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 */ 16 17 /* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30 #include <sys/param.h> 31 #include <sys/systm.h> 32 #include <sys/buf.h> 33 #include <sys/conf.h> 34 #include <sys/devicestat.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/dsched.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 #include <vm/vm_pager.h> 57 #include <vm/swap_pager.h> 58 59 #include <sys/buf2.h> 60 #include <sys/thread2.h> 61 #include <sys/spinlock2.h> 62 #include <sys/mplock2.h> 63 #include <vm/vm_page2.h> 64 65 #include "opt_ddb.h" 66 #ifdef DDB 67 #include <ddb/ddb.h> 68 #endif 69 70 /* 71 * Buffer queues. 72 */ 73 enum bufq_type { 74 BQUEUE_NONE, /* not on any queue */ 75 BQUEUE_LOCKED, /* locked buffers */ 76 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 77 BQUEUE_DIRTY, /* B_DELWRI buffers */ 78 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */ 79 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 80 BQUEUE_EMPTY, /* empty buffer headers */ 81 82 BUFFER_QUEUES /* number of buffer queues */ 83 }; 84 85 typedef enum bufq_type bufq_type_t; 86 87 #define BD_WAKE_SIZE 16384 88 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1) 89 90 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 91 static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin); 92 static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin); 93 94 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 95 96 struct buf *buf; /* buffer header pool */ 97 98 static void vfs_clean_pages(struct buf *bp); 99 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m); 100 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m); 101 static void vfs_vmio_release(struct buf *bp); 102 static int flushbufqueues(bufq_type_t q); 103 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit); 104 105 static void bd_signal(int totalspace); 106 static void buf_daemon(void); 107 static void buf_daemon_hw(void); 108 109 /* 110 * bogus page -- for I/O to/from partially complete buffers 111 * this is a temporary solution to the problem, but it is not 112 * really that bad. it would be better to split the buffer 113 * for input in the case of buffers partially already in memory, 114 * but the code is intricate enough already. 115 */ 116 vm_page_t bogus_page; 117 118 /* 119 * These are all static, but make the ones we export globals so we do 120 * not need to use compiler magic. 121 */ 122 int bufspace; /* locked by buffer_map */ 123 int maxbufspace; 124 static int bufmallocspace; /* atomic ops */ 125 int maxbufmallocspace, lobufspace, hibufspace; 126 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 127 static int lorunningspace; 128 static int hirunningspace; 129 static int runningbufreq; /* locked by bufcspin */ 130 static int dirtybufspace; /* locked by bufcspin */ 131 static int dirtybufcount; /* locked by bufcspin */ 132 static int dirtybufspacehw; /* locked by bufcspin */ 133 static int dirtybufcounthw; /* locked by bufcspin */ 134 static int runningbufspace; /* locked by bufcspin */ 135 static int runningbufcount; /* locked by bufcspin */ 136 int lodirtybufspace; 137 int hidirtybufspace; 138 static int getnewbufcalls; 139 static int getnewbufrestarts; 140 static int recoverbufcalls; 141 static int needsbuffer; /* locked by bufcspin */ 142 static int bd_request; /* locked by bufcspin */ 143 static int bd_request_hw; /* locked by bufcspin */ 144 static u_int bd_wake_ary[BD_WAKE_SIZE]; 145 static u_int bd_wake_index; 146 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 147 static int debug_commit; 148 149 static struct thread *bufdaemon_td; 150 static struct thread *bufdaemonhw_td; 151 static u_int lowmempgallocs; 152 static u_int lowmempgfails; 153 154 /* 155 * Sysctls for operational control of the buffer cache. 156 */ 157 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 158 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 159 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 160 "High watermark used to trigger explicit flushing of dirty buffers"); 161 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 162 "Minimum amount of buffer space required for active I/O"); 163 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 164 "Maximum amount of buffer space to usable for active I/O"); 165 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 166 "Page allocations done during periods of very low free memory"); 167 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 168 "Page allocations which failed during periods of very low free memory"); 169 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 170 "Recycle pages to active or inactive queue transition pt 0-64"); 171 /* 172 * Sysctls determining current state of the buffer cache. 173 */ 174 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 175 "Total number of buffers in buffer cache"); 176 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 177 "Pending bytes of dirty buffers (all)"); 178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 179 "Pending bytes of dirty buffers (heavy weight)"); 180 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 181 "Pending number of dirty buffers"); 182 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 183 "Pending number of dirty buffers (heavy weight)"); 184 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 185 "I/O bytes currently in progress due to asynchronous writes"); 186 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 187 "I/O buffers currently in progress due to asynchronous writes"); 188 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 189 "Hard limit on maximum amount of memory usable for buffer space"); 190 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 191 "Soft limit on maximum amount of memory usable for buffer space"); 192 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 193 "Minimum amount of memory to reserve for system buffer space"); 194 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 195 "Amount of memory available for buffers"); 196 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 197 0, "Maximum amount of memory reserved for buffers using malloc"); 198 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 199 "Amount of memory left for buffers using malloc-scheme"); 200 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 201 "New buffer header acquisition requests"); 202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 203 0, "New buffer header acquisition restarts"); 204 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 205 "Recover VM space in an emergency"); 206 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 207 "Buffer acquisition restarts due to fragmented buffer map"); 208 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 209 "Amount of time KVA space was deallocated in an arbitrary buffer"); 210 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 211 "Amount of time buffer re-use operations were successful"); 212 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, ""); 213 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 214 "sizeof(struct buf)"); 215 216 char *buf_wmesg = BUF_WMESG; 217 218 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 219 #define VFS_BIO_NEED_UNUSED02 0x02 220 #define VFS_BIO_NEED_UNUSED04 0x04 221 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 222 223 /* 224 * bufspacewakeup: 225 * 226 * Called when buffer space is potentially available for recovery. 227 * getnewbuf() will block on this flag when it is unable to free 228 * sufficient buffer space. Buffer space becomes recoverable when 229 * bp's get placed back in the queues. 230 */ 231 static __inline void 232 bufspacewakeup(void) 233 { 234 /* 235 * If someone is waiting for BUF space, wake them up. Even 236 * though we haven't freed the kva space yet, the waiting 237 * process will be able to now. 238 */ 239 spin_lock(&bufcspin); 240 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 241 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 242 spin_unlock(&bufcspin); 243 wakeup(&needsbuffer); 244 } else { 245 spin_unlock(&bufcspin); 246 } 247 } 248 249 /* 250 * runningbufwakeup: 251 * 252 * Accounting for I/O in progress. 253 * 254 */ 255 static __inline void 256 runningbufwakeup(struct buf *bp) 257 { 258 int totalspace; 259 int limit; 260 261 if ((totalspace = bp->b_runningbufspace) != 0) { 262 spin_lock(&bufcspin); 263 runningbufspace -= totalspace; 264 --runningbufcount; 265 bp->b_runningbufspace = 0; 266 267 /* 268 * see waitrunningbufspace() for limit test. 269 */ 270 limit = hirunningspace * 3 / 6; 271 if (runningbufreq && runningbufspace <= limit) { 272 runningbufreq = 0; 273 spin_unlock(&bufcspin); 274 wakeup(&runningbufreq); 275 } else { 276 spin_unlock(&bufcspin); 277 } 278 bd_signal(totalspace); 279 } 280 } 281 282 /* 283 * bufcountwakeup: 284 * 285 * Called when a buffer has been added to one of the free queues to 286 * account for the buffer and to wakeup anyone waiting for free buffers. 287 * This typically occurs when large amounts of metadata are being handled 288 * by the buffer cache ( else buffer space runs out first, usually ). 289 * 290 * MPSAFE 291 */ 292 static __inline void 293 bufcountwakeup(void) 294 { 295 spin_lock(&bufcspin); 296 if (needsbuffer) { 297 needsbuffer &= ~VFS_BIO_NEED_ANY; 298 spin_unlock(&bufcspin); 299 wakeup(&needsbuffer); 300 } else { 301 spin_unlock(&bufcspin); 302 } 303 } 304 305 /* 306 * waitrunningbufspace() 307 * 308 * If runningbufspace exceeds 4/6 hirunningspace we block until 309 * runningbufspace drops to 3/6 hirunningspace. We also block if another 310 * thread blocked here in order to be fair, even if runningbufspace 311 * is now lower than the limit. 312 * 313 * The caller may be using this function to block in a tight loop, we 314 * must block while runningbufspace is greater than at least 315 * hirunningspace * 3 / 6. 316 */ 317 void 318 waitrunningbufspace(void) 319 { 320 int limit = hirunningspace * 4 / 6; 321 322 if (runningbufspace > limit || runningbufreq) { 323 spin_lock(&bufcspin); 324 while (runningbufspace > limit || runningbufreq) { 325 runningbufreq = 1; 326 ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0); 327 } 328 spin_unlock(&bufcspin); 329 } 330 } 331 332 /* 333 * buf_dirty_count_severe: 334 * 335 * Return true if we have too many dirty buffers. 336 */ 337 int 338 buf_dirty_count_severe(void) 339 { 340 return (runningbufspace + dirtybufspace >= hidirtybufspace || 341 dirtybufcount >= nbuf / 2); 342 } 343 344 /* 345 * Return true if the amount of running I/O is severe and BIOQ should 346 * start bursting. 347 */ 348 int 349 buf_runningbufspace_severe(void) 350 { 351 return (runningbufspace >= hirunningspace * 4 / 6); 352 } 353 354 /* 355 * vfs_buf_test_cache: 356 * 357 * Called when a buffer is extended. This function clears the B_CACHE 358 * bit if the newly extended portion of the buffer does not contain 359 * valid data. 360 * 361 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 362 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 363 * them while a clean buffer was present. 364 */ 365 static __inline__ 366 void 367 vfs_buf_test_cache(struct buf *bp, 368 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 369 vm_page_t m) 370 { 371 if (bp->b_flags & B_CACHE) { 372 int base = (foff + off) & PAGE_MASK; 373 if (vm_page_is_valid(m, base, size) == 0) 374 bp->b_flags &= ~B_CACHE; 375 } 376 } 377 378 /* 379 * bd_speedup() 380 * 381 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 382 * low water mark. 383 * 384 * MPSAFE 385 */ 386 static __inline__ 387 void 388 bd_speedup(void) 389 { 390 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2) 391 return; 392 393 if (bd_request == 0 && 394 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 || 395 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 396 spin_lock(&bufcspin); 397 bd_request = 1; 398 spin_unlock(&bufcspin); 399 wakeup(&bd_request); 400 } 401 if (bd_request_hw == 0 && 402 (dirtybufspacehw > lodirtybufspace / 2 || 403 dirtybufcounthw >= nbuf / 2)) { 404 spin_lock(&bufcspin); 405 bd_request_hw = 1; 406 spin_unlock(&bufcspin); 407 wakeup(&bd_request_hw); 408 } 409 } 410 411 /* 412 * bd_heatup() 413 * 414 * Get the buf_daemon heated up when the number of running and dirty 415 * buffers exceeds the mid-point. 416 * 417 * Return the total number of dirty bytes past the second mid point 418 * as a measure of how much excess dirty data there is in the system. 419 * 420 * MPSAFE 421 */ 422 int 423 bd_heatup(void) 424 { 425 int mid1; 426 int mid2; 427 int totalspace; 428 429 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 430 431 totalspace = runningbufspace + dirtybufspace; 432 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 433 bd_speedup(); 434 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 435 if (totalspace >= mid2) 436 return(totalspace - mid2); 437 } 438 return(0); 439 } 440 441 /* 442 * bd_wait() 443 * 444 * Wait for the buffer cache to flush (totalspace) bytes worth of 445 * buffers, then return. 446 * 447 * Regardless this function blocks while the number of dirty buffers 448 * exceeds hidirtybufspace. 449 * 450 * MPSAFE 451 */ 452 void 453 bd_wait(int totalspace) 454 { 455 u_int i; 456 int count; 457 458 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 459 return; 460 461 while (totalspace > 0) { 462 bd_heatup(); 463 if (totalspace > runningbufspace + dirtybufspace) 464 totalspace = runningbufspace + dirtybufspace; 465 count = totalspace / BKVASIZE; 466 if (count >= BD_WAKE_SIZE) 467 count = BD_WAKE_SIZE - 1; 468 469 spin_lock(&bufcspin); 470 i = (bd_wake_index + count) & BD_WAKE_MASK; 471 ++bd_wake_ary[i]; 472 473 /* 474 * This is not a strict interlock, so we play a bit loose 475 * with locking access to dirtybufspace* 476 */ 477 tsleep_interlock(&bd_wake_ary[i], 0); 478 spin_unlock(&bufcspin); 479 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz); 480 481 totalspace = runningbufspace + dirtybufspace - hidirtybufspace; 482 } 483 } 484 485 /* 486 * bd_signal() 487 * 488 * This function is called whenever runningbufspace or dirtybufspace 489 * is reduced. Track threads waiting for run+dirty buffer I/O 490 * complete. 491 * 492 * MPSAFE 493 */ 494 static void 495 bd_signal(int totalspace) 496 { 497 u_int i; 498 499 if (totalspace > 0) { 500 if (totalspace > BKVASIZE * BD_WAKE_SIZE) 501 totalspace = BKVASIZE * BD_WAKE_SIZE; 502 spin_lock(&bufcspin); 503 while (totalspace > 0) { 504 i = bd_wake_index++; 505 i &= BD_WAKE_MASK; 506 if (bd_wake_ary[i]) { 507 bd_wake_ary[i] = 0; 508 spin_unlock(&bufcspin); 509 wakeup(&bd_wake_ary[i]); 510 spin_lock(&bufcspin); 511 } 512 totalspace -= BKVASIZE; 513 } 514 spin_unlock(&bufcspin); 515 } 516 } 517 518 /* 519 * BIO tracking support routines. 520 * 521 * Release a ref on a bio_track. Wakeup requests are atomically released 522 * along with the last reference so bk_active will never wind up set to 523 * only 0x80000000. 524 * 525 * MPSAFE 526 */ 527 static 528 void 529 bio_track_rel(struct bio_track *track) 530 { 531 int active; 532 int desired; 533 534 /* 535 * Shortcut 536 */ 537 active = track->bk_active; 538 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 539 return; 540 541 /* 542 * Full-on. Note that the wait flag is only atomically released on 543 * the 1->0 count transition. 544 * 545 * We check for a negative count transition using bit 30 since bit 31 546 * has a different meaning. 547 */ 548 for (;;) { 549 desired = (active & 0x7FFFFFFF) - 1; 550 if (desired) 551 desired |= active & 0x80000000; 552 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 553 if (desired & 0x40000000) 554 panic("bio_track_rel: bad count: %p\n", track); 555 if (active & 0x80000000) 556 wakeup(track); 557 break; 558 } 559 active = track->bk_active; 560 } 561 } 562 563 /* 564 * Wait for the tracking count to reach 0. 565 * 566 * Use atomic ops such that the wait flag is only set atomically when 567 * bk_active is non-zero. 568 * 569 * MPSAFE 570 */ 571 int 572 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 573 { 574 int active; 575 int desired; 576 int error; 577 578 /* 579 * Shortcut 580 */ 581 if (track->bk_active == 0) 582 return(0); 583 584 /* 585 * Full-on. Note that the wait flag may only be atomically set if 586 * the active count is non-zero. 587 * 588 * NOTE: We cannot optimize active == desired since a wakeup could 589 * clear active prior to our tsleep_interlock(). 590 */ 591 error = 0; 592 while ((active = track->bk_active) != 0) { 593 cpu_ccfence(); 594 desired = active | 0x80000000; 595 tsleep_interlock(track, slp_flags); 596 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 597 error = tsleep(track, slp_flags | PINTERLOCKED, 598 "trwait", slp_timo); 599 if (error) 600 break; 601 } 602 } 603 return (error); 604 } 605 606 /* 607 * bufinit: 608 * 609 * Load time initialisation of the buffer cache, called from machine 610 * dependant initialization code. 611 */ 612 void 613 bufinit(void) 614 { 615 struct buf *bp; 616 vm_offset_t bogus_offset; 617 int i; 618 619 /* next, make a null set of free lists */ 620 for (i = 0; i < BUFFER_QUEUES; i++) 621 TAILQ_INIT(&bufqueues[i]); 622 623 /* finally, initialize each buffer header and stick on empty q */ 624 for (i = 0; i < nbuf; i++) { 625 bp = &buf[i]; 626 bzero(bp, sizeof *bp); 627 bp->b_flags = B_INVAL; /* we're just an empty header */ 628 bp->b_cmd = BUF_CMD_DONE; 629 bp->b_qindex = BQUEUE_EMPTY; 630 initbufbio(bp); 631 xio_init(&bp->b_xio); 632 buf_dep_init(bp); 633 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist); 634 } 635 636 /* 637 * maxbufspace is the absolute maximum amount of buffer space we are 638 * allowed to reserve in KVM and in real terms. The absolute maximum 639 * is nominally used by buf_daemon. hibufspace is the nominal maximum 640 * used by most other processes. The differential is required to 641 * ensure that buf_daemon is able to run when other processes might 642 * be blocked waiting for buffer space. 643 * 644 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 645 * this may result in KVM fragmentation which is not handled optimally 646 * by the system. 