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