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 void repurposebuf(struct buf *bp, int size); 110 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj, 111 vm_pindex_t pg, int deficit); 112 113 static void bd_signal(long totalspace); 114 static void buf_daemon(void); 115 static void buf_daemon_hw(void); 116 117 /* 118 * bogus page -- for I/O to/from partially complete buffers 119 * this is a temporary solution to the problem, but it is not 120 * really that bad. it would be better to split the buffer 121 * for input in the case of buffers partially already in memory, 122 * but the code is intricate enough already. 123 */ 124 vm_page_t bogus_page; 125 126 /* 127 * These are all static, but make the ones we export globals so we do 128 * not need to use compiler magic. 129 */ 130 long bufspace; /* atomic ops */ 131 long maxbufspace; 132 static long bufmallocspace; /* atomic ops */ 133 long maxbufmallocspace, lobufspace, hibufspace; 134 static long lorunningspace; 135 static long hirunningspace; 136 static long dirtykvaspace; /* atomic */ 137 long dirtybufspace; /* atomic (global for systat) */ 138 static long dirtybufcount; /* atomic */ 139 static long dirtybufspacehw; /* atomic */ 140 static long dirtybufcounthw; /* atomic */ 141 static long runningbufspace; /* atomic */ 142 static long runningbufcount; /* atomic */ 143 static long repurposedspace; 144 long lodirtybufspace; 145 long hidirtybufspace; 146 static int getnewbufcalls; 147 static int recoverbufcalls; 148 static int needsbuffer; /* atomic */ 149 static int runningbufreq; /* atomic */ 150 static int bd_request; /* atomic */ 151 static int bd_request_hw; /* atomic */ 152 static u_int bd_wake_ary[BD_WAKE_SIZE]; 153 static u_int bd_wake_index; 154 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 155 static int debug_commit; 156 static int debug_bufbio; 157 static long bufcache_bw = 200 * 1024 * 1024; 158 static long bufcache_bw_accum; 159 static int bufcache_bw_ticks; 160 161 static struct thread *bufdaemon_td; 162 static struct thread *bufdaemonhw_td; 163 static u_int lowmempgallocs; 164 static u_int lowmempgfails; 165 static u_int flushperqueue = 1024; 166 static int repurpose_enable; 167 168 /* 169 * Sysctls for operational control of the buffer cache. 170 */ 171 SYSCTL_UINT(_vfs, OID_AUTO, flushperqueue, CTLFLAG_RW, &flushperqueue, 0, 172 "Number of buffers to flush from each per-cpu queue"); 173 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 174 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 175 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 176 "High watermark used to trigger explicit flushing of dirty buffers"); 177 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 178 "Minimum amount of buffer space required for active I/O"); 179 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 180 "Maximum amount of buffer space to usable for active I/O"); 181 SYSCTL_LONG(_vfs, OID_AUTO, bufcache_bw, CTLFLAG_RW, &bufcache_bw, 0, 182 "Buffer-cache -> VM page cache transfer bandwidth"); 183 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 184 "Page allocations done during periods of very low free memory"); 185 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 186 "Page allocations which failed during periods of very low free memory"); 187 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 188 "Recycle pages to active or inactive queue transition pt 0-64"); 189 SYSCTL_UINT(_vfs, OID_AUTO, repurpose_enable, CTLFLAG_RW, &repurpose_enable, 0, 190 "Enable buffer cache VM repurposing for high-I/O"); 191 /* 192 * Sysctls determining current state of the buffer cache. 193 */ 194 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 195 "Total number of buffers in buffer cache"); 196 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0, 197 "KVA reserved by dirty buffers (all)"); 198 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 199 "Pending bytes of dirty buffers (all)"); 200 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 201 "Pending bytes of dirty buffers (heavy weight)"); 202 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 203 "Pending number of dirty buffers"); 204 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 205 "Pending number of dirty buffers (heavy weight)"); 206 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 207 "I/O bytes currently in progress due to asynchronous writes"); 208 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 209 "I/O buffers currently in progress due to asynchronous writes"); 210 SYSCTL_LONG(_vfs, OID_AUTO, repurposedspace, CTLFLAG_RD, &repurposedspace, 0, 211 "Buffer-cache memory repurposed in-place"); 212 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 213 "Hard limit on maximum amount of memory