647 */ 648 maxbufspace = nbuf * BKVASIZE; 649 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 650 lobufspace = hibufspace - MAXBSIZE; 651 652 lorunningspace = 512 * 1024; 653 /* hirunningspace -- see below */ 654 655 /* 656 * Limit the amount of malloc memory since it is wired permanently 657 * into the kernel space. Even though this is accounted for in 658 * the buffer allocation, we don't want the malloced region to grow 659 * uncontrolled. The malloc scheme improves memory utilization 660 * significantly on average (small) directories. 661 */ 662 maxbufmallocspace = hibufspace / 20; 663 664 /* 665 * Reduce the chance of a deadlock occuring by limiting the number 666 * of delayed-write dirty buffers we allow to stack up. 667 * 668 * We don't want too much actually queued to the device at once 669 * (XXX this needs to be per-mount!), because the buffers will 670 * wind up locked for a very long period of time while the I/O 671 * drains. 672 */ 673 hidirtybufspace = hibufspace / 2; /* dirty + running */ 674 hirunningspace = hibufspace / 16; /* locked & queued to device */ 675 if (hirunningspace < 1024 * 1024) 676 hirunningspace = 1024 * 1024; 677 678 dirtybufspace = 0; 679 dirtybufspacehw = 0; 680 681 lodirtybufspace = hidirtybufspace / 2; 682 683 /* 684 * Maximum number of async ops initiated per buf_daemon loop. This is 685 * somewhat of a hack at the moment, we really need to limit ourselves 686 * based on the number of bytes of I/O in-transit that were initiated 687 * from buf_daemon. 688 */ 689 690 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 691 vm_object_hold(&kernel_object); 692 bogus_page = vm_page_alloc(&kernel_object, 693 (bogus_offset >> PAGE_SHIFT), 694 VM_ALLOC_NORMAL); 695 vm_object_drop(&kernel_object); 696 vmstats.v_wire_count++; 697 698 } 699 700 /* 701 * Initialize the embedded bio structures, typically used by 702 * deprecated code which tries to allocate its own struct bufs. 703 */ 704 void 705 initbufbio(struct buf *bp) 706 { 707 bp->b_bio1.bio_buf = bp; 708 bp->b_bio1.bio_prev = NULL; 709 bp->b_bio1.bio_offset = NOOFFSET; 710 bp->b_bio1.bio_next = &bp->b_bio2; 711 bp->b_bio1.bio_done = NULL; 712 bp->b_bio1.bio_flags = 0; 713 714 bp->b_bio2.bio_buf = bp; 715 bp->b_bio2.bio_prev = &bp->b_bio1; 716 bp->b_bio2.bio_offset = NOOFFSET; 717 bp->b_bio2.bio_next = NULL; 718 bp->b_bio2.bio_done = NULL; 719 bp->b_bio2.bio_flags = 0; 720 721 BUF_LOCKINIT(bp); 722 } 723 724 /* 725 * Reinitialize the embedded bio structures as well as any additional 726 * translation cache layers. 727 */ 728 void 729 reinitbufbio(struct buf *bp) 730 { 731 struct bio *bio; 732 733 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 734 bio->bio_done = NULL; 735 bio->bio_offset = NOOFFSET; 736 } 737 } 738 739 /* 740 * Undo the effects of an initbufbio(). 741 */ 742 void 743 uninitbufbio(struct buf *bp) 744 { 745 dsched_exit_buf(bp); 746 BUF_LOCKFREE(bp); 747 } 748 749 /* 750 * Push another BIO layer onto an existing BIO and return it. The new 751 * BIO layer may already exist, holding cached translation data. 752 */ 753 struct bio * 754 push_bio(struct bio *bio) 755 { 756 struct bio *nbio; 757 758 if ((nbio = bio->bio_next) == NULL) { 759 int index = bio - &bio->bio_buf->b_bio_array[0]; 760 if (index >= NBUF_BIO - 1) { 761 panic("push_bio: too many layers bp %p\n", 762 bio->bio_buf); 763 } 764 nbio = &bio->bio_buf->b_bio_array[index + 1]; 765 bio->bio_next = nbio; 766 nbio->bio_prev = bio; 767 nbio->bio_buf = bio->bio_buf; 768 nbio->bio_offset = NOOFFSET; 769 nbio->bio_done = NULL; 770 nbio->bio_next = NULL; 771 } 772 KKASSERT(nbio->bio_done == NULL); 773 return(nbio); 774 } 775 776 /* 777 * Pop a BIO translation layer, returning the previous layer. The 778 * must have been previously pushed. 779 */ 780 struct bio * 781 pop_bio(struct bio *bio) 782 { 783 return(bio->bio_prev); 784 } 785 786 void 787 clearbiocache(struct bio *bio) 788 { 789 while (bio) { 790 bio->bio_offset = NOOFFSET; 791 bio = bio->bio_next; 792 } 793 } 794 795 /* 796 * bfreekva: 797 * 798 * Free the KVA allocation for buffer 'bp'. 799 * 800 * Must be called from a critical section as this is the only locking for 801 * buffer_map. 802 * 803 * Since this call frees up buffer space, we call bufspacewakeup(). 804 * 805 * MPALMOSTSAFE 806 */ 807 static void 808 bfreekva(struct buf *bp) 809 { 810 int count; 811 812 if (bp->b_kvasize) { 813 ++buffreekvacnt; 814 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 815 vm_map_lock(&buffer_map); 816 bufspace -= bp->b_kvasize; 817 vm_map_delete(&buffer_map, 818 (vm_offset_t) bp->b_kvabase, 819 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 820 &count 821 ); 822 vm_map_unlock(&buffer_map); 823 vm_map_entry_release(count); 824 bp->b_kvasize = 0; 825 bp->b_kvabase = NULL; 826 bufspacewakeup(); 827 } 828 } 829 830 /* 831 * bremfree: 832 * 833 * Remove the buffer from the appropriate free list. 834 */ 835 static __inline void 836 _bremfree(struct buf *bp) 837 { 838 if (bp->b_qindex != BQUEUE_NONE) { 839 KASSERT(BUF_REFCNTNB(bp) == 1, 840 ("bremfree: bp %p not locked",bp)); 841 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 842 bp->b_qindex = BQUEUE_NONE; 843 } else { 844 if (BUF_REFCNTNB(bp) <= 1) 845 panic("bremfree: removing a buffer not on a queue"); 846 } 847 } 848 849 void 850 bremfree(struct buf *bp) 851 { 852 spin_lock(&bufqspin); 853 _bremfree(bp); 854 spin_unlock(&bufqspin); 855 } 856 857 static void 858 bremfree_locked(struct buf *bp) 859 { 860 _bremfree(bp); 861 } 862 863 /* 864 * bread: 865 * 866 * Get a buffer with the specified data. Look in the cache first. We 867 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 868 * is set, the buffer is valid and we do not have to do anything ( see 869 * getblk() ). 870 * 871 */ 872 int 873 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp) 874 { 875 return (breadn(vp, loffset, size, NULL, NULL, 0, bpp)); 876 } 877 878 /* 879 * This version of bread issues any required I/O asyncnronously and 880 * makes a callback on completion. 881 * 882 * The callback must check whether BIO_DONE is set in the bio and issue 883 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 884 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 885 */ 886 void 887 breadcb(struct vnode *vp, off_t loffset, int size, 888 void (*func)(struct bio *), void *arg) 889 { 890 struct buf *bp; 891 892 bp = getblk(vp, loffset, size, 0, 0); 893 894 /* if not found in cache, do some I/O */ 895 if ((bp->b_flags & B_CACHE) == 0) { 896 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 897 bp->b_cmd = BUF_CMD_READ; 898 bp->b_bio1.bio_done = func; 899 bp->b_bio1.bio_caller_info1.ptr = arg; 900 vfs_busy_pages(vp, bp); 901 BUF_KERNPROC(bp); 902 vn_strategy(vp, &bp->b_bio1); 903 } else if (func) { 904 /* 905 * Since we are issuing the callback synchronously it cannot 906 * race the BIO_DONE, so no need for atomic ops here. 907 */ 908 /*bp->b_bio1.bio_done = func;*/ 909 bp->b_bio1.bio_caller_info1.ptr = arg; 910 bp->b_bio1.bio_flags |= BIO_DONE; 911 func(&bp->b_bio1); 912 } else { 913 bqrelse(bp); 914 } 915 } 916 917 /* 918 * breadn: 919 * 920 * Operates like bread, but also starts asynchronous I/O on 921 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 922 * to initiating I/O . If B_CACHE is set, the buffer is valid 923 * and we do not have to do anything. 924 * 925 */ 926 int 927 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset, 928 int *rabsize, int cnt, struct buf **bpp) 929 { 930 struct buf *bp, *rabp; 931 int i; 932 int rv = 0, readwait = 0; 933 934 *bpp = bp = getblk(vp, loffset, size, 0, 0); 935 936 /* if not found in cache, do some I/O */ 937 if ((bp->b_flags & B_CACHE) == 0) { 938 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 939 bp->b_cmd = BUF_CMD_READ; 940 bp->b_bio1.bio_done = biodone_sync; 941 bp->b_bio1.bio_flags |= BIO_SYNC; 942 vfs_busy_pages(vp, bp); 943 vn_strategy(vp, &bp->b_bio1); 944 ++readwait; 945 } 946 947 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 948 if (inmem(vp, *raoffset)) 949 continue; 950 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 951 952 if ((rabp->b_flags & B_CACHE) == 0) { 953 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 954 rabp->b_cmd = BUF_CMD_READ; 955 vfs_busy_pages(vp, rabp); 956 BUF_KERNPROC(rabp); 957 vn_strategy(vp, &rabp->b_bio1); 958 } else { 959 brelse(rabp); 960 } 961 } 962 if (readwait) 963 rv = biowait(&bp->b_bio1, "biord"); 964 return (rv); 965 } 966 967 /* 968 * bwrite: 969 * 970 * Synchronous write, waits for completion. 971 * 972 * Write, release buffer on completion. (Done by iodone 973 * if async). Do not bother writing anything if the buffer 974 * is invalid. 975 * 976 * Note that we set B_CACHE here, indicating that buffer is 977 * fully valid and thus cacheable. This is true even of NFS 978 * now so we set it generally. This could be set either here 979 * or in biodone() since the I/O is synchronous. We put it 980 * here. 981 */ 982 int 983 bwrite(struct buf *bp) 984 { 985 int error; 986 987 if (bp->b_flags & B_INVAL) { 988 brelse(bp); 989 return (0); 990 } 991 if (BUF_REFCNTNB(bp) == 0) 992 panic("bwrite: buffer is not busy???"); 993 994 /* Mark the buffer clean */ 995 bundirty(bp); 996 997 bp->b_flags &= ~(B_ERROR | B_EINTR); 998 bp->b_flags |= B_CACHE; 999 bp->b_cmd = BUF_CMD_WRITE; 1000 bp->b_bio1.bio_done = biodone_sync; 1001 bp->b_bio1.bio_flags |= BIO_SYNC; 1002 vfs_busy_pages(bp->b_vp, bp); 1003 1004 /* 1005 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1006 * valid for vnode-backed buffers. 1007 */ 1008 bsetrunningbufspace(bp, bp->b_bufsize); 1009 vn_strategy(bp->b_vp, &bp->b_bio1); 1010 error = biowait(&bp->b_bio1, "biows"); 1011 brelse(bp); 1012 1013 return (error); 1014 } 1015 1016 /* 1017 * bawrite: 1018 * 1019 * Asynchronous write. Start output on a buffer, but do not wait for 1020 * it to complete. The buffer is released when the output completes. 1021 * 1022 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1023 * B_INVAL buffers. Not us. 1024 */ 1025 void 1026 bawrite(struct buf *bp) 1027 { 1028 if (bp->b_flags & B_INVAL) { 1029 brelse(bp); 1030 return; 1031 } 1032 if (BUF_REFCNTNB(bp) == 0) 1033 panic("bwrite: buffer is not busy???"); 1034 1035 /* Mark the buffer clean */ 1036 bundirty(bp); 1037 1038 bp->b_flags &= ~(B_ERROR | B_EINTR); 1039 bp->b_flags |= B_CACHE; 1040 bp->b_cmd = BUF_CMD_WRITE; 1041 KKASSERT(bp->b_bio1.bio_done == NULL); 1042 vfs_busy_pages(bp->b_vp, bp); 1043 1044 /* 1045 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1046 * valid for vnode-backed buffers. 1047 */ 1048 bsetrunningbufspace(bp, bp->b_bufsize); 1049 BUF_KERNPROC(bp); 1050 vn_strategy(bp->b_vp, &bp->b_bio1); 1051 } 1052 1053 /* 1054 * bowrite: 1055 * 1056 * Ordered write. Start output on a buffer, and flag it so that the 1057 * device will write it in the order it was queued. The buffer is 1058 * released when the output completes. bwrite() ( or the VOP routine 1059 * anyway ) is responsible for handling B_INVAL buffers. 1060 */ 1061 int 1062 bowrite(struct buf *bp) 1063 { 1064 bp->b_flags |= B_ORDERED; 1065 bawrite(bp); 1066 return (0); 1067 } 1068 1069 /* 1070 * bdwrite: 1071 * 1072 * Delayed write. (Buffer is marked dirty). Do not bother writing 1073 * anything if the buffer is marked invalid. 1074 * 1075 * Note that since the buffer must be completely valid, we can safely 1076 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1077 * biodone() in order to prevent getblk from writing the buffer 1078 * out synchronously. 1079 */ 1080 void 1081 bdwrite(struct buf *bp) 1082 { 1083 if (BUF_REFCNTNB(bp) == 0) 1084 panic("bdwrite: buffer is not busy"); 1085 1086 if (bp->b_flags & B_INVAL) { 1087 brelse(bp); 1088 return; 1089 } 1090 bdirty(bp); 1091 1092 if (dsched_is_clear_buf_priv(bp)) 1093 dsched_new_buf(bp); 1094 1095 /* 1096 * Set B_CACHE, indicating that the buffer is fully valid. This is 1097 * true even of NFS now. 1098 */ 1099 bp->b_flags |= B_CACHE; 1100 1101 /* 1102 * This bmap keeps the system from needing to do the bmap later, 1103 * perhaps when the system is attempting to do a sync. Since it 1104 * is likely that the indirect block -- or whatever other datastructure 1105 * that the filesystem needs is still in memory now, it is a good 1106 * thing to do this. Note also, that if the pageout daemon is 1107 * requesting a sync -- there might not be enough memory to do 1108 * the bmap then... So, this is important to do. 1109 */ 1110 if (bp->b_bio2.bio_offset == NOOFFSET) { 1111 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1112 NULL, NULL, BUF_CMD_WRITE); 1113 } 1114 1115 /* 1116 * Because the underlying pages may still be mapped and 1117 * writable trying to set the dirty buffer (b_dirtyoff/end) 1118 * range here will be inaccurate. 1119 * 1120 * However, we must still clean the pages to satisfy the 1121 * vnode_pager and pageout daemon, so theythink the pages 1122 * have been "cleaned". What has really occured is that 1123 * they've been earmarked for later writing by the buffer 1124 * cache. 1125 * 1126 * So we get the b_dirtyoff/end update but will not actually 1127 * depend on it (NFS that is) until the pages are busied for 1128 * writing later on. 1129 */ 1130 vfs_clean_pages(bp); 1131 bqrelse(bp); 1132 1133 /* 1134 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1135 * due to the softdep code. 1136 */ 1137 } 1138 1139 /* 1140 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1141 * This is used by tmpfs. 1142 * 1143 * It is important for any VFS using this routine to NOT use it for 1144 * IO_SYNC or IO_ASYNC operations which occur when the system really 1145 * wants to flush VM pages to backing store. 1146 */ 1147 void 1148 buwrite(struct buf *bp) 1149 { 1150 vm_page_t m; 1151 int i; 1152 1153 /* 1154 * Only works for VMIO buffers. If the buffer is already 1155 * marked for delayed-write we can't avoid the bdwrite(). 1156 */ 1157 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1158 bdwrite(bp); 1159 return; 1160 } 1161 1162 /* 1163 * Set valid & dirty. 1164 */ 1165 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1166 m = bp->b_xio.xio_pages[i]; 1167 vfs_dirty_one_page(bp, i, m); 1168 } 1169 bqrelse(bp); 1170 } 1171 1172 /* 1173 * bdirty: 1174 * 1175 * Turn buffer into delayed write request by marking it B_DELWRI. 1176 * B_RELBUF and B_NOCACHE must be cleared. 1177 * 1178 * We reassign the buffer to itself to properly update it in the 1179 * dirty/clean lists. 1180 * 1181 * Must be called from a critical section. 1182 * The buffer must be on BQUEUE_NONE. 1183 */ 1184 void 1185 bdirty(struct buf *bp) 1186 { 1187 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1188 if (bp->b_flags & B_NOCACHE) { 1189 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1190 bp->b_flags &= ~B_NOCACHE; 1191 } 1192 if (bp->b_flags & B_INVAL) { 1193 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1194 } 1195 bp->b_flags &= ~B_RELBUF; 1196 1197 if ((bp->b_flags & B_DELWRI) == 0) { 1198 lwkt_gettoken(&bp->b_vp->v_token); 1199 bp->b_flags |= B_DELWRI; 1200 reassignbuf(bp); 1201 lwkt_reltoken(&bp->b_vp->v_token); 1202 1203 spin_lock(&bufcspin); 1204 ++dirtybufcount; 1205 dirtybufspace += bp->b_bufsize; 1206 if (bp->b_flags & B_HEAVY) { 1207 ++dirtybufcounthw; 1208 dirtybufspacehw += bp->b_bufsize; 1209 } 1210 spin_unlock(&bufcspin); 1211 1212 bd_heatup(); 1213 } 1214 } 1215 1216 /* 1217 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1218 * needs to be flushed with a different buf_daemon thread to avoid 1219 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1220 */ 1221 void 1222 bheavy(struct buf *bp) 1223 { 1224 if ((bp->b_flags & B_HEAVY) == 0) { 1225 bp->b_flags |= B_HEAVY; 1226 if (bp->b_flags & B_DELWRI) { 1227 spin_lock(&bufcspin); 1228 ++dirtybufcounthw; 1229 dirtybufspacehw += bp->b_bufsize; 1230 spin_unlock(&bufcspin); 1231 } 1232 } 1233 } 1234 1235 /* 1236 * bundirty: 1237 * 1238 * Clear B_DELWRI for buffer. 1239 * 1240 * Must be called from a critical section. 1241 * 1242 * The buffer is typically on BQUEUE_NONE but there is one case in 1243 * brelse() that calls this function after placing the buffer on 1244 * a different queue. 1245 * 1246 * MPSAFE 1247 */ 1248 void 1249 bundirty(struct buf *bp) 1250 { 1251 if (bp->b_flags & B_DELWRI) { 1252 lwkt_gettoken(&bp->b_vp->v_token); 1253 bp->b_flags &= ~B_DELWRI; 1254 reassignbuf(bp); 1255 lwkt_reltoken(&bp->b_vp->v_token); 1256 1257 spin_lock(&bufcspin); 1258 --dirtybufcount; 1259 dirtybufspace -= bp->b_bufsize; 1260 if (bp->b_flags & B_HEAVY) { 1261 --dirtybufcounthw; 1262 dirtybufspacehw -= bp->b_bufsize; 1263 } 1264 spin_unlock(&bufcspin); 1265 1266 bd_signal(bp->b_bufsize); 1267 } 1268 /* 1269 * Since it is now being written, we can clear its deferred write flag. 1270 */ 1271 bp->b_flags &= ~B_DEFERRED; 1272 } 1273 1274 /* 1275 * Set the b_runningbufspace field, used to track how much I/O is 1276 * in progress at any given moment. 1277 */ 1278 void 1279 bsetrunningbufspace(struct buf *bp, int bytes) 1280 { 1281 bp->b_runningbufspace = bytes; 1282 if (bytes) { 1283 spin_lock(&bufcspin); 1284 runningbufspace += bytes; 1285 ++runningbufcount; 1286 spin_unlock(&bufcspin); 1287 } 1288 } 1289 1290 /* 1291 * brelse: 1292 * 1293 * Release a busy buffer and, if requested, free its resources. The 1294 * buffer will be stashed in the appropriate bufqueue[] allowing it 1295 * to be accessed later as a cache entity or reused for other purposes. 1296 * 1297 * MPALMOSTSAFE 1298 */ 1299 void 1300 brelse(struct buf *bp) 1301 { 1302 #ifdef INVARIANTS 1303 int saved_flags = bp->b_flags; 1304 #endif 1305 1306 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1307 1308 /* 1309 * If B_NOCACHE is set we are being asked to destroy the buffer and 1310 * its backing store. Clear B_DELWRI. 1311 * 1312 * B_NOCACHE is set in two cases: (1) when the caller really wants 1313 * to destroy the buffer and backing store and (2) when the caller 1314 * wants to destroy the buffer and backing store after a write 1315 * completes. 1316 */ 1317 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1318 bundirty(bp); 1319 } 1320 1321 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1322 /* 1323 * A re-dirtied buffer is only subject to destruction 1324 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1325 */ 1326 /* leave buffer intact */ 1327 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1328 (bp->b_bufsize <= 0)) { 1329 /* 1330 * Either a failed read or we were asked to free or not 1331 * cache the buffer. This path is reached with B_DELWRI 1332 * set only if B_INVAL is already set. B_NOCACHE governs 1333 * backing store destruction. 