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. Note that man buf(9) doesn't reflect 28 * the actual buf/bio implementation in DragonFly. 29 */ 30 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/buf.h> 34 #include <sys/conf.h> 35 #include <sys/devicestat.h> 36 #include <sys/eventhandler.h> 37 #include <sys/lock.h> 38 #include <sys/malloc.h> 39 #include <sys/mount.h> 40 #include <sys/kernel.h> 41 #include <sys/kthread.h> 42 #include <sys/proc.h> 43 #include <sys/reboot.h> 44 #include <sys/resourcevar.h> 45 #include <sys/sysctl.h> 46 #include <sys/vmmeter.h> 47 #include <sys/vnode.h> 48 #include <sys/dsched.h> 49 #include <vm/vm.h> 50 #include <vm/vm_param.h> 51 #include <vm/vm_kern.h> 52 #include <vm/vm_pageout.h> 53 #include <vm/vm_page.h> 54 #include <vm/vm_object.h> 55 #include <vm/vm_extern.h> 56 #include <vm/vm_map.h> 57 #include <vm/vm_pager.h> 58 #include <vm/swap_pager.h> 59 60 #include <sys/buf2.h> 61 #include <sys/thread2.h> 62 #include <sys/spinlock2.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_EMPTY, /* empty buffer headers */ 80 81 BUFFER_QUEUES /* number of buffer queues */ 82 }; 83 84 typedef enum bufq_type bufq_type_t; 85 86 #define BD_WAKE_SIZE 16384 87 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1) 88 89 TAILQ_HEAD(bqueues, buf); 90 91 struct bufpcpu { 92 struct spinlock spin; 93 struct bqueues bufqueues[BUFFER_QUEUES]; 94 } __cachealign; 95 96 struct bufpcpu bufpcpu[MAXCPU]; 97 98 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 99 100 struct buf *buf; /* buffer header pool */ 101 102 static void vfs_clean_pages(struct buf *bp); 103 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m); 104 #if 0 105 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m); 106 #endif 107 static void vfs_vmio_release(struct buf *bp); 108 static int flushbufqueues(struct buf *marker, bufq_type_t q); 109 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj, 110 vm_pindex_t pg, int deficit); 111 112 static void bd_signal(long totalspace); 113 static void buf_daemon(void); 114 static void buf_daemon_hw(void); 115 116 /* 117 * bogus page -- for I/O to/from partially complete buffers 118 * this is a temporary solution to the problem, but it is not 119 * really that bad. it would be better to split the buffer 120 * for input in the case of buffers partially already in memory, 121 * but the code is intricate enough already. 122 */ 123 vm_page_t bogus_page; 124 125 /* 126 * These are all static, but make the ones we export globals so we do 127 * not need to use compiler magic. 128 */ 129 long bufspace; /* atomic ops */ 130 long maxbufspace; 131 static long bufmallocspace; /* atomic ops */ 132 long maxbufmallocspace, lobufspace, hibufspace; 133 static long lorunningspace; 134 static long hirunningspace; 135 static long dirtykvaspace; /* atomic */ 136 long dirtybufspace; /* atomic (global for systat) */ 137 static long dirtybufcount; /* atomic */ 138 static long dirtybufspacehw; /* atomic */ 139 static long dirtybufcounthw; /* atomic */ 140 static long runningbufspace; /* atomic */ 141 static long runningbufcount; /* atomic */ 142 long lodirtybufspace; 143 long hidirtybufspace; 144 static int getnewbufcalls; 145 static int recoverbufcalls; 146 static int needsbuffer; /* atomic */ 147 static int runningbufreq; /* atomic */ 148 static int bd_request; /* atomic */ 149 static int bd_request_hw; /* atomic */ 150 static u_int bd_wake_ary[BD_WAKE_SIZE]; 151 static u_int bd_wake_index; 152 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 153 static int debug_commit; 154 static int debug_bufbio; 155 static int debug_kvabio; 156 static long bufcache_bw = 200 * 1024 * 1024; 157 158 static struct thread *bufdaemon_td; 159 static struct thread *bufdaemonhw_td; 160 static u_int lowmempgallocs; 161 static u_int lowmempgfails; 162 static u_int flushperqueue = 1024; 163 164 /* 165 * Sysctls for operational control of the buffer cache. 166 */ 167 SYSCTL_UINT(_vfs, OID_AUTO, flushperqueue, CTLFLAG_RW, &flushperqueue, 0, 168 "Number of buffers to flush from each per-cpu queue"); 169 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 170 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 171 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 172 "High watermark used to trigger explicit flushing of dirty buffers"); 173 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 174 "Minimum amount of buffer space required for active I/O"); 175 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 176 "Maximum amount of buffer space to usable for active I/O"); 177 SYSCTL_LONG(_vfs, OID_AUTO, bufcache_bw, CTLFLAG_RW, &bufcache_bw, 0, 178 "Buffer-cache -> VM page cache transfer bandwidth"); 179 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 180 "Page allocations done during periods of very low free memory"); 181 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 182 "Page allocations which failed during periods of very low free memory"); 183 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 184 "Recycle pages to active or inactive queue transition pt 0-64"); 185 /* 186 * Sysctls determining current state of the buffer cache. 187 */ 188 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 189 "Total number of buffers in buffer cache"); 190 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0, 191 "KVA reserved by dirty buffers (all)"); 192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 193 "Pending bytes of dirty buffers (all)"); 194 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 195 "Pending bytes of dirty buffers (heavy weight)"); 196 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 197 "Pending number of dirty buffers"); 198 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 199 "Pending number of dirty buffers (heavy weight)"); 200 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 201 "I/O bytes currently in progress due to asynchronous writes"); 202 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 203 "I/O buffers currently in progress due to asynchronous writes"); 204 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 205 "Hard limit on maximum amount of memory usable for buffer space"); 206 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 207 "Soft limit on maximum amount of memory usable for buffer space"); 208 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 209 "Minimum amount of memory to reserve for system buffer space"); 210 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 211 "Amount of memory available for buffers"); 212 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 213 0, "Maximum amount of memory reserved for buffers using malloc"); 214 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 215 "Amount of memory left for buffers using malloc-scheme"); 216 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 217 "New buffer header acquisition requests"); 218 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 219 "Recover VM space in an emergency"); 220 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, ""); 221 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, ""); 222 SYSCTL_INT(_vfs, OID_AUTO, debug_kvabio, CTLFLAG_RW, &debug_kvabio, 0, ""); 223 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 224 "sizeof(struct buf)"); 225 226 char *buf_wmesg = BUF_WMESG; 227 228 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 229 #define VFS_BIO_NEED_UNUSED02 0x02 230 #define VFS_BIO_NEED_UNUSED04 0x04 231 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 232 233 /* 234 * Called when buffer space is potentially available for recovery. 235 * getnewbuf() will block on this flag when it is unable to free 236 * sufficient buffer space. Buffer space becomes recoverable when 237 * bp's get placed back in the queues. 238 */ 239 static __inline void 240 bufspacewakeup(void) 241 { 242 /* 243 * If someone is waiting for BUF space, wake them up. Even 244 * though we haven't freed the kva space yet, the waiting 245 * process will be able to now. 246 */ 247 for (;;) { 248 int flags = needsbuffer; 249 cpu_ccfence(); 250 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0) 251 break; 252 if (atomic_cmpset_int(&needsbuffer, flags, 253 flags & ~VFS_BIO_NEED_BUFSPACE)) { 254 wakeup(&needsbuffer); 255 break; 256 } 257 /* retry */ 258 } 259 } 260 261 /* 262 * runningbufwakeup: 263 * 264 * Accounting for I/O in progress. 265 * 266 */ 267 static __inline void 268 runningbufwakeup(struct buf *bp) 269 { 270 long totalspace; 271 long flags; 272 273 if ((totalspace = bp->b_runningbufspace) != 0) { 274 atomic_add_long(&runningbufspace, -totalspace); 275 atomic_add_long(&runningbufcount, -1); 276 bp->b_runningbufspace = 0; 277 278 /* 279 * see waitrunningbufspace() for limit test. 280 */ 281 for (;;) { 282 flags = runningbufreq; 283 cpu_ccfence(); 284 if (flags == 0) 285 break; 286 if (atomic_cmpset_int(&runningbufreq, flags, 0)) { 287 wakeup(&runningbufreq); 288 break; 289 } 290 /* retry */ 291 } 292 bd_signal(totalspace); 293 } 294 } 295 296 /* 297 * bufcountwakeup: 298 * 299 * Called when a buffer has been added to one of the free queues to 300 * account for the buffer and to wakeup anyone waiting for free buffers. 301 * This typically occurs when large amounts of metadata are being handled 302 * by the buffer cache ( else buffer space runs out first, usually ). 303 */ 304 static __inline void 305 bufcountwakeup(void) 306 { 307 long flags; 308 309 for (;;) { 310 flags = needsbuffer; 311 if (flags == 0) 312 break; 313 if (atomic_cmpset_int(&needsbuffer, flags, 314 (flags & ~VFS_BIO_NEED_ANY))) { 315 wakeup(&needsbuffer); 316 break; 317 } 318 /* retry */ 319 } 320 } 321 322 /* 323 * waitrunningbufspace() 324 * 325 * If runningbufspace exceeds 4/6 hirunningspace we block until 326 * runningbufspace drops to 3/6 hirunningspace. We also block if another 327 * thread blocked here in order to be fair, even if runningbufspace 328 * is now lower than the limit. 329 * 330 * The caller may be using this function to block in a tight loop, we 331 * must block while runningbufspace is greater than at least 332 * hirunningspace * 3 / 6. 333 */ 334 void 335 waitrunningbufspace(void) 336 { 337 long limit = hirunningspace * 4 / 6; 338 long flags; 339 340 while (runningbufspace > limit || runningbufreq) { 341 tsleep_interlock(&runningbufreq, 0); 342 flags = atomic_fetchadd_int(&runningbufreq, 1); 343 if (runningbufspace > limit || flags) 344 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz); 345 } 346 } 347 348 /* 349 * buf_dirty_count_severe: 350 * 351 * Return true if we have too many dirty buffers. 352 */ 353 int 354 buf_dirty_count_severe(void) 355 { 356 return (runningbufspace + dirtykvaspace >= hidirtybufspace || 357 dirtybufcount >= nbuf / 2); 358 } 359 360 /* 361 * Return true if the amount of running I/O is severe and BIOQ should 362 * start bursting. 363 */ 364 int 365 buf_runningbufspace_severe(void) 366 { 367 return (runningbufspace >= hirunningspace * 4 / 6); 368 } 369 370 /* 371 * vfs_buf_test_cache: 372 * 373 * Called when a buffer is extended. This function clears the B_CACHE 374 * bit if the newly extended portion of the buffer does not contain 375 * valid data. 376 * 377 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 378 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 379 * them while a clean buffer was present. 380 */ 381 static __inline__ 382 void 383 vfs_buf_test_cache(struct buf *bp, 384 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 385 vm_page_t m) 386 { 387 if (bp->b_flags & B_CACHE) { 388 int base = (foff + off) & PAGE_MASK; 389 if (vm_page_is_valid(m, base, size) == 0) 390 bp->b_flags &= ~B_CACHE; 391 } 392 } 393 394 /* 395 * bd_speedup() 396 * 397 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 398 * low water mark. 399 */ 400 static __inline__ 401 void 402 bd_speedup(void) 403 { 404 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2) 405 return; 406 407 if (bd_request == 0 && 408 (dirtykvaspace > lodirtybufspace / 2 || 409 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 410 if (atomic_fetchadd_int(&bd_request, 1) == 0) 411 wakeup(&bd_request); 412 } 413 if (bd_request_hw == 0 && 414 (dirtykvaspace > lodirtybufspace / 2 || 415 dirtybufcounthw >= nbuf / 2)) { 416 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0) 417 wakeup(&bd_request_hw); 418 } 419 } 420 421 /* 422 * bd_heatup() 423 * 424 * Get the buf_daemon heated up when the number of running and dirty 425 * buffers exceeds the mid-point. 426 * 427 * Return the total number of dirty bytes past the second mid point 428 * as a measure of how much excess dirty data there is in the system. 429 */ 430 long 431 bd_heatup(void) 432 { 433 long mid1; 434 long mid2; 435 long totalspace; 436 437 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 438 439 totalspace = runningbufspace + dirtykvaspace; 440 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 441 bd_speedup(); 442 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 443 if (totalspace >= mid2) 444 return(totalspace - mid2); 445 } 446 return(0); 447 } 448 449 /* 450 * bd_wait() 451 * 452 * Wait for the buffer cache to flush (totalspace) bytes worth of 453 * buffers, then return. 454 * 455 * Regardless this function blocks while the number of dirty buffers 456 * exceeds hidirtybufspace. 457 */ 458 void 459 bd_wait(long totalspace) 460 { 461 u_int i; 462 u_int j; 463 u_int mi; 464 int count; 465 466 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 467 return; 468 469 while (totalspace > 0) { 470 bd_heatup(); 471 472 /* 473 * Order is important. Suppliers adjust bd_wake_index after 474 * updating runningbufspace/dirtykvaspace. We want to fetch 475 * bd_wake_index before accessing. Any error should thus 476 * be in our favor. 477 */ 478 i = atomic_fetchadd_int(&bd_wake_index, 0); 479 if (totalspace > runningbufspace + dirtykvaspace) 480 totalspace = runningbufspace + dirtykvaspace; 481 count = totalspace / MAXBSIZE; 482 if (count >= BD_WAKE_SIZE / 2) 483 count = BD_WAKE_SIZE / 2; 484 i = i + count; 485 mi = i & BD_WAKE_MASK; 486 487 /* 488 * This is not a strict interlock, so we play a bit loose 489 * with locking access to dirtybufspace*. We have to re-check 490 * bd_wake_index to ensure that it hasn't passed us. 491 */ 492 tsleep_interlock(&bd_wake_ary[mi], 0); 493 atomic_add_int(&bd_wake_ary[mi], 1); 494 j = atomic_fetchadd_int(&bd_wake_index, 0); 495 if ((int)(i - j) >= 0) 496 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz); 497 498 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace; 499 } 500 } 501 502 /* 503 * bd_signal() 504 * 505 * This function is called whenever runningbufspace or dirtykvaspace 506 * is reduced. Track threads waiting for run+dirty buffer I/O 507 * complete. 508 */ 509 static void 510 bd_signal(long totalspace) 511 { 512 u_int i; 513 514 if (totalspace > 0) { 515 if (totalspace > MAXBSIZE * BD_WAKE_SIZE) 516 totalspace = MAXBSIZE * BD_WAKE_SIZE; 517 while (totalspace > 0) { 518 i = atomic_fetchadd_int(&bd_wake_index, 1); 519 i &= BD_WAKE_MASK; 520 if (atomic_readandclear_int(&bd_wake_ary[i])) 521 wakeup(&bd_wake_ary[i]); 522 totalspace -= MAXBSIZE; 523 } 524 } 525 } 526 527 /* 528 * BIO tracking support routines. 529 * 530 * Release a ref on a bio_track. Wakeup requests are atomically released 531 * along with the last reference so bk_active will never wind up set to 532 * only 0x80000000. 533 */ 534 static 535 void 536 bio_track_rel(struct bio_track *track) 537 { 538 int active; 539 int desired; 540 541 /* 542 * Shortcut 543 */ 544 active = track->bk_active; 545 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 546 return; 547 548 /* 549 * Full-on. Note that the wait flag is only atomically released on 550 * the 1->0 count transition. 551 * 552 * We check for a negative count transition using bit 30 since bit 31 553 * has a different meaning. 554 */ 555 for (;;) { 556 desired = (active & 0x7FFFFFFF) - 1; 557 if (desired) 558 desired |= active & 0x80000000; 559 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 560 if (desired & 0x40000000) 561 panic("bio_track_rel: bad count: %p", track); 562 if (active & 0x80000000) 563 wakeup(track); 564 break; 565 } 566 active = track->bk_active; 567 } 568 } 569 570 /* 571 * Wait for the tracking count to reach 0. 572 * 573 * Use atomic ops such that the wait flag is only set atomically when 574 * bk_active is non-zero. 575 */ 576 int 577 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 578 { 579 int active; 580 int desired; 581 int error; 582 583 /* 584 * Shortcut 585 */ 586 if (track->bk_active == 0) 587 return(0); 588 589 /* 590 * Full-on. Note that the wait flag may only be atomically set if 591 * the active count is non-zero. 592 * 593 * NOTE: We cannot optimize active == desired since a wakeup could 594 * clear active prior to our tsleep_interlock(). 