usable for buffer space"); 214 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 215 "Soft limit on maximum amount of memory usable for buffer space"); 216 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 217 "Minimum amount of memory to reserve for system buffer space"); 218 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 219 "Amount of memory available for buffers"); 220 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 221 0, "Maximum amount of memory reserved for buffers using malloc"); 222 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 223 "Amount of memory left for buffers using malloc-scheme"); 224 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 225 "New buffer header acquisition requests"); 226 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 227 "Recover VM space in an emergency"); 228 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, ""); 229 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, ""); 230 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 231 "sizeof(struct buf)"); 232 233 char *buf_wmesg = BUF_WMESG; 234 235 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 236 #define VFS_BIO_NEED_UNUSED02 0x02 237 #define VFS_BIO_NEED_UNUSED04 0x04 238 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 239 240 /* 241 * Called when buffer space is potentially available for recovery. 242 * getnewbuf() will block on this flag when it is unable to free 243 * sufficient buffer space. Buffer space becomes recoverable when 244 * bp's get placed back in the queues. 245 */ 246 static __inline void 247 bufspacewakeup(void) 248 { 249 /* 250 * If someone is waiting for BUF space, wake them up. Even 251 * though we haven't freed the kva space yet, the waiting 252 * process will be able to now. 253 */ 254 for (;;) { 255 int flags = needsbuffer; 256 cpu_ccfence(); 257 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0) 258 break; 259 if (atomic_cmpset_int(&needsbuffer, flags, 260 flags & ~VFS_BIO_NEED_BUFSPACE)) { 261 wakeup(&needsbuffer); 262 break; 263 } 264 /* retry */ 265 } 266 } 267 268 /* 269 * runningbufwakeup: 270 * 271 * Accounting for I/O in progress. 272 * 273 */ 274 static __inline void 275 runningbufwakeup(struct buf *bp) 276 { 277 long totalspace; 278 long flags; 279 280 if ((totalspace = bp->b_runningbufspace) != 0) { 281 atomic_add_long(&runningbufspace, -totalspace); 282 atomic_add_long(&runningbufcount, -1); 283 bp->b_runningbufspace = 0; 284 285 /* 286 * see waitrunningbufspace() for limit test. 287 */ 288 for (;;) { 289 flags = runningbufreq; 290 cpu_ccfence(); 291 if (flags == 0) 292 break; 293 if (atomic_cmpset_int(&runningbufreq, flags, 0)) { 294 wakeup(&runningbufreq); 295 break; 296 } 297 /* retry */ 298 } 299 bd_signal(totalspace); 300 } 301 } 302 303 /* 304 * bufcountwakeup: 305 * 306 * Called when a buffer has been added to one of the free queues to 307 * account for the buffer and to wakeup anyone waiting for free buffers. 308 * This typically occurs when large amounts of metadata are being handled 309 * by the buffer cache ( else buffer space runs out first, usually ). 310 */ 311 static __inline void 312 bufcountwakeup(void) 313 { 314 long flags; 315 316 for (;;) { 317 flags = needsbuffer; 318 if (flags == 0) 319 break; 320 if (atomic_cmpset_int(&needsbuffer, flags, 321 (flags & ~VFS_BIO_NEED_ANY))) { 322 wakeup(&needsbuffer); 323 break; 324 } 325 /* retry */ 326 } 327 } 328 329 /* 330 * waitrunningbufspace() 331 * 332 * If runningbufspace exceeds 4/6 hirunningspace we block until 333 * runningbufspace drops to 3/6 hirunningspace. We also block if another 334 * thread blocked here in order to be fair, even if runningbufspace 335 * is now lower than the limit. 336 * 337 * The caller may be using this function to block in a tight loop, we 338 * must block while runningbufspace is greater than at least 339 * hirunningspace * 3 / 6. 340 */ 341 void 342 waitrunningbufspace(void) 343 { 344 long limit = hirunningspace * 4 / 6; 345 long flags; 346 347 while (runningbufspace > limit || runningbufreq) { 348 tsleep_interlock(&runningbufreq, 0); 349 flags = atomic_fetchadd_int(&runningbufreq, 1); 350 if (runningbufspace > limit || flags) 351 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz); 352 } 353 } 354 355 /* 356 * buf_dirty_count_severe: 357 * 358 * Return true if we have too many dirty buffers. 359 */ 360 int 361 buf_dirty_count_severe(void) 362 { 363 return (runningbufspace + dirtykvaspace >= hidirtybufspace || 364 dirtybufcount >= nbuf / 2); 365 } 366 367 /* 368 * Return true if the amount of running I/O is severe and BIOQ should 369 * start bursting. 