1334 * 1335 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1336 * buffer cannot be immediately freed. 1337 */ 1338 bp->b_flags |= B_INVAL; 1339 if (LIST_FIRST(&bp->b_dep) != NULL) 1340 buf_deallocate(bp); 1341 if (bp->b_flags & B_DELWRI) { 1342 spin_lock(&bufcspin); 1343 --dirtybufcount; 1344 dirtybufspace -= bp->b_bufsize; 1345 if (bp->b_flags & B_HEAVY) { 1346 --dirtybufcounthw; 1347 dirtybufspacehw -= bp->b_bufsize; 1348 } 1349 spin_unlock(&bufcspin); 1350 1351 bd_signal(bp->b_bufsize); 1352 } 1353 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1354 } 1355 1356 /* 1357 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set, 1358 * or if b_refs is non-zero. 1359 * 1360 * If vfs_vmio_release() is called with either bit set, the 1361 * underlying pages may wind up getting freed causing a previous 1362 * write (bdwrite()) to get 'lost' because pages associated with 1363 * a B_DELWRI bp are marked clean. Pages associated with a 1364 * B_LOCKED buffer may be mapped by the filesystem. 1365 * 1366 * If we want to release the buffer ourselves (rather then the 1367 * originator asking us to release it), give the originator a 1368 * chance to countermand the release by setting B_LOCKED. 1369 * 1370 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1371 * if B_DELWRI is set. 1372 * 1373 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1374 * on pages to return pages to the VM page queues. 1375 */ 1376 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) { 1377 bp->b_flags &= ~B_RELBUF; 1378 } else if (vm_page_count_severe()) { 1379 if (LIST_FIRST(&bp->b_dep) != NULL) 1380 buf_deallocate(bp); /* can set B_LOCKED */ 1381 if (bp->b_flags & (B_DELWRI | B_LOCKED)) 1382 bp->b_flags &= ~B_RELBUF; 1383 else 1384 bp->b_flags |= B_RELBUF; 1385 } 1386 1387 /* 1388 * Make sure b_cmd is clear. It may have already been cleared by 1389 * biodone(). 1390 * 1391 * At this point destroying the buffer is governed by the B_INVAL 1392 * or B_RELBUF flags. 1393 */ 1394 bp->b_cmd = BUF_CMD_DONE; 1395 dsched_exit_buf(bp); 1396 1397 /* 1398 * VMIO buffer rundown. Make sure the VM page array is restored 1399 * after an I/O may have replaces some of the pages with bogus pages 1400 * in order to not destroy dirty pages in a fill-in read. 1401 * 1402 * Note that due to the code above, if a buffer is marked B_DELWRI 1403 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1404 * B_INVAL may still be set, however. 1405 * 1406 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1407 * but not the backing store. B_NOCACHE will destroy the backing 1408 * store. 1409 * 1410 * Note that dirty NFS buffers contain byte-granular write ranges 1411 * and should not be destroyed w/ B_INVAL even if the backing store 1412 * is left intact. 1413 */ 1414 if (bp->b_flags & B_VMIO) { 1415 /* 1416 * Rundown for VMIO buffers which are not dirty NFS buffers. 1417 */ 1418 int i, j, resid; 1419 vm_page_t m; 1420 off_t foff; 1421 vm_pindex_t poff; 1422 vm_object_t obj; 1423 struct vnode *vp; 1424 1425 vp = bp->b_vp; 1426 1427 /* 1428 * Get the base offset and length of the buffer. Note that 1429 * in the VMIO case if the buffer block size is not 1430 * page-aligned then b_data pointer may not be page-aligned. 1431 * But our b_xio.xio_pages array *IS* page aligned. 1432 * 1433 * block sizes less then DEV_BSIZE (usually 512) are not 1434 * supported due to the page granularity bits (m->valid, 1435 * m->dirty, etc...). 1436 * 1437 * See man buf(9) for more information 1438 */ 1439 1440 resid = bp->b_bufsize; 1441 foff = bp->b_loffset; 1442 1443 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1444 m = bp->b_xio.xio_pages[i]; 1445 vm_page_flag_clear(m, PG_ZERO); 1446 /* 1447 * If we hit a bogus page, fixup *all* of them 1448 * now. Note that we left these pages wired 1449 * when we removed them so they had better exist, 1450 * and they cannot be ripped out from under us so 1451 * no critical section protection is necessary. 1452 */ 1453 if (m == bogus_page) { 1454 obj = vp->v_object; 1455 poff = OFF_TO_IDX(bp->b_loffset); 1456 1457 vm_object_hold(obj); 1458 for (j = i; j < bp->b_xio.xio_npages; j++) { 1459 vm_page_t mtmp; 1460 1461 mtmp = bp->b_xio.xio_pages[j]; 1462 if (mtmp == bogus_page) { 1463 mtmp = vm_page_lookup(obj, poff + j); 1464 if (!mtmp) { 1465 panic("brelse: page missing"); 1466 } 1467 bp->b_xio.xio_pages[j] = mtmp; 1468 } 1469 } 1470 bp->b_flags &= ~B_HASBOGUS; 1471 vm_object_drop(obj); 1472 1473 if ((bp->b_flags & B_INVAL) == 0) { 1474 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1475 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1476 } 1477 m = bp->b_xio.xio_pages[i]; 1478 } 1479 1480 /* 1481 * Invalidate the backing store if B_NOCACHE is set 1482 * (e.g. used with vinvalbuf()). If this is NFS 1483 * we impose a requirement that the block size be 1484 * a multiple of PAGE_SIZE and create a temporary 1485 * hack to basically invalidate the whole page. The 1486 * problem is that NFS uses really odd buffer sizes 1487 * especially when tracking piecemeal writes and 1488 * it also vinvalbuf()'s a lot, which would result 1489 * in only partial page validation and invalidation 1490 * here. If the file page is mmap()'d, however, 1491 * all the valid bits get set so after we invalidate 1492 * here we would end up with weird m->valid values 1493 * like 0xfc. nfs_getpages() can't handle this so 1494 * we clear all the valid bits for the NFS case 1495 * instead of just some of them. 1496 * 1497 * The real bug is the VM system having to set m->valid 1498 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1499 * itself is an artifact of the whole 512-byte 1500 * granular mess that exists to support odd block 1501 * sizes and UFS meta-data block sizes (e.g. 6144). 1502 * A complete rewrite is required. 1503 * 1504 * XXX 1505 */ 1506 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1507 int poffset = foff & PAGE_MASK; 1508 int presid; 1509 1510 presid = PAGE_SIZE - poffset; 1511 if (bp->b_vp->v_tag == VT_NFS && 1512 bp->b_vp->v_type == VREG) { 1513 ; /* entire page */ 1514 } else if (presid > resid) { 1515 presid = resid; 1516 } 1517 KASSERT(presid >= 0, ("brelse: extra page")); 1518 vm_page_set_invalid(m, poffset, presid); 1519 1520 /* 1521 * Also make sure any swap cache is removed 1522 * as it is now stale (HAMMER in particular 1523 * uses B_NOCACHE to deal with buffer 1524 * aliasing). 1525 */ 1526 swap_pager_unswapped(m); 1527 } 1528 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1529 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1530 } 1531 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1532 vfs_vmio_release(bp); 1533 } else { 1534 /* 1535 * Rundown for non-VMIO buffers. 1536 */ 1537 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1538 if (bp->b_bufsize) 1539 allocbuf(bp, 0); 1540 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1541 if (bp->b_vp) 1542 brelvp(bp); 1543 } 1544 } 1545 1546 if (bp->b_qindex != BQUEUE_NONE) 1547 panic("brelse: free buffer onto another queue???"); 1548 if (BUF_REFCNTNB(bp) > 1) { 1549 /* Temporary panic to verify exclusive locking */ 1550 /* This panic goes away when we allow shared refs */ 1551 panic("brelse: multiple refs"); 1552 /* NOT REACHED */ 1553 return; 1554 } 1555 1556 /* 1557 * Figure out the correct queue to place the cleaned up buffer on. 1558 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1559 * disassociated from their vnode. 1560 */ 1561 spin_lock(&bufqspin); 1562 if (bp->b_flags & B_LOCKED) { 1563 /* 1564 * Buffers that are locked are placed in the locked queue 1565 * immediately, regardless of their state. 1566 */ 1567 bp->b_qindex = BQUEUE_LOCKED; 1568 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1569 } else if (bp->b_bufsize == 0) { 1570 /* 1571 * Buffers with no memory. Due to conditionals near the top 1572 * of brelse() such buffers should probably already be 1573 * marked B_INVAL and disassociated from their vnode. 1574 */ 1575 bp->b_flags |= B_INVAL; 1576 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1577 KKASSERT((bp->b_flags & B_HASHED) == 0); 1578 if (bp->b_kvasize) { 1579 bp->b_qindex = BQUEUE_EMPTYKVA; 1580 } else { 1581 bp->b_qindex = BQUEUE_EMPTY; 1582 } 1583 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1584 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1585 /* 1586 * Buffers with junk contents. Again these buffers had better 1587 * already be disassociated from their vnode. 1588 */ 1589 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1590 KKASSERT((bp->b_flags & B_HASHED) == 0); 1591 bp->b_flags |= B_INVAL; 1592 bp->b_qindex = BQUEUE_CLEAN; 1593 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1594 } else { 1595 /* 1596 * Remaining buffers. These buffers are still associated with 1597 * their vnode. 1598 */ 1599 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1600 case B_DELWRI: 1601 bp->b_qindex = BQUEUE_DIRTY; 1602 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1603 break; 1604 case B_DELWRI | B_HEAVY: 1605 bp->b_qindex = BQUEUE_DIRTY_HW; 1606 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp, 1607 b_freelist); 1608 break; 1609 default: 1610 /* 1611 * NOTE: Buffers are always placed at the end of the 1612 * queue. If B_AGE is not set the buffer will cycle 1613 * through the queue twice. 1614 */ 1615 bp->b_qindex = BQUEUE_CLEAN; 1616 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1617 break; 1618 } 1619 } 1620 spin_unlock(&bufqspin); 1621 1622 /* 1623 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1624 * on the correct queue. 1625 */ 1626 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1627 bundirty(bp); 1628 1629 /* 1630 * The bp is on an appropriate queue unless locked. If it is not 1631 * locked or dirty we can wakeup threads waiting for buffer space. 1632 * 1633 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1634 * if B_INVAL is set ). 1635 */ 1636 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1637 bufcountwakeup(); 1638 1639 /* 1640 * Something we can maybe free or reuse 1641 */ 1642 if (bp->b_bufsize || bp->b_kvasize) 1643 bufspacewakeup(); 1644 1645 /* 1646 * Clean up temporary flags and unlock the buffer. 1647 */ 1648 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT); 1649 BUF_UNLOCK(bp); 1650 } 1651 1652 /* 1653 * bqrelse: 1654 * 1655 * Release a buffer back to the appropriate queue but do not try to free 1656 * it. The buffer is expected to be used again soon. 1657 * 1658 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1659 * biodone() to requeue an async I/O on completion. It is also used when 1660 * known good buffers need to be requeued but we think we may need the data 1661 * again soon. 1662 * 1663 * XXX we should be able to leave the B_RELBUF hint set on completion. 1664 * 1665 * MPSAFE 1666 */ 1667 void 1668 bqrelse(struct buf *bp) 1669 { 1670 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1671 1672 if (bp->b_qindex != BQUEUE_NONE) 1673 panic("bqrelse: free buffer onto another queue???"); 1674 if (BUF_REFCNTNB(bp) > 1) { 1675 /* do not release to free list */ 1676 panic("bqrelse: multiple refs"); 1677 return; 1678 } 1679 1680 buf_act_advance(bp); 1681 1682 spin_lock(&bufqspin); 1683 if (bp->b_flags & B_LOCKED) { 1684 /* 1685 * Locked buffers are released to the locked queue. However, 1686 * if the buffer is dirty it will first go into the dirty 1687 * queue and later on after the I/O completes successfully it 1688 * will be released to the locked queue. 1689 */ 1690 bp->b_qindex = BQUEUE_LOCKED; 1691 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1692 } else if (bp->b_flags & B_DELWRI) { 1693 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1694 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1695 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1696 } else if (vm_page_count_severe()) { 1697 /* 1698 * We are too low on memory, we have to try to free the 1699 * buffer (most importantly: the wired pages making up its 1700 * backing store) *now*. 1701 */ 1702 spin_unlock(&bufqspin); 1703 brelse(bp); 1704 return; 1705 } else { 1706 bp->b_qindex = BQUEUE_CLEAN; 1707 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1708 } 1709 spin_unlock(&bufqspin); 1710 1711 if ((bp->b_flags & B_LOCKED) == 0 && 1712 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1713 bufcountwakeup(); 1714 } 1715 1716 /* 1717 * Something we can maybe free or reuse. 1718 */ 1719 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1720 bufspacewakeup(); 1721 1722 /* 1723 * Final cleanup and unlock. Clear bits that are only used while a 1724 * buffer is actively locked. 1725 */ 1726 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF); 1727 dsched_exit_buf(bp); 1728 BUF_UNLOCK(bp); 1729 } 1730 1731 /* 1732 * Hold a buffer, preventing it from being reused. This will prevent 1733 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL 1734 * operations. If a B_INVAL operation occurs the buffer will remain held 1735 * but the underlying pages may get ripped out. 1736 * 1737 * These functions are typically used in VOP_READ/VOP_WRITE functions 1738 * to hold a buffer during a copyin or copyout, preventing deadlocks 1739 * or recursive lock panics when read()/write() is used over mmap()'d 1740 * space. 1741 * 1742 * NOTE: bqhold() requires that the buffer be locked at the time of the 1743 * hold. bqdrop() has no requirements other than the buffer having 1744 * previously been held. 1745 */ 1746 void 1747 bqhold(struct buf *bp) 1748 { 1749 atomic_add_int(&bp->b_refs, 1); 1750 } 1751 1752 void 1753 bqdrop(struct buf *bp) 1754 { 1755 KKASSERT(bp->b_refs > 0); 1756 atomic_add_int(&bp->b_refs, -1); 1757 } 1758 1759 /* 1760 * vfs_vmio_release: 1761 * 1762 * Return backing pages held by the buffer 'bp' back to the VM system 1763 * if possible. The pages are freed if they are no longer valid or 1764 * attempt to free if it was used for direct I/O otherwise they are 1765 * sent to the page cache. 1766 * 1767 * Pages that were marked busy are left alone and skipped. 1768 * 1769 * The KVA mapping (b_data) for the underlying pages is removed by 1770 * this function. 1771 */ 1772 static void 1773 vfs_vmio_release(struct buf *bp) 1774 { 1775 int i; 1776 vm_page_t m; 1777 1778 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1779 m = bp->b_xio.xio_pages[i]; 1780 bp->b_xio.xio_pages[i] = NULL; 1781 1782 vm_page_busy_wait(m, FALSE, "vmiopg"); 1783 1784 /* 1785 * The VFS is telling us this is not a meta-data buffer 1786 * even if it is backed by a block device. 1787 */ 1788 if (bp->b_flags & B_NOTMETA) 1789 vm_page_flag_set(m, PG_NOTMETA); 1790 1791 /* 1792 * This is a very important bit of code. We try to track 1793 * VM page use whether the pages are wired into the buffer 1794 * cache or not. While wired into the buffer cache the 1795 * bp tracks the act_count. 1796 * 1797 * We can choose to place unwired pages on the inactive 1798 * queue (0) or active queue (1). If we place too many 1799 * on the active queue the queue will cycle the act_count 1800 * on pages we'd like to keep, just from single-use pages 1801 * (such as when doing a tar-up or file scan). 1802 */ 1803 if (bp->b_act_count < vm_cycle_point) 1804 vm_page_unwire(m, 0); 1805 else 1806 vm_page_unwire(m, 1); 1807 1808 /* 1809 * We don't mess with busy pages, it is the responsibility 1810 * of the process that busied the pages to deal with them. 1811 * 1812 * However, the caller may have marked the page invalid and 1813 * we must still make sure the page is no longer mapped. 1814 */ 1815 if ((m->flags & PG_BUSY) || (m->busy != 0)) { 1816 vm_page_protect(m, VM_PROT_NONE); 1817 vm_page_wakeup(m); 1818 continue; 1819 } 1820 1821 if (m->wire_count == 0) { 1822 vm_page_flag_clear(m, PG_ZERO); 1823 /* 1824 * Might as well free the page if we can and it has 1825 * no valid data. We also free the page if the 1826 * buffer was used for direct I/O. 1827 */ 1828 #if 0 1829 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1830 m->hold_count == 0) { 1831 vm_page_protect(m, VM_PROT_NONE); 1832 vm_page_free(m); 1833 } else 1834 #endif 1835 /* 1836 * Cache the page if we are really low on free 1837 * pages. 1838 * 1839 * Also bypass the active and inactive queues 1840 * if B_NOTMETA is set. This flag is set by HAMMER 1841 * on a regular file buffer when double buffering 1842 * is enabled or on a block device buffer representing 1843 * file data when double buffering is not enabled. 1844 * The flag prevents two copies of the same data from 1845 * being cached for long periods of time. 1846 */ 1847 if (bp->b_flags & B_DIRECT) { 1848 vm_page_wakeup(m); 1849 vm_page_try_to_free(m); 1850 } else if ((bp->b_flags & B_NOTMETA) || 1851 vm_page_count_severe()) { 1852 m->act_count = bp->b_act_count; 1853 vm_page_wakeup(m); 1854 vm_page_try_to_cache(m); 1855 } else { 1856 m->act_count = bp->b_act_count; 1857 vm_page_wakeup(m); 1858 } 1859 } else { 1860 vm_page_wakeup(m); 1861 } 1862 } 1863 1864 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1865 bp->b_xio.xio_npages); 1866 if (bp->b_bufsize) { 1867 bufspacewakeup(); 1868 bp->b_bufsize = 0; 1869 } 1870 bp->b_xio.xio_npages = 0; 1871 bp->b_flags &= ~B_VMIO; 1872 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1873 if (bp->b_vp) 1874 brelvp(bp); 1875 } 1876 1877 /* 1878 * vfs_bio_awrite: 1879 * 1880 * Implement clustered async writes for clearing out B_DELWRI buffers. 1881 * This is much better then the old way of writing only one buffer at 1882 * a time. Note that we may not be presented with the buffers in the 1883 * correct order, so we search for the cluster in both directions. 1884 * 1885 * The buffer is locked on call. 1886 */ 1887 int 1888 vfs_bio_awrite(struct buf *bp) 1889 { 1890 int i; 1891 int j; 1892 off_t loffset = bp->b_loffset; 1893 struct vnode *vp = bp->b_vp; 1894 int nbytes; 1895 struct buf *bpa; 1896 int nwritten; 1897 int size; 1898 1899 /* 1900 * right now we support clustered writing only to regular files. If 1901 * we find a clusterable block we could be in the middle of a cluster 1902 * rather then at the beginning. 1903 * 1904 * NOTE: b_bio1 contains the logical loffset and is aliased 1905 * to b_loffset. b_bio2 contains the translated block number. 