595 */ 596 error = 0; 597 while ((active = track->bk_active) != 0) { 598 cpu_ccfence(); 599 desired = active | 0x80000000; 600 tsleep_interlock(track, slp_flags); 601 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 602 error = tsleep(track, slp_flags | PINTERLOCKED, 603 "trwait", slp_timo); 604 if (error) 605 break; 606 } 607 } 608 return (error); 609 } 610 611 /* 612 * bufinit: 613 * 614 * Load time initialisation of the buffer cache, called from machine 615 * dependant initialization code. 616 */ 617 static 618 void 619 bufinit(void *dummy __unused) 620 { 621 struct bufpcpu *pcpu; 622 struct buf *bp; 623 vm_offset_t bogus_offset; 624 int i; 625 int j; 626 long n; 627 628 /* next, make a null set of free lists */ 629 for (i = 0; i < ncpus; ++i) { 630 pcpu = &bufpcpu[i]; 631 spin_init(&pcpu->spin, "bufinit"); 632 for (j = 0; j < BUFFER_QUEUES; j++) 633 TAILQ_INIT(&pcpu->bufqueues[j]); 634 } 635 636 /* 637 * Finally, initialize each buffer header and stick on empty q. 638 * Each buffer gets its own KVA reservation. 639 */ 640 i = 0; 641 pcpu = &bufpcpu[i]; 642 643 for (n = 0; n < nbuf; n++) { 644 bp = &buf[n]; 645 bzero(bp, sizeof *bp); 646 bp->b_flags = B_INVAL; /* we're just an empty header */ 647 bp->b_cmd = BUF_CMD_DONE; 648 bp->b_qindex = BQUEUE_EMPTY; 649 bp->b_qcpu = i; 650 bp->b_kvabase = (void *)(vm_map_min(&buffer_map) + 651 MAXBSIZE * n); 652 bp->b_kvasize = MAXBSIZE; 653 initbufbio(bp); 654 xio_init(&bp->b_xio); 655 buf_dep_init(bp); 656 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 657 bp, b_freelist); 658 659 i = (i + 1) % ncpus; 660 pcpu = &bufpcpu[i]; 661 } 662 663 /* 664 * maxbufspace is the absolute maximum amount of buffer space we are 665 * allowed to reserve in KVM and in real terms. The absolute maximum 666 * is nominally used by buf_daemon. hibufspace is the nominal maximum 667 * used by most other processes. The differential is required to 668 * ensure that buf_daemon is able to run when other processes might 669 * be blocked waiting for buffer space. 670 * 671 * Calculate hysteresis (lobufspace, hibufspace). Don't make it 672 * too large or we might lockup a cpu for too long a period of 673 * time in our tight loop. 674 */ 675 maxbufspace = nbuf * NBUFCALCSIZE; 676 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 677 lobufspace = hibufspace * 7 / 8; 678 if (hibufspace - lobufspace > 64 * 1024 * 1024) 679 lobufspace = hibufspace - 64 * 1024 * 1024; 680 if (lobufspace > hibufspace - MAXBSIZE) 681 lobufspace = hibufspace - MAXBSIZE; 682 683 lorunningspace = 512 * 1024; 684 /* hirunningspace -- see below */ 685 686 /* 687 * Limit the amount of malloc memory since it is wired permanently 688 * into the kernel space. Even though this is accounted for in 689 * the buffer allocation, we don't want the malloced region to grow 690 * uncontrolled. The malloc scheme improves memory utilization 691 * significantly on average (small) directories. 692 */ 693 maxbufmallocspace = hibufspace / 20; 694 695 /* 696 * Reduce the chance of a deadlock occuring by limiting the number 697 * of delayed-write dirty buffers we allow to stack up. 698 * 699 * We don't want too much actually queued to the device at once 700 * (XXX this needs to be per-mount!), because the buffers will 701 * wind up locked for a very long period of time while the I/O 702 * drains. 703 */ 704 hidirtybufspace = hibufspace / 2; /* dirty + running */ 705 hirunningspace = hibufspace / 16; /* locked & queued to device */ 706 if (hirunningspace < 1024 * 1024) 707 hirunningspace = 1024 * 1024; 708 709 dirtykvaspace = 0; 710 dirtybufspace = 0; 711 dirtybufspacehw = 0; 712 713 lodirtybufspace = hidirtybufspace / 2; 714 715 /* 716 * Maximum number of async ops initiated per buf_daemon loop. This is 717 * somewhat of a hack at the moment, we really need to limit ourselves 718 * based on the number of bytes of I/O in-transit that were initiated 719 * from buf_daemon. 720 */ 721 722 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE, 723 VM_SUBSYS_BOGUS); 724 vm_object_hold(&kernel_object); 725 bogus_page = vm_page_alloc(&kernel_object, 726 (bogus_offset >> PAGE_SHIFT), 727 VM_ALLOC_NORMAL); 728 vm_object_drop(&kernel_object); 729 vmstats.v_wire_count++; 730 731 } 732 733 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL); 734 735 /* 736 * Initialize the embedded bio structures, typically used by 737 * deprecated code which tries to allocate its own struct bufs. 738 */ 739 void 740 initbufbio(struct buf *bp) 741 { 742 bp->b_bio1.bio_buf = bp; 743 bp->b_bio1.bio_prev = NULL; 744 bp->b_bio1.bio_offset = NOOFFSET; 745 bp->b_bio1.bio_next = &bp->b_bio2; 746 bp->b_bio1.bio_done = NULL; 747 bp->b_bio1.bio_flags = 0; 748 749 bp->b_bio2.bio_buf = bp; 750 bp->b_bio2.bio_prev = &bp->b_bio1; 751 bp->b_bio2.bio_offset = NOOFFSET; 752 bp->b_bio2.bio_next = NULL; 753 bp->b_bio2.bio_done = NULL; 754 bp->b_bio2.bio_flags = 0; 755 756 BUF_LOCKINIT(bp); 757 } 758 759 /* 760 * Reinitialize the embedded bio structures as well as any additional 761 * translation cache layers. 762 */ 763 void 764 reinitbufbio(struct buf *bp) 765 { 766 struct bio *bio; 767 768 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 769 bio->bio_done = NULL; 770 bio->bio_offset = NOOFFSET; 771 } 772 } 773 774 /* 775 * Undo the effects of an initbufbio(). 776 */ 777 void 778 uninitbufbio(struct buf *bp) 779 { 780 dsched_buf_exit(bp); 781 BUF_LOCKFREE(bp); 782 } 783 784 /* 785 * Push another BIO layer onto an existing BIO and return it. The new 786 * BIO layer may already exist, holding cached translation data. 787 */ 788 struct bio * 789 push_bio(struct bio *bio) 790 { 791 struct bio *nbio; 792 793 if ((nbio = bio->bio_next) == NULL) { 794 int index = bio - &bio->bio_buf->b_bio_array[0]; 795 if (index >= NBUF_BIO - 1) { 796 panic("push_bio: too many layers %d for bp %p", 797 index, bio->bio_buf); 798 } 799 nbio = &bio->bio_buf->b_bio_array[index + 1]; 800 bio->bio_next = nbio; 801 nbio->bio_prev = bio; 802 nbio->bio_buf = bio->bio_buf; 803 nbio->bio_offset = NOOFFSET; 804 nbio->bio_done = NULL; 805 nbio->bio_next = NULL; 806 } 807 KKASSERT(nbio->bio_done == NULL); 808 return(nbio); 809 } 810 811 /* 812 * Pop a BIO translation layer, returning the previous layer. The 813 * must have been previously pushed. 814 */ 815 struct bio * 816 pop_bio(struct bio *bio) 817 { 818 return(bio->bio_prev); 819 } 820 821 void 822 clearbiocache(struct bio *bio) 823 { 824 while (bio) { 825 bio->bio_offset = NOOFFSET; 826 bio = bio->bio_next; 827 } 828 } 829 830 /* 831 * Remove the buffer from the appropriate free list. 832 * (caller must be locked) 833 */ 834 static __inline void 835 _bremfree(struct buf *bp) 836 { 837 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 838 839 if (bp->b_qindex != BQUEUE_NONE) { 840 KASSERT(BUF_LOCKINUSE(bp), ("bremfree: bp %p not locked", bp)); 841 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist); 842 bp->b_qindex = BQUEUE_NONE; 843 } else { 844 if (!BUF_LOCKINUSE(bp)) 845 panic("bremfree: removing a buffer not on a queue"); 846 } 847 } 848 849 /* 850 * bremfree() - must be called with a locked buffer 851 */ 852 void 853 bremfree(struct buf *bp) 854 { 855 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 856 857 spin_lock(&pcpu->spin); 858 _bremfree(bp); 859 spin_unlock(&pcpu->spin); 860 } 861 862 /* 863 * bremfree_locked - must be called with pcpu->spin locked 864 */ 865 static void 866 bremfree_locked(struct buf *bp) 867 { 868 _bremfree(bp); 869 } 870 871 /* 872 * This version of bread issues any required I/O asyncnronously and 873 * makes a callback on completion. 874 * 875 * The callback must check whether BIO_DONE is set in the bio and issue 876 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 877 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 878 */ 879 void 880 breadcb(struct vnode *vp, off_t loffset, int size, int bflags, 881 void (*func)(struct bio *), void *arg) 882 { 883 struct buf *bp; 884 885 bp = getblk(vp, loffset, size, 0, 0); 886 887 /* if not found in cache, do some I/O */ 888 if ((bp->b_flags & B_CACHE) == 0) { 889 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA); 890 bp->b_flags |= bflags; 891 bp->b_cmd = BUF_CMD_READ; 892 bp->b_bio1.bio_done = func; 893 bp->b_bio1.bio_caller_info1.ptr = arg; 894 vfs_busy_pages(vp, bp); 895 BUF_KERNPROC(bp); 896 vn_strategy(vp, &bp->b_bio1); 897 } else if (func) { 898 /* 899 * Since we are issuing the callback synchronously it cannot 900 * race the BIO_DONE, so no need for atomic ops here. 901 */ 902 /*bp->b_bio1.bio_done = func;*/ 903 bp->b_bio1.bio_caller_info1.ptr = arg; 904 bp->b_bio1.bio_flags |= BIO_DONE; 905 func(&bp->b_bio1); 906 } else { 907 bqrelse(bp); 908 } 909 } 910 911 /* 912 * breadnx() - Terminal function for bread() and breadn(). 913 * 914 * This function will start asynchronous I/O on read-ahead blocks as well 915 * as satisfy the primary request. 916 * 917 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is 918 * set, the buffer is valid and we do not have to do anything. 919 */ 920 int 921 breadnx(struct vnode *vp, off_t loffset, int size, int bflags, 922 off_t *raoffset, int *rabsize, 923 int cnt, struct buf **bpp) 924 { 925 struct buf *bp, *rabp; 926 int i; 927 int rv = 0, readwait = 0; 928 int blkflags = (bflags & B_KVABIO) ? GETBLK_KVABIO : 0; 929 930 if (*bpp) 931 bp = *bpp; 932 else 933 *bpp = bp = getblk(vp, loffset, size, blkflags, 0); 934 935 /* if not found in cache, do some I/O */ 936 if ((bp->b_flags & B_CACHE) == 0) { 937 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA); 938 bp->b_flags |= bflags; 939 bp->b_cmd = BUF_CMD_READ; 940 bp->b_bio1.bio_done = biodone_sync; 941 bp->b_bio1.bio_flags |= BIO_SYNC; 942 vfs_busy_pages(vp, bp); 943 vn_strategy(vp, &bp->b_bio1); 944 ++readwait; 945 } 946 947 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 948 if (inmem(vp, *raoffset)) 949 continue; 950 rabp = getblk(vp, *raoffset, *rabsize, GETBLK_KVABIO, 0); 951 952 if ((rabp->b_flags & B_CACHE) == 0) { 953 rabp->b_flags &= ~(B_ERROR | B_EINTR | 954 B_INVAL | B_NOTMETA); 955 rabp->b_flags |= (bflags & ~B_KVABIO); 956 rabp->b_cmd = BUF_CMD_READ; 957 vfs_busy_pages(vp, rabp); 958 BUF_KERNPROC(rabp); 959 vn_strategy(vp, &rabp->b_bio1); 960 } else { 961 brelse(rabp); 962 } 963 } 964 if (readwait) 965 rv = biowait(&bp->b_bio1, "biord"); 966 return (rv); 967 } 968 969 /* 970 * bwrite: 971 * 972 * Synchronous write, waits for completion. 973 * 974 * Write, release buffer on completion. (Done by iodone 975 * if async). Do not bother writing anything if the buffer 976 * is invalid. 977 * 978 * Note that we set B_CACHE here, indicating that buffer is 979 * fully valid and thus cacheable. This is true even of NFS 980 * now so we set it generally. This could be set either here 981 * or in biodone() since the I/O is synchronous. We put it 982 * here. 983 */ 984 int 985 bwrite(struct buf *bp) 986 { 987 int error; 988 989 if (bp->b_flags & B_INVAL) { 990 brelse(bp); 991 return (0); 992 } 993 if (BUF_LOCKINUSE(bp) == 0) 994 panic("bwrite: buffer is not busy???"); 995 996 /* 997 * NOTE: We no longer mark the buffer clear prior to the vn_strategy() 998 * call because it will remove the buffer from the vnode's 999 * dirty buffer list prematurely and possibly cause filesystem 1000 * checks to race buffer flushes. This is now handled in 1001 * bpdone(). 1002 * 1003 * bundirty(bp); REMOVED 1004 */ 1005 1006 bp->b_flags &= ~(B_ERROR | B_EINTR); 1007 bp->b_flags |= B_CACHE; 1008 bp->b_cmd = BUF_CMD_WRITE; 1009 bp->b_bio1.bio_done = biodone_sync; 1010 bp->b_bio1.bio_flags |= BIO_SYNC; 1011 vfs_busy_pages(bp->b_vp, bp); 1012 1013 /* 1014 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1015 * valid for vnode-backed buffers. 1016 */ 1017 bsetrunningbufspace(bp, bp->b_bufsize); 1018 vn_strategy(bp->b_vp, &bp->b_bio1); 1019 error = biowait(&bp->b_bio1, "biows"); 1020 brelse(bp); 1021 1022 return (error); 1023 } 1024 1025 /* 1026 * bawrite: 1027 * 1028 * Asynchronous write. Start output on a buffer, but do not wait for 1029 * it to complete. The buffer is released when the output completes. 1030 * 1031 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1032 * B_INVAL buffers. Not us. 1033 */ 1034 void 1035 bawrite(struct buf *bp) 1036 { 1037 if (bp->b_flags & B_INVAL) { 1038 brelse(bp); 1039 return; 1040 } 1041 if (BUF_LOCKINUSE(bp) == 0) 1042 panic("bawrite: buffer is not busy???"); 1043 1044 /* 1045 * NOTE: We no longer mark the buffer clear prior to the vn_strategy() 1046 * call because it will remove the buffer from the vnode's 1047 * dirty buffer list prematurely and possibly cause filesystem 1048 * checks to race buffer flushes. This is now handled in 1049 * bpdone(). 1050 * 1051 * bundirty(bp); REMOVED 1052 */ 1053 bp->b_flags &= ~(B_ERROR | B_EINTR); 1054 bp->b_flags |= B_CACHE; 1055 bp->b_cmd = BUF_CMD_WRITE; 1056 KKASSERT(bp->b_bio1.bio_done == NULL); 1057 vfs_busy_pages(bp->b_vp, bp); 1058 1059 /* 1060 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1061 * valid for vnode-backed buffers. 1062 */ 1063 bsetrunningbufspace(bp, bp->b_bufsize); 1064 BUF_KERNPROC(bp); 1065 vn_strategy(bp->b_vp, &bp->b_bio1); 1066 } 1067 1068 /* 1069 * bdwrite: 1070 * 1071 * Delayed write. (Buffer is marked dirty). Do not bother writing 1072 * anything if the buffer is marked invalid. 1073 * 1074 * Note that since the buffer must be completely valid, we can safely 1075 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1076 * biodone() in order to prevent getblk from writing the buffer 1077 * out synchronously. 1078 */ 1079 void 1080 bdwrite(struct buf *bp) 1081 { 1082 if (BUF_LOCKINUSE(bp) == 0) 1083 panic("bdwrite: buffer is not busy"); 1084 1085 if (bp->b_flags & B_INVAL) { 1086 brelse(bp); 1087 return; 1088 } 1089 bdirty(bp); 1090 1091 dsched_buf_enter(bp); /* might stack */ 1092 1093 /* 1094 * Set B_CACHE, indicating that the buffer is fully valid. This is 1095 * true even of NFS now. 1096 */ 1097 bp->b_flags |= B_CACHE; 1098 1099 /* 1100 * This bmap keeps the system from needing to do the bmap later, 1101 * perhaps when the system is attempting to do a sync. Since it 1102 * is likely that the indirect block -- or whatever other datastructure 1103 * that the filesystem needs is still in memory now, it is a good 1104 * thing to do this. Note also, that if the pageout daemon is 1105 * requesting a sync -- there might not be enough memory to do 1106 * the bmap then... So, this is important to do. 1107 */ 1108 if (bp->b_bio2.bio_offset == NOOFFSET) { 1109 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1110 NULL, NULL, BUF_CMD_WRITE); 1111 } 1112 1113 /* 1114 * Because the underlying pages may still be mapped and 1115 * writable trying to set the dirty buffer (b_dirtyoff/end) 1116 * range here will be inaccurate. 1117 * 1118 * However, we must still clean the pages to satisfy the 1119 * vnode_pager and pageout daemon, so they think the pages 1120 * have been "cleaned". What has really occured is that 1121 * they've been earmarked for later writing by the buffer 1122 * cache. 1123 * 1124 * So we get the b_dirtyoff/end update but will not actually 1125 * depend on it (NFS that is) until the pages are busied for 1126 * writing later on. 1127 */ 1128 vfs_clean_pages(bp); 1129 bqrelse(bp); 1130 1131 /* 1132 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1133 * due to the softdep code. 1134 */ 1135 } 1136 1137 /* 1138 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1139 * This is used by tmpfs. 1140 * 1141 * It is important for any VFS using this routine to NOT use it for 1142 * IO_SYNC or IO_ASYNC operations which occur when the system really 1143 * wants to flush VM pages to backing store. 1144 */ 1145 void 1146 buwrite(struct buf *bp) 1147 { 1148 vm_page_t m; 1149 int i; 1150 1151 /* 1152 * Only works for VMIO buffers. If the buffer is already 1153 * marked for delayed-write we can't avoid the bdwrite(). 1154 */ 1155 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1156 bdwrite(bp); 1157 return; 1158 } 1159 1160 /* 1161 * Mark as needing a commit. 1162 */ 1163 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1164 m = bp->b_xio.xio_pages[i]; 1165 vm_page_need_commit(m); 1166 } 1167 bqrelse(bp); 1168 } 1169 1170 /* 1171 * bdirty: 1172 * 1173 * Turn buffer into delayed write request by marking it B_DELWRI. 1174 * B_RELBUF and B_NOCACHE must be cleared. 1175 * 1176 * We reassign the buffer to itself to properly update it in the 1177 * dirty/clean lists. 1178 * 1179 * Must be called from a critical section. 