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