1906 */ 1907 if ((vp->v_type == VREG) && 1908 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1909 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1910 1911 size = vp->v_mount->mnt_stat.f_iosize; 1912 1913 for (i = size; i < MAXPHYS; i += size) { 1914 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) && 1915 BUF_REFCNT(bpa) == 0 && 1916 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1917 (B_DELWRI | B_CLUSTEROK)) && 1918 (bpa->b_bufsize == size)) { 1919 if ((bpa->b_bio2.bio_offset == NOOFFSET) || 1920 (bpa->b_bio2.bio_offset != 1921 bp->b_bio2.bio_offset + i)) 1922 break; 1923 } else { 1924 break; 1925 } 1926 } 1927 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) { 1928 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) && 1929 BUF_REFCNT(bpa) == 0 && 1930 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1931 (B_DELWRI | B_CLUSTEROK)) && 1932 (bpa->b_bufsize == size)) { 1933 if ((bpa->b_bio2.bio_offset == NOOFFSET) || 1934 (bpa->b_bio2.bio_offset != 1935 bp->b_bio2.bio_offset - j)) 1936 break; 1937 } else { 1938 break; 1939 } 1940 } 1941 j -= size; 1942 nbytes = (i + j); 1943 1944 /* 1945 * this is a possible cluster write 1946 */ 1947 if (nbytes != size) { 1948 BUF_UNLOCK(bp); 1949 nwritten = cluster_wbuild(vp, size, 1950 loffset - j, nbytes); 1951 return nwritten; 1952 } 1953 } 1954 1955 /* 1956 * default (old) behavior, writing out only one block 1957 * 1958 * XXX returns b_bufsize instead of b_bcount for nwritten? 1959 */ 1960 nwritten = bp->b_bufsize; 1961 bremfree(bp); 1962 bawrite(bp); 1963 1964 return nwritten; 1965 } 1966 1967 /* 1968 * getnewbuf: 1969 * 1970 * Find and initialize a new buffer header, freeing up existing buffers 1971 * in the bufqueues as necessary. The new buffer is returned locked. 1972 * 1973 * Important: B_INVAL is not set. If the caller wishes to throw the 1974 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1975 * 1976 * We block if: 1977 * We have insufficient buffer headers 1978 * We have insufficient buffer space 1979 * buffer_map is too fragmented ( space reservation fails ) 1980 * If we have to flush dirty buffers ( but we try to avoid this ) 1981 * 1982 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1983 * Instead we ask the buf daemon to do it for us. We attempt to 1984 * avoid piecemeal wakeups of the pageout daemon. 1985 * 1986 * MPALMOSTSAFE 1987 */ 1988 struct buf * 1989 getnewbuf(int blkflags, int slptimeo, int size, int maxsize) 1990 { 1991 struct buf *bp; 1992 struct buf *nbp; 1993 int defrag = 0; 1994 int nqindex; 1995 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1996 static int flushingbufs; 1997 1998 /* 1999 * We can't afford to block since we might be holding a vnode lock, 2000 * which may prevent system daemons from running. We deal with 2001 * low-memory situations by proactively returning memory and running 2002 * async I/O rather then sync I/O. 2003 */ 2004 2005 ++getnewbufcalls; 2006 --getnewbufrestarts; 2007 restart: 2008 ++getnewbufrestarts; 2009 2010 /* 2011 * Setup for scan. If we do not have enough free buffers, 2012 * we setup a degenerate case that immediately fails. Note 2013 * that if we are specially marked process, we are allowed to 2014 * dip into our reserves. 2015 * 2016 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2017 * 2018 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2019 * However, there are a number of cases (defragging, reusing, ...) 2020 * where we cannot backup. 2021 */ 2022 nqindex = BQUEUE_EMPTYKVA; 2023 spin_lock(&bufqspin); 2024 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]); 2025 2026 if (nbp == NULL) { 2027 /* 2028 * If no EMPTYKVA buffers and we are either 2029 * defragging or reusing, locate a CLEAN buffer 2030 * to free or reuse. If bufspace useage is low 2031 * skip this step so we can allocate a new buffer. 2032 */ 2033 if (defrag || bufspace >= lobufspace) { 2034 nqindex = BQUEUE_CLEAN; 2035 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 2036 } 2037 2038 /* 2039 * If we could not find or were not allowed to reuse a 2040 * CLEAN buffer, check to see if it is ok to use an EMPTY 2041 * buffer. We can only use an EMPTY buffer if allocating 2042 * its KVA would not otherwise run us out of buffer space. 2043 */ 2044 if (nbp == NULL && defrag == 0 && 2045 bufspace + maxsize < hibufspace) { 2046 nqindex = BQUEUE_EMPTY; 2047 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]); 2048 } 2049 } 2050 2051 /* 2052 * Run scan, possibly freeing data and/or kva mappings on the fly 2053 * depending. 2054 * 2055 * WARNING! bufqspin is held! 2056 */ 2057 while ((bp = nbp) != NULL) { 2058 int qindex = nqindex; 2059 2060 nbp = TAILQ_NEXT(bp, b_freelist); 2061 2062 /* 2063 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2064 * cycles through the queue twice before being selected. 2065 */ 2066 if (qindex == BQUEUE_CLEAN && 2067 (bp->b_flags & B_AGE) == 0 && nbp) { 2068 bp->b_flags |= B_AGE; 2069 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist); 2070 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist); 2071 continue; 2072 } 2073 2074 /* 2075 * Calculate next bp ( we can only use it if we do not block 2076 * or do other fancy things ). 2077 */ 2078 if (nbp == NULL) { 2079 switch(qindex) { 2080 case BQUEUE_EMPTY: 2081 nqindex = BQUEUE_EMPTYKVA; 2082 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]))) 2083 break; 2084 /* fall through */ 2085 case BQUEUE_EMPTYKVA: 2086 nqindex = BQUEUE_CLEAN; 2087 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]))) 2088 break; 2089 /* fall through */ 2090 case BQUEUE_CLEAN: 2091 /* 2092 * nbp is NULL. 2093 */ 2094 break; 2095 } 2096 } 2097 2098 /* 2099 * Sanity Checks 2100 */ 2101 KASSERT(bp->b_qindex == qindex, 2102 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2103 2104 /* 2105 * Note: we no longer distinguish between VMIO and non-VMIO 2106 * buffers. 2107 */ 2108 KASSERT((bp->b_flags & B_DELWRI) == 0, 2109 ("delwri buffer %p found in queue %d", bp, qindex)); 2110 2111 /* 2112 * Do not try to reuse a buffer with a non-zero b_refs. 2113 * This is an unsynchronized test. A synchronized test 2114 * is also performed after we lock the buffer. 2115 */ 2116 if (bp->b_refs) 2117 continue; 2118 2119 /* 2120 * If we are defragging then we need a buffer with 2121 * b_kvasize != 0. XXX this situation should no longer 2122 * occur, if defrag is non-zero the buffer's b_kvasize 2123 * should also be non-zero at this point. XXX 2124 */ 2125 if (defrag && bp->b_kvasize == 0) { 2126 kprintf("Warning: defrag empty buffer %p\n", bp); 2127 continue; 2128 } 2129 2130 /* 2131 * Start freeing the bp. This is somewhat involved. nbp 2132 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 2133 * on the clean list must be disassociated from their 2134 * current vnode. Buffers on the empty[kva] lists have 2135 * already been disassociated. 2136 * 2137 * b_refs is checked after locking along with queue changes. 2138 * We must check here to deal with zero->nonzero transitions 2139 * made by the owner of the buffer lock, which is used by 2140 * VFS's to hold the buffer while issuing an unlocked 2141 * uiomove()s. We cannot invalidate the buffer's pages 2142 * for this case. Once we successfully lock a buffer the 2143 * only 0->1 transitions of b_refs will occur via findblk(). 2144 * 2145 * We must also check for queue changes after successful 2146 * locking as the current lock holder may dispose of the 2147 * buffer and change its queue. 2148 */ 2149 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2150 spin_unlock(&bufqspin); 2151 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2152 goto restart; 2153 } 2154 if (bp->b_qindex != qindex || bp->b_refs) { 2155 spin_unlock(&bufqspin); 2156 BUF_UNLOCK(bp); 2157 goto restart; 2158 } 2159 bremfree_locked(bp); 2160 spin_unlock(&bufqspin); 2161 2162 /* 2163 * Dependancies must be handled before we disassociate the 2164 * vnode. 2165 * 2166 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2167 * be immediately disassociated. HAMMER then becomes 2168 * responsible for releasing the buffer. 2169 * 2170 * NOTE: bufqspin is UNLOCKED now. 2171 */ 2172 if (LIST_FIRST(&bp->b_dep) != NULL) { 2173 buf_deallocate(bp); 2174 if (bp->b_flags & B_LOCKED) { 2175 bqrelse(bp); 2176 goto restart; 2177 } 2178 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2179 } 2180 2181 if (qindex == BQUEUE_CLEAN) { 2182 if (bp->b_flags & B_VMIO) 2183 vfs_vmio_release(bp); 2184 if (bp->b_vp) 2185 brelvp(bp); 2186 } 2187 2188 /* 2189 * NOTE: nbp is now entirely invalid. We can only restart 2190 * the scan from this point on. 2191 * 2192 * Get the rest of the buffer freed up. b_kva* is still 2193 * valid after this operation. 2194 */ 2195 KASSERT(bp->b_vp == NULL, 2196 ("bp3 %p flags %08x vnode %p qindex %d " 2197 "unexpectededly still associated!", 2198 bp, bp->b_flags, bp->b_vp, qindex)); 2199 KKASSERT((bp->b_flags & B_HASHED) == 0); 2200 2201 /* 2202 * critical section protection is not required when 2203 * scrapping a buffer's contents because it is already 2204 * wired. 2205 */ 2206 if (bp->b_bufsize) 2207 allocbuf(bp, 0); 2208 2209 bp->b_flags = B_BNOCLIP; 2210 bp->b_cmd = BUF_CMD_DONE; 2211 bp->b_vp = NULL; 2212 bp->b_error = 0; 2213 bp->b_resid = 0; 2214 bp->b_bcount = 0; 2215 bp->b_xio.xio_npages = 0; 2216 bp->b_dirtyoff = bp->b_dirtyend = 0; 2217 bp->b_act_count = ACT_INIT; 2218 reinitbufbio(bp); 2219 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2220 buf_dep_init(bp); 2221 if (blkflags & GETBLK_BHEAVY) 2222 bp->b_flags |= B_HEAVY; 2223 2224 /* 2225 * If we are defragging then free the buffer. 2226 */ 2227 if (defrag) { 2228 bp->b_flags |= B_INVAL; 2229 bfreekva(bp); 2230 brelse(bp); 2231 defrag = 0; 2232 goto restart; 2233 } 2234 2235 /* 2236 * If we are overcomitted then recover the buffer and its 2237 * KVM space. This occurs in rare situations when multiple 2238 * processes are blocked in getnewbuf() or allocbuf(). 2239 */ 2240 if (bufspace >= hibufspace) 2241 flushingbufs = 1; 2242 if (flushingbufs && bp->b_kvasize != 0) { 2243 bp->b_flags |= B_INVAL; 2244 bfreekva(bp); 2245 brelse(bp); 2246 goto restart; 2247 } 2248 if (bufspace < lobufspace) 2249 flushingbufs = 0; 2250 2251 /* 2252 * b_refs can transition to a non-zero value while we hold 2253 * the buffer locked due to a findblk(). Our brelvp() above 2254 * interlocked any future possible transitions due to 2255 * findblk()s. 2256 * 2257 * If we find b_refs to be non-zero we can destroy the 2258 * buffer's contents but we cannot yet reuse the buffer. 2259 */ 2260 if (bp->b_refs) { 2261 bp->b_flags |= B_INVAL; 2262 bfreekva(bp); 2263 brelse(bp); 2264 goto restart; 2265 } 2266 break; 2267 /* NOT REACHED, bufqspin not held */ 2268 } 2269 2270 /* 2271 * If we exhausted our list, sleep as appropriate. We may have to 2272 * wakeup various daemons and write out some dirty buffers. 2273 * 2274 * Generally we are sleeping due to insufficient buffer space. 2275 * 2276 * NOTE: bufqspin is held if bp is NULL, else it is not held. 2277 */ 2278 if (bp == NULL) { 2279 int flags; 2280 char *waitmsg; 2281 2282 spin_unlock(&bufqspin); 2283 if (defrag) { 2284 flags = VFS_BIO_NEED_BUFSPACE; 2285 waitmsg = "nbufkv"; 2286 } else if (bufspace >= hibufspace) { 2287 waitmsg = "nbufbs"; 2288 flags = VFS_BIO_NEED_BUFSPACE; 2289 } else { 2290 waitmsg = "newbuf"; 2291 flags = VFS_BIO_NEED_ANY; 2292 } 2293 2294 bd_speedup(); /* heeeelp */ 2295 spin_lock(&bufcspin); 2296 needsbuffer |= flags; 2297 while (needsbuffer & flags) { 2298 if (ssleep(&needsbuffer, &bufcspin, 2299 slpflags, waitmsg, slptimeo)) { 2300 spin_unlock(&bufcspin); 2301 return (NULL); 2302 } 2303 } 2304 spin_unlock(&bufcspin); 2305 } else { 2306 /* 2307 * We finally have a valid bp. We aren't quite out of the 2308 * woods, we still have to reserve kva space. In order 2309 * to keep fragmentation sane we only allocate kva in 2310 * BKVASIZE chunks. 2311 * 2312 * (bufqspin is not held) 2313 */ 2314 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2315 2316 if (maxsize != bp->b_kvasize) { 2317 vm_offset_t addr = 0; 2318 int count; 2319 2320 bfreekva(bp); 2321 2322 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2323 vm_map_lock(&buffer_map); 2324 2325 if (vm_map_findspace(&buffer_map, 2326 vm_map_min(&buffer_map), maxsize, 2327 maxsize, 0, &addr)) { 2328 /* 2329 * Uh oh. Buffer map is too fragmented. We 2330 * must defragment the map. 2331 */ 2332 vm_map_unlock(&buffer_map); 2333 vm_map_entry_release(count); 2334 ++bufdefragcnt; 2335 defrag = 1; 2336 bp->b_flags |= B_INVAL; 2337 brelse(bp); 2338 goto restart; 2339 } 2340 if (addr) { 2341 vm_map_insert(&buffer_map, &count, 2342 NULL, 0, 2343 addr, addr + maxsize, 2344 VM_MAPTYPE_NORMAL, 2345 VM_PROT_ALL, VM_PROT_ALL, 2346 MAP_NOFAULT); 2347 2348 bp->b_kvabase = (caddr_t) addr; 2349 bp->b_kvasize = maxsize; 2350 bufspace += bp->b_kvasize; 2351 ++bufreusecnt; 2352 } 2353 vm_map_unlock(&buffer_map); 2354 vm_map_entry_release(count); 2355 } 2356 bp->b_data = bp->b_kvabase; 2357 } 2358 return(bp); 2359 } 2360 2361 /* 2362 * This routine is called in an emergency to recover VM pages from the 2363 * buffer cache by cashing in clean buffers. The idea is to recover 2364 * enough pages to be able to satisfy a stuck bio_page_alloc(). 2365 * 2366 * MPSAFE 2367 */ 2368 static int 2369 recoverbufpages(void) 2370 { 2371 struct buf *bp; 2372 int bytes = 0; 2373 2374 ++recoverbufcalls; 2375 2376 spin_lock(&bufqspin); 2377 while (bytes < MAXBSIZE) { 2378 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 2379 if (bp == NULL) 2380 break; 2381 2382 /* 2383 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2384 * cycles through the queue twice before being selected. 2385 */ 2386 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) { 2387 bp->b_flags |= B_AGE; 2388 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 2389 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], 2390 bp, b_freelist); 2391 continue; 2392 } 2393 2394 /* 2395 * Sanity Checks 2396 */ 2397 KKASSERT(bp->b_qindex == BQUEUE_CLEAN); 2398 KKASSERT((bp->b_flags & B_DELWRI) == 0); 2399 2400 /* 2401 * Start freeing the bp. This is somewhat involved. 2402 * 2403 * Buffers on the clean list must be disassociated from 2404 * their current vnode 2405 */ 2406 2407 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2408 kprintf("recoverbufpages: warning, locked buf %p, " 2409 "race corrected\n", 2410 bp); 2411 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100); 2412 continue; 2413 } 2414 if (bp->b_qindex != BQUEUE_CLEAN) { 2415 kprintf("recoverbufpages: warning, BUF_LOCK blocked " 2416 "unexpectedly on buf %p index %d, race " 2417 "corrected\n", 2418 bp, bp->b_qindex); 2419 BUF_UNLOCK(bp); 2420 continue; 2421 } 2422 bremfree_locked(bp); 2423 spin_unlock(&bufqspin); 2424 2425 /* 2426 * Dependancies must be handled before we disassociate the 2427 * vnode. 2428 * 2429 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2430 * be immediately disassociated. HAMMER then becomes 2431 * responsible for releasing the buffer. 2432 */ 2433 if (LIST_FIRST(&bp->b_dep) != NULL) { 2434 buf_deallocate(bp); 2435 if (bp->b_flags & B_LOCKED) { 2436 bqrelse(bp); 2437 spin_lock(&bufqspin); 2438 continue; 2439 } 2440 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2441 } 2442 2443 bytes += bp->b_bufsize; 2444 2445 if (bp->b_flags & B_VMIO) { 2446 bp->b_flags |= B_DIRECT; /* try to free pages */ 2447 vfs_vmio_release(bp); 2448 } 2449 if (bp->b_vp) 2450 brelvp(bp); 2451 2452 KKASSERT(bp->b_vp == NULL); 2453 KKASSERT((bp->b_flags & B_HASHED) == 0); 2454 2455 /* 2456 * critical section protection is not required when 2457 * scrapping a buffer's contents because it is already 2458 * wired. 2459 */ 2460 if (bp->b_bufsize) 2461 allocbuf(bp, 0); 2462 2463 bp->b_flags = B_BNOCLIP; 2464 bp->b_cmd = BUF_CMD_DONE; 2465 bp->b_vp = NULL; 2466 bp->b_error = 0; 2467 bp->b_resid = 0; 2468 bp->b_bcount = 0; 2469 bp->b_xio.xio_npages = 0; 2470 bp->b_dirtyoff = bp->b_dirtyend = 0; 2471 reinitbufbio(bp); 2472 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2473 buf_dep_init(bp); 2474 bp->b_flags |= B_INVAL; 2475 /* bfreekva(bp); */ 2476 brelse(bp); 2477 spin_lock(&bufqspin); 2478 } 2479 spin_unlock(&bufqspin); 2480 return(bytes); 2481 } 2482 2483 /* 2484 * buf_daemon: 2485 * 2486 * Buffer flushing daemon. Buffers are normally flushed by the 2487 * update daemon but if it cannot keep up this process starts to 2488 * take the load in an attempt to prevent getnewbuf() from blocking. 2489 * 2490 * Once a flush is initiated it does not stop until the number 2491 * of buffers falls below lodirtybuffers, but we will wake up anyone 2492 * waiting at the mid-point. 2493 */ 2494 2495 static struct kproc_desc buf_kp = { 2496 "bufdaemon", 2497 buf_daemon, 2498 &bufdaemon_td 2499 }; 2500 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2501 kproc_start, &buf_kp) 2502 2503 static struct kproc_desc bufhw_kp = { 2504 "bufdaemon_hw", 2505 buf_daemon_hw, 2506 &bufdaemonhw_td 2507 }; 2508 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2509 kproc_start, &bufhw_kp) 2510 2511 /* 2512 * MPSAFE thread 2513 */ 2514 static void 2515 buf_daemon(void) 2516 { 2517 int limit; 2518 2519 /* 2520 * This process needs to be suspended prior to shutdown sync. 2521 */ 2522 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2523 bufdaemon_td, SHUTDOWN_PRI_LAST); 2524 curthread->td_flags |= TDF_SYSTHREAD; 2525 2526 /* 2527 * This process is allowed to take the buffer cache to the limit 2528 */ 2529 for (;;) { 2530 kproc_suspend_loop(); 2531 2532 /* 2533 * Do the flush as long as the number of dirty buffers 2534 * (including those running) exceeds lodirtybufspace. 2535 * 2536 * When flushing limit running I/O to hirunningspace 2537 * Do the flush. Limit the amount of in-transit I/O we 2538 * allow to build up, otherwise we would completely saturate 2539 * the I/O system. Wakeup any waiting processes before we 2540 * normally would so they can run in parallel with our drain. 2541 * 2542 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2543 * but because we split the operation into two threads we 2544 * have to cut it in half for each thread. 