1180 * The buffer must be on BQUEUE_NONE. 1181 */ 1182 void 1183 bdirty(struct buf *bp) 1184 { 1185 KASSERT(bp->b_qindex == BQUEUE_NONE, 1186 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1187 if (bp->b_flags & B_NOCACHE) { 1188 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1189 bp->b_flags &= ~B_NOCACHE; 1190 } 1191 if (bp->b_flags & B_INVAL) { 1192 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1193 } 1194 bp->b_flags &= ~B_RELBUF; 1195 1196 if ((bp->b_flags & B_DELWRI) == 0) { 1197 lwkt_gettoken(&bp->b_vp->v_token); 1198 bp->b_flags |= B_DELWRI; 1199 reassignbuf(bp); 1200 lwkt_reltoken(&bp->b_vp->v_token); 1201 1202 atomic_add_long(&dirtybufcount, 1); 1203 atomic_add_long(&dirtykvaspace, bp->b_kvasize); 1204 atomic_add_long(&dirtybufspace, bp->b_bufsize); 1205 if (bp->b_flags & B_HEAVY) { 1206 atomic_add_long(&dirtybufcounthw, 1); 1207 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1208 } 1209 bd_heatup(); 1210 } 1211 } 1212 1213 /* 1214 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1215 * needs to be flushed with a different buf_daemon thread to avoid 1216 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1217 */ 1218 void 1219 bheavy(struct buf *bp) 1220 { 1221 if ((bp->b_flags & B_HEAVY) == 0) { 1222 bp->b_flags |= B_HEAVY; 1223 if (bp->b_flags & B_DELWRI) { 1224 atomic_add_long(&dirtybufcounthw, 1); 1225 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1226 } 1227 } 1228 } 1229 1230 /* 1231 * bundirty: 1232 * 1233 * Clear B_DELWRI for buffer. 1234 * 1235 * Must be called from a critical section. 1236 * 1237 * The buffer is typically on BQUEUE_NONE but there is one case in 1238 * brelse() that calls this function after placing the buffer on 1239 * a different queue. 1240 */ 1241 void 1242 bundirty(struct buf *bp) 1243 { 1244 if (bp->b_flags & B_DELWRI) { 1245 lwkt_gettoken(&bp->b_vp->v_token); 1246 bp->b_flags &= ~B_DELWRI; 1247 reassignbuf(bp); 1248 lwkt_reltoken(&bp->b_vp->v_token); 1249 1250 atomic_add_long(&dirtybufcount, -1); 1251 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1252 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1253 if (bp->b_flags & B_HEAVY) { 1254 atomic_add_long(&dirtybufcounthw, -1); 1255 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize); 1256 } 1257 bd_signal(bp->b_bufsize); 1258 } 1259 /* 1260 * Since it is now being written, we can clear its deferred write flag. 1261 */ 1262 bp->b_flags &= ~B_DEFERRED; 1263 } 1264 1265 /* 1266 * Set the b_runningbufspace field, used to track how much I/O is 1267 * in progress at any given moment. 1268 */ 1269 void 1270 bsetrunningbufspace(struct buf *bp, int bytes) 1271 { 1272 bp->b_runningbufspace = bytes; 1273 if (bytes) { 1274 atomic_add_long(&runningbufspace, bytes); 1275 atomic_add_long(&runningbufcount, 1); 1276 } 1277 } 1278 1279 /* 1280 * brelse: 1281 * 1282 * Release a busy buffer and, if requested, free its resources. The 1283 * buffer will be stashed in the appropriate bufqueue[] allowing it 1284 * to be accessed later as a cache entity or reused for other purposes. 1285 */ 1286 void 1287 brelse(struct buf *bp) 1288 { 1289 struct bufpcpu *pcpu; 1290 #ifdef INVARIANTS 1291 int saved_flags = bp->b_flags; 1292 #endif 1293 1294 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1295 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1296 1297 /* 1298 * If B_NOCACHE is set we are being asked to destroy the buffer and 1299 * its backing store. Clear B_DELWRI. 1300 * 1301 * B_NOCACHE is set in two cases: (1) when the caller really wants 1302 * to destroy the buffer and backing store and (2) when the caller 1303 * wants to destroy the buffer and backing store after a write 1304 * completes. 1305 */ 1306 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1307 bundirty(bp); 1308 } 1309 1310 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1311 /* 1312 * A re-dirtied buffer is only subject to destruction 1313 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1314 */ 1315 /* leave buffer intact */ 1316 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1317 (bp->b_bufsize <= 0)) { 1318 /* 1319 * Either a failed read or we were asked to free or not 1320 * cache the buffer. This path is reached with B_DELWRI 1321 * set only if B_INVAL is already set. B_NOCACHE governs 1322 * backing store destruction. 1323 * 1324 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1325 * buffer cannot be immediately freed. 1326 */ 1327 bp->b_flags |= B_INVAL; 1328 if (LIST_FIRST(&bp->b_dep) != NULL) 1329 buf_deallocate(bp); 1330 if (bp->b_flags & B_DELWRI) { 1331 atomic_add_long(&dirtybufcount, -1); 1332 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1333 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1334 if (bp->b_flags & B_HEAVY) { 1335 atomic_add_long(&dirtybufcounthw, -1); 1336 atomic_add_long(&dirtybufspacehw, 1337 -bp->b_bufsize); 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_min(0)) { 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_buf_exit(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 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 if ((bp->b_flags & B_HASBOGUS) == 0) 1452 panic("brelse: bp %p corrupt bogus", bp); 1453 mtmp = vm_page_lookup(obj, poff + j); 1454 if (!mtmp) 1455 panic("brelse: bp %p page %d missing", bp, j); 1456 bp->b_xio.xio_pages[j] = mtmp; 1457 } 1458 } 1459 vm_object_drop(obj); 1460 1461 if ((bp->b_flags & B_HASBOGUS) || 1462 (bp->b_flags & B_INVAL) == 0) { 1463 pmap_qenter_noinval( 1464 trunc_page((vm_offset_t)bp->b_data), 1465 bp->b_xio.xio_pages, 1466 bp->b_xio.xio_npages); 1467 bp->b_flags &= ~B_HASBOGUS; 1468 bp->b_flags |= B_KVABIO; 1469 bkvareset(bp); 1470 } 1471 m = bp->b_xio.xio_pages[i]; 1472 } 1473 1474 /* 1475 * Invalidate the backing store if B_NOCACHE is set 1476 * (e.g. used with vinvalbuf()). If this is NFS 1477 * we impose a requirement that the block size be 1478 * a multiple of PAGE_SIZE and create a temporary 1479 * hack to basically invalidate the whole page. The 1480 * problem is that NFS uses really odd buffer sizes 1481 * especially when tracking piecemeal writes and 1482 * it also vinvalbuf()'s a lot, which would result 1483 * in only partial page validation and invalidation 1484 * here. If the file page is mmap()'d, however, 1485 * all the valid bits get set so after we invalidate 1486 * here we would end up with weird m->valid values 1487 * like 0xfc. nfs_getpages() can't handle this so 1488 * we clear all the valid bits for the NFS case 1489 * instead of just some of them. 1490 * 1491 * The real bug is the VM system having to set m->valid 1492 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1493 * itself is an artifact of the whole 512-byte 1494 * granular mess that exists to support odd block 1495 * sizes and UFS meta-data block sizes (e.g. 6144). 1496 * A complete rewrite is required. 1497 * 1498 * XXX 1499 */ 1500 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1501 int poffset = foff & PAGE_MASK; 1502 int presid; 1503 1504 presid = PAGE_SIZE - poffset; 1505 if (bp->b_vp->v_tag == VT_NFS && 1506 bp->b_vp->v_type == VREG) { 1507 ; /* entire page */ 1508 } else if (presid > resid) { 1509 presid = resid; 1510 } 1511 KASSERT(presid >= 0, ("brelse: extra page")); 1512 vm_page_set_invalid(m, poffset, presid); 1513 1514 /* 1515 * Also make sure any swap cache is removed 1516 * as it is now stale (HAMMER in particular 1517 * uses B_NOCACHE to deal with buffer 1518 * aliasing). 1519 */ 1520 swap_pager_unswapped(m); 1521 } 1522 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1523 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1524 } 1525 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1526 vfs_vmio_release(bp); 1527 } else { 1528 /* 1529 * Rundown for non-VMIO buffers. 1530 */ 1531 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1532 if (bp->b_bufsize) 1533 allocbuf(bp, 0); 1534 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1535 if (bp->b_vp) 1536 brelvp(bp); 1537 } 1538 } 1539 1540 if (bp->b_qindex != BQUEUE_NONE) 1541 panic("brelse: free buffer onto another queue???"); 1542 1543 /* 1544 * Figure out the correct queue to place the cleaned up buffer on. 1545 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1546 * disassociated from their vnode. 1547 * 1548 * Return the buffer to its original pcpu area 1549 */ 1550 pcpu = &bufpcpu[bp->b_qcpu]; 1551 spin_lock(&pcpu->spin); 1552 1553 if (bp->b_flags & B_LOCKED) { 1554 /* 1555 * Buffers that are locked are placed in the locked queue 1556 * immediately, regardless of their state. 1557 */ 1558 bp->b_qindex = BQUEUE_LOCKED; 1559 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1560 bp, b_freelist); 1561 } else if (bp->b_bufsize == 0) { 1562 /* 1563 * Buffers with no memory. Due to conditionals near the top 1564 * of brelse() such buffers should probably already be 1565 * marked B_INVAL and disassociated from their vnode. 1566 */ 1567 bp->b_flags |= B_INVAL; 1568 KASSERT(bp->b_vp == NULL, 1569 ("bp1 %p flags %08x/%08x vnode %p " 1570 "unexpectededly still associated!", 1571 bp, saved_flags, bp->b_flags, bp->b_vp)); 1572 KKASSERT((bp->b_flags & B_HASHED) == 0); 1573 bp->b_qindex = BQUEUE_EMPTY; 1574 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1575 bp, b_freelist); 1576 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1577 /* 1578 * Buffers with junk contents. Again these buffers had better 1579 * already be disassociated from their vnode. 1580 */ 1581 KASSERT(bp->b_vp == NULL, 1582 ("bp2 %p flags %08x/%08x vnode %p unexpectededly " 1583 "still associated!", 1584 bp, saved_flags, bp->b_flags, bp->b_vp)); 1585 KKASSERT((bp->b_flags & B_HASHED) == 0); 1586 bp->b_flags |= B_INVAL; 1587 bp->b_qindex = BQUEUE_CLEAN; 1588 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1589 bp, b_freelist); 1590 } else { 1591 /* 1592 * Remaining buffers. These buffers are still associated with 1593 * their vnode. 1594 */ 1595 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1596 case B_DELWRI: 1597 bp->b_qindex = BQUEUE_DIRTY; 1598 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1599 bp, b_freelist); 1600 break; 1601 case B_DELWRI | B_HEAVY: 1602 bp->b_qindex = BQUEUE_DIRTY_HW; 1603 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1604 bp, b_freelist); 1605 break; 1606 default: 1607 /* 1608 * NOTE: Buffers are always placed at the end of the 1609 * queue. If B_AGE is not set the buffer will cycle 1610 * through the queue twice. 1611 */ 1612 bp->b_qindex = BQUEUE_CLEAN; 1613 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1614 bp, b_freelist); 1615 break; 1616 } 1617 } 1618 spin_unlock(&pcpu->spin); 1619 1620 /* 1621 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1622 * on the correct queue but we have not yet unlocked it. 1623 */ 1624 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1625 bundirty(bp); 1626 1627 /* 1628 * The bp is on an appropriate queue unless locked. If it is not 1629 * locked or dirty we can wakeup threads waiting for buffer space. 1630 * 1631 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1632 * if B_INVAL is set ). 1633 */ 1634 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1635 bufcountwakeup(); 1636 1637 /* 1638 * Something we can maybe free or reuse 1639 */ 1640 if (bp->b_bufsize || bp->b_kvasize) 1641 bufspacewakeup(); 1642 1643 /* 1644 * Clean up temporary flags and unlock the buffer. 1645 */ 1646 bp->b_flags &= ~(B_NOCACHE | B_RELBUF | B_DIRECT); 1647 BUF_UNLOCK(bp); 1648 } 1649 1650 /* 1651 * bqrelse: 1652 * 1653 * Release a buffer back to the appropriate queue but do not try to free 1654 * it. The buffer is expected to be used again soon. 1655 * 1656 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1657 * biodone() to requeue an async I/O on completion. It is also used when 1658 * known good buffers need to be requeued but we think we may need the data 1659 * again soon. 1660 * 1661 * XXX we should be able to leave the B_RELBUF hint set on completion. 1662 */ 1663 void 1664 bqrelse(struct buf *bp) 1665 { 1666 struct bufpcpu *pcpu; 1667 1668 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1669 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1670 1671 if (bp->b_qindex != BQUEUE_NONE) 1672 panic("bqrelse: free buffer onto another queue???"); 1673 1674 buf_act_advance(bp); 1675 1676 pcpu = &bufpcpu[bp->b_qcpu]; 1677 spin_lock(&pcpu->spin); 1678 1679 if (bp->b_flags & B_LOCKED) { 1680 /* 1681 * Locked buffers are released to the locked queue. However, 1682 * if the buffer is dirty it will first go into the dirty 1683 * queue and later on after the I/O completes successfully it 1684 * will be released to the locked queue. 1685 */ 1686 bp->b_qindex = BQUEUE_LOCKED; 1687 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1688 bp, b_freelist); 1689 } else if (bp->b_flags & B_DELWRI) { 1690 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1691 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1692 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1693 bp, b_freelist); 1694 } else if (vm_page_count_min(0)) { 1695 /* 1696 * We are too low on memory, we have to try to free the 1697 * buffer (most importantly: the wired pages making up its 1698 * backing store) *now*. 1699 */ 1700 spin_unlock(&pcpu->spin); 1701 brelse(bp); 1702 return; 1703 } else { 1704 bp->b_qindex = BQUEUE_CLEAN; 1705 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1706 bp, b_freelist); 1707 } 1708 spin_unlock(&pcpu->spin); 1709 1710 /* 1711 * We have now placed the buffer on the proper queue, but have yet 1712 * to unlock it. 1713 */ 1714 if ((bp->b_flags & B_LOCKED) == 0 && 1715 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1716 bufcountwakeup(); 1717 } 1718 1719 /* 1720 * Something we can maybe free or reuse. 1721 */ 1722 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1723 bufspacewakeup(); 1724 1725 /* 1726 * Final cleanup and unlock. Clear bits that are only used while a 1727 * buffer is actively locked. 1728 */ 1729 bp->b_flags &= ~(B_NOCACHE | B_RELBUF); 1730 dsched_buf_exit(bp); 1731 BUF_UNLOCK(bp); 1732 } 1733 1734 /* 1735 * Hold a buffer, preventing it from being reused. This will prevent 1736 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL 1737 * operations. If a B_INVAL operation occurs the buffer will remain held 1738 * but the underlying pages may get ripped out. 1739 * 1740 * These functions are typically used in VOP_READ/VOP_WRITE functions 1741 * to hold a buffer during a copyin or copyout, preventing deadlocks 1742 * or recursive lock panics when read()/write() is used over mmap()'d 1743 * space. 1744 * 1745 * NOTE: bqhold() requires that the buffer be locked at the time of the 1746 * hold. bqdrop() has no requirements other than the buffer having 1747 * previously been held. 1748 */ 1749 void 1750 bqhold(struct buf *bp) 1751 { 1752 atomic_add_int(&bp->b_refs, 1); 1753 } 1754 1755 void 1756 bqdrop(struct buf *bp) 1757 { 1758 KKASSERT(bp->b_refs > 0); 1759 atomic_add_int(&bp->b_refs, -1); 1760 } 1761 1762 /* 1763 * Return backing pages held by the buffer 'bp' back to the VM system. 1764 * This routine is called when the bp is invalidated, released, or 1765 * reused. 1766 * 1767 * The KVA mapping (b_data) for the underlying pages is removed by 1768 * this function. 1769 * 1770 * WARNING! This routine is integral to the low memory critical path 1771 * when a buffer is B_RELBUF'd. If the system has a severe page 1772 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE 1773 * queues so they can be reused in the current pageout daemon 1774 * pass. 1775 */ 1776 static void 1777 vfs_vmio_release(struct buf *bp) 1778 { 1779 int i; 1780 vm_page_t m; 1781 1782 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1783 m = bp->b_xio.xio_pages[i]; 1784 bp->b_xio.xio_pages[i] = NULL; 1785 1786 /* 1787 * We need to own the page in order to safely unwire it. 1788 */ 1789 vm_page_busy_wait(m, FALSE, "vmiopg"); 1790 1791 /* 1792 * The VFS is telling us this is not a meta-data buffer 1793 * even if it is backed by a block device. 1794 */ 1795 if (bp->b_flags & B_NOTMETA) 1796 vm_page_flag_set(m, PG_NOTMETA); 1797 1798 /* 1799 * This is a very important bit of code. We try to track 1800 * VM page use whether the pages are wired into the buffer 1801 * cache or not. While wired into the buffer cache the 1802 * bp tracks the act_count. 1803 * 1804 * We can choose to place unwired pages on the inactive 1805 * queue (0) or active queue (1). If we place too many 1806 * on the active queue the queue will cycle the act_count 1807 * on pages we'd like to keep, just from single-use pages 1808 * (such as when doing a tar-up or file scan). 1809 */ 1810 if (bp->b_act_count < vm_cycle_point) 1811 vm_page_unwire(m, 0); 1812 else 1813 vm_page_unwire(m, 1); 1814 1815 /* 1816 * If the wire_count has dropped to 0 we may need to take 1817 * further action before unbusying the page. 1818 * 1819 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us. 1820 */ 1821 if (m->wire_count == 0) { 1822 if (bp->b_flags & B_DIRECT) { 1823 /* 1824 * Attempt to free the page if B_DIRECT is 1825 * set, the caller does not desire the page 1826 * to be cached. 