4767 /* 4768 * Decrement the operations in progress counter and terminate the 4769 * I/O if this was the last bit. 4770 */ 4771 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4772 mbp->b_resid = 0; 4773 if (stats) 4774 devstat_end_transaction_buf(stats, mbp); 4775 biodone(mbio); 4776 } 4777 } 4778 4779 /* 4780 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4781 * the mbio from being biodone()'d while we are still adding sub-bios to 4782 * it. 4783 */ 4784 void 4785 nestiobuf_init(struct bio *bio) 4786 { 4787 bio->bio_driver_info = (void *)1; 4788 } 4789 4790 /* 4791 * The BIOs added to the nestedio have already been started, remove the 4792 * count that placeheld our mbio and biodone() it if the count would 4793 * transition to 0. 4794 */ 4795 void 4796 nestiobuf_start(struct bio *mbio) 4797 { 4798 struct buf *mbp = mbio->bio_buf; 4799 4800 /* 4801 * Decrement the operations in progress counter and terminate the 4802 * I/O if this was the last bit. 4803 */ 4804 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4805 if (mbp->b_flags & B_ERROR) 4806 mbp->b_resid = mbp->b_bcount; 4807 else 4808 mbp->b_resid = 0; 4809 biodone(mbio); 4810 } 4811 } 4812 4813 /* 4814 * Set an intermediate error prior to calling nestiobuf_start() 4815 */ 4816 void 4817 nestiobuf_error(struct bio *mbio, int error) 4818 { 4819 struct buf *mbp = mbio->bio_buf; 4820 4821 if (error) { 4822 mbp->b_error = error; 4823 atomic_set_int(&mbp->b_flags, B_ERROR); 4824 } 4825 } 4826 4827 /* 4828 * nestiobuf_add: setup a "nested" buffer. 4829 * 4830 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4831 * => 'bp' should be a buffer allocated by getiobuf. 4832 * => 'offset' is a byte offset in the master buffer. 4833 * => 'size' is a size in bytes of this nested buffer. 4834 */ 4835 void 4836 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4837 { 4838 struct buf *mbp = mbio->bio_buf; 4839 struct vnode *vp = mbp->b_vp; 4840 4841 KKASSERT(mbp->b_bcount >= offset + size); 4842 4843 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4844 4845 /* kernel needs to own the lock for it to be released in biodone */ 4846 BUF_KERNPROC(bp); 4847 bp->b_vp = vp; 4848 bp->b_cmd = mbp->b_cmd; 4849 bp->b_bio1.bio_done = nestiobuf_iodone; 4850 bp->b_data = (char *)mbp->b_data + offset; 4851 bp->b_resid = bp->b_bcount = size; 4852 bp->b_bufsize = bp->b_bcount; 4853 4854 bp->b_bio1.bio_track = NULL; 4855 bp->b_bio1.bio_caller_info1.ptr = mbio; 4856 bp->b_bio1.bio_caller_info2.ptr = stats; 4857 } 4858 4859 /* 4860 * print out statistics from the current status of the buffer pool 4861 * this can be toggeled by the system control option debug.syncprt 4862 */ 4863 #ifdef DEBUG 4864 void 4865 vfs_bufstats(void) 4866 { 4867 int i, j, count; 4868 struct buf *bp; 4869 struct bqueues *dp; 4870 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 4871 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 4872 4873 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 4874 count = 0; 4875 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4876 counts[j] = 0; 4877 4878 spin_lock(&bufqspin); 4879 TAILQ_FOREACH(bp, dp, b_freelist) { 4880 counts[bp->b_bufsize/PAGE_SIZE]++; 4881 count++; 4882 } 4883 spin_unlock(&bufqspin); 4884 4885 kprintf("%s: total-%d", bname[i], count); 4886 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4887 if (counts[j] != 0) 4888 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]); 4889 kprintf("\n"); 4890 } 4891 } 4892 #endif 4893 4894 #ifdef DDB 4895 4896 DB_SHOW_COMMAND(buffer, db_show_buffer) 4897 { 4898 /* get args */ 4899 struct buf *bp = (struct buf *)addr; 4900 4901 if (!have_addr) { 4902 db_printf("usage: show buffer <addr>\n"); 4903 return; 4904 } 4905 4906 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4907 db_printf("b_cmd = %d\n", bp->b_cmd); 4908 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4909 "b_resid = %d\n, b_data = %p, " 4910 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4911 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4912 bp->b_data, 4913 (long long)bp->b_bio2.bio_offset, 4914 (long long)(bp->b_bio2.bio_next ? 4915 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4916 if (bp->b_xio.xio_npages) { 4917 int i; 4918 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4919 bp->b_xio.xio_npages); 4920 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4921 vm_page_t m; 4922 m = bp->b_xio.xio_pages[i]; 4923 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4924 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4925 if ((i + 1) < bp->b_xio.xio_npages) 4926 db_printf(","); 4927 } 4928 db_printf("\n"); 4929 } 4930 } 4931 #endif /* DDB */ 4932