2545 */ 2546 waitrunningbufspace(); 2547 limit = lodirtybufspace / 2; 2548 while (runningbufspace + dirtybufspace > limit || 2549 dirtybufcount - dirtybufcounthw >= nbuf / 2) { 2550 if (flushbufqueues(BQUEUE_DIRTY) == 0) 2551 break; 2552 if (runningbufspace < hirunningspace) 2553 continue; 2554 waitrunningbufspace(); 2555 } 2556 2557 /* 2558 * We reached our low water mark, reset the 2559 * request and sleep until we are needed again. 2560 * The sleep is just so the suspend code works. 2561 */ 2562 spin_lock(&bufcspin); 2563 if (bd_request == 0) 2564 ssleep(&bd_request, &bufcspin, 0, "psleep", hz); 2565 bd_request = 0; 2566 spin_unlock(&bufcspin); 2567 } 2568 } 2569 2570 /* 2571 * MPSAFE thread 2572 */ 2573 static void 2574 buf_daemon_hw(void) 2575 { 2576 int limit; 2577 2578 /* 2579 * This process needs to be suspended prior to shutdown sync. 2580 */ 2581 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2582 bufdaemonhw_td, SHUTDOWN_PRI_LAST); 2583 curthread->td_flags |= TDF_SYSTHREAD; 2584 2585 /* 2586 * This process is allowed to take the buffer cache to the limit 2587 */ 2588 for (;;) { 2589 kproc_suspend_loop(); 2590 2591 /* 2592 * Do the flush. Limit the amount of in-transit I/O we 2593 * allow to build up, otherwise we would completely saturate 2594 * the I/O system. Wakeup any waiting processes before we 2595 * normally would so they can run in parallel with our drain. 2596 * 2597 * Once we decide to flush push the queued I/O up to 2598 * hirunningspace in order to trigger bursting by the bioq 2599 * subsystem. 2600 * 2601 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2602 * but because we split the operation into two threads we 2603 * have to cut it in half for each thread. 2604 */ 2605 waitrunningbufspace(); 2606 limit = lodirtybufspace / 2; 2607 while (runningbufspace + dirtybufspacehw > limit || 2608 dirtybufcounthw >= nbuf / 2) { 2609 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0) 2610 break; 2611 if (runningbufspace < hirunningspace) 2612 continue; 2613 waitrunningbufspace(); 2614 } 2615 2616 /* 2617 * We reached our low water mark, reset the 2618 * request and sleep until we are needed again. 2619 * The sleep is just so the suspend code works. 2620 */ 2621 spin_lock(&bufcspin); 2622 if (bd_request_hw == 0) 2623 ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz); 2624 bd_request_hw = 0; 2625 spin_unlock(&bufcspin); 2626 } 2627 } 2628 2629 /* 2630 * flushbufqueues: 2631 * 2632 * Try to flush a buffer in the dirty queue. We must be careful to 2633 * free up B_INVAL buffers instead of write them, which NFS is 2634 * particularly sensitive to. 2635 * 2636 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate 2637 * that we really want to try to get the buffer out and reuse it 2638 * due to the write load on the machine. 2639 * 2640 * We must lock the buffer in order to check its validity before we 2641 * can mess with its contents. bufqspin isn't enough. 2642 */ 2643 static int 2644 flushbufqueues(bufq_type_t q) 2645 { 2646 struct buf *bp; 2647 int r = 0; 2648 int spun; 2649 2650 spin_lock(&bufqspin); 2651 spun = 1; 2652 2653 bp = TAILQ_FIRST(&bufqueues[q]); 2654 while (bp) { 2655 if ((bp->b_flags & B_DELWRI) == 0) { 2656 kprintf("Unexpected clean buffer %p\n", bp); 2657 bp = TAILQ_NEXT(bp, b_freelist); 2658 continue; 2659 } 2660 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2661 bp = TAILQ_NEXT(bp, b_freelist); 2662 continue; 2663 } 2664 KKASSERT(bp->b_qindex == q); 2665 2666 /* 2667 * Must recheck B_DELWRI after successfully locking 2668 * the buffer. 2669 */ 2670 if ((bp->b_flags & B_DELWRI) == 0) { 2671 BUF_UNLOCK(bp); 2672 bp = TAILQ_NEXT(bp, b_freelist); 2673 continue; 2674 } 2675 2676 if (bp->b_flags & B_INVAL) { 2677 _bremfree(bp); 2678 spin_unlock(&bufqspin); 2679 spun = 0; 2680 brelse(bp); 2681 ++r; 2682 break; 2683 } 2684 2685 spin_unlock(&bufqspin); 2686 lwkt_yield(); 2687 spun = 0; 2688 2689 if (LIST_FIRST(&bp->b_dep) != NULL && 2690 (bp->b_flags & B_DEFERRED) == 0 && 2691 buf_countdeps(bp, 0)) { 2692 spin_lock(&bufqspin); 2693 spun = 1; 2694 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist); 2695 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist); 2696 bp->b_flags |= B_DEFERRED; 2697 BUF_UNLOCK(bp); 2698 bp = TAILQ_FIRST(&bufqueues[q]); 2699 continue; 2700 } 2701 2702 /* 2703 * If the buffer has a dependancy, buf_checkwrite() must 2704 * also return 0 for us to be able to initate the write. 2705 * 2706 * If the buffer is flagged B_ERROR it may be requeued 2707 * over and over again, we try to avoid a live lock. 2708 * 2709 * NOTE: buf_checkwrite is MPSAFE. 2710 */ 2711 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2712 bremfree(bp); 2713 brelse(bp); 2714 } else if (bp->b_flags & B_ERROR) { 2715 tsleep(bp, 0, "bioer", 1); 2716 bp->b_flags &= ~B_AGE; 2717 vfs_bio_awrite(bp); 2718 } else { 2719 bp->b_flags |= B_AGE; 2720 vfs_bio_awrite(bp); 2721 } 2722 ++r; 2723 break; 2724 } 2725 if (spun) 2726 spin_unlock(&bufqspin); 2727 return (r); 2728 } 2729 2730 /* 2731 * inmem: 2732 * 2733 * Returns true if no I/O is needed to access the associated VM object. 2734 * This is like findblk except it also hunts around in the VM system for 2735 * the data. 2736 * 2737 * Note that we ignore vm_page_free() races from interrupts against our 2738 * lookup, since if the caller is not protected our return value will not 2739 * be any more valid then otherwise once we exit the critical section. 2740 */ 2741 int 2742 inmem(struct vnode *vp, off_t loffset) 2743 { 2744 vm_object_t obj; 2745 vm_offset_t toff, tinc, size; 2746 vm_page_t m; 2747 int res = 1; 2748 2749 if (findblk(vp, loffset, FINDBLK_TEST)) 2750 return 1; 2751 if (vp->v_mount == NULL) 2752 return 0; 2753 if ((obj = vp->v_object) == NULL) 2754 return 0; 2755 2756 size = PAGE_SIZE; 2757 if (size > vp->v_mount->mnt_stat.f_iosize) 2758 size = vp->v_mount->mnt_stat.f_iosize; 2759 2760 vm_object_hold(obj); 2761 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2762 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2763 if (m == NULL) { 2764 res = 0; 2765 break; 2766 } 2767 tinc = size; 2768 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2769 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2770 if (vm_page_is_valid(m, 2771 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2772 res = 0; 2773 break; 2774 } 2775 } 2776 vm_object_drop(obj); 2777 return (res); 2778 } 2779 2780 /* 2781 * findblk: 2782 * 2783 * Locate and return the specified buffer. Unless flagged otherwise, 2784 * a locked buffer will be returned if it exists or NULL if it does not. 2785 * 2786 * findblk()'d buffers are still on the bufqueues and if you intend 2787 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2788 * and possibly do other stuff to it. 2789 * 2790 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2791 * for locking the buffer and ensuring that it remains 2792 * the desired buffer after locking. 2793 * 2794 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2795 * to acquire the lock we return NULL, even if the 2796 * buffer exists. 2797 * 2798 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2799 * reuse by getnewbuf() but does not prevent 2800 * disassociation (B_INVAL). Used to avoid deadlocks 2801 * against random (vp,loffset)s due to reassignment. 2802 * 2803 * (0) - Lock the buffer blocking. 2804 * 2805 * MPSAFE 2806 */ 2807 struct buf * 2808 findblk(struct vnode *vp, off_t loffset, int flags) 2809 { 2810 struct buf *bp; 2811 int lkflags; 2812 2813 lkflags = LK_EXCLUSIVE; 2814 if (flags & FINDBLK_NBLOCK) 2815 lkflags |= LK_NOWAIT; 2816 2817 for (;;) { 2818 /* 2819 * Lookup. Ref the buf while holding v_token to prevent 2820 * reuse (but does not prevent diassociation). 2821 */ 2822 lwkt_gettoken(&vp->v_token); 2823 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2824 if (bp == NULL) { 2825 lwkt_reltoken(&vp->v_token); 2826 return(NULL); 2827 } 2828 bqhold(bp); 2829 lwkt_reltoken(&vp->v_token); 2830 2831 /* 2832 * If testing only break and return bp, do not lock. 2833 */ 2834 if (flags & FINDBLK_TEST) 2835 break; 2836 2837 /* 2838 * Lock the buffer, return an error if the lock fails. 2839 * (only FINDBLK_NBLOCK can cause the lock to fail). 2840 */ 2841 if (BUF_LOCK(bp, lkflags)) { 2842 atomic_subtract_int(&bp->b_refs, 1); 2843 /* bp = NULL; not needed */ 2844 return(NULL); 2845 } 2846 2847 /* 2848 * Revalidate the locked buf before allowing it to be 2849 * returned. 2850 */ 2851 if (bp->b_vp == vp && bp->b_loffset == loffset) 2852 break; 2853 atomic_subtract_int(&bp->b_refs, 1); 2854 BUF_UNLOCK(bp); 2855 } 2856 2857 /* 2858 * Success 2859 */ 2860 if ((flags & FINDBLK_REF) == 0) 2861 atomic_subtract_int(&bp->b_refs, 1); 2862 return(bp); 2863 } 2864 2865 /* 2866 * getcacheblk: 2867 * 2868 * Similar to getblk() except only returns the buffer if it is 2869 * B_CACHE and requires no other manipulation. Otherwise NULL 2870 * is returned. 2871 * 2872 * If B_RAM is set the buffer might be just fine, but we return 2873 * NULL anyway because we want the code to fall through to the 2874 * cluster read. Otherwise read-ahead breaks. 2875 * 2876 * If blksize is 0 the buffer cache buffer must already be fully 2877 * cached. 2878 * 2879 * If blksize is non-zero getblk() will be used, allowing a buffer 2880 * to be reinstantiated from its VM backing store. The buffer must 2881 * still be fully cached after reinstantiation to be returned. 2882 */ 2883 struct buf * 2884 getcacheblk(struct vnode *vp, off_t loffset, int blksize) 2885 { 2886 struct buf *bp; 2887 2888 if (blksize) { 2889 bp = getblk(vp, loffset, blksize, 0, 0); 2890 if (bp) { 2891 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2892 B_CACHE) { 2893 bp->b_flags &= ~B_AGE; 2894 } else { 2895 brelse(bp); 2896 bp = NULL; 2897 } 2898 } 2899 } else { 2900 bp = findblk(vp, loffset, 0); 2901 if (bp) { 2902 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2903 B_CACHE) { 2904 bp->b_flags &= ~B_AGE; 2905 bremfree(bp); 2906 } else { 2907 BUF_UNLOCK(bp); 2908 bp = NULL; 2909 } 2910 } 2911 } 2912 return (bp); 2913 } 2914 2915 /* 2916 * getblk: 2917 * 2918 * Get a block given a specified block and offset into a file/device. 2919 * B_INVAL may or may not be set on return. The caller should clear 2920 * B_INVAL prior to initiating a READ. 2921 * 2922 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2923 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2924 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2925 * without doing any of those things the system will likely believe 2926 * the buffer to be valid (especially if it is not B_VMIO), and the 2927 * next getblk() will return the buffer with B_CACHE set. 2928 * 2929 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2930 * an existing buffer. 2931 * 2932 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2933 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2934 * and then cleared based on the backing VM. If the previous buffer is 2935 * non-0-sized but invalid, B_CACHE will be cleared. 2936 * 2937 * If getblk() must create a new buffer, the new buffer is returned with 2938 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2939 * case it is returned with B_INVAL clear and B_CACHE set based on the 2940 * backing VM. 2941 * 2942 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2943 * B_CACHE bit is clear. 2944 * 2945 * What this means, basically, is that the caller should use B_CACHE to 2946 * determine whether the buffer is fully valid or not and should clear 2947 * B_INVAL prior to issuing a read. If the caller intends to validate 2948 * the buffer by loading its data area with something, the caller needs 2949 * to clear B_INVAL. If the caller does this without issuing an I/O, 2950 * the caller should set B_CACHE ( as an optimization ), else the caller 2951 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2952 * a write attempt or if it was a successfull read. If the caller 2953 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2954 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2955 * 2956 * getblk flags: 2957 * 2958 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2959 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2960 * 2961 * MPALMOSTSAFE 2962 */ 2963 struct buf * 2964 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2965 { 2966 struct buf *bp; 2967 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2968 int error; 2969 int lkflags; 2970 2971 if (size > MAXBSIZE) 2972 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2973 if (vp->v_object == NULL) 2974 panic("getblk: vnode %p has no object!", vp); 2975 2976 loop: 2977 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2978 /* 2979 * The buffer was found in the cache, but we need to lock it. 2980 * We must acquire a ref on the bp to prevent reuse, but 2981 * this will not prevent disassociation (brelvp()) so we 2982 * must recheck (vp,loffset) after acquiring the lock. 2983 * 2984 * Without the ref the buffer could potentially be reused 2985 * before we acquire the lock and create a deadlock 2986 * situation between the thread trying to reuse the buffer 2987 * and us due to the fact that we would wind up blocking 2988 * on a random (vp,loffset). 2989 */ 2990 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2991 if (blkflags & GETBLK_NOWAIT) { 2992 bqdrop(bp); 2993 return(NULL); 2994 } 2995 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2996 if (blkflags & GETBLK_PCATCH) 2997 lkflags |= LK_PCATCH; 2998 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2999 if (error) { 3000 bqdrop(bp); 3001 if (error == ENOLCK) 3002 goto loop; 3003 return (NULL); 3004 } 3005 /* buffer may have changed on us */ 3006 } 3007 bqdrop(bp); 3008 3009 /* 3010 * Once the buffer has been locked, make sure we didn't race 3011 * a buffer recyclement. Buffers that are no longer hashed 3012 * will have b_vp == NULL, so this takes care of that check 3013 * as well. 3014 */ 3015 if (bp->b_vp != vp || bp->b_loffset != loffset) { 3016 kprintf("Warning buffer %p (vp %p loffset %lld) " 3017 "was recycled\n", 3018 bp, vp, (long long)loffset); 3019 BUF_UNLOCK(bp); 3020 goto loop; 3021 } 3022 3023 /* 3024 * If SZMATCH any pre-existing buffer must be of the requested 3025 * size or NULL is returned. The caller absolutely does not 3026 * want getblk() to bwrite() the buffer on a size mismatch. 3027 */ 3028 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 3029 BUF_UNLOCK(bp); 3030 return(NULL); 3031 } 3032 3033 /* 3034 * All vnode-based buffers must be backed by a VM object. 3035 */ 3036 KKASSERT(bp->b_flags & B_VMIO); 3037 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3038 bp->b_flags &= ~B_AGE; 3039 3040 /* 3041 * Make sure that B_INVAL buffers do not have a cached 3042 * block number translation. 3043 */ 3044 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 3045 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 3046 " did not have cleared bio_offset cache\n", 3047 bp, vp, (long long)loffset); 3048 clearbiocache(&bp->b_bio2); 3049 } 3050 3051 /* 3052 * The buffer is locked. B_CACHE is cleared if the buffer is 3053 * invalid. 3054 */ 3055 if (bp->b_flags & B_INVAL) 3056 bp->b_flags &= ~B_CACHE; 3057 bremfree(bp); 3058 3059 /* 3060 * Any size inconsistancy with a dirty buffer or a buffer 3061 * with a softupdates dependancy must be resolved. Resizing 3062 * the buffer in such circumstances can lead to problems. 3063 * 3064 * Dirty or dependant buffers are written synchronously. 3065 * Other types of buffers are simply released and 3066 * reconstituted as they may be backed by valid, dirty VM 3067 * pages (but not marked B_DELWRI). 3068 * 3069 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 3070 * and may be left over from a prior truncation (and thus 3071 * no longer represent the actual EOF point), so we 3072 * definitely do not want to B_NOCACHE the backing store. 3073 */ 3074 if (size != bp->b_bcount) { 3075 if (bp->b_flags & B_DELWRI) { 3076 bp->b_flags |= B_RELBUF; 3077 bwrite(bp); 3078 } else if (LIST_FIRST(&bp->b_dep)) { 3079 bp->b_flags |= B_RELBUF; 3080 bwrite(bp); 3081 } else { 3082 bp->b_flags |= B_RELBUF; 3083 brelse(bp); 3084 } 3085 goto loop; 3086 } 3087 KKASSERT(size <= bp->b_kvasize); 3088 KASSERT(bp->b_loffset != NOOFFSET, 3089 ("getblk: no buffer offset")); 3090 3091 /* 3092 * A buffer with B_DELWRI set and B_CACHE clear must 3093 * be committed before we can return the buffer in 3094 * order to prevent the caller from issuing a read 3095 * ( due to B_CACHE not being set ) and overwriting 3096 * it. 3097 * 3098 * Most callers, including NFS and FFS, need this to 3099 * operate properly either because they assume they 3100 * can issue a read if B_CACHE is not set, or because 3101 * ( for example ) an uncached B_DELWRI might loop due 3102 * to softupdates re-dirtying the buffer. In the latter 3103 * case, B_CACHE is set after the first write completes, 3104 * preventing further loops. 3105 * 3106 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3107 * above while extending the buffer, we cannot allow the 3108 * buffer to remain with B_CACHE set after the write 3109 * completes or it will represent a corrupt state. To 3110 * deal with this we set B_NOCACHE to scrap the buffer 3111 * after the write. 3112 * 3113 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3114 * I'm not even sure this state is still possible 3115 * now that getblk() writes out any dirty buffers 3116 * on size changes. 3117 * 3118 * We might be able to do something fancy, like setting 3119 * B_CACHE in bwrite() except if B_DELWRI is already set, 3120 * so the below call doesn't set B_CACHE, but that gets real 3121 * confusing. This is much easier. 3122 */ 3123 3124 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3125 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3126 "and CACHE clear, b_flags %08x\n", 3127 bp, (intmax_t)bp->b_loffset, bp->b_flags); 3128 bp->b_flags |= B_NOCACHE; 3129 bwrite(bp); 3130 goto loop; 3131 } 3132 } else { 3133 /* 3134 * Buffer is not in-core, create new buffer. The buffer 3135 * returned by getnewbuf() is locked. Note that the returned 3136 * buffer is also considered valid (not marked B_INVAL). 