1827 */ 1828 vm_page_wakeup(m); 1829 vm_page_try_to_free(m); 1830 } else if ((bp->b_flags & B_NOTMETA) || 1831 vm_page_count_min(0)) { 1832 /* 1833 * Attempt to move the page to PQ_CACHE 1834 * if B_NOTMETA is set. This flag is set 1835 * by HAMMER to remove one of the two pages 1836 * present when double buffering is enabled. 1837 * 1838 * Attempt to move the page to PQ_CACHE 1839 * If we have a severe page deficit. This 1840 * will cause buffer cache operations related 1841 * to pageouts to recycle the related pages 1842 * in order to avoid a low memory deadlock. 1843 */ 1844 m->act_count = bp->b_act_count; 1845 vm_page_try_to_cache(m); 1846 } else { 1847 /* 1848 * Nominal case, leave the page on the 1849 * queue the original unwiring placed it on 1850 * (active or inactive). 1851 */ 1852 m->act_count = bp->b_act_count; 1853 vm_page_wakeup(m); 1854 } 1855 } else { 1856 vm_page_wakeup(m); 1857 } 1858 } 1859 1860 /* 1861 * Zero out the pmap pte's for the mapping, but don't bother 1862 * invalidating the TLB. The range will be properly invalidating 1863 * when new pages are entered into the mapping. 1864 * 1865 * This in particular reduces tmpfs tear-down overhead and reduces 1866 * buffer cache re-use overhead (one invalidation sequence instead 1867 * of two per re-use). 1868 */ 1869 pmap_qremove_noinval(trunc_page((vm_offset_t) bp->b_data), 1870 bp->b_xio.xio_npages); 1871 CPUMASK_ASSZERO(bp->b_cpumask); 1872 if (bp->b_bufsize) { 1873 atomic_add_long(&bufspace, -bp->b_bufsize); 1874 bp->b_bufsize = 0; 1875 bufspacewakeup(); 1876 } 1877 bp->b_xio.xio_npages = 0; 1878 bp->b_flags &= ~B_VMIO; 1879 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1880 if (bp->b_vp) 1881 brelvp(bp); 1882 } 1883 1884 /* 1885 * Find and initialize a new buffer header, freeing up existing buffers 1886 * in the bufqueues as necessary. The new buffer is returned locked. 1887 * 1888 * Important: B_INVAL is not set. If the caller wishes to throw the 1889 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1890 * 1891 * We block if: 1892 * We have insufficient buffer headers 1893 * We have insufficient buffer space 1894 * 1895 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1896 * Instead we ask the buf daemon to do it for us. We attempt to 1897 * avoid piecemeal wakeups of the pageout daemon. 1898 */ 1899 struct buf * 1900 getnewbuf(int blkflags, int slptimeo, int size, int maxsize) 1901 { 1902 struct bufpcpu *pcpu; 1903 struct buf *bp; 1904 struct buf *nbp; 1905 int nqindex; 1906 int nqcpu; 1907 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1908 int maxloops = 200000; 1909 int restart_reason = 0; 1910 struct buf *restart_bp = NULL; 1911 static char flushingbufs[MAXCPU]; 1912 char *flushingp; 1913 1914 /* 1915 * We can't afford to block since we might be holding a vnode lock, 1916 * which may prevent system daemons from running. We deal with 1917 * low-memory situations by proactively returning memory and running 1918 * async I/O rather then sync I/O. 1919 */ 1920 1921 ++getnewbufcalls; 1922 nqcpu = mycpu->gd_cpuid; 1923 flushingp = &flushingbufs[nqcpu]; 1924 restart: 1925 if (bufspace < lobufspace) 1926 *flushingp = 0; 1927 1928 if (debug_bufbio && --maxloops == 0) 1929 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)", 1930 mycpu->gd_cpuid, restart_reason, restart_bp); 1931 1932 /* 1933 * Setup for scan. If we do not have enough free buffers, 1934 * we setup a degenerate case that immediately fails. Note 1935 * that if we are specially marked process, we are allowed to 1936 * dip into our reserves. 1937 * 1938 * The scanning sequence is nominally: EMPTY->CLEAN 1939 */ 1940 pcpu = &bufpcpu[nqcpu]; 1941 spin_lock(&pcpu->spin); 1942 1943 /* 1944 * Prime the scan for this cpu. Locate the first buffer to 1945 * check. If we are flushing buffers we must skip the 1946 * EMPTY queue. 1947 */ 1948 nqindex = BQUEUE_EMPTY; 1949 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]); 1950 if (nbp == NULL || *flushingp) { 1951 nqindex = BQUEUE_CLEAN; 1952 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]); 1953 } 1954 1955 /* 1956 * Run scan, possibly freeing data and/or kva mappings on the fly, 1957 * depending. 1958 * 1959 * WARNING! spin is held! 1960 */ 1961 while ((bp = nbp) != NULL) { 1962 int qindex = nqindex; 1963 1964 nbp = TAILQ_NEXT(bp, b_freelist); 1965 1966 /* 1967 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 1968 * cycles through the queue twice before being selected. 1969 */ 1970 if (qindex == BQUEUE_CLEAN && 1971 (bp->b_flags & B_AGE) == 0 && nbp) { 1972 bp->b_flags |= B_AGE; 1973 TAILQ_REMOVE(&pcpu->bufqueues[qindex], 1974 bp, b_freelist); 1975 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex], 1976 bp, b_freelist); 1977 continue; 1978 } 1979 1980 /* 1981 * Calculate next bp ( we can only use it if we do not block 1982 * or do other fancy things ). 1983 */ 1984 if (nbp == NULL) { 1985 switch(qindex) { 1986 case BQUEUE_EMPTY: 1987 nqindex = BQUEUE_CLEAN; 1988 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]))) 1989 break; 1990 /* fall through */ 1991 case BQUEUE_CLEAN: 1992 /* 1993 * nbp is NULL. 1994 */ 1995 break; 1996 } 1997 } 1998 1999 /* 2000 * Sanity Checks 2001 */ 2002 KASSERT(bp->b_qindex == qindex, 2003 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2004 2005 /* 2006 * Note: we no longer distinguish between VMIO and non-VMIO 2007 * buffers. 2008 */ 2009 KASSERT((bp->b_flags & B_DELWRI) == 0, 2010 ("delwri buffer %p found in queue %d", bp, qindex)); 2011 2012 /* 2013 * Do not try to reuse a buffer with a non-zero b_refs. 2014 * This is an unsynchronized test. A synchronized test 2015 * is also performed after we lock the buffer. 2016 */ 2017 if (bp->b_refs) 2018 continue; 2019 2020 /* 2021 * Start freeing the bp. This is somewhat involved. nbp 2022 * remains valid only for BQUEUE_EMPTY bp's. Buffers 2023 * on the clean list must be disassociated from their 2024 * current vnode. Buffers on the empty lists have 2025 * already been disassociated. 2026 * 2027 * b_refs is checked after locking along with queue changes. 2028 * We must check here to deal with zero->nonzero transitions 2029 * made by the owner of the buffer lock, which is used by 2030 * VFS's to hold the buffer while issuing an unlocked 2031 * uiomove()s. We cannot invalidate the buffer's pages 2032 * for this case. Once we successfully lock a buffer the 2033 * only 0->1 transitions of b_refs will occur via findblk(). 2034 * 2035 * We must also check for queue changes after successful 2036 * locking as the current lock holder may dispose of the 2037 * buffer and change its queue. 2038 */ 2039 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2040 spin_unlock(&pcpu->spin); 2041 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2042 restart_reason = 1; 2043 restart_bp = bp; 2044 goto restart; 2045 } 2046 if (bp->b_qindex != qindex || bp->b_refs) { 2047 spin_unlock(&pcpu->spin); 2048 BUF_UNLOCK(bp); 2049 restart_reason = 2; 2050 restart_bp = bp; 2051 goto restart; 2052 } 2053 bremfree_locked(bp); 2054 spin_unlock(&pcpu->spin); 2055 2056 /* 2057 * Dependancies must be handled before we disassociate the 2058 * vnode. 2059 * 2060 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2061 * be immediately disassociated. HAMMER then becomes 2062 * responsible for releasing the buffer. 2063 * 2064 * NOTE: spin is UNLOCKED now. 2065 */ 2066 if (LIST_FIRST(&bp->b_dep) != NULL) { 2067 buf_deallocate(bp); 2068 if (bp->b_flags & B_LOCKED) { 2069 bqrelse(bp); 2070 restart_reason = 3; 2071 restart_bp = bp; 2072 goto restart; 2073 } 2074 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2075 } 2076 2077 /* 2078 * CLEAN buffers have content or associations that must be 2079 * cleaned out if not repurposing. 2080 */ 2081 if (qindex == BQUEUE_CLEAN) { 2082 if (bp->b_flags & B_VMIO) 2083 vfs_vmio_release(bp); 2084 if (bp->b_vp) 2085 brelvp(bp); 2086 } 2087 2088 /* 2089 * NOTE: nbp is now entirely invalid. We can only restart 2090 * the scan from this point on. 2091 * 2092 * Get the rest of the buffer freed up. b_kva* is still 2093 * valid after this operation. 2094 */ 2095 KASSERT(bp->b_vp == NULL, 2096 ("bp3 %p flags %08x vnode %p qindex %d " 2097 "unexpectededly still associated!", 2098 bp, bp->b_flags, bp->b_vp, qindex)); 2099 KKASSERT((bp->b_flags & B_HASHED) == 0); 2100 2101 if (bp->b_bufsize) 2102 allocbuf(bp, 0); 2103 2104 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) { 2105 kprintf("getnewbuf: caught bug vp queue " 2106 "%p/%08x qidx %d\n", 2107 bp, bp->b_flags, qindex); 2108 brelvp(bp); 2109 } 2110 bp->b_flags = B_BNOCLIP; 2111 bp->b_cmd = BUF_CMD_DONE; 2112 bp->b_vp = NULL; 2113 bp->b_error = 0; 2114 bp->b_resid = 0; 2115 bp->b_bcount = 0; 2116 bp->b_xio.xio_npages = 0; 2117 bp->b_dirtyoff = bp->b_dirtyend = 0; 2118 bp->b_act_count = ACT_INIT; 2119 reinitbufbio(bp); 2120 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2121 buf_dep_init(bp); 2122 if (blkflags & GETBLK_BHEAVY) 2123 bp->b_flags |= B_HEAVY; 2124 2125 if (bufspace >= hibufspace) 2126 *flushingp = 1; 2127 if (bufspace < lobufspace) 2128 *flushingp = 0; 2129 if (*flushingp) { 2130 bp->b_flags |= B_INVAL; 2131 brelse(bp); 2132 restart_reason = 5; 2133 restart_bp = bp; 2134 goto restart; 2135 } 2136 2137 /* 2138 * b_refs can transition to a non-zero value while we hold 2139 * the buffer locked due to a findblk(). Our brelvp() above 2140 * interlocked any future possible transitions due to 2141 * findblk()s. 2142 * 2143 * If we find b_refs to be non-zero we can destroy the 2144 * buffer's contents but we cannot yet reuse the buffer. 2145 */ 2146 if (bp->b_refs) { 2147 bp->b_flags |= B_INVAL; 2148 brelse(bp); 2149 restart_reason = 6; 2150 restart_bp = bp; 2151 2152 goto restart; 2153 } 2154 2155 /* 2156 * We found our buffer! 2157 */ 2158 break; 2159 } 2160 2161 /* 2162 * If we exhausted our list, iterate other cpus. If that fails, 2163 * sleep as appropriate. We may have to wakeup various daemons 2164 * and write out some dirty buffers. 2165 * 2166 * Generally we are sleeping due to insufficient buffer space. 2167 * 2168 * NOTE: spin is held if bp is NULL, else it is not held. 2169 */ 2170 if (bp == NULL) { 2171 int flags; 2172 char *waitmsg; 2173 2174 spin_unlock(&pcpu->spin); 2175 2176 nqcpu = (nqcpu + 1) % ncpus; 2177 if (nqcpu != mycpu->gd_cpuid) { 2178 restart_reason = 7; 2179 restart_bp = bp; 2180 goto restart; 2181 } 2182 2183 if (bufspace >= hibufspace) { 2184 waitmsg = "bufspc"; 2185 flags = VFS_BIO_NEED_BUFSPACE; 2186 } else { 2187 waitmsg = "newbuf"; 2188 flags = VFS_BIO_NEED_ANY; 2189 } 2190 2191 bd_speedup(); /* heeeelp */ 2192 atomic_set_int(&needsbuffer, flags); 2193 while (needsbuffer & flags) { 2194 int value; 2195 2196 tsleep_interlock(&needsbuffer, 0); 2197 value = atomic_fetchadd_int(&needsbuffer, 0); 2198 if (value & flags) { 2199 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags, 2200 waitmsg, slptimeo)) { 2201 return (NULL); 2202 } 2203 } 2204 } 2205 } else { 2206 /* 2207 * We finally have a valid bp. Reset b_data. 2208 * 2209 * (spin is not held) 2210 */ 2211 bp->b_data = bp->b_kvabase; 2212 } 2213 return(bp); 2214 } 2215 2216 /* 2217 * buf_daemon: 2218 * 2219 * Buffer flushing daemon. Buffers are normally flushed by the 2220 * update daemon but if it cannot keep up this process starts to 2221 * take the load in an attempt to prevent getnewbuf() from blocking. 2222 * 2223 * Once a flush is initiated it does not stop until the number 2224 * of buffers falls below lodirtybuffers, but we will wake up anyone 2225 * waiting at the mid-point. 2226 */ 2227 static struct kproc_desc buf_kp = { 2228 "bufdaemon", 2229 buf_daemon, 2230 &bufdaemon_td 2231 }; 2232 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2233 kproc_start, &buf_kp); 2234 2235 static struct kproc_desc bufhw_kp = { 2236 "bufdaemon_hw", 2237 buf_daemon_hw, 2238 &bufdaemonhw_td 2239 }; 2240 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2241 kproc_start, &bufhw_kp); 2242 2243 static void 2244 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long), 2245 int *bd_req) 2246 { 2247 long limit; 2248 struct buf *marker; 2249 2250 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO); 2251 marker->b_flags |= B_MARKER; 2252 marker->b_qindex = BQUEUE_NONE; 2253 marker->b_qcpu = 0; 2254 2255 /* 2256 * This process needs to be suspended prior to shutdown sync. 2257 */ 2258 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2259 td, SHUTDOWN_PRI_LAST); 2260 curthread->td_flags |= TDF_SYSTHREAD; 2261 2262 /* 2263 * This process is allowed to take the buffer cache to the limit 2264 */ 2265 for (;;) { 2266 kproc_suspend_loop(); 2267 2268 /* 2269 * Do the flush as long as the number of dirty buffers 2270 * (including those running) exceeds lodirtybufspace. 2271 * 2272 * When flushing limit running I/O to hirunningspace 2273 * Do the flush. Limit the amount of in-transit I/O we 2274 * allow to build up, otherwise we would completely saturate 2275 * the I/O system. Wakeup any waiting processes before we 2276 * normally would so they can run in parallel with our drain. 2277 * 2278 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2279 * but because we split the operation into two threads we 2280 * have to cut it in half for each thread. 2281 */ 2282 waitrunningbufspace(); 2283 limit = lodirtybufspace / 2; 2284 while (buf_limit_fn(limit)) { 2285 if (flushbufqueues(marker, queue) == 0) 2286 break; 2287 if (runningbufspace < hirunningspace) 2288 continue; 2289 waitrunningbufspace(); 2290 } 2291 2292 /* 2293 * We reached our low water mark, reset the 2294 * request and sleep until we are needed again. 2295 * The sleep is just so the suspend code works. 2296 */ 2297 tsleep_interlock(bd_req, 0); 2298 if (atomic_swap_int(bd_req, 0) == 0) 2299 tsleep(bd_req, PINTERLOCKED, "psleep", hz); 2300 } 2301 /* NOT REACHED */ 2302 /*kfree(marker, M_BIOBUF);*/ 2303 } 2304 2305 static int 2306 buf_daemon_limit(long limit) 2307 { 2308 return (runningbufspace + dirtykvaspace > limit || 2309 dirtybufcount - dirtybufcounthw >= nbuf / 2); 2310 } 2311 2312 static int 2313 buf_daemon_hw_limit(long limit) 2314 { 2315 return (runningbufspace + dirtykvaspace > limit || 2316 dirtybufcounthw >= nbuf / 2); 2317 } 2318 2319 static void 2320 buf_daemon(void) 2321 { 2322 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit, 2323 &bd_request); 2324 } 2325 2326 static void 2327 buf_daemon_hw(void) 2328 { 2329 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit, 2330 &bd_request_hw); 2331 } 2332 2333 /* 2334 * Flush up to (flushperqueue) buffers in the dirty queue. Each cpu has a 2335 * localized version of the queue. Each call made to this function iterates 2336 * to another cpu. It is desireable to flush several buffers from the same 2337 * cpu's queue at once, as these are likely going to be linear. 2338 * 2339 * We must be careful to free up B_INVAL buffers instead of write them, which 2340 * NFS is particularly sensitive to. 2341 * 2342 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate that we 2343 * really want to try to get the buffer out and reuse it due to the write 2344 * load on the machine. 2345 * 2346 * We must lock the buffer in order to check its validity before we can mess 2347 * with its contents. spin isn't enough. 2348 */ 2349 static int 2350 flushbufqueues(struct buf *marker, bufq_type_t q) 2351 { 2352 struct bufpcpu *pcpu; 2353 struct buf *bp; 2354 int r = 0; 2355 u_int loops = flushperqueue; 2356 int lcpu = marker->b_qcpu; 2357 2358 KKASSERT(marker->b_qindex == BQUEUE_NONE); 2359 KKASSERT(marker->b_flags & B_MARKER); 2360 2361 again: 2362 /* 2363 * Spinlock needed to perform operations on the queue and may be 2364 * held through a non-blocking BUF_LOCK(), but cannot be held when 2365 * BUF_UNLOCK()ing or through any other major operation. 2366 */ 2367 pcpu = &bufpcpu[marker->b_qcpu]; 2368 spin_lock(&pcpu->spin); 2369 marker->b_qindex = q; 2370 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist); 2371 bp = marker; 2372 2373 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) { 2374 /* 2375 * NOTE: spinlock is always held at the top of the loop 2376 */ 2377 if (bp->b_flags & B_MARKER) 2378 continue; 2379 if ((bp->b_flags & B_DELWRI) == 0) { 2380 kprintf("Unexpected clean buffer %p\n", bp); 2381 continue; 2382 } 2383 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 2384 continue; 2385 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q); 2386 2387 /* 2388 * Once the buffer is locked we will have no choice but to 2389 * unlock the spinlock around a later BUF_UNLOCK and re-set 2390 * bp = marker when looping. Move the marker now to make 2391 * things easier. 