3137 * 3138 * Calculating the offset for the I/O requires figuring out 3139 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3140 * the mount's f_iosize otherwise. If the vnode does not 3141 * have an associated mount we assume that the passed size is 3142 * the block size. 3143 * 3144 * Note that vn_isdisk() cannot be used here since it may 3145 * return a failure for numerous reasons. Note that the 3146 * buffer size may be larger then the block size (the caller 3147 * will use block numbers with the proper multiple). Beware 3148 * of using any v_* fields which are part of unions. In 3149 * particular, in DragonFly the mount point overloading 3150 * mechanism uses the namecache only and the underlying 3151 * directory vnode is not a special case. 3152 */ 3153 int bsize, maxsize; 3154 3155 if (vp->v_type == VBLK || vp->v_type == VCHR) 3156 bsize = DEV_BSIZE; 3157 else if (vp->v_mount) 3158 bsize = vp->v_mount->mnt_stat.f_iosize; 3159 else 3160 bsize = size; 3161 3162 maxsize = size + (loffset & PAGE_MASK); 3163 maxsize = imax(maxsize, bsize); 3164 3165 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3166 if (bp == NULL) { 3167 if (slpflags || slptimeo) 3168 return NULL; 3169 goto loop; 3170 } 3171 3172 /* 3173 * Atomically insert the buffer into the hash, so that it can 3174 * be found by findblk(). 3175 * 3176 * If bgetvp() returns non-zero a collision occured, and the 3177 * bp will not be associated with the vnode. 3178 * 3179 * Make sure the translation layer has been cleared. 3180 */ 3181 bp->b_loffset = loffset; 3182 bp->b_bio2.bio_offset = NOOFFSET; 3183 /* bp->b_bio2.bio_next = NULL; */ 3184 3185 if (bgetvp(vp, bp, size)) { 3186 bp->b_flags |= B_INVAL; 3187 brelse(bp); 3188 goto loop; 3189 } 3190 3191 /* 3192 * All vnode-based buffers must be backed by a VM object. 3193 */ 3194 KKASSERT(vp->v_object != NULL); 3195 bp->b_flags |= B_VMIO; 3196 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3197 3198 allocbuf(bp, size); 3199 } 3200 KKASSERT(dsched_is_clear_buf_priv(bp)); 3201 return (bp); 3202 } 3203 3204 /* 3205 * regetblk(bp) 3206 * 3207 * Reacquire a buffer that was previously released to the locked queue, 3208 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3209 * set B_LOCKED (which handles the acquisition race). 3210 * 3211 * To this end, either B_LOCKED must be set or the dependancy list must be 3212 * non-empty. 3213 * 3214 * MPSAFE 3215 */ 3216 void 3217 regetblk(struct buf *bp) 3218 { 3219 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3220 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3221 bremfree(bp); 3222 } 3223 3224 /* 3225 * geteblk: 3226 * 3227 * Get an empty, disassociated buffer of given size. The buffer is 3228 * initially set to B_INVAL. 3229 * 3230 * critical section protection is not required for the allocbuf() 3231 * call because races are impossible here. 3232 * 3233 * MPALMOSTSAFE 3234 */ 3235 struct buf * 3236 geteblk(int size) 3237 { 3238 struct buf *bp; 3239 int maxsize; 3240 3241 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3242 3243 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 3244 ; 3245 allocbuf(bp, size); 3246 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3247 KKASSERT(dsched_is_clear_buf_priv(bp)); 3248 return (bp); 3249 } 3250 3251 3252 /* 3253 * allocbuf: 3254 * 3255 * This code constitutes the buffer memory from either anonymous system 3256 * memory (in the case of non-VMIO operations) or from an associated 3257 * VM object (in the case of VMIO operations). This code is able to 3258 * resize a buffer up or down. 3259 * 3260 * Note that this code is tricky, and has many complications to resolve 3261 * deadlock or inconsistant data situations. Tread lightly!!! 3262 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3263 * the caller. Calling this code willy nilly can result in the loss of 3264 * data. 3265 * 3266 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3267 * B_CACHE for the non-VMIO case. 3268 * 3269 * This routine does not need to be called from a critical section but you 3270 * must own the buffer. 3271 * 3272 * MPSAFE 3273 */ 3274 int 3275 allocbuf(struct buf *bp, int size) 3276 { 3277 int newbsize, mbsize; 3278 int i; 3279 3280 if (BUF_REFCNT(bp) == 0) 3281 panic("allocbuf: buffer not busy"); 3282 3283 if (bp->b_kvasize < size) 3284 panic("allocbuf: buffer too small"); 3285 3286 if ((bp->b_flags & B_VMIO) == 0) { 3287 caddr_t origbuf; 3288 int origbufsize; 3289 /* 3290 * Just get anonymous memory from the kernel. Don't 3291 * mess with B_CACHE. 3292 */ 3293 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3294 if (bp->b_flags & B_MALLOC) 3295 newbsize = mbsize; 3296 else 3297 newbsize = round_page(size); 3298 3299 if (newbsize < bp->b_bufsize) { 3300 /* 3301 * Malloced buffers are not shrunk 3302 */ 3303 if (bp->b_flags & B_MALLOC) { 3304 if (newbsize) { 3305 bp->b_bcount = size; 3306 } else { 3307 kfree(bp->b_data, M_BIOBUF); 3308 if (bp->b_bufsize) { 3309 atomic_subtract_int(&bufmallocspace, bp->b_bufsize); 3310 bufspacewakeup(); 3311 bp->b_bufsize = 0; 3312 } 3313 bp->b_data = bp->b_kvabase; 3314 bp->b_bcount = 0; 3315 bp->b_flags &= ~B_MALLOC; 3316 } 3317 return 1; 3318 } 3319 vm_hold_free_pages( 3320 bp, 3321 (vm_offset_t) bp->b_data + newbsize, 3322 (vm_offset_t) bp->b_data + bp->b_bufsize); 3323 } else if (newbsize > bp->b_bufsize) { 3324 /* 3325 * We only use malloced memory on the first allocation. 3326 * and revert to page-allocated memory when the buffer 3327 * grows. 3328 */ 3329 if ((bufmallocspace < maxbufmallocspace) && 3330 (bp->b_bufsize == 0) && 3331 (mbsize <= PAGE_SIZE/2)) { 3332 3333 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3334 bp->b_bufsize = mbsize; 3335 bp->b_bcount = size; 3336 bp->b_flags |= B_MALLOC; 3337 atomic_add_int(&bufmallocspace, mbsize); 3338 return 1; 3339 } 3340 origbuf = NULL; 3341 origbufsize = 0; 3342 /* 3343 * If the buffer is growing on its other-than-first 3344 * allocation, then we revert to the page-allocation 3345 * scheme. 3346 */ 3347 if (bp->b_flags & B_MALLOC) { 3348 origbuf = bp->b_data; 3349 origbufsize = bp->b_bufsize; 3350 bp->b_data = bp->b_kvabase; 3351 if (bp->b_bufsize) { 3352 atomic_subtract_int(&bufmallocspace, 3353 bp->b_bufsize); 3354 bufspacewakeup(); 3355 bp->b_bufsize = 0; 3356 } 3357 bp->b_flags &= ~B_MALLOC; 3358 newbsize = round_page(newbsize); 3359 } 3360 vm_hold_load_pages( 3361 bp, 3362 (vm_offset_t) bp->b_data + bp->b_bufsize, 3363 (vm_offset_t) bp->b_data + newbsize); 3364 if (origbuf) { 3365 bcopy(origbuf, bp->b_data, origbufsize); 3366 kfree(origbuf, M_BIOBUF); 3367 } 3368 } 3369 } else { 3370 vm_page_t m; 3371 int desiredpages; 3372 3373 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3374 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3375 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3376 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3377 3378 if (bp->b_flags & B_MALLOC) 3379 panic("allocbuf: VMIO buffer can't be malloced"); 3380 /* 3381 * Set B_CACHE initially if buffer is 0 length or will become 3382 * 0-length. 3383 */ 3384 if (size == 0 || bp->b_bufsize == 0) 3385 bp->b_flags |= B_CACHE; 3386 3387 if (newbsize < bp->b_bufsize) { 3388 /* 3389 * DEV_BSIZE aligned new buffer size is less then the 3390 * DEV_BSIZE aligned existing buffer size. Figure out 3391 * if we have to remove any pages. 3392 */ 3393 if (desiredpages < bp->b_xio.xio_npages) { 3394 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3395 /* 3396 * the page is not freed here -- it 3397 * is the responsibility of 3398 * vnode_pager_setsize 3399 */ 3400 m = bp->b_xio.xio_pages[i]; 3401 KASSERT(m != bogus_page, 3402 ("allocbuf: bogus page found")); 3403 vm_page_busy_wait(m, TRUE, "biodep"); 3404 bp->b_xio.xio_pages[i] = NULL; 3405 vm_page_unwire(m, 0); 3406 vm_page_wakeup(m); 3407 } 3408 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3409 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3410 bp->b_xio.xio_npages = desiredpages; 3411 } 3412 } else if (size > bp->b_bcount) { 3413 /* 3414 * We are growing the buffer, possibly in a 3415 * byte-granular fashion. 3416 */ 3417 struct vnode *vp; 3418 vm_object_t obj; 3419 vm_offset_t toff; 3420 vm_offset_t tinc; 3421 3422 /* 3423 * Step 1, bring in the VM pages from the object, 3424 * allocating them if necessary. We must clear 3425 * B_CACHE if these pages are not valid for the 3426 * range covered by the buffer. 3427 * 3428 * critical section protection is required to protect 3429 * against interrupts unbusying and freeing pages 3430 * between our vm_page_lookup() and our 3431 * busycheck/wiring call. 3432 */ 3433 vp = bp->b_vp; 3434 obj = vp->v_object; 3435 3436 vm_object_hold(obj); 3437 while (bp->b_xio.xio_npages < desiredpages) { 3438 vm_page_t m; 3439 vm_pindex_t pi; 3440 int error; 3441 3442 pi = OFF_TO_IDX(bp->b_loffset) + 3443 bp->b_xio.xio_npages; 3444 3445 /* 3446 * Blocking on m->busy might lead to a 3447 * deadlock: 3448 * 3449 * vm_fault->getpages->cluster_read->allocbuf 3450 */ 3451 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3452 &error); 3453 if (error) { 3454 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3455 continue; 3456 } 3457 if (m == NULL) { 3458 /* 3459 * note: must allocate system pages 3460 * since blocking here could intefere 3461 * with paging I/O, no matter which 3462 * process we are. 3463 */ 3464 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages); 3465 if (m) { 3466 vm_page_wire(m); 3467 vm_page_flag_clear(m, PG_ZERO); 3468 vm_page_wakeup(m); 3469 bp->b_flags &= ~B_CACHE; 3470 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3471 ++bp->b_xio.xio_npages; 3472 } 3473 continue; 3474 } 3475 3476 /* 3477 * We found a page and were able to busy it. 3478 */ 3479 vm_page_flag_clear(m, PG_ZERO); 3480 vm_page_wire(m); 3481 vm_page_wakeup(m); 3482 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3483 ++bp->b_xio.xio_npages; 3484 if (bp->b_act_count < m->act_count) 3485 bp->b_act_count = m->act_count; 3486 } 3487 vm_object_drop(obj); 3488 3489 /* 3490 * Step 2. We've loaded the pages into the buffer, 3491 * we have to figure out if we can still have B_CACHE 3492 * set. Note that B_CACHE is set according to the 3493 * byte-granular range ( bcount and size ), not the 3494 * aligned range ( newbsize ). 3495 * 3496 * The VM test is against m->valid, which is DEV_BSIZE 3497 * aligned. Needless to say, the validity of the data 3498 * needs to also be DEV_BSIZE aligned. Note that this 3499 * fails with NFS if the server or some other client 3500 * extends the file's EOF. If our buffer is resized, 3501 * B_CACHE may remain set! XXX 3502 */ 3503 3504 toff = bp->b_bcount; 3505 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3506 3507 while ((bp->b_flags & B_CACHE) && toff < size) { 3508 vm_pindex_t pi; 3509 3510 if (tinc > (size - toff)) 3511 tinc = size - toff; 3512 3513 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3514 PAGE_SHIFT; 3515 3516 vfs_buf_test_cache( 3517 bp, 3518 bp->b_loffset, 3519 toff, 3520 tinc, 3521 bp->b_xio.xio_pages[pi] 3522 ); 3523 toff += tinc; 3524 tinc = PAGE_SIZE; 3525 } 3526 3527 /* 3528 * Step 3, fixup the KVM pmap. Remember that 3529 * bp->b_data is relative to bp->b_loffset, but 3530 * bp->b_loffset may be offset into the first page. 3531 */ 3532 3533 bp->b_data = (caddr_t) 3534 trunc_page((vm_offset_t)bp->b_data); 3535 pmap_qenter( 3536 (vm_offset_t)bp->b_data, 3537 bp->b_xio.xio_pages, 3538 bp->b_xio.xio_npages 3539 ); 3540 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3541 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3542 } 3543 } 3544 3545 /* adjust space use on already-dirty buffer */ 3546 if (bp->b_flags & B_DELWRI) { 3547 spin_lock(&bufcspin); 3548 dirtybufspace += newbsize - bp->b_bufsize; 3549 if (bp->b_flags & B_HEAVY) 3550 dirtybufspacehw += newbsize - bp->b_bufsize; 3551 spin_unlock(&bufcspin); 3552 } 3553 if (newbsize < bp->b_bufsize) 3554 bufspacewakeup(); 3555 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3556 bp->b_bcount = size; /* requested buffer size */ 3557 return 1; 3558 } 3559 3560 /* 3561 * biowait: 3562 * 3563 * Wait for buffer I/O completion, returning error status. B_EINTR 3564 * is converted into an EINTR error but not cleared (since a chain 3565 * of biowait() calls may occur). 3566 * 3567 * On return bpdone() will have been called but the buffer will remain 3568 * locked and will not have been brelse()'d. 3569 * 3570 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3571 * likely still in progress on return. 3572 * 3573 * NOTE! This operation is on a BIO, not a BUF. 3574 * 3575 * NOTE! BIO_DONE is cleared by vn_strategy() 3576 * 3577 * MPSAFE 3578 */ 3579 static __inline int 3580 _biowait(struct bio *bio, const char *wmesg, int to) 3581 { 3582 struct buf *bp = bio->bio_buf; 3583 u_int32_t flags; 3584 u_int32_t nflags; 3585 int error; 3586 3587 KKASSERT(bio == &bp->b_bio1); 3588 for (;;) { 3589 flags = bio->bio_flags; 3590 if (flags & BIO_DONE) 3591 break; 3592 nflags = flags | BIO_WANT; 3593 tsleep_interlock(bio, 0); 3594 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3595 if (wmesg) 3596 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3597 else if (bp->b_cmd == BUF_CMD_READ) 3598 error = tsleep(bio, PINTERLOCKED, "biord", to); 3599 else 3600 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3601 if (error) { 3602 kprintf("tsleep error biowait %d\n", error); 3603 return (error); 3604 } 3605 } 3606 } 3607 3608 /* 3609 * Finish up. 3610 */ 3611 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3612 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3613 if (bp->b_flags & B_EINTR) 3614 return (EINTR); 3615 if (bp->b_flags & B_ERROR) 3616 return (bp->b_error ? bp->b_error : EIO); 3617 return (0); 3618 } 3619 3620 int 3621 biowait(struct bio *bio, const char *wmesg) 3622 { 3623 return(_biowait(bio, wmesg, 0)); 3624 } 3625 3626 int 3627 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3628 { 3629 return(_biowait(bio, wmesg, to)); 3630 } 3631 3632 /* 3633 * This associates a tracking count with an I/O. vn_strategy() and 3634 * dev_dstrategy() do this automatically but there are a few cases 3635 * where a vnode or device layer is bypassed when a block translation 3636 * is cached. In such cases bio_start_transaction() may be called on 3637 * the bypassed layers so the system gets an I/O in progress indication 3638 * for those higher layers. 3639 */ 3640 void 3641 bio_start_transaction(struct bio *bio, struct bio_track *track) 3642 { 3643 bio->bio_track = track; 3644 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3645 dsched_new_buf(bio->bio_buf); 3646 bio_track_ref(track); 3647 } 3648 3649 /* 3650 * Initiate I/O on a vnode. 3651 * 3652 * SWAPCACHE OPERATION: 3653 * 3654 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3655 * devfs also uses b_vp for fake buffers so we also have to check 3656 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3657 * underlying block device. The swap assignments are related to the 3658 * buffer cache buffer's b_vp, not the passed vp. 3659 * 3660 * The passed vp == bp->b_vp only in the case where the strategy call 3661 * is made on the vp itself for its own buffers (a regular file or 3662 * block device vp). The filesystem usually then re-calls vn_strategy() 3663 * after translating the request to an underlying device. 3664 * 3665 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3666 * underlying buffer cache buffers. 3667 * 3668 * We can only deal with page-aligned buffers at the moment, because 3669 * we can't tell what the real dirty state for pages straddling a buffer 3670 * are. 3671 * 3672 * In order to call swap_pager_strategy() we must provide the VM object 3673 * and base offset for the underlying buffer cache pages so it can find 3674 * the swap blocks. 3675 */ 3676 void 3677 vn_strategy(struct vnode *vp, struct bio *bio) 3678 { 3679 struct bio_track *track; 3680 struct buf *bp = bio->bio_buf; 3681 3682 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3683 3684 /* 3685 * Set when an I/O is issued on the bp. Cleared by consumers 3686 * (aka HAMMER), allowing the consumer to determine if I/O had 3687 * actually occurred. 3688 */ 3689 bp->b_flags |= B_IODEBUG; 3690 3691 /* 3692 * Handle the swap cache intercept. 3693 */ 3694 if (vn_cache_strategy(vp, bio)) 3695 return; 3696 3697 /* 3698 * Otherwise do the operation through the filesystem 3699 */ 3700 if (bp->b_cmd == BUF_CMD_READ) 3701 track = &vp->v_track_read; 3702 else 3703 track = &vp->v_track_write; 3704 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3705 bio->bio_track = track; 3706 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3707 dsched_new_buf(bio->bio_buf); 3708 bio_track_ref(track); 3709 vop_strategy(*vp->v_ops, vp, bio); 3710 } 3711 3712 static void vn_cache_strategy_callback(struct bio *bio); 3713 3714 int 3715 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3716 { 3717 struct buf *bp = bio->bio_buf; 3718 struct bio *nbio; 3719 vm_object_t object; 3720 vm_page_t m; 3721 int i; 3722 3723 /* 3724 * Is this buffer cache buffer suitable for reading from 3725 * the swap cache? 3726 */ 3727 if (vm_swapcache_read_enable == 0 || 3728 bp->b_cmd != BUF_CMD_READ || 3729 ((bp->b_flags & B_CLUSTER) == 0 && 3730 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3731 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3732 (bp->b_bcount & PAGE_MASK) != 0) { 3733 return(0); 3734 } 3735 3736 /* 3737 * Figure out the original VM object (it will match the underlying 3738 * VM pages). Note that swap cached data uses page indices relative 3739 * to that object, not relative to bio->bio_offset. 3740 */ 3741 if (bp->b_flags & B_CLUSTER) 3742 object = vp->v_object; 3743 else 3744 object = bp->b_vp->v_object; 3745 3746 /* 3747 * In order to be able to use the swap cache all underlying VM 3748 * pages must be marked as such, and we can't have any bogus pages. 3749 */ 3750 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3751 m = bp->b_xio.xio_pages[i]; 3752 if ((m->flags & PG_SWAPPED) == 0) 3753 break; 3754 if (m == bogus_page) 3755 break; 3756 } 3757 3758 /* 3759 * If we are good then issue the I/O using swap_pager_strategy(). 3760 * 3761 * We can only do this if the buffer actually supports object-backed 3762 * I/O. If it doesn't npages will be 0. 3763 */ 3764 if (i && i == bp->b_xio.xio_npages) { 3765 m = bp->b_xio.xio_pages[0]; 3766 nbio = push_bio(bio); 3767 nbio->bio_done = vn_cache_strategy_callback; 3768 nbio->bio_offset = ptoa(m->pindex); 3769 KKASSERT(m->object == object); 3770 swap_pager_strategy(object, nbio); 3771 return(1); 3772 } 3773 return(0); 3774 } 3775 3776 /* 3777 * This is a bit of a hack but since the vn_cache_strategy() function can 3778 * override a VFS's strategy function we must make sure that the bio, which 3779 * is probably bio2, doesn't leak an unexpected offset value back to the 3780 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3781 * bio went through its own file strategy function and the the bio2 offset 3782 * is a cached disk offset when, in fact, it isn't. 3783 */ 3784 static void 3785 vn_cache_strategy_callback(struct bio *bio) 3786 { 3787 bio->bio_offset = NOOFFSET; 3788 biodone(pop_bio(bio)); 3789 } 3790 3791 /* 3792 * bpdone: 3793 * 3794 * Finish I/O on a buffer after all BIOs have been processed. 3795 * Called when the bio chain is exhausted or by biowait. If called 3796 * by biowait, elseit is typically 0. 3797 * 3798 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3799 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3800 * assuming B_INVAL is clear. 3801 * 3802 * For the VMIO case, we set B_CACHE if the op was a read and no 3803 * read error occured, or if the op was a write. B_CACHE is never 3804 * set if the buffer is invalid or otherwise uncacheable. 3805 * 3806 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3807 * initiator to leave B_INVAL set to brelse the buffer out of existance 3808 * in the biodone routine. 3809 */ 3810 void 3811 bpdone(struct buf *bp, int elseit) 3812 { 3813 buf_cmd_t cmd; 3814 3815 KASSERT(BUF_REFCNTNB(bp) > 0, 3816 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3817 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3818 ("biodone: bp %p already done!", bp)); 3819 3820 /* 3821 * No more BIOs are left. All completion functions have been dealt 3822 * with, now we clean up the buffer. 3823 */ 3824 cmd = bp->b_cmd; 3825 bp->b_cmd = BUF_CMD_DONE; 3826 3827 /* 3828 * Only reads and writes are processed past this point. 3829 */ 3830 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3831 if (cmd == BUF_CMD_FREEBLKS) 3832 bp->b_flags |= B_NOCACHE; 3833 if (elseit) 3834 brelse(bp); 3835 return; 3836 } 3837 3838 /* 3839 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3840 * a lot worse. XXX - move this above the clearing of b_cmd 3841 */ 3842 if (LIST_FIRST(&bp->b_dep) != NULL) 3843 buf_complete(bp); /* MPSAFE */ 3844 3845 /* 3846 * A failed write must re-dirty the buffer unless B_INVAL 3847 * was set. Only applicable to normal buffers (with VPs). 3848 * vinum buffers may not have a vp. 3849 */ 3850 if (cmd == BUF_CMD_WRITE && 3851 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3852 bp->b_flags &= ~B_NOCACHE; 3853 if (bp->b_vp) 3854 bdirty(bp); 3855 } 3856 3857 if (bp->b_flags & B_VMIO) { 3858 int i; 3859 vm_ooffset_t foff; 3860 vm_page_t m; 3861 vm_object_t obj; 3862 int iosize; 3863 struct vnode *vp = bp->b_vp; 3864 3865 obj = vp->v_object; 3866 3867 #if defined(VFS_BIO_DEBUG) 3868 if (vp->v_auxrefs == 0) 3869 panic("biodone: zero vnode hold count"); 3870 if ((vp->v_flag & VOBJBUF) == 0) 3871 panic("biodone: vnode is not setup for merged cache"); 3872 #endif 3873 3874 foff = bp->b_loffset; 3875 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 3876 KASSERT(obj != NULL, ("biodone: missing VM object")); 3877 3878 #if defined(VFS_BIO_DEBUG) 3879 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3880 kprintf("biodone: paging in progress(%d) < " 3881 "bp->b_xio.xio_npages(%d)\n", 3882 obj->paging_in_progress, 3883 bp->b_xio.xio_npages); 3884 } 3885 #endif 3886 3887 /* 3888 * Set B_CACHE if the op was a normal read and no error 3889 * occured. B_CACHE is set for writes in the b*write() 3890 * routines. 3891 */ 3892 iosize = bp->b_bcount - bp->b_resid; 3893 if (cmd == BUF_CMD_READ && 3894 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3895 bp->b_flags |= B_CACHE; 3896 } 3897 3898 vm_object_hold(obj); 3899 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3900 int bogusflag = 0; 3901 int resid; 3902 3903 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3904 if (resid > iosize) 3905 resid = iosize; 3906 3907 /* 3908 * cleanup bogus pages, restoring the originals. Since 3909 * the originals should still be wired, we don't have 3910 * to worry about interrupt/freeing races destroying 3911 * the VM object association. 3912 */ 3913 m = bp->b_xio.xio_pages[i]; 3914 if (m == bogus_page) { 3915 bogusflag = 1; 3916 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3917 if (m == NULL) 3918 panic("biodone: page disappeared"); 3919 bp->b_xio.xio_pages[i] = m; 3920 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3921 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3922 } 3923 #if defined(VFS_BIO_DEBUG) 3924 if (OFF_TO_IDX(foff) != m->pindex) { 3925 kprintf("biodone: foff(%lu)/m->pindex(%ld) " 3926 "mismatch\n", 3927 (unsigned long)foff, (long)m->pindex); 3928 } 3929 #endif 3930 3931 /* 3932 * In the write case, the valid and clean bits are 3933 * already changed correctly (see bdwrite()), so we 3934 * only need to do this here in the read case. 3935 */ 3936 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3937 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3938 vfs_clean_one_page(bp, i, m); 3939 } 3940 vm_page_flag_clear(m, PG_ZERO); 3941 3942 /* 3943 * when debugging new filesystems or buffer I/O 3944 * methods, this is the most common error that pops 3945 * up. if you see this, you have not set the page 3946 * busy flag correctly!!! 3947 */ 3948 if (m->busy == 0) { 3949 kprintf("biodone: page busy < 0, " 3950 "pindex: %d, foff: 0x(%x,%x), " 3951 "resid: %d, index: %d\n", 3952 (int) m->pindex, (int)(foff >> 32), 3953 (int) foff & 0xffffffff, resid, i); 3954 if (!vn_isdisk(vp, NULL)) 3955 kprintf(" iosize: %ld, loffset: %lld, " 3956 "flags: 0x%08x, npages: %d\n", 3957 bp->b_vp->v_mount->mnt_stat.f_iosize, 3958 (long long)bp->b_loffset, 3959 bp->b_flags, bp->b_xio.xio_npages); 3960 else 3961 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3962 (long long)bp->b_loffset, 3963 bp->b_flags, bp->b_xio.xio_npages); 3964 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3965 m->valid, m->dirty, m->wire_count); 3966 panic("biodone: page busy < 0"); 3967 } 3968 vm_page_io_finish(m); 3969 vm_page_wakeup(m); 3970 vm_object_pip_wakeup(obj); 3971 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3972 iosize -= resid; 3973 } 3974 bp->b_flags &= ~B_HASBOGUS; 3975 vm_object_drop(obj); 3976 } 3977 3978 /* 3979 * Finish up by releasing the buffer. There are no more synchronous 3980 * or asynchronous completions, those were handled by bio_done 3981 * callbacks. 3982 */ 3983 if (elseit) { 3984 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3985 brelse(bp); 3986 else 3987 bqrelse(bp); 3988 } 3989 } 3990 3991 /* 3992 * Normal biodone. 3993 */ 3994 void 3995 biodone(struct bio *bio) 3996 { 3997 struct buf *bp = bio->bio_buf; 3998 3999 runningbufwakeup(bp); 4000 4001 /* 4002 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 4003 */ 4004 while (bio) { 4005 biodone_t *done_func; 4006 struct bio_track *track; 4007 4008 /* 4009 * BIO tracking. Most but not all BIOs are tracked. 4010 */ 4011 if ((track = bio->bio_track) != NULL) { 4012 bio_track_rel(track); 4013 bio->bio_track = NULL; 4014 } 4015 4016 /* 4017 * A bio_done function terminates the loop. The function 4018 * will be responsible for any further chaining and/or 4019 * buffer management. 4020 * 4021 * WARNING! The done function can deallocate the buffer! 4022 */ 4023 if ((done_func = bio->bio_done) != NULL) { 4024 bio->bio_done = NULL; 4025 done_func(bio); 4026 return; 4027 } 4028 bio = bio->bio_prev; 4029 } 4030 4031 /* 4032 * If we've run out of bio's do normal [a]synchronous completion. 4033 */ 4034 bpdone(bp, 1); 4035 } 4036 4037 /* 4038 * Synchronous biodone - this terminates a synchronous BIO. 4039 * 4040 * bpdone() is called with elseit=FALSE, leaving the buffer completed 4041 * but still locked. The caller must brelse() the buffer after waiting 4042 * for completion. 4043 */ 4044 void 4045 biodone_sync(struct bio *bio) 4046 { 4047 struct buf *bp = bio->bio_buf; 4048 int flags; 4049 int nflags; 4050 4051 KKASSERT(bio == &bp->b_bio1); 4052 bpdone(bp, 0); 4053 4054 for (;;) { 4055 flags = bio->bio_flags; 4056 nflags = (flags | BIO_DONE) & ~BIO_WANT; 4057 4058 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 4059 if (flags & BIO_WANT) 4060 wakeup(bio); 4061 break; 4062 } 4063 } 4064 } 4065 4066 /* 4067 * vfs_unbusy_pages: 4068 * 4069 * This routine is called in lieu of iodone in the case of 4070 * incomplete I/O. This keeps the busy status for pages 4071 * consistant. 4072 */ 4073 void 4074 vfs_unbusy_pages(struct buf *bp) 4075 { 4076 int i; 4077 4078 runningbufwakeup(bp); 4079 4080 if (bp->b_flags & B_VMIO) { 4081 struct vnode *vp = bp->b_vp; 4082 vm_object_t obj; 4083 4084 obj = vp->v_object; 4085 vm_object_hold(obj); 4086 4087 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4088 vm_page_t m = bp->b_xio.xio_pages[i]; 4089 4090 /* 4091 * When restoring bogus changes the original pages 4092 * should still be wired, so we are in no danger of 4093 * losing the object association and do not need 4094 * critical section protection particularly. 4095 */ 4096 if (m == bogus_page) { 4097 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 4098 if (!m) { 4099 panic("vfs_unbusy_pages: page missing"); 4100 } 4101 bp->b_xio.xio_pages[i] = m; 4102 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4103 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4104 } 4105 vm_page_busy_wait(m, FALSE, "bpdpgw"); 4106 vm_page_flag_clear(m, PG_ZERO); 4107 vm_page_io_finish(m); 4108 vm_page_wakeup(m); 4109 vm_object_pip_wakeup(obj); 4110 } 4111 bp->b_flags &= ~B_HASBOGUS; 4112 vm_object_drop(obj); 4113 } 4114 } 4115 4116 /* 4117 * vfs_busy_pages: 4118 * 4119 * This routine is called before a device strategy routine. 4120 * It is used to tell the VM system that paging I/O is in 4121 * progress, and treat the pages associated with the buffer 4122 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4123 * flag is handled to make sure that the object doesn't become 4124 * inconsistant. 4125 * 4126 * Since I/O has not been initiated yet, certain buffer flags 4127 * such as B_ERROR or B_INVAL may be in an inconsistant state 4128 * and should be ignored. 4129 * 4130 * MPSAFE 4131 */ 4132 void 4133 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4134 { 4135 int i, bogus; 4136 struct lwp *lp = curthread->td_lwp; 4137 4138 /* 4139 * The buffer's I/O command must already be set. If reading, 4140 * B_CACHE must be 0 (double check against callers only doing 4141 * I/O when B_CACHE is 0). 4142 */ 4143 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4144 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4145 4146 if (bp->b_flags & B_VMIO) { 4147 vm_object_t obj; 4148 4149 obj = vp->v_object; 4150 KASSERT(bp->b_loffset != NOOFFSET, 4151 ("vfs_busy_pages: no buffer offset")); 4152 4153 /* 4154 * Busy all the pages. We have to busy them all at once 4155 * to avoid deadlocks. 4156 */ 4157 retry: 4158 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4159 vm_page_t m = bp->b_xio.xio_pages[i]; 4160 4161 if (vm_page_busy_try(m, FALSE)) { 4162 vm_page_sleep_busy(m, FALSE, "vbpage"); 4163 while (--i >= 0) 4164 vm_page_wakeup(bp->b_xio.xio_pages[i]); 4165 goto retry; 4166 } 4167 } 4168 4169 /* 4170 * Setup for I/O, soft-busy the page right now because 4171 * the next loop may block. 4172 */ 4173 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4174 vm_page_t m = bp->b_xio.xio_pages[i]; 4175 4176 vm_page_flag_clear(m, PG_ZERO); 4177 if ((bp->b_flags & B_CLUSTER) == 0) { 4178 vm_object_pip_add(obj, 1); 4179 vm_page_io_start(m); 4180 } 4181 } 4182 4183 /* 4184 * Adjust protections for I/O and do bogus-page mapping. 4185 * Assume that vm_page_protect() can block (it can block 4186 * if VM_PROT_NONE, don't take any chances regardless). 4187 * 4188 * In particular note that for writes we must incorporate 4189 * page dirtyness from the VM system into the buffer's 4190 * dirty range. 4191 * 4192 * For reads we theoretically must incorporate page dirtyness 4193 * from the VM system to determine if the page needs bogus 4194 * replacement, but we shortcut the test by simply checking 4195 * that all m->valid bits are set, indicating that the page 4196 * is fully valid and does not need to be re-read. For any 4197 * VM system dirtyness the page will also be fully valid 4198 * since it was mapped at one point. 4199 */ 4200 bogus = 0; 4201 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4202 vm_page_t m = bp->b_xio.xio_pages[i]; 4203 4204 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4205 if (bp->b_cmd == BUF_CMD_WRITE) { 4206 /* 4207 * When readying a vnode-backed buffer for 4208 * a write we must zero-fill any invalid 4209 * portions of the backing VM pages, mark 4210 * it valid and clear related dirty bits. 4211 * 4212 * vfs_clean_one_page() incorporates any 4213 * VM dirtyness and updates the b_dirtyoff 4214 * range (after we've made the page RO). 4215 * 4216 * It is also expected that the pmap modified 4217 * bit has already been cleared by the 4218 * vm_page_protect(). We may not be able 4219 * to clear all dirty bits for a page if it 4220 * was also memory mapped (NFS). 4221 * 4222 * Finally be sure to unassign any swap-cache 4223 * backing store as it is now stale. 4224 */ 4225 vm_page_protect(m, VM_PROT_READ); 4226 vfs_clean_one_page(bp, i, m); 4227 swap_pager_unswapped(m); 4228 } else if (m->valid == VM_PAGE_BITS_ALL) { 4229 /* 4230 * When readying a vnode-backed buffer for 4231 * read we must replace any dirty pages with 4232 * a bogus page so dirty data is not destroyed 4233 * when filling gaps. 4234 * 4235 * To avoid testing whether the page is 4236 * dirty we instead test that the page was 4237 * at some point mapped (m->valid fully 4238 * valid) with the understanding that 4239 * this also covers the dirty case. 4240 */ 4241 bp->b_xio.xio_pages[i] = bogus_page; 4242 bp->b_flags |= B_HASBOGUS; 4243 bogus++; 4244 } else if (m->valid & m->dirty) { 4245 /* 4246 * This case should not occur as partial 4247 * dirtyment can only happen if the buffer 4248 * is B_CACHE, and this code is not entered 4249 * if the buffer is B_CACHE. 4250 */ 4251 kprintf("Warning: vfs_busy_pages - page not " 4252 "fully valid! loff=%jx bpf=%08x " 4253 "idx=%d val=%02x dir=%02x\n", 4254 (intmax_t)bp->b_loffset, bp->b_flags, 4255 i, m->valid, m->dirty); 4256 vm_page_protect(m, VM_PROT_NONE); 4257 } else { 4258 /* 4259 * The page is not valid and can be made 4260 * part of the read. 4261 */ 4262 vm_page_protect(m, VM_PROT_NONE); 4263 } 4264 vm_page_wakeup(m); 4265 } 4266 if (bogus) { 4267 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4268 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4269 } 4270 } 4271 4272 /* 4273 * This is the easiest place to put the process accounting for the I/O 4274 * for now. 4275 */ 4276 if (lp != NULL) { 4277 if (bp->b_cmd == BUF_CMD_READ) 4278 lp->lwp_ru.ru_inblock++; 4279 else 4280 lp->lwp_ru.ru_oublock++; 4281 } 4282 } 4283 4284 /* 4285 * Tell the VM system that the pages associated with this buffer 4286 * are clean. This is used for delayed writes where the data is 4287 * going to go to disk eventually without additional VM intevention. 4288 * 4289 * NOTE: While we only really need to clean through to b_bcount, we 4290 * just go ahead and clean through to b_bufsize. 4291 */ 4292 static void 4293 vfs_clean_pages(struct buf *bp) 4294 { 4295 vm_page_t m; 4296 int i; 4297 4298 if ((bp->b_flags & B_VMIO) == 0) 4299 return; 4300 4301 KASSERT(bp->b_loffset != NOOFFSET, 4302 ("vfs_clean_pages: no buffer offset")); 4303 4304 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4305 m = bp->b_xio.xio_pages[i]; 4306 vfs_clean_one_page(bp, i, m); 4307 } 4308 } 4309 4310 /* 4311 * vfs_clean_one_page: 4312 * 4313 * Set the valid bits and clear the dirty bits in a page within a 4314 * buffer. The range is restricted to the buffer's size and the 4315 * buffer's logical offset might index into the first page. 4316 * 4317 * The caller has busied or soft-busied the page and it is not mapped, 4318 * test and incorporate the dirty bits into b_dirtyoff/end before 4319 * clearing them. Note that we need to clear the pmap modified bits 4320 * after determining the the page was dirty, vm_page_set_validclean() 4321 * does not do it for us. 4322 * 4323 * This routine is typically called after a read completes (dirty should 4324 * be zero in that case as we are not called on bogus-replace pages), 4325 * or before a write is initiated. 4326 */ 4327 static void 4328 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4329 { 4330 int bcount; 4331 int xoff; 4332 int soff; 4333 int eoff; 4334 4335 /* 4336 * Calculate offset range within the page but relative to buffer's 4337 * loffset. loffset might be offset into the first page. 4338 */ 4339 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4340 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4341 4342 if (pageno == 0) { 4343 soff = xoff; 4344 eoff = PAGE_SIZE; 4345 } else { 4346 soff = (pageno << PAGE_SHIFT); 4347 eoff = soff + PAGE_SIZE; 4348 } 4349 if (eoff > bcount) 4350 eoff = bcount; 4351 if (soff >= eoff) 4352 return; 4353 4354 /* 4355 * Test dirty bits and adjust b_dirtyoff/end. 4356 * 4357 * If dirty pages are incorporated into the bp any prior 4358 * B_NEEDCOMMIT state (NFS) must be cleared because the 4359 * caller has not taken into account the new dirty data. 4360 * 4361 * If the page was memory mapped the dirty bits might go beyond the 4362 * end of the buffer, but we can't really make the assumption that 4363 * a file EOF straddles the buffer (even though this is the case for 4364 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4365 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4366 * This also saves some console spam. 4367 * 4368 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4369 * NFS can handle huge commits but not huge writes. 