2392 */ 2393 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2394 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist); 2395 2396 /* 2397 * Must recheck B_DELWRI after successfully locking 2398 * the buffer. 2399 */ 2400 if ((bp->b_flags & B_DELWRI) == 0) { 2401 spin_unlock(&pcpu->spin); 2402 BUF_UNLOCK(bp); 2403 spin_lock(&pcpu->spin); 2404 bp = marker; 2405 continue; 2406 } 2407 2408 /* 2409 * Remove the buffer from its queue. We still own the 2410 * spinlock here. 2411 */ 2412 _bremfree(bp); 2413 2414 /* 2415 * Disposing of an invalid buffer counts as a flush op 2416 */ 2417 if (bp->b_flags & B_INVAL) { 2418 spin_unlock(&pcpu->spin); 2419 brelse(bp); 2420 goto doloop; 2421 } 2422 2423 /* 2424 * Release the spinlock for the more complex ops we 2425 * are now going to do. 2426 */ 2427 spin_unlock(&pcpu->spin); 2428 lwkt_yield(); 2429 2430 /* 2431 * This is a bit messy 2432 */ 2433 if (LIST_FIRST(&bp->b_dep) != NULL && 2434 (bp->b_flags & B_DEFERRED) == 0 && 2435 buf_countdeps(bp, 0)) { 2436 spin_lock(&pcpu->spin); 2437 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist); 2438 bp->b_qindex = q; 2439 bp->b_flags |= B_DEFERRED; 2440 spin_unlock(&pcpu->spin); 2441 BUF_UNLOCK(bp); 2442 spin_lock(&pcpu->spin); 2443 bp = marker; 2444 continue; 2445 } 2446 2447 /* 2448 * spinlock not held here. 2449 * 2450 * If the buffer has a dependancy, buf_checkwrite() must 2451 * also return 0 for us to be able to initate the write. 2452 * 2453 * If the buffer is flagged B_ERROR it may be requeued 2454 * over and over again, we try to avoid a live lock. 2455 */ 2456 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2457 brelse(bp); 2458 } else if (bp->b_flags & B_ERROR) { 2459 tsleep(bp, 0, "bioer", 1); 2460 bp->b_flags &= ~B_AGE; 2461 cluster_awrite(bp); 2462 } else { 2463 bp->b_flags |= B_AGE | B_KVABIO; 2464 cluster_awrite(bp); 2465 } 2466 /* bp invalid but needs to be NULL-tested if we break out */ 2467 doloop: 2468 spin_lock(&pcpu->spin); 2469 ++r; 2470 if (--loops == 0) 2471 break; 2472 bp = marker; 2473 } 2474 /* bp is invalid here but can be NULL-tested to advance */ 2475 2476 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2477 marker->b_qindex = BQUEUE_NONE; 2478 spin_unlock(&pcpu->spin); 2479 2480 /* 2481 * Advance the marker to be fair. 2482 */ 2483 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus; 2484 if (bp == NULL) { 2485 if (marker->b_qcpu != lcpu) 2486 goto again; 2487 } 2488 2489 return (r); 2490 } 2491 2492 /* 2493 * inmem: 2494 * 2495 * Returns true if no I/O is needed to access the associated VM object. 2496 * This is like findblk except it also hunts around in the VM system for 2497 * the data. 2498 * 2499 * Note that we ignore vm_page_free() races from interrupts against our 2500 * lookup, since if the caller is not protected our return value will not 2501 * be any more valid then otherwise once we exit the critical section. 2502 */ 2503 int 2504 inmem(struct vnode *vp, off_t loffset) 2505 { 2506 vm_object_t obj; 2507 vm_offset_t toff, tinc, size; 2508 vm_page_t m; 2509 int res = 1; 2510 2511 if (findblk(vp, loffset, FINDBLK_TEST)) 2512 return 1; 2513 if (vp->v_mount == NULL) 2514 return 0; 2515 if ((obj = vp->v_object) == NULL) 2516 return 0; 2517 2518 size = PAGE_SIZE; 2519 if (size > vp->v_mount->mnt_stat.f_iosize) 2520 size = vp->v_mount->mnt_stat.f_iosize; 2521 2522 vm_object_hold(obj); 2523 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2524 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2525 if (m == NULL) { 2526 res = 0; 2527 break; 2528 } 2529 tinc = size; 2530 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2531 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2532 if (vm_page_is_valid(m, 2533 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2534 res = 0; 2535 break; 2536 } 2537 } 2538 vm_object_drop(obj); 2539 return (res); 2540 } 2541 2542 /* 2543 * findblk: 2544 * 2545 * Locate and return the specified buffer. Unless flagged otherwise, 2546 * a locked buffer will be returned if it exists or NULL if it does not. 2547 * 2548 * findblk()'d buffers are still on the bufqueues and if you intend 2549 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2550 * and possibly do other stuff to it. 2551 * 2552 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2553 * for locking the buffer and ensuring that it remains 2554 * the desired buffer after locking. 2555 * 2556 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2557 * to acquire the lock we return NULL, even if the 2558 * buffer exists. 2559 * 2560 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2561 * reuse by getnewbuf() but does not prevent 2562 * disassociation (B_INVAL). Used to avoid deadlocks 2563 * against random (vp,loffset)s due to reassignment. 2564 * 2565 * FINDBLK_KVABIO - Only applicable when returning a locked buffer. 2566 * Indicates that the caller supports B_KVABIO. 2567 * 2568 * (0) - Lock the buffer blocking. 2569 */ 2570 struct buf * 2571 findblk(struct vnode *vp, off_t loffset, int flags) 2572 { 2573 struct buf *bp; 2574 int lkflags; 2575 2576 lkflags = LK_EXCLUSIVE; 2577 if (flags & FINDBLK_NBLOCK) 2578 lkflags |= LK_NOWAIT; 2579 2580 for (;;) { 2581 /* 2582 * Lookup. Ref the buf while holding v_token to prevent 2583 * reuse (but does not prevent diassociation). 2584 */ 2585 lwkt_gettoken_shared(&vp->v_token); 2586 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2587 if (bp == NULL) { 2588 lwkt_reltoken(&vp->v_token); 2589 return(NULL); 2590 } 2591 bqhold(bp); 2592 lwkt_reltoken(&vp->v_token); 2593 2594 /* 2595 * If testing only break and return bp, do not lock. 2596 */ 2597 if (flags & FINDBLK_TEST) 2598 break; 2599 2600 /* 2601 * Lock the buffer, return an error if the lock fails. 2602 * (only FINDBLK_NBLOCK can cause the lock to fail). 2603 */ 2604 if (BUF_LOCK(bp, lkflags)) { 2605 atomic_subtract_int(&bp->b_refs, 1); 2606 /* bp = NULL; not needed */ 2607 return(NULL); 2608 } 2609 2610 /* 2611 * Revalidate the locked buf before allowing it to be 2612 * returned. 2613 * 2614 * B_KVABIO is only set/cleared when locking. When 2615 * clearing B_KVABIO, we must ensure that the buffer 2616 * is synchronized to all cpus. 2617 */ 2618 if (bp->b_vp == vp && bp->b_loffset == loffset) { 2619 if (flags & FINDBLK_KVABIO) 2620 bp->b_flags |= B_KVABIO; 2621 else 2622 bkvasync_all(bp); 2623 break; 2624 } 2625 atomic_subtract_int(&bp->b_refs, 1); 2626 BUF_UNLOCK(bp); 2627 } 2628 2629 /* 2630 * Success 2631 */ 2632 if ((flags & FINDBLK_REF) == 0) 2633 atomic_subtract_int(&bp->b_refs, 1); 2634 return(bp); 2635 } 2636 2637 /* 2638 * getcacheblk: 2639 * 2640 * Similar to getblk() except only returns the buffer if it is 2641 * B_CACHE and requires no other manipulation. Otherwise NULL 2642 * is returned. NULL is also returned if GETBLK_NOWAIT is set 2643 * and the getblk() would block. 2644 * 2645 * If B_RAM is set the buffer might be just fine, but we return 2646 * NULL anyway because we want the code to fall through to the 2647 * cluster read to issue more read-aheads. Otherwise read-ahead breaks. 2648 * 2649 * If blksize is 0 the buffer cache buffer must already be fully 2650 * cached. 2651 * 2652 * If blksize is non-zero getblk() will be used, allowing a buffer 2653 * to be reinstantiated from its VM backing store. The buffer must 2654 * still be fully cached after reinstantiation to be returned. 2655 */ 2656 struct buf * 2657 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags) 2658 { 2659 struct buf *bp; 2660 int fndflags = 0; 2661 2662 if (blkflags & GETBLK_NOWAIT) 2663 fndflags |= FINDBLK_NBLOCK; 2664 if (blkflags & GETBLK_KVABIO) 2665 fndflags |= FINDBLK_KVABIO; 2666 2667 if (blksize) { 2668 bp = getblk(vp, loffset, blksize, blkflags, 0); 2669 if (bp) { 2670 if ((bp->b_flags & (B_INVAL | B_CACHE)) == B_CACHE) { 2671 bp->b_flags &= ~B_AGE; 2672 if (bp->b_flags & B_RAM) { 2673 bqrelse(bp); 2674 bp = NULL; 2675 } 2676 } else { 2677 brelse(bp); 2678 bp = NULL; 2679 } 2680 } 2681 } else { 2682 bp = findblk(vp, loffset, fndflags); 2683 if (bp) { 2684 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2685 B_CACHE) { 2686 bp->b_flags &= ~B_AGE; 2687 bremfree(bp); 2688 } else { 2689 BUF_UNLOCK(bp); 2690 bp = NULL; 2691 } 2692 } 2693 } 2694 return (bp); 2695 } 2696 2697 /* 2698 * getblk: 2699 * 2700 * Get a block given a specified block and offset into a file/device. 2701 * B_INVAL may or may not be set on return. The caller should clear 2702 * B_INVAL prior to initiating a READ. 2703 * 2704 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2705 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2706 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2707 * without doing any of those things the system will likely believe 2708 * the buffer to be valid (especially if it is not B_VMIO), and the 2709 * next getblk() will return the buffer with B_CACHE set. 2710 * 2711 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2712 * an existing buffer. 2713 * 2714 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2715 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2716 * and then cleared based on the backing VM. If the previous buffer is 2717 * non-0-sized but invalid, B_CACHE will be cleared. 2718 * 2719 * If getblk() must create a new buffer, the new buffer is returned with 2720 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2721 * case it is returned with B_INVAL clear and B_CACHE set based on the 2722 * backing VM. 2723 * 2724 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2725 * B_CACHE bit is clear. 2726 * 2727 * What this means, basically, is that the caller should use B_CACHE to 2728 * determine whether the buffer is fully valid or not and should clear 2729 * B_INVAL prior to issuing a read. If the caller intends to validate 2730 * the buffer by loading its data area with something, the caller needs 2731 * to clear B_INVAL. If the caller does this without issuing an I/O, 2732 * the caller should set B_CACHE ( as an optimization ), else the caller 2733 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2734 * a write attempt or if it was a successfull read. If the caller 2735 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2736 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2737 * 2738 * getblk flags: 2739 * 2740 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2741 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2742 */ 2743 struct buf * 2744 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2745 { 2746 struct buf *bp; 2747 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2748 int error; 2749 int lkflags; 2750 2751 if (size > MAXBSIZE) 2752 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2753 if (vp->v_object == NULL) 2754 panic("getblk: vnode %p has no object!", vp); 2755 2756 /* 2757 * NOTE: findblk does not try to resolve KVABIO in REF-only mode. 2758 * we still have to handle that ourselves. 2759 */ 2760 loop: 2761 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2762 /* 2763 * The buffer was found in the cache, but we need to lock it. 2764 * We must acquire a ref on the bp to prevent reuse, but 2765 * this will not prevent disassociation (brelvp()) so we 2766 * must recheck (vp,loffset) after acquiring the lock. 2767 * 2768 * Without the ref the buffer could potentially be reused 2769 * before we acquire the lock and create a deadlock 2770 * situation between the thread trying to reuse the buffer 2771 * and us due to the fact that we would wind up blocking 2772 * on a random (vp,loffset). 2773 */ 2774 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2775 if (blkflags & GETBLK_NOWAIT) { 2776 bqdrop(bp); 2777 return(NULL); 2778 } 2779 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2780 if (blkflags & GETBLK_PCATCH) 2781 lkflags |= LK_PCATCH; 2782 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2783 if (error) { 2784 bqdrop(bp); 2785 if (error == ENOLCK) 2786 goto loop; 2787 return (NULL); 2788 } 2789 /* buffer may have changed on us */ 2790 } 2791 bqdrop(bp); 2792 2793 /* 2794 * Once the buffer has been locked, make sure we didn't race 2795 * a buffer recyclement. Buffers that are no longer hashed 2796 * will have b_vp == NULL, so this takes care of that check 2797 * as well. 2798 */ 2799 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2800 #if 0 2801 kprintf("Warning buffer %p (vp %p loffset %lld) " 2802 "was recycled\n", 2803 bp, vp, (long long)loffset); 2804 #endif 2805 BUF_UNLOCK(bp); 2806 goto loop; 2807 } 2808 2809 /* 2810 * If SZMATCH any pre-existing buffer must be of the requested 2811 * size or NULL is returned. The caller absolutely does not 2812 * want getblk() to bwrite() the buffer on a size mismatch. 2813 */ 2814 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2815 BUF_UNLOCK(bp); 2816 return(NULL); 2817 } 2818 2819 /* 2820 * All vnode-based buffers must be backed by a VM object. 2821 * 2822 * Set B_KVABIO for any incidental work, we will fix it 2823 * up later. 2824 */ 2825 KKASSERT(bp->b_flags & B_VMIO); 2826 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2827 bp->b_flags &= ~B_AGE; 2828 bp->b_flags |= B_KVABIO; 2829 2830 /* 2831 * Make sure that B_INVAL buffers do not have a cached 2832 * block number translation. 2833 */ 2834 if ((bp->b_flags & B_INVAL) && 2835 (bp->b_bio2.bio_offset != NOOFFSET)) { 2836 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2837 " did not have cleared bio_offset cache\n", 2838 bp, vp, (long long)loffset); 2839 clearbiocache(&bp->b_bio2); 2840 } 2841 2842 /* 2843 * The buffer is locked. B_CACHE is cleared if the buffer is 2844 * invalid. 2845 * 2846 * After the bremfree(), disposals must use b[q]relse(). 2847 */ 2848 if (bp->b_flags & B_INVAL) 2849 bp->b_flags &= ~B_CACHE; 2850 bremfree(bp); 2851 2852 /* 2853 * Any size inconsistancy with a dirty buffer or a buffer 2854 * with a softupdates dependancy must be resolved. Resizing 2855 * the buffer in such circumstances can lead to problems. 2856 * 2857 * Dirty or dependant buffers are written synchronously. 2858 * Other types of buffers are simply released and 2859 * reconstituted as they may be backed by valid, dirty VM 2860 * pages (but not marked B_DELWRI). 2861 * 2862 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2863 * and may be left over from a prior truncation (and thus 2864 * no longer represent the actual EOF point), so we 2865 * definitely do not want to B_NOCACHE the backing store. 2866 */ 2867 if (size != bp->b_bcount) { 2868 if (bp->b_flags & B_DELWRI) { 2869 bp->b_flags |= B_RELBUF; 2870 bwrite(bp); 2871 } else if (LIST_FIRST(&bp->b_dep)) { 2872 bp->b_flags |= B_RELBUF; 2873 bwrite(bp); 2874 } else { 2875 bp->b_flags |= B_RELBUF; 2876 brelse(bp); 2877 } 2878 goto loop; 2879 } 2880 KKASSERT(size <= bp->b_kvasize); 2881 KASSERT(bp->b_loffset != NOOFFSET, 2882 ("getblk: no buffer offset")); 2883 2884 /* 2885 * A buffer with B_DELWRI set and B_CACHE clear must 2886 * be committed before we can return the buffer in 2887 * order to prevent the caller from issuing a read 2888 * ( due to B_CACHE not being set ) and overwriting 2889 * it. 2890 * 2891 * Most callers, including NFS and FFS, need this to 2892 * operate properly either because they assume they 2893 * can issue a read if B_CACHE is not set, or because 2894 * ( for example ) an uncached B_DELWRI might loop due 2895 * to softupdates re-dirtying the buffer. In the latter 2896 * case, B_CACHE is set after the first write completes, 2897 * preventing further loops. 2898 * 2899 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2900 * above while extending the buffer, we cannot allow the 2901 * buffer to remain with B_CACHE set after the write 2902 * completes or it will represent a corrupt state. To 2903 * deal with this we set B_NOCACHE to scrap the buffer 2904 * after the write. 2905 * 2906 * XXX Should this be B_RELBUF instead of B_NOCACHE? 2907 * I'm not even sure this state is still possible 2908 * now that getblk() writes out any dirty buffers 2909 * on size changes. 2910 * 2911 * We might be able to do something fancy, like setting 2912 * B_CACHE in bwrite() except if B_DELWRI is already set, 2913 * so the below call doesn't set B_CACHE, but that gets real 2914 * confusing. This is much easier. 2915 */ 2916 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2917 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 2918 "and CACHE clear, b_flags %08x\n", 2919 bp, (uintmax_t)bp->b_loffset, bp->b_flags); 2920 bp->b_flags |= B_NOCACHE; 2921 bwrite(bp); 2922 goto loop; 2923 } 2924 } else { 2925 /* 2926 * Buffer is not in-core, create new buffer. The buffer 2927 * returned by getnewbuf() is locked. Note that the returned 2928 * buffer is also considered valid (not marked B_INVAL). 2929 * 2930 * Calculating the offset for the I/O requires figuring out 2931 * the block size. We use DEV_BSIZE for VBLK or VCHR and 2932 * the mount's f_iosize otherwise. If the vnode does not 2933 * have an associated mount we assume that the passed size is 2934 * the block size. 2935 * 2936 * Note that vn_isdisk() cannot be used here since it may 2937 * return a failure for numerous reasons. Note that the 2938 * buffer size may be larger then the block size (the caller 2939 * will use block numbers with the proper multiple). Beware 2940 * of using any v_* fields which are part of unions. In 2941 * particular, in DragonFly the mount point overloading 2942 * mechanism uses the namecache only and the underlying 2943 * directory vnode is not a special case. 2944 */ 2945 int bsize, maxsize; 2946 2947 if (vp->v_type == VBLK || vp->v_type == VCHR) 2948 bsize = DEV_BSIZE; 2949 else if (vp->v_mount) 2950 bsize = vp->v_mount->mnt_stat.f_iosize; 2951 else 2952 bsize = size; 2953 2954 maxsize = size + (loffset & PAGE_MASK); 2955 maxsize = imax(maxsize, bsize); 2956 2957 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 2958 if (bp == NULL) { 2959 if (slpflags || slptimeo) 2960 return NULL; 2961 goto loop; 2962 } 2963 2964 /* 2965 * Atomically insert the buffer into the hash, so that it can 2966 * be found by findblk(). 2967 * 2968 * If bgetvp() returns non-zero a collision occured, and the 2969 * bp will not be associated with the vnode. 2970 * 2971 * Make sure the translation layer has been cleared. 2972 */ 2973 bp->b_loffset = loffset; 2974 bp->b_bio2.bio_offset = NOOFFSET; 2975 /* bp->b_bio2.bio_next = NULL; */ 2976 2977 if (bgetvp(vp, bp, size)) { 2978 bp->b_flags |= B_INVAL; 2979 brelse(bp); 2980 goto loop; 2981 } 2982 2983 /* 2984 * All vnode-based buffers must be backed by a VM object. 2985 * 2986 * Set B_KVABIO for incidental work 2987 */ 2988 KKASSERT(vp->v_object != NULL); 2989 bp->b_flags |= B_VMIO | B_KVABIO; 2990 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2991 2992 allocbuf(bp, size); 2993 } 2994 2995 /* 2996 * Do the nasty smp broadcast (if the buffer needs it) when KVABIO 2997 * is not supported. 2998 */ 2999 if (bp && (blkflags & GETBLK_KVABIO) == 0) { 3000 bkvasync_all(bp); 3001 } 3002 return (bp); 3003 } 3004 3005 /* 3006 * regetblk(bp) 3007 * 3008 * Reacquire a buffer that was previously released to the locked queue, 3009 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3010 * set B_LOCKED (which handles the acquisition race). 3011 * 3012 * To this end, either B_LOCKED must be set or the dependancy list must be 3013 * non-empty. 3014 */ 3015 void 3016 regetblk(struct buf *bp) 3017 { 3018 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3019 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3020 bremfree(bp); 3021 } 3022 3023 /* 3024 * allocbuf: 3025 * 3026 * This code constitutes the buffer memory from either anonymous system 3027 * memory (in the case of non-VMIO operations) or from an associated 3028 * VM object (in the case of VMIO operations). This code is able to 3029 * resize a buffer up or down. 3030 * 3031 * Note that this code is tricky, and has many complications to resolve 3032 * deadlock or inconsistant data situations. Tread lightly!!! 3033 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3034 * the caller. Calling this code willy nilly can result in the loss of 3035 * data. 3036 * 3037 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3038 * B_CACHE for the non-VMIO case. 3039 * 3040 * This routine does not need to be called from a critical section but you 3041 * must own the buffer. 3042 */ 3043 void 3044 allocbuf(struct buf *bp, int size) 3045 { 3046 vm_page_t m; 3047 int newbsize; 3048 int desiredpages; 3049 int i; 3050 3051 if (BUF_LOCKINUSE(bp) == 0) 3052 panic("allocbuf: buffer not busy"); 3053 3054 if (bp->b_kvasize < size) 3055 panic("allocbuf: buffer too small"); 3056 3057 KKASSERT(bp->b_flags & B_VMIO); 3058 3059 newbsize = roundup2(size, DEV_BSIZE); 3060 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3061 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3062 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3063 3064 /* 3065 * Set B_CACHE initially if buffer is 0 length or will become 3066 * 0-length. 3067 */ 3068 if (size == 0 || bp->b_bufsize == 0) 3069 bp->b_flags |= B_CACHE; 3070 3071 if (newbsize < bp->b_bufsize) { 3072 /* 3073 * DEV_BSIZE aligned new buffer size is less then the 3074 * DEV_BSIZE aligned existing buffer size. Figure out 3075 * if we have to remove any pages. 3076 */ 3077 if (desiredpages < bp->b_xio.xio_npages) { 3078 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3079 /* 3080 * the page is not freed here -- it 3081 * is the responsibility of 3082 * vnode_pager_setsize 3083 */ 3084 m = bp->b_xio.xio_pages[i]; 3085 KASSERT(m != bogus_page, 3086 ("allocbuf: bogus page found")); 3087 vm_page_busy_wait(m, TRUE, "biodep"); 3088 bp->b_xio.xio_pages[i] = NULL; 3089 vm_page_unwire(m, 0); 3090 vm_page_wakeup(m); 3091 } 3092 pmap_qremove_noinval((vm_offset_t) 3093 trunc_page((vm_offset_t)bp->b_data) + 3094 (desiredpages << PAGE_SHIFT), 3095 (bp->b_xio.xio_npages - desiredpages)); 3096 bp->b_xio.xio_npages = desiredpages; 3097 3098 /* 3099 * Don't bother invalidating the pmap changes 3100 * (which wastes global SMP invalidation IPIs) 3101 * when setting the size to 0. This case occurs 3102 * when called via getnewbuf() during buffer 3103 * recyclement. 3104 */ 3105 if (desiredpages == 0) { 3106 CPUMASK_ASSZERO(bp->b_cpumask); 3107 } else { 3108 bkvareset(bp); 3109 } 3110 } 3111 } else if (size > bp->b_bcount) { 3112 /* 3113 * We are growing the buffer, possibly in a 3114 * byte-granular fashion. 3115 */ 3116 struct vnode *vp; 3117 vm_object_t obj; 3118 vm_offset_t toff; 3119 vm_offset_t tinc; 3120 3121 /* 3122 * Step 1, bring in the VM pages from the object, 3123 * allocating them if necessary. We must clear 3124 * B_CACHE if these pages are not valid for the 3125 * range covered by the buffer. 3126 */ 3127 vp = bp->b_vp; 3128 obj = vp->v_object; 3129 3130 vm_object_hold(obj); 3131 while (bp->b_xio.xio_npages < desiredpages) { 3132 vm_page_t m; 3133 vm_pindex_t pi; 3134 int error; 3135 3136 pi = OFF_TO_IDX(bp->b_loffset) + 3137 bp->b_xio.xio_npages; 3138 3139 /* 3140 * Blocking on m->busy_count might lead to a 3141 * deadlock: 3142 * 3143 * vm_fault->getpages->cluster_read->allocbuf 3144 */ 3145 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3146 &error); 3147 if (error) { 3148 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3149 continue; 3150 } 3151 if (m == NULL) { 3152 /* 3153 * note: must allocate system pages 3154 * since blocking here could intefere 3155 * with paging I/O, no matter which 3156 * process we are. 3157 */ 3158 m = bio_page_alloc(bp, obj, pi, 3159 desiredpages - 3160 bp->b_xio.xio_npages); 3161 if (m) { 3162 vm_page_wire(m); 3163 vm_page_wakeup(m); 3164 bp->b_flags &= ~B_CACHE; 3165 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3166 ++bp->b_xio.xio_npages; 3167 } 3168 continue; 3169 } 3170 3171 /* 3172 * We found a page and were able to busy it. 3173 */ 3174 vm_page_wire(m); 3175 vm_page_wakeup(m); 3176 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3177 ++bp->b_xio.xio_npages; 3178 if (bp->b_act_count < m->act_count) 3179 bp->b_act_count = m->act_count; 3180 } 3181 vm_object_drop(obj); 3182 3183 /* 3184 * Step 2. We've loaded the pages into the buffer, 3185 * we have to figure out if we can still have B_CACHE 3186 * set. Note that B_CACHE is set according to the 3187 * byte-granular range ( bcount and size ), not the 3188 * aligned range ( newbsize ). 3189 * 3190 * The VM test is against m->valid, which is DEV_BSIZE 3191 * aligned. Needless to say, the validity of the data 3192 * needs to also be DEV_BSIZE aligned. Note that this 3193 * fails with NFS if the server or some other client 3194 * extends the file's EOF. If our buffer is resized, 3195 * B_CACHE may remain set! XXX 3196 */ 3197 3198 toff = bp->b_bcount; 3199 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3200 3201 while ((bp->b_flags & B_CACHE) && toff < size) { 3202 vm_pindex_t pi; 3203 3204 if (tinc > (size - toff)) 3205 tinc = size - toff; 3206 3207 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3208 PAGE_SHIFT; 3209 3210 vfs_buf_test_cache( 3211 bp, 3212 bp->b_loffset, 3213 toff, 3214 tinc, 3215 bp->b_xio.xio_pages[pi] 3216 ); 3217 toff += tinc; 3218 tinc = PAGE_SIZE; 3219 } 3220 3221 /* 3222 * Step 3, fixup the KVM pmap. Remember that 3223 * bp->b_data is relative to bp->b_loffset, but 3224 * bp->b_loffset may be offset into the first page. 3225 */ 3226 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 3227 pmap_qenter_noinval((vm_offset_t)bp->b_data, 3228 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3229 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3230 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3231 bkvareset(bp); 3232 } 3233 atomic_add_long(&bufspace, newbsize - bp->b_bufsize); 3234 3235 /* adjust space use on already-dirty buffer */ 3236 if (bp->b_flags & B_DELWRI) { 3237 /* dirtykvaspace unchanged */ 3238 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize); 3239 if (bp->b_flags & B_HEAVY) { 3240 atomic_add_long(&dirtybufspacehw, 3241 newbsize - bp->b_bufsize); 3242 } 3243 } 3244 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3245 bp->b_bcount = size; /* requested buffer size */ 3246 bufspacewakeup(); 3247 } 3248 3249 /* 3250 * biowait: 3251 * 3252 * Wait for buffer I/O completion, returning error status. B_EINTR 3253 * is converted into an EINTR error but not cleared (since a chain 3254 * of biowait() calls may occur). 3255 * 3256 * On return bpdone() will have been called but the buffer will remain 3257 * locked and will not have been brelse()'d. 3258 * 3259 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3260 * likely still in progress on return. 3261 * 3262 * NOTE! This operation is on a BIO, not a BUF. 3263 * 3264 * NOTE! BIO_DONE is cleared by vn_strategy() 3265 */ 3266 static __inline int 3267 _biowait(struct bio *bio, const char *wmesg, int to) 3268 { 3269 struct buf *bp = bio->bio_buf; 3270 u_int32_t flags; 3271 u_int32_t nflags; 3272 int error; 3273 3274 KKASSERT(bio == &bp->b_bio1); 3275 for (;;) { 3276 flags = bio->bio_flags; 3277 if (flags & BIO_DONE) 3278 break; 3279 nflags = flags | BIO_WANT; 3280 tsleep_interlock(bio, 0); 3281 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3282 if (wmesg) 3283 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3284 else if (bp->b_cmd == BUF_CMD_READ) 3285 error = tsleep(bio, PINTERLOCKED, "biord", to); 3286 else 3287 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3288 if (error) { 3289 kprintf("tsleep error biowait %d\n", error); 3290 return (error); 3291 } 3292 } 3293 } 3294 3295 /* 3296 * Finish up. 3297 */ 3298 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3299 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3300 if (bp->b_flags & B_EINTR) 3301 return (EINTR); 3302 if (bp->b_flags & B_ERROR) 3303 return (bp->b_error ? bp->b_error : EIO); 3304 return (0); 3305 } 3306 3307 int 3308 biowait(struct bio *bio, const char *wmesg) 3309 { 3310 return(_biowait(bio, wmesg, 0)); 3311 } 3312 3313 int 3314 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3315 { 3316 return(_biowait(bio, wmesg, to)); 3317 } 3318 3319 /* 3320 * This associates a tracking count with an I/O. vn_strategy() and 3321 * dev_dstrategy() do this automatically but there are a few cases 3322 * where a vnode or device layer is bypassed when a block translation 3323 * is cached. In such cases bio_start_transaction() may be called on 3324 * the bypassed layers so the system gets an I/O in progress indication 3325 * for those higher layers. 3326 */ 3327 void 3328 bio_start_transaction(struct bio *bio, struct bio_track *track) 3329 { 3330 bio->bio_track = track; 3331 bio_track_ref(track); 3332 dsched_buf_enter(bio->bio_buf); /* might stack */ 3333 } 3334 3335 /* 3336 * Initiate I/O on a vnode. 3337 * 3338 * SWAPCACHE OPERATION: 3339 * 3340 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3341 * devfs also uses b_vp for fake buffers so we also have to check 3342 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3343 * underlying block device. The swap assignments are related to the 3344 * buffer cache buffer's b_vp, not the passed vp. 3345 * 3346 * The passed vp == bp->b_vp only in the case where the strategy call 3347 * is made on the vp itself for its own buffers (a regular file or 3348 * block device vp). The filesystem usually then re-calls vn_strategy() 3349 * after translating the request to an underlying device. 3350 * 3351 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3352 * underlying buffer cache buffers. 3353 * 3354 * We can only deal with page-aligned buffers at the moment, because 3355 * we can't tell what the real dirty state for pages straddling a buffer 3356 * are. 3357 * 3358 * In order to call swap_pager_strategy() we must provide the VM object 3359 * and base offset for the underlying buffer cache pages so it can find 3360 * the swap blocks. 3361 */ 3362 void 3363 vn_strategy(struct vnode *vp, struct bio *bio) 3364 { 3365 struct bio_track *track; 3366 struct buf *bp = bio->bio_buf; 3367 3368 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3369 3370 /* 3371 * Set when an I/O is issued on the bp. Cleared by consumers 3372 * (aka HAMMER), allowing the consumer to determine if I/O had 3373 * actually occurred. 3374 */ 3375 bp->b_flags |= B_IOISSUED; 3376 3377 /* 3378 * Handle the swapcache intercept. 3379 * 3380 * NOTE: The swapcache itself always supports KVABIO and will 3381 * do the right thing if its underlying devices do not. 3382 */ 3383 if (vn_cache_strategy(vp, bio)) 3384 return; 3385 3386 /* 3387 * If the vnode does not support KVABIO and the buffer is using 3388 * KVABIO, we must synchronize b_data to all cpus before dispatching. 3389 */ 3390 if ((vp->v_flag & VKVABIO) == 0 && (bp->b_flags & B_KVABIO)) 3391 bkvasync_all(bp); 3392 3393 /* 3394 * Otherwise do the operation through the filesystem 3395 */ 3396 if (bp->b_cmd == BUF_CMD_READ) 3397 track = &vp->v_track_read; 3398 else 3399 track = &vp->v_track_write; 3400 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3401 bio->bio_track = track; 3402 bio_track_ref(track); 3403 dsched_buf_enter(bp); /* might stack */ 3404 vop_strategy(*vp->v_ops, vp, bio); 3405 } 3406 3407 /* 3408 * vn_cache_strategy() 3409 * 3410 * NOTE: This function supports the KVABIO API wherein b_data might not 3411 * be synchronized to the current cpu. 3412 */ 3413 static void vn_cache_strategy_callback(struct bio *bio); 3414 3415 int 3416 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3417 { 3418 struct buf *bp = bio->bio_buf; 3419 struct bio *nbio; 3420 vm_object_t object; 3421 vm_page_t m; 3422 int i; 3423 3424 /* 3425 * Stop using swapcache if paniced, dumping, or dumped 3426 */ 3427 if (panicstr || dumping) 3428 return(0); 3429 3430 /* 3431 * Is this buffer cache buffer suitable for reading from 3432 * the swap cache? 3433 */ 3434 if (vm_swapcache_read_enable == 0 || 3435 bp->b_cmd != BUF_CMD_READ || 3436 ((bp->b_flags & B_CLUSTER) == 0 && 3437 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3438 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3439 (bp->b_bcount & PAGE_MASK) != 0) { 3440 return(0); 3441 } 3442 3443 /* 3444 * Figure out the original VM object (it will match the underlying 3445 * VM pages). Note that swap cached data uses page indices relative 3446 * to that object, not relative to bio->bio_offset. 3447 */ 3448 if (bp->b_flags & B_CLUSTER) 3449 object = vp->v_object; 3450 else 3451 object = bp->b_vp->v_object; 3452 3453 /* 3454 * In order to be able to use the swap cache all underlying VM 3455 * pages must be marked as such, and we can't have any bogus pages. 3456 */ 3457 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3458 m = bp->b_xio.xio_pages[i]; 3459 if ((m->flags & PG_SWAPPED) == 0) 3460 break; 3461 if (m == bogus_page) 3462 break; 3463 } 3464 3465 /* 3466 * If we are good then issue the I/O using swap_pager_strategy(). 3467 * 3468 * We can only do this if the buffer actually supports object-backed 3469 * I/O. If it doesn't npages will be 0. 3470 */ 3471 if (i && i == bp->b_xio.xio_npages) { 3472 m = bp->b_xio.xio_pages[0]; 3473 nbio = push_bio(bio); 3474 nbio->bio_done = vn_cache_strategy_callback; 3475 nbio->bio_offset = ptoa(m->pindex); 3476 KKASSERT(m->object == object); 3477 swap_pager_strategy(object, nbio); 3478 return(1); 3479 } 3480 return(0); 3481 } 3482 3483 /* 3484 * This is a bit of a hack but since the vn_cache_strategy() function can 3485 * override a VFS's strategy function we must make sure that the bio, which 3486 * is probably bio2, doesn't leak an unexpected offset value back to the 3487 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3488 * bio went through its own file strategy function and the the bio2 offset 3489 * is a cached disk offset when, in fact, it isn't. 3490 */ 3491 static void 3492 vn_cache_strategy_callback(struct bio *bio) 3493 { 3494 bio->bio_offset = NOOFFSET; 3495 biodone(pop_bio(bio)); 3496 } 3497 3498 /* 3499 * bpdone: 3500 * 3501 * Finish I/O on a buffer after all BIOs have been processed. 