4370 */ 4371 vm_page_test_dirty(m); 4372 if (m->dirty) { 4373 if ((bp->b_flags & B_NEEDCOMMIT) && 4374 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4375 if (debug_commit) 4376 kprintf("Warning: vfs_clean_one_page: bp %p " 4377 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4378 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4379 "doff/end %d %d\n", 4380 bp, (intmax_t)bp->b_loffset, bp->b_bcount, 4381 bp->b_flags, bp->b_cmd, 4382 m->valid, m->dirty, xoff, soff, eoff, 4383 bp->b_dirtyoff, bp->b_dirtyend); 4384 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4385 if (debug_commit) 4386 print_backtrace(-1); 4387 } 4388 /* 4389 * Only clear the pmap modified bits if ALL the dirty bits 4390 * are set, otherwise the system might mis-clear portions 4391 * of a page. 4392 */ 4393 if (m->dirty == VM_PAGE_BITS_ALL && 4394 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4395 pmap_clear_modify(m); 4396 } 4397 if (bp->b_dirtyoff > soff - xoff) 4398 bp->b_dirtyoff = soff - xoff; 4399 if (bp->b_dirtyend < eoff - xoff) 4400 bp->b_dirtyend = eoff - xoff; 4401 } 4402 4403 /* 4404 * Set related valid bits, clear related dirty bits. 4405 * Does not mess with the pmap modified bit. 4406 * 4407 * WARNING! We cannot just clear all of m->dirty here as the 4408 * buffer cache buffers may use a DEV_BSIZE'd aligned 4409 * block size, or have an odd size (e.g. NFS at file EOF). 4410 * The putpages code can clear m->dirty to 0. 4411 * 4412 * If a VOP_WRITE generates a buffer cache buffer which 4413 * covers the same space as mapped writable pages the 4414 * buffer flush might not be able to clear all the dirty 4415 * bits and still require a putpages from the VM system 4416 * to finish it off. 4417 * 4418 * WARNING! vm_page_set_validclean() currently assumes vm_token 4419 * is held. The page might not be busied (bdwrite() case). 4420 * XXX remove this comment once we've validated that this 4421 * is no longer an issue. 4422 */ 4423 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4424 } 4425 4426 /* 4427 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4428 * The page data is assumed to be valid (there is no zeroing here). 4429 */ 4430 static void 4431 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4432 { 4433 int bcount; 4434 int xoff; 4435 int soff; 4436 int eoff; 4437 4438 /* 4439 * Calculate offset range within the page but relative to buffer's 4440 * loffset. loffset might be offset into the first page. 4441 */ 4442 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4443 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4444 4445 if (pageno == 0) { 4446 soff = xoff; 4447 eoff = PAGE_SIZE; 4448 } else { 4449 soff = (pageno << PAGE_SHIFT); 4450 eoff = soff + PAGE_SIZE; 4451 } 4452 if (eoff > bcount) 4453 eoff = bcount; 4454 if (soff >= eoff) 4455 return; 4456 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4457 } 4458 4459 /* 4460 * vfs_bio_clrbuf: 4461 * 4462 * Clear a buffer. This routine essentially fakes an I/O, so we need 4463 * to clear B_ERROR and B_INVAL. 4464 * 4465 * Note that while we only theoretically need to clear through b_bcount, 4466 * we go ahead and clear through b_bufsize. 4467 */ 4468 4469 void 4470 vfs_bio_clrbuf(struct buf *bp) 4471 { 4472 int i, mask = 0; 4473 caddr_t sa, ea; 4474 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4475 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4476 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4477 (bp->b_loffset & PAGE_MASK) == 0) { 4478 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4479 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4480 bp->b_resid = 0; 4481 return; 4482 } 4483 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4484 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4485 bzero(bp->b_data, bp->b_bufsize); 4486 bp->b_xio.xio_pages[0]->valid |= mask; 4487 bp->b_resid = 0; 4488 return; 4489 } 4490 } 4491 sa = bp->b_data; 4492 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4493 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4494 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4495 ea = (caddr_t)(vm_offset_t)ulmin( 4496 (u_long)(vm_offset_t)ea, 4497 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4498 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4499 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4500 continue; 4501 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4502 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4503 bzero(sa, ea - sa); 4504 } 4505 } else { 4506 for (; sa < ea; sa += DEV_BSIZE, j++) { 4507 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4508 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4509 bzero(sa, DEV_BSIZE); 4510 } 4511 } 4512 bp->b_xio.xio_pages[i]->valid |= mask; 4513 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4514 } 4515 bp->b_resid = 0; 4516 } else { 4517 clrbuf(bp); 4518 } 4519 } 4520 4521 /* 4522 * vm_hold_load_pages: 4523 * 4524 * Load pages into the buffer's address space. The pages are 4525 * allocated from the kernel object in order to reduce interference 4526 * with the any VM paging I/O activity. The range of loaded 4527 * pages will be wired. 4528 * 4529 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4530 * retrieve the full range (to - from) of pages. 4531 * 4532 * MPSAFE 4533 */ 4534 void 4535 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4536 { 4537 vm_offset_t pg; 4538 vm_page_t p; 4539 int index; 4540 4541 to = round_page(to); 4542 from = round_page(from); 4543 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4544 4545 pg = from; 4546 while (pg < to) { 4547 /* 4548 * Note: must allocate system pages since blocking here 4549 * could intefere with paging I/O, no matter which 4550 * process we are. 4551 */ 4552 vm_object_hold(&kernel_object); 4553 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT, 4554 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4555 vm_object_drop(&kernel_object); 4556 if (p) { 4557 vm_page_wire(p); 4558 p->valid = VM_PAGE_BITS_ALL; 4559 vm_page_flag_clear(p, PG_ZERO); 4560 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 4561 bp->b_xio.xio_pages[index] = p; 4562 vm_page_wakeup(p); 4563 4564 pg += PAGE_SIZE; 4565 ++index; 4566 } 4567 } 4568 bp->b_xio.xio_npages = index; 4569 } 4570 4571 /* 4572 * Allocate pages for a buffer cache buffer. 4573 * 4574 * Under extremely severe memory conditions even allocating out of the 4575 * system reserve can fail. If this occurs we must allocate out of the 4576 * interrupt reserve to avoid a deadlock with the pageout daemon. 4577 * 4578 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf). 4579 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock 4580 * against the pageout daemon if pages are not freed from other sources. 4581 * 4582 * If NULL is returned the caller is expected to retry (typically check if 4583 * the page already exists on retry before trying to allocate one). 4584 */ 4585 static 4586 vm_page_t 4587 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit) 4588 { 4589 vm_page_t p; 4590 4591 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4592 4593 /* 4594 * Try a normal allocation, allow use of system reserve. 4595 */ 4596 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 4597 VM_ALLOC_NULL_OK); 4598 if (p) 4599 return(p); 4600 4601 /* 4602 * The normal allocation failed and we clearly have a page 4603 * deficit. Try to reclaim some clean VM pages directly 4604 * from the buffer cache. 4605 */ 4606 vm_pageout_deficit += deficit; 4607 recoverbufpages(); 4608 4609 /* 4610 * We may have blocked, the caller will know what to do if the 4611 * page now exists. 4612 */ 4613 if (vm_page_lookup(obj, pg)) { 4614 return(NULL); 4615 } 4616 4617 /* 4618 * Only system threads can use the interrupt reserve 4619 */ 4620 if ((curthread->td_flags & TDF_SYSTHREAD) == 0) { 4621 vm_wait(hz); 4622 return(NULL); 4623 } 4624 4625 4626 /* 4627 * Allocate and allow use of the interrupt reserve. 4628 * 4629 * If after all that we still can't allocate a VM page we are 4630 * in real trouble, but we slog on anyway hoping that the system 4631 * won't deadlock. 4632 */ 4633 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 4634 VM_ALLOC_INTERRUPT | VM_ALLOC_NULL_OK); 4635 if (p) { 4636 if (vm_page_count_severe()) { 4637 ++lowmempgallocs; 4638 vm_wait(hz / 20 + 1); 4639 } 4640 } else if (vm_page_lookup(obj, pg) == NULL) { 4641 kprintf("bio_page_alloc: Memory exhausted during bufcache " 4642 "page allocation\n"); 4643 ++lowmempgfails; 4644 vm_wait(hz); 4645 } 4646 return(p); 4647 } 4648 4649 /* 4650 * vm_hold_free_pages: 4651 * 4652 * Return pages associated with the buffer back to the VM system. 4653 * 4654 * The range of pages underlying the buffer's address space will 4655 * be unmapped and un-wired. 4656 * 4657 * MPSAFE 4658 */ 4659 void 4660 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4661 { 4662 vm_offset_t pg; 4663 vm_page_t p; 4664 int index, newnpages; 4665 4666 from = round_page(from); 4667 to = round_page(to); 4668 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4669 newnpages = index; 4670 4671 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4672 p = bp->b_xio.xio_pages[index]; 4673 if (p && (index < bp->b_xio.xio_npages)) { 4674 if (p->busy) { 4675 kprintf("vm_hold_free_pages: doffset: %lld, " 4676 "loffset: %lld\n", 4677 (long long)bp->b_bio2.bio_offset, 4678 (long long)bp->b_loffset); 4679 } 4680 bp->b_xio.xio_pages[index] = NULL; 4681 pmap_kremove(pg); 4682 vm_page_busy_wait(p, FALSE, "vmhldpg"); 4683 vm_page_unwire(p, 0); 4684 vm_page_free(p); 4685 } 4686 } 4687 bp->b_xio.xio_npages = newnpages; 4688 } 4689 4690 /* 4691 * vmapbuf: 4692 * 4693 * Map a user buffer into KVM via a pbuf. On return the buffer's 4694 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4695 * initialized. 4696 */ 4697 int 4698 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4699 { 4700 caddr_t addr; 4701 vm_offset_t va; 4702 vm_page_t m; 4703 int vmprot; 4704 int error; 4705 int pidx; 4706 int i; 4707 4708 /* 4709 * bp had better have a command and it better be a pbuf. 4710 */ 4711 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4712 KKASSERT(bp->b_flags & B_PAGING); 4713 KKASSERT(bp->b_kvabase); 4714 4715 if (bytes < 0) 4716 return (-1); 4717 4718 /* 4719 * Map the user data into KVM. Mappings have to be page-aligned. 4720 */ 4721 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4722 pidx = 0; 4723 4724 vmprot = VM_PROT_READ; 4725 if (bp->b_cmd == BUF_CMD_READ) 4726 vmprot |= VM_PROT_WRITE; 4727 4728 while (addr < udata + bytes) { 4729 /* 4730 * Do the vm_fault if needed; do the copy-on-write thing 4731 * when reading stuff off device into memory. 4732 * 4733 * vm_fault_page*() returns a held VM page. 4734 */ 4735 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4736 va = trunc_page(va); 4737 4738 m = vm_fault_page_quick(va, vmprot, &error); 4739 if (m == NULL) { 4740 for (i = 0; i < pidx; ++i) { 4741 vm_page_unhold(bp->b_xio.xio_pages[i]); 4742 bp->b_xio.xio_pages[i] = NULL; 4743 } 4744 return(-1); 4745 } 4746 bp->b_xio.xio_pages[pidx] = m; 4747 addr += PAGE_SIZE; 4748 ++pidx; 4749 } 4750 4751 /* 4752 * Map the page array and set the buffer fields to point to 4753 * the mapped data buffer. 4754 */ 4755 if (pidx > btoc(MAXPHYS)) 4756 panic("vmapbuf: mapped more than MAXPHYS"); 4757 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4758 4759 bp->b_xio.xio_npages = pidx; 4760 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4761 bp->b_bcount = bytes; 4762 bp->b_bufsize = bytes; 4763 return(0); 4764 } 4765 4766 /* 4767 * vunmapbuf: 4768 * 4769 * Free the io map PTEs associated with this IO operation. 4770 * We also invalidate the TLB entries and restore the original b_addr. 4771 */ 4772 void 4773 vunmapbuf(struct buf *bp) 4774 { 4775 int pidx; 4776 int npages; 4777 4778 KKASSERT(bp->b_flags & B_PAGING); 4779 4780 npages = bp->b_xio.xio_npages; 4781 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4782 for (pidx = 0; pidx < npages; ++pidx) { 4783 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4784 bp->b_xio.xio_pages[pidx] = NULL; 4785 } 4786 bp->b_xio.xio_npages = 0; 4787 bp->b_data = bp->b_kvabase; 4788 } 4789 4790 /* 4791 * Scan all buffers in the system and issue the callback. 4792 */ 4793 int 4794 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4795 { 4796 int count = 0; 4797 int error; 4798 int n; 4799 4800 for (n = 0; n < nbuf; ++n) { 4801 if ((error = callback(&buf[n], info)) < 0) { 4802 count = error; 4803 break; 4804 } 4805 count += error; 4806 } 4807 return (count); 4808 } 4809 4810 /* 4811 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4812 * completion to the master buffer. 4813 */ 4814 static void 4815 nestiobuf_iodone(struct bio *bio) 4816 { 4817 struct bio *mbio; 4818 struct buf *mbp, *bp; 4819 struct devstat *stats; 4820 int error; 4821 int donebytes; 4822 4823 bp = bio->bio_buf; 4824 mbio = bio->bio_caller_info1.ptr; 4825 stats = bio->bio_caller_info2.ptr; 4826 mbp = mbio->bio_buf; 4827 4828 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4829 KKASSERT(mbp != bp); 4830 4831 error = bp->b_error; 4832 if (bp->b_error == 0 && 4833 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4834 /* 4835 * Not all got transfered, raise an error. We have no way to 4836 * propagate these conditions to mbp. 4837 */ 4838 error = EIO; 4839 } 4840 4841 donebytes = bp->b_bufsize; 4842 4843 relpbuf(bp, NULL); 4844 4845 nestiobuf_done(mbio, donebytes, error, stats); 4846 } 4847 4848 void 4849 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4850 { 4851 struct buf *mbp; 4852 4853 mbp = mbio->bio_buf; 4854 4855 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4856 4857 /* 4858 * If an error occured, propagate it to the master buffer. 4859 * 4860 * Several biodone()s may wind up running concurrently so 4861 * use an atomic op to adjust b_flags. 4862 */ 4863 if (error) { 4864 mbp->b_error = error; 4865 atomic_set_int(&mbp->b_flags, B_ERROR); 4866 } 4867 4868 /* 4869 * Decrement the operations in progress counter and terminate the 4870 * I/O if this was the last bit. 4871 */ 4872 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4873 mbp->b_resid = 0; 4874 if (stats) 4875 devstat_end_transaction_buf(stats, mbp); 4876 biodone(mbio); 4877 } 4878 } 4879 4880 /* 4881 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4882 * the mbio from being biodone()'d while we are still adding sub-bios to 4883 * it. 4884 */ 4885 void 4886 nestiobuf_init(struct bio *bio) 4887 { 4888 bio->bio_driver_info = (void *)1; 4889 } 4890 4891 /* 4892 * The BIOs added to the nestedio have already been started, remove the 4893 * count that placeheld our mbio and biodone() it if the count would 4894 * transition to 0. 4895 */ 4896 void 4897 nestiobuf_start(struct bio *mbio) 4898 { 4899 struct buf *mbp = mbio->bio_buf; 4900 4901 /* 4902 * Decrement the operations in progress counter and terminate the 4903 * I/O if this was the last bit. 4904 */ 4905 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4906 if (mbp->b_flags & B_ERROR) 4907 mbp->b_resid = mbp->b_bcount; 4908 else 4909 mbp->b_resid = 0; 4910 biodone(mbio); 4911 } 4912 } 4913 4914 /* 4915 * Set an intermediate error prior to calling nestiobuf_start() 4916 */ 4917 void 4918 nestiobuf_error(struct bio *mbio, int error) 4919 { 4920 struct buf *mbp = mbio->bio_buf; 4921 4922 if (error) { 4923 mbp->b_error = error; 4924 atomic_set_int(&mbp->b_flags, B_ERROR); 4925 } 4926 } 4927 4928 /* 4929 * nestiobuf_add: setup a "nested" buffer. 4930 * 4931 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4932 * => 'bp' should be a buffer allocated by getiobuf. 4933 * => 'offset' is a byte offset in the master buffer. 4934 * => 'size' is a size in bytes of this nested buffer. 4935 */ 4936 void 4937 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4938 { 4939 struct buf *mbp = mbio->bio_buf; 4940 struct vnode *vp = mbp->b_vp; 4941 4942 KKASSERT(mbp->b_bcount >= offset + size); 4943 4944 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4945 4946 /* kernel needs to own the lock for it to be released in biodone */ 4947 BUF_KERNPROC(bp); 4948 bp->b_vp = vp; 4949 bp->b_cmd = mbp->b_cmd; 4950 bp->b_bio1.bio_done = nestiobuf_iodone; 4951 bp->b_data = (char *)mbp->b_data + offset; 4952 bp->b_resid = bp->b_bcount = size; 4953 bp->b_bufsize = bp->b_bcount; 4954 4955 bp->b_bio1.bio_track = NULL; 4956 bp->b_bio1.bio_caller_info1.ptr = mbio; 4957 bp->b_bio1.bio_caller_info2.ptr = stats; 4958 } 4959 4960 /* 4961 * print out statistics from the current status of the buffer pool 4962 * this can be toggeled by the system control option debug.syncprt 4963 */ 4964 #ifdef DEBUG 4965 void 4966 vfs_bufstats(void) 4967 { 4968 int i, j, count; 4969 struct buf *bp; 4970 struct bqueues *dp; 4971 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 4972 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 4973 4974 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 4975 count = 0; 4976 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4977 counts[j] = 0; 4978 4979 spin_lock(&bufqspin); 4980 TAILQ_FOREACH(bp, dp, b_freelist) { 4981 counts[bp->b_bufsize/PAGE_SIZE]++; 4982 count++; 4983 } 4984 spin_unlock(&bufqspin); 4985 4986 kprintf("%s: total-%d", bname[i], count); 4987 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4988 if (counts[j] != 0) 4989 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]); 4990 kprintf("\n"); 4991 } 4992 } 4993 #endif 4994 4995 #ifdef DDB 4996 4997 DB_SHOW_COMMAND(buffer, db_show_buffer) 4998 { 4999 /* get args */ 5000 struct buf *bp = (struct buf *)addr; 5001 5002 if (!have_addr) { 5003 db_printf("usage: show buffer <addr>\n"); 5004 return; 5005 } 5006 5007 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 5008 db_printf("b_cmd = %d\n", bp->b_cmd); 5009 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 5010 "b_resid = %d\n, b_data = %p, " 5011 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 5012 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5013 bp->b_data, 5014 (long long)bp->b_bio2.bio_offset, 5015 (long long)(bp->b_bio2.bio_next ? 5016 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 5017 if (bp->b_xio.xio_npages) { 5018 int i; 5019 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 5020 bp->b_xio.xio_npages); 5021 for (i = 0; i < bp->b_xio.xio_npages; i++) { 5022 vm_page_t m; 5023 m = bp->b_xio.xio_pages[i]; 5024 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 5025 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 5026 if ((i + 1) < bp->b_xio.xio_npages) 5027 db_printf(","); 5028 } 5029 db_printf("\n"); 5030 } 5031 } 5032 #endif /* DDB */ 5033