3502 * Called when the bio chain is exhausted or by biowait. If called 3503 * by biowait, elseit is typically 0. 3504 * 3505 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3506 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3507 * assuming B_INVAL is clear. 3508 * 3509 * For the VMIO case, we set B_CACHE if the op was a read and no 3510 * read error occured, or if the op was a write. B_CACHE is never 3511 * set if the buffer is invalid or otherwise uncacheable. 3512 * 3513 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3514 * initiator to leave B_INVAL set to brelse the buffer out of existance 3515 * in the biodone routine. 3516 * 3517 * bpdone is responsible for calling bundirty() on the buffer after a 3518 * successful write. We previously did this prior to initiating the 3519 * write under the assumption that the buffer might be dirtied again 3520 * while the write was in progress, however doing it before-hand creates 3521 * a race condition prior to the call to vn_strategy() where the 3522 * filesystem may not be aware that a dirty buffer is present. 3523 * It should not be possible for the buffer or its underlying pages to 3524 * be redirtied prior to bpdone()'s unbusying of the underlying VM 3525 * pages. 3526 */ 3527 void 3528 bpdone(struct buf *bp, int elseit) 3529 { 3530 buf_cmd_t cmd; 3531 3532 KASSERT(BUF_LOCKINUSE(bp), ("bpdone: bp %p not busy", bp)); 3533 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3534 ("bpdone: bp %p already done!", bp)); 3535 3536 /* 3537 * No more BIOs are left. All completion functions have been dealt 3538 * with, now we clean up the buffer. 3539 */ 3540 cmd = bp->b_cmd; 3541 bp->b_cmd = BUF_CMD_DONE; 3542 3543 /* 3544 * Only reads and writes are processed past this point. 3545 */ 3546 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3547 if (cmd == BUF_CMD_FREEBLKS) 3548 bp->b_flags |= B_NOCACHE; 3549 if (elseit) 3550 brelse(bp); 3551 return; 3552 } 3553 3554 /* 3555 * A failed write must re-dirty the buffer unless B_INVAL 3556 * was set. 3557 * 3558 * A successful write must clear the dirty flag. This is done after 3559 * the write to ensure that the buffer remains on the vnode's dirty 3560 * list for filesystem interlocks / checks until the write is actually 3561 * complete. HAMMER2 is sensitive to this issue. 3562 * 3563 * Only applicable to normal buffers (with VPs). vinum buffers may 3564 * not have a vp. 3565 * 3566 * Must be done prior to calling buf_complete() as the callback might 3567 * re-dirty the buffer. 3568 */ 3569 if (cmd == BUF_CMD_WRITE) { 3570 if ((bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3571 bp->b_flags &= ~B_NOCACHE; 3572 if (bp->b_vp) 3573 bdirty(bp); 3574 } else { 3575 if (bp->b_vp) 3576 bundirty(bp); 3577 } 3578 } 3579 3580 /* 3581 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3582 * a lot worse. XXX - move this above the clearing of b_cmd 3583 */ 3584 if (LIST_FIRST(&bp->b_dep) != NULL) 3585 buf_complete(bp); 3586 3587 if (bp->b_flags & B_VMIO) { 3588 int i; 3589 vm_ooffset_t foff; 3590 vm_page_t m; 3591 vm_object_t obj; 3592 int iosize; 3593 struct vnode *vp = bp->b_vp; 3594 3595 obj = vp->v_object; 3596 3597 #if defined(VFS_BIO_DEBUG) 3598 if (vp->v_auxrefs == 0) 3599 panic("bpdone: zero vnode hold count"); 3600 if ((vp->v_flag & VOBJBUF) == 0) 3601 panic("bpdone: vnode is not setup for merged cache"); 3602 #endif 3603 3604 foff = bp->b_loffset; 3605 KASSERT(foff != NOOFFSET, ("bpdone: no buffer offset")); 3606 KASSERT(obj != NULL, ("bpdone: missing VM object")); 3607 3608 #if defined(VFS_BIO_DEBUG) 3609 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3610 kprintf("bpdone: paging in progress(%d) < " 3611 "bp->b_xio.xio_npages(%d)\n", 3612 obj->paging_in_progress, 3613 bp->b_xio.xio_npages); 3614 } 3615 #endif 3616 3617 /* 3618 * Set B_CACHE if the op was a normal read and no error 3619 * occured. B_CACHE is set for writes in the b*write() 3620 * routines. 3621 */ 3622 iosize = bp->b_bcount - bp->b_resid; 3623 if (cmd == BUF_CMD_READ && 3624 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3625 bp->b_flags |= B_CACHE; 3626 } 3627 3628 vm_object_hold(obj); 3629 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3630 int resid; 3631 int isbogus; 3632 3633 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3634 if (resid > iosize) 3635 resid = iosize; 3636 3637 /* 3638 * cleanup bogus pages, restoring the originals. Since 3639 * the originals should still be wired, we don't have 3640 * to worry about interrupt/freeing races destroying 3641 * the VM object association. 3642 */ 3643 m = bp->b_xio.xio_pages[i]; 3644 if (m == bogus_page) { 3645 if ((bp->b_flags & B_HASBOGUS) == 0) 3646 panic("bpdone: bp %p corrupt bogus", bp); 3647 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3648 if (m == NULL) 3649 panic("bpdone: page disappeared"); 3650 bp->b_xio.xio_pages[i] = m; 3651 isbogus = 1; 3652 } else { 3653 isbogus = 0; 3654 } 3655 #if defined(VFS_BIO_DEBUG) 3656 if (OFF_TO_IDX(foff) != m->pindex) { 3657 kprintf("bpdone: foff(%lu)/m->pindex(%ld) " 3658 "mismatch\n", 3659 (unsigned long)foff, (long)m->pindex); 3660 } 3661 #endif 3662 3663 /* 3664 * In the write case, the valid and clean bits are 3665 * already changed correctly (see bdwrite()), so we 3666 * only need to do this here in the read case. 3667 */ 3668 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3669 if (cmd == BUF_CMD_READ && isbogus == 0 && resid > 0) 3670 vfs_clean_one_page(bp, i, m); 3671 3672 /* 3673 * when debugging new filesystems or buffer I/O 3674 * methods, this is the most common error that pops 3675 * up. if you see this, you have not set the page 3676 * busy flag correctly!!! 3677 */ 3678 if ((m->busy_count & PBUSY_MASK) == 0) { 3679 kprintf("bpdone: page busy < 0, " 3680 "pindex: %d, foff: 0x(%x,%x), " 3681 "resid: %d, index: %d\n", 3682 (int) m->pindex, (int)(foff >> 32), 3683 (int) foff & 0xffffffff, resid, i); 3684 if (!vn_isdisk(vp, NULL)) 3685 kprintf(" iosize: %ld, loffset: %lld, " 3686 "flags: 0x%08x, npages: %d\n", 3687 bp->b_vp->v_mount->mnt_stat.f_iosize, 3688 (long long)bp->b_loffset, 3689 bp->b_flags, bp->b_xio.xio_npages); 3690 else 3691 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3692 (long long)bp->b_loffset, 3693 bp->b_flags, bp->b_xio.xio_npages); 3694 kprintf(" valid: 0x%x, dirty: 0x%x, " 3695 "wired: %d\n", 3696 m->valid, m->dirty, 3697 m->wire_count); 3698 panic("bpdone: page busy < 0"); 3699 } 3700 vm_page_io_finish(m); 3701 vm_page_wakeup(m); 3702 vm_object_pip_wakeup(obj); 3703 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3704 iosize -= resid; 3705 } 3706 if (bp->b_flags & B_HASBOGUS) { 3707 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data), 3708 bp->b_xio.xio_pages, 3709 bp->b_xio.xio_npages); 3710 bp->b_flags &= ~B_HASBOGUS; 3711 bkvareset(bp); 3712 } 3713 vm_object_drop(obj); 3714 } 3715 3716 /* 3717 * Finish up by releasing the buffer. There are no more synchronous 3718 * or asynchronous completions, those were handled by bio_done 3719 * callbacks. 3720 */ 3721 if (elseit) { 3722 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3723 brelse(bp); 3724 else 3725 bqrelse(bp); 3726 } 3727 } 3728 3729 /* 3730 * Normal biodone. 3731 */ 3732 void 3733 biodone(struct bio *bio) 3734 { 3735 struct buf *bp = bio->bio_buf; 3736 3737 runningbufwakeup(bp); 3738 3739 /* 3740 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3741 */ 3742 while (bio) { 3743 biodone_t *done_func; 3744 struct bio_track *track; 3745 3746 /* 3747 * BIO tracking. Most but not all BIOs are tracked. 3748 */ 3749 if ((track = bio->bio_track) != NULL) { 3750 bio_track_rel(track); 3751 bio->bio_track = NULL; 3752 } 3753 3754 /* 3755 * A bio_done function terminates the loop. The function 3756 * will be responsible for any further chaining and/or 3757 * buffer management. 3758 * 3759 * WARNING! The done function can deallocate the buffer! 3760 */ 3761 if ((done_func = bio->bio_done) != NULL) { 3762 bio->bio_done = NULL; 3763 done_func(bio); 3764 return; 3765 } 3766 bio = bio->bio_prev; 3767 } 3768 3769 /* 3770 * If we've run out of bio's do normal [a]synchronous completion. 3771 */ 3772 bpdone(bp, 1); 3773 } 3774 3775 /* 3776 * Synchronous biodone - this terminates a synchronous BIO. 3777 * 3778 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3779 * but still locked. The caller must brelse() the buffer after waiting 3780 * for completion. 3781 */ 3782 void 3783 biodone_sync(struct bio *bio) 3784 { 3785 struct buf *bp = bio->bio_buf; 3786 int flags; 3787 int nflags; 3788 3789 KKASSERT(bio == &bp->b_bio1); 3790 bpdone(bp, 0); 3791 3792 for (;;) { 3793 flags = bio->bio_flags; 3794 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3795 3796 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3797 if (flags & BIO_WANT) 3798 wakeup(bio); 3799 break; 3800 } 3801 } 3802 } 3803 3804 /* 3805 * vfs_unbusy_pages: 3806 * 3807 * This routine is called in lieu of iodone in the case of 3808 * incomplete I/O. This keeps the busy status for pages 3809 * consistant. 3810 */ 3811 void 3812 vfs_unbusy_pages(struct buf *bp) 3813 { 3814 int i; 3815 3816 runningbufwakeup(bp); 3817 3818 if (bp->b_flags & B_VMIO) { 3819 struct vnode *vp = bp->b_vp; 3820 vm_object_t obj; 3821 3822 obj = vp->v_object; 3823 vm_object_hold(obj); 3824 3825 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3826 vm_page_t m = bp->b_xio.xio_pages[i]; 3827 3828 /* 3829 * When restoring bogus changes the original pages 3830 * should still be wired, so we are in no danger of 3831 * losing the object association and do not need 3832 * critical section protection particularly. 3833 */ 3834 if (m == bogus_page) { 3835 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 3836 if (!m) { 3837 panic("vfs_unbusy_pages: page missing"); 3838 } 3839 bp->b_xio.xio_pages[i] = m; 3840 } 3841 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3842 vm_page_io_finish(m); 3843 vm_page_wakeup(m); 3844 vm_object_pip_wakeup(obj); 3845 } 3846 if (bp->b_flags & B_HASBOGUS) { 3847 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data), 3848 bp->b_xio.xio_pages, 3849 bp->b_xio.xio_npages); 3850 bp->b_flags &= ~B_HASBOGUS; 3851 bkvareset(bp); 3852 } 3853 vm_object_drop(obj); 3854 } 3855 } 3856 3857 /* 3858 * vfs_busy_pages: 3859 * 3860 * This routine is called before a device strategy routine. 3861 * It is used to tell the VM system that paging I/O is in 3862 * progress, and treat the pages associated with the buffer 3863 * almost as being PBUSY_LOCKED. Also the object 'paging_in_progress' 3864 * flag is handled to make sure that the object doesn't become 3865 * inconsistant. 3866 * 3867 * Since I/O has not been initiated yet, certain buffer flags 3868 * such as B_ERROR or B_INVAL may be in an inconsistant state 3869 * and should be ignored. 3870 */ 3871 void 3872 vfs_busy_pages(struct vnode *vp, struct buf *bp) 3873 { 3874 int i, bogus; 3875 struct lwp *lp = curthread->td_lwp; 3876 3877 /* 3878 * The buffer's I/O command must already be set. If reading, 3879 * B_CACHE must be 0 (double check against callers only doing 3880 * I/O when B_CACHE is 0). 3881 */ 3882 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3883 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 3884 3885 if (bp->b_flags & B_VMIO) { 3886 vm_object_t obj; 3887 3888 obj = vp->v_object; 3889 KASSERT(bp->b_loffset != NOOFFSET, 3890 ("vfs_busy_pages: no buffer offset")); 3891 3892 /* 3893 * Busy all the pages. We have to busy them all at once 3894 * to avoid deadlocks. 3895 */ 3896 retry: 3897 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3898 vm_page_t m = bp->b_xio.xio_pages[i]; 3899 3900 if (vm_page_busy_try(m, FALSE)) { 3901 vm_page_sleep_busy(m, FALSE, "vbpage"); 3902 while (--i >= 0) 3903 vm_page_wakeup(bp->b_xio.xio_pages[i]); 3904 goto retry; 3905 } 3906 } 3907 3908 /* 3909 * Setup for I/O, soft-busy the page right now because 3910 * the next loop may block. 3911 */ 3912 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3913 vm_page_t m = bp->b_xio.xio_pages[i]; 3914 3915 if ((bp->b_flags & B_CLUSTER) == 0) { 3916 vm_object_pip_add(obj, 1); 3917 vm_page_io_start(m); 3918 } 3919 } 3920 3921 /* 3922 * Adjust protections for I/O and do bogus-page mapping. 3923 * Assume that vm_page_protect() can block (it can block 3924 * if VM_PROT_NONE, don't take any chances regardless). 3925 * 3926 * In particular note that for writes we must incorporate 3927 * page dirtyness from the VM system into the buffer's 3928 * dirty range. 3929 * 3930 * For reads we theoretically must incorporate page dirtyness 3931 * from the VM system to determine if the page needs bogus 3932 * replacement, but we shortcut the test by simply checking 3933 * that all m->valid bits are set, indicating that the page 3934 * is fully valid and does not need to be re-read. For any 3935 * VM system dirtyness the page will also be fully valid 3936 * since it was mapped at one point. 3937 */ 3938 bogus = 0; 3939 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3940 vm_page_t m = bp->b_xio.xio_pages[i]; 3941 3942 if (bp->b_cmd == BUF_CMD_WRITE) { 3943 /* 3944 * When readying a vnode-backed buffer for 3945 * a write we must zero-fill any invalid 3946 * portions of the backing VM pages, mark 3947 * it valid and clear related dirty bits. 3948 * 3949 * vfs_clean_one_page() incorporates any 3950 * VM dirtyness and updates the b_dirtyoff 3951 * range (after we've made the page RO). 3952 * 3953 * It is also expected that the pmap modified 3954 * bit has already been cleared by the 3955 * vm_page_protect(). We may not be able 3956 * to clear all dirty bits for a page if it 3957 * was also memory mapped (NFS). 3958 * 3959 * Finally be sure to unassign any swap-cache 3960 * backing store as it is now stale. 3961 */ 3962 vm_page_protect(m, VM_PROT_READ); 3963 vfs_clean_one_page(bp, i, m); 3964 swap_pager_unswapped(m); 3965 } else if (m->valid == VM_PAGE_BITS_ALL) { 3966 /* 3967 * When readying a vnode-backed buffer for 3968 * read we must replace any dirty pages with 3969 * a bogus page so dirty data is not destroyed 3970 * when filling gaps. 3971 * 3972 * To avoid testing whether the page is 3973 * dirty we instead test that the page was 3974 * at some point mapped (m->valid fully 3975 * valid) with the understanding that 3976 * this also covers the dirty case. 3977 */ 3978 bp->b_xio.xio_pages[i] = bogus_page; 3979 bp->b_flags |= B_HASBOGUS; 3980 bogus++; 3981 } else if (m->valid & m->dirty) { 3982 /* 3983 * This case should not occur as partial 3984 * dirtyment can only happen if the buffer 3985 * is B_CACHE, and this code is not entered 3986 * if the buffer is B_CACHE. 3987 */ 3988 kprintf("Warning: vfs_busy_pages - page not " 3989 "fully valid! loff=%jx bpf=%08x " 3990 "idx=%d val=%02x dir=%02x\n", 3991 (uintmax_t)bp->b_loffset, bp->b_flags, 3992 i, m->valid, m->dirty); 3993 vm_page_protect(m, VM_PROT_NONE); 3994 } else { 3995 /* 3996 * The page is not valid and can be made 3997 * part of the read. 3998 */ 3999 vm_page_protect(m, VM_PROT_NONE); 4000 } 4001 vm_page_wakeup(m); 4002 } 4003 if (bogus) { 4004 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data), 4005 bp->b_xio.xio_pages, 4006 bp->b_xio.xio_npages); 4007 bkvareset(bp); 4008 } 4009 } 4010 4011 /* 4012 * This is the easiest place to put the process accounting for the I/O 4013 * for now. 4014 */ 4015 if (lp != NULL) { 4016 if (bp->b_cmd == BUF_CMD_READ) 4017 lp->lwp_ru.ru_inblock++; 4018 else 4019 lp->lwp_ru.ru_oublock++; 4020 } 4021 } 4022 4023 /* 4024 * Tell the VM system that the pages associated with this buffer 4025 * are clean. This is used for delayed writes where the data is 4026 * going to go to disk eventually without additional VM intevention. 4027 * 4028 * NOTE: While we only really need to clean through to b_bcount, we 4029 * just go ahead and clean through to b_bufsize. 4030 */ 4031 static void 4032 vfs_clean_pages(struct buf *bp) 4033 { 4034 vm_page_t m; 4035 int i; 4036 4037 if ((bp->b_flags & B_VMIO) == 0) 4038 return; 4039 4040 KASSERT(bp->b_loffset != NOOFFSET, 4041 ("vfs_clean_pages: no buffer offset")); 4042 4043 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4044 m = bp->b_xio.xio_pages[i]; 4045 vfs_clean_one_page(bp, i, m); 4046 } 4047 } 4048 4049 /* 4050 * vfs_clean_one_page: 4051 * 4052 * Set the valid bits and clear the dirty bits in a page within a 4053 * buffer. The range is restricted to the buffer's size and the 4054 * buffer's logical offset might index into the first page. 4055 * 4056 * The caller has busied or soft-busied the page and it is not mapped, 4057 * test and incorporate the dirty bits into b_dirtyoff/end before 4058 * clearing them. Note that we need to clear the pmap modified bits 4059 * after determining the the page was dirty, vm_page_set_validclean() 4060 * does not do it for us. 4061 * 4062 * This routine is typically called after a read completes (dirty should 4063 * be zero in that case as we are not called on bogus-replace pages), 4064 * or before a write is initiated. 4065 */ 4066 static void 4067 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4068 { 4069 int bcount; 4070 int xoff; 4071 int soff; 4072 int eoff; 4073 4074 /* 4075 * Calculate offset range within the page but relative to buffer's 4076 * loffset. loffset might be offset into the first page. 4077 */ 4078 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4079 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4080 4081 if (pageno == 0) { 4082 soff = xoff; 4083 eoff = PAGE_SIZE; 4084 } else { 4085 soff = (pageno << PAGE_SHIFT); 4086 eoff = soff + PAGE_SIZE; 4087 } 4088 if (eoff > bcount) 4089 eoff = bcount; 4090 if (soff >= eoff) 4091 return; 4092 4093 /* 4094 * Test dirty bits and adjust b_dirtyoff/end. 4095 * 4096 * If dirty pages are incorporated into the bp any prior 4097 * B_NEEDCOMMIT state (NFS) must be cleared because the 4098 * caller has not taken into account the new dirty data. 4099 * 4100 * If the page was memory mapped the dirty bits might go beyond the 4101 * end of the buffer, but we can't really make the assumption that 4102 * a file EOF straddles the buffer (even though this is the case for 4103 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4104 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4105 * This also saves some console spam. 4106 * 4107 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4108 * NFS can handle huge commits but not huge writes. 4109 */ 4110 vm_page_test_dirty(m); 4111 if (m->dirty) { 4112 if ((bp->b_flags & B_NEEDCOMMIT) && 4113 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4114 if (debug_commit) 4115 kprintf("Warning: vfs_clean_one_page: bp %p " 4116 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4117 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4118 "doff/end %d %d\n", 4119 bp, (uintmax_t)bp->b_loffset, bp->b_bcount, 4120 bp->b_flags, bp->b_cmd, 4121 m->valid, m->dirty, xoff, soff, eoff, 4122 bp->b_dirtyoff, bp->b_dirtyend); 4123 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4124 if (debug_commit) 4125 print_backtrace(-1); 4126 } 4127 /* 4128 * Only clear the pmap modified bits if ALL the dirty bits 4129 * are set, otherwise the system might mis-clear portions 4130 * of a page. 4131 */ 4132 if (m->dirty == VM_PAGE_BITS_ALL && 4133 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4134 pmap_clear_modify(m); 4135 } 4136 if (bp->b_dirtyoff > soff - xoff) 4137 bp->b_dirtyoff = soff - xoff; 4138 if (bp->b_dirtyend < eoff - xoff) 4139 bp->b_dirtyend = eoff - xoff; 4140 } 4141 4142 /* 4143 * Set related valid bits, clear related dirty bits. 4144 * Does not mess with the pmap modified bit. 4145 * 4146 * WARNING! We cannot just clear all of m->dirty here as the 4147 * buffer cache buffers may use a DEV_BSIZE'd aligned 4148 * block size, or have an odd size (e.g. NFS at file EOF). 4149 * The putpages code can clear m->dirty to 0. 4150 * 4151 * If a VOP_WRITE generates a buffer cache buffer which 4152 * covers the same space as mapped writable pages the 4153 * buffer flush might not be able to clear all the dirty 4154 * bits and still require a putpages from the VM system 4155 * to finish it off. 4156 * 4157 * WARNING! vm_page_set_validclean() currently assumes vm_token 4158 * is held. The page might not be busied (bdwrite() case). 4159 * XXX remove this comment once we've validated that this 4160 * is no longer an issue. 4161 */ 4162 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4163 } 4164 4165 #if 0 4166 /* 4167 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4168 * The page data is assumed to be valid (there is no zeroing here). 4169 */ 4170 static void 4171 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4172 { 4173 int bcount; 4174 int xoff; 4175 int soff; 4176 int eoff; 4177 4178 /* 4179 * Calculate offset range within the page but relative to buffer's 4180 * loffset. loffset might be offset into the first page. 4181 */ 4182 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4183 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4184 4185 if (pageno == 0) { 4186 soff = xoff; 4187 eoff = PAGE_SIZE; 4188 } else { 4189 soff = (pageno << PAGE_SHIFT); 4190 eoff = soff + PAGE_SIZE; 4191 } 4192 if (eoff > bcount) 4193 eoff = bcount; 4194 if (soff >= eoff) 4195 return; 4196 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4197 } 4198 #endif 4199 4200 /* 4201 * vfs_bio_clrbuf: 4202 * 4203 * Clear a buffer. This routine essentially fakes an I/O, so we need 4204 * to clear B_ERROR and B_INVAL. 4205 * 4206 * Note that while we only theoretically need to clear through b_bcount, 4207 * we go ahead and clear through b_bufsize. 4208 */ 4209 void 4210 vfs_bio_clrbuf(struct buf *bp) 4211 { 4212 int i, mask = 0; 4213 caddr_t sa, ea; 4214 KKASSERT(bp->b_flags & B_VMIO); 4215 4216 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4217 bkvasync(bp); 4218 4219 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4220 (bp->b_loffset & PAGE_MASK) == 0) { 4221 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4222 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4223 bp->b_resid = 0; 4224 return; 4225 } 4226 if ((bp->b_xio.xio_pages[0]->valid & mask) == 0) { 4227 bzero(bp->b_data, bp->b_bufsize); 4228 bp->b_xio.xio_pages[0]->valid |= mask; 4229 bp->b_resid = 0; 4230 return; 4231 } 4232 } 4233 sa = bp->b_data; 4234 for(i = 0; i < bp->b_xio.xio_npages; i++, sa=ea) { 4235 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4236 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4237 ea = (caddr_t)(vm_offset_t)ulmin( 4238 (u_long)(vm_offset_t)ea, 4239 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4240 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4241 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4242 continue; 4243 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4244 bzero(sa, ea - sa); 4245 } else { 4246 for (; sa < ea; sa += DEV_BSIZE, j++) { 4247 if ((bp->b_xio.xio_pages[i]->valid & 4248 (1<<j)) == 0) { 4249 bzero(sa, DEV_BSIZE); 4250 } 4251 } 4252 } 4253 bp->b_xio.xio_pages[i]->valid |= mask; 4254 } 4255 bp->b_resid = 0; 4256 } 4257 4258 /* 4259 * Allocate a page for a buffer cache buffer. 4260 * 4261 * If NULL is returned the caller is expected to retry (typically check if 4262 * the page already exists on retry before trying to allocate one). 4263 * 4264 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This 4265 * function will use the system reserve with the hope that the page 4266 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller 4267 * is done with the buffer. 4268 * 4269 * NOTE! However, TMPFS is a special case because flushing a dirty buffer 4270 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon 4271 * is capable of retiring pages (to swap). For TMPFS we don't dig 4272 * into the system reserve because doing so could stall out pretty 4273 * much every process running on the system. 4274 */ 4275 static 4276 vm_page_t 4277 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit) 4278 { 4279 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK; 4280 vm_page_t p; 4281 4282 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4283 4284 /* 4285 * Try a normal allocation first. 4286 */ 4287 p = vm_page_alloc(obj, pg, vmflags); 4288 if (p) 4289 return(p); 4290 if (vm_page_lookup(obj, pg)) 4291 return(NULL); 4292 vm_pageout_deficit += deficit; 4293 4294 /* 4295 * Try again, digging into the system reserve. 4296 * 4297 * Trying to recover pages from the buffer cache here can deadlock 4298 * against other threads trying to busy underlying pages so we 4299 * depend on the code in brelse() and bqrelse() to free/cache the 4300 * underlying buffer cache pages when memory is low. 4301 */ 4302 if (curthread->td_flags & TDF_SYSTHREAD) 4303 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT; 4304 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS) 4305 vmflags |= 0; 4306 else 4307 vmflags |= VM_ALLOC_SYSTEM; 4308 4309 /*recoverbufpages();*/ 4310 p = vm_page_alloc(obj, pg, vmflags); 4311 if (p) 4312 return(p); 4313 if (vm_page_lookup(obj, pg)) 4314 return(NULL); 4315 4316 /* 4317 * Wait for memory to free up and try again 4318 */ 4319 if (vm_page_count_severe()) 4320 ++lowmempgallocs; 4321 vm_wait(hz / 20 + 1); 4322 4323 p = vm_page_alloc(obj, pg, vmflags); 4324 if (p) 4325 return(p); 4326 if (vm_page_lookup(obj, pg)) 4327 return(NULL); 4328 4329 /* 4330 * Ok, now we are really in trouble. 4331 */ 4332 if (bootverbose) { 4333 static struct krate biokrate = { .freq = 1 }; 4334 krateprintf(&biokrate, 4335 "Warning: bio_page_alloc: memory exhausted " 4336 "during buffer cache page allocation from %s\n", 4337 curthread->td_comm); 4338 } 4339 if (curthread->td_flags & TDF_SYSTHREAD) 4340 vm_wait(hz / 20 + 1); 4341 else 4342 vm_wait(hz / 2 + 1); 4343 return (NULL); 4344 } 4345 4346 /* 4347 * The buffer's mapping has changed. Adjust the buffer's memory 4348 * synchronization. The caller is the exclusive holder of the buffer 4349 * and has set or cleared B_KVABIO according to preference. 4350 * 4351 * WARNING! If the caller is using B_KVABIO mode, this function will 4352 * not map the data to the current cpu. The caller must also 4353 * call bkvasync(bp). 4354 */ 4355 void 4356 bkvareset(struct buf *bp) 4357 { 4358 if (bp->b_flags & B_KVABIO) { 4359 CPUMASK_ASSZERO(bp->b_cpumask); 4360 } else { 4361 CPUMASK_ORMASK(bp->b_cpumask, smp_active_mask); 4362 smp_invltlb(); 4363 cpu_invltlb(); 4364 } 4365 } 4366 4367 /* 4368 * The buffer will be used by the caller on the caller's cpu, synchronize 4369 * its data to the current cpu. 4370 * 4371 * If B_KVABIO is not set, the buffer is already fully synchronized. 4372 */ 4373 void 4374 bkvasync(struct buf *bp) 4375 { 4376 int cpuid = mycpu->gd_cpuid; 4377 char *bdata; 4378 4379 if ((bp->b_flags & B_KVABIO) && 4380 CPUMASK_TESTBIT(bp->b_cpumask, cpuid) == 0) { 4381 bdata = bp->b_data; 4382 while (bdata < bp->b_data + bp->b_bufsize) { 4383 cpu_invlpg(bdata); 4384 bdata += PAGE_SIZE - 4385 ((intptr_t)bdata & PAGE_MASK); 4386 } 4387 ATOMIC_CPUMASK_ORBIT(bp->b_cpumask, cpuid); 4388 } 4389 } 4390 4391 /* 4392 * The buffer will be used by a subsystem that does not understand 4393 * the KVABIO API. Make sure its data is synchronized to all cpus. 4394 * 4395 * If B_KVABIO is not set, the buffer is already fully synchronized. 4396 * 4397 * NOTE! This is the only safe way to clear B_KVABIO on a buffer. 4398 */ 4399 void 4400 bkvasync_all(struct buf *bp) 4401 { 4402 if (debug_kvabio > 0) { 4403 --debug_kvabio; 4404 print_backtrace(10); 4405 } 4406 4407 if ((bp->b_flags & B_KVABIO) && 4408 CPUMASK_CMPMASKNEQ(bp->b_cpumask, smp_active_mask)) { 4409 smp_invltlb(); 4410 cpu_invltlb(); 4411 ATOMIC_CPUMASK_ORMASK(bp->b_cpumask, smp_active_mask); 4412 } 4413 bp->b_flags &= ~B_KVABIO; 4414 } 4415 4416 /* 4417 * Scan all buffers in the system and issue the callback. 4418 */ 4419 int 4420 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4421 { 4422 int count = 0; 4423 int error; 4424 long n; 4425 4426 for (n = 0; n < nbuf; ++n) { 4427 if ((error = callback(&buf[n], info)) < 0) { 4428 count = error; 4429 break; 4430 } 4431 count += error; 4432 } 4433 return (count); 4434 } 4435 4436 /* 4437 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4438 * completion to the master buffer. 4439 */ 4440 static void 4441 nestiobuf_iodone(struct bio *bio) 4442 { 4443 struct bio *mbio; 4444 struct buf *mbp, *bp; 4445 struct devstat *stats; 4446 int error; 4447 int donebytes; 4448 4449 bp = bio->bio_buf; 4450 mbio = bio->bio_caller_info1.ptr; 4451 stats = bio->bio_caller_info2.ptr; 4452 mbp = mbio->bio_buf; 4453 4454 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4455 KKASSERT(mbp != bp); 4456 4457 error = bp->b_error; 4458 if (bp->b_error == 0 && 4459 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4460 /* 4461 * Not all got transfered, raise an error. We have no way to 4462 * propagate these conditions to mbp. 4463 */ 4464 error = EIO; 4465 } 4466 4467 donebytes = bp->b_bufsize; 4468 4469 relpbuf(bp, NULL); 4470 4471 nestiobuf_done(mbio, donebytes, error, stats); 4472 } 4473 4474 void 4475 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4476 { 4477 struct buf *mbp; 4478 4479 mbp = mbio->bio_buf; 4480 4481 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4482 4483 /* 4484 * If an error occured, propagate it to the master buffer. 4485 * 4486 * Several biodone()s may wind up running concurrently so 4487 * use an atomic op to adjust b_flags. 4488 */ 4489 if (error) { 4490 mbp->b_error = error; 4491 atomic_set_int(&mbp->b_flags, B_ERROR); 4492 } 4493 4494 /* 4495 * Decrement the operations in progress counter and terminate the 4496 * I/O if this was the last bit. 4497 */ 4498 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4499 mbp->b_resid = 0; 4500 if (stats) 4501 devstat_end_transaction_buf(stats, mbp); 4502 biodone(mbio); 4503 } 4504 } 4505 4506 /* 4507 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4508 * the mbio from being biodone()'d while we are still adding sub-bios to 4509 * it. 4510 */ 4511 void 4512 nestiobuf_init(struct bio *bio) 4513 { 4514 bio->bio_driver_info = (void *)1; 4515 } 4516 4517 /* 4518 * The BIOs added to the nestedio have already been started, remove the 4519 * count that placeheld our mbio and biodone() it if the count would 4520 * transition to 0. 4521 */ 4522 void 4523 nestiobuf_start(struct bio *mbio) 4524 { 4525 struct buf *mbp = mbio->bio_buf; 4526 4527 /* 4528 * Decrement the operations in progress counter and terminate the 4529 * I/O if this was the last bit. 4530 */ 4531 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4532 if (mbp->b_flags & B_ERROR) 4533 mbp->b_resid = mbp->b_bcount; 4534 else 4535 mbp->b_resid = 0; 4536 biodone(mbio); 4537 } 4538 } 4539 4540 /* 4541 * Set an intermediate error prior to calling nestiobuf_start() 4542 */ 4543 void 4544 nestiobuf_error(struct bio *mbio, int error) 4545 { 4546 struct buf *mbp = mbio->bio_buf; 4547 4548 if (error) { 4549 mbp->b_error = error; 4550 atomic_set_int(&mbp->b_flags, B_ERROR); 4551 } 4552 } 4553 4554 /* 4555 * nestiobuf_add: setup a "nested" buffer. 4556 * 4557 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4558 * => 'bp' should be a buffer allocated by getiobuf. 4559 * => 'offset' is a byte offset in the master buffer. 4560 * => 'size' is a size in bytes of this nested buffer. 4561 */ 4562 void 4563 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4564 { 4565 struct buf *mbp = mbio->bio_buf; 4566 struct vnode *vp = mbp->b_vp; 4567 4568 KKASSERT(mbp->b_bcount >= offset + size); 4569 4570 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4571 4572 /* kernel needs to own the lock for it to be released in biodone */ 4573 BUF_KERNPROC(bp); 4574 bp->b_vp = vp; 4575 bp->b_cmd = mbp->b_cmd; 4576 bp->b_bio1.bio_done = nestiobuf_iodone; 4577 bp->b_data = (char *)mbp->b_data + offset; 4578 bp->b_resid = bp->b_bcount = size; 4579 bp->b_bufsize = bp->b_bcount; 4580 4581 bp->b_bio1.bio_track = NULL; 4582 bp->b_bio1.bio_caller_info1.ptr = mbio; 4583 bp->b_bio1.bio_caller_info2.ptr = stats; 4584 } 4585 4586 #ifdef DDB 4587 4588 DB_SHOW_COMMAND(buffer, db_show_buffer) 4589 { 4590 /* get args */ 4591 struct buf *bp = (struct buf *)addr; 4592 4593 if (!have_addr) { 4594 db_printf("usage: show buffer <addr>\n"); 4595 return; 4596 } 4597 4598 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4599 db_printf("b_cmd = %d\n", bp->b_cmd); 4600 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4601 "b_resid = %d\n, b_data = %p, " 4602 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4603 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4604 bp->b_data, 4605 (long long)bp->b_bio2.bio_offset, 4606 (long long)(bp->b_bio2.bio_next ? 4607 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4608 if (bp->b_xio.xio_npages) { 4609 int i; 4610 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4611 bp->b_xio.xio_npages); 4612 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4613 vm_page_t m; 4614 m = bp->b_xio.xio_pages[i]; 4615 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4616 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4617 if ((i + 1) < bp->b_xio.xio_npages) 4618 db_printf(","); 4619 } 4620 db_printf("\n"); 4621 } 4622 } 4623 #endif /* DDB */ 4624