370 */ 371 int 372 buf_runningbufspace_severe(void) 373 { 374 return (runningbufspace >= hirunningspace * 4 / 6); 375 } 376 377 /* 378 * vfs_buf_test_cache: 379 * 380 * Called when a buffer is extended. This function clears the B_CACHE 381 * bit if the newly extended portion of the buffer does not contain 382 * valid data. 383 * 384 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 385 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 386 * them while a clean buffer was present. 387 */ 388 static __inline__ 389 void 390 vfs_buf_test_cache(struct buf *bp, 391 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 392 vm_page_t m) 393 { 394 if (bp->b_flags & B_CACHE) { 395 int base = (foff + off) & PAGE_MASK; 396 if (vm_page_is_valid(m, base, size) == 0) 397 bp->b_flags &= ~B_CACHE; 398 } 399 } 400 401 /* 402 * bd_speedup() 403 * 404 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 405 * low water mark. 406 */ 407 static __inline__ 408 void 409 bd_speedup(void) 410 { 411 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2) 412 return; 413 414 if (bd_request == 0 && 415 (dirtykvaspace > lodirtybufspace / 2 || 416 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 417 if (atomic_fetchadd_int(&bd_request, 1) == 0) 418 wakeup(&bd_request); 419 } 420 if (bd_request_hw == 0 && 421 (dirtykvaspace > lodirtybufspace / 2 || 422 dirtybufcounthw >= nbuf / 2)) { 423 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0) 424 wakeup(&bd_request_hw); 425 } 426 } 427 428 /* 429 * bd_heatup() 430 * 431 * Get the buf_daemon heated up when the number of running and dirty 432 * buffers exceeds the mid-point. 433 * 434 * Return the total number of dirty bytes past the second mid point 435 * as a measure of how much excess dirty data there is in the system. 436 */ 437 long 438 bd_heatup(void) 439 { 440 long mid1; 441 long mid2; 442 long totalspace; 443 444 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 445 446 totalspace = runningbufspace + dirtykvaspace; 447 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 448 bd_speedup(); 449 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 450 if (totalspace >= mid2) 451 return(totalspace - mid2); 452 } 453 return(0); 454 } 455 456 /* 457 * bd_wait() 458 * 459 * Wait for the buffer cache to flush (totalspace) bytes worth of 460 * buffers, then return. 461 * 462 * Regardless this function blocks while the number of dirty buffers 463 * exceeds hidirtybufspace. 464 */ 465 void 466 bd_wait(long totalspace) 467 { 468 u_int i; 469 u_int j; 470 u_int mi; 471 int count; 472 473 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 474 return; 475 476 while (totalspace > 0) { 477 bd_heatup(); 478 479 /* 480 * Order is important. Suppliers adjust bd_wake_index after 481 * updating runningbufspace/dirtykvaspace. We want to fetch 482 * bd_wake_index before accessing. Any error should thus 483 * be in our favor. 484 */ 485 i = atomic_fetchadd_int(&bd_wake_index, 0); 486 if (totalspace > runningbufspace + dirtykvaspace) 487 totalspace = runningbufspace + dirtykvaspace; 488 count = totalspace / MAXBSIZE; 489 if (count >= BD_WAKE_SIZE / 2) 490 count = BD_WAKE_SIZE / 2; 491 i = i + count; 492 mi = i & BD_WAKE_MASK; 493 494 /* 495 * This is not a strict interlock, so we play a bit loose 496 * with locking access to dirtybufspace*. We have to re-check 497 * bd_wake_index to ensure that it hasn't passed us. 498 */ 499 tsleep_interlock(&bd_wake_ary[mi], 0); 500 atomic_add_int(&bd_wake_ary[mi], 1); 501 j = atomic_fetchadd_int(&bd_wake_index, 0); 502 if ((int)(i - j) >= 0) 503 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz); 504 505 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace; 506 } 507 } 508 509 /* 510 * bd_signal() 511 * 512 * This function is called whenever runningbufspace or dirtykvaspace 513 * is reduced. Track threads waiting for run+dirty buffer I/O 514 * complete. 515 */ 516 static void 517 bd_signal(long totalspace) 518 { 519 u_int i; 520 521 if (totalspace > 0) { 522 if (totalspace > MAXBSIZE * BD_WAKE_SIZE) 523 totalspace = MAXBSIZE * BD_WAKE_SIZE; 524 while (totalspace > 0) { 525 i = atomic_fetchadd_int(&bd_wake_index, 1); 526 i &= BD_WAKE_MASK; 527 if (atomic_readandclear_int(&bd_wake_ary[i])) 528 wakeup(&bd_wake_ary[i]); 529 totalspace -= MAXBSIZE; 530 } 531 } 532 } 533 534 /* 535 * BIO tracking support routines. 536 * 537 * Release a ref on a bio_track. Wakeup requests are atomically released 538 * along with the last reference so bk_active will never wind up set to 539 * only 0x80000000. 540 */ 541 static 542 void 543 bio_track_rel(struct bio_track *track) 544 { 545 int active; 546 int desired; 547 548 /* 549 * Shortcut 550 */ 551 active = track->bk_active; 552 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 553 return; 554 555 /* 556 * Full-on. Note that the wait flag is only atomically released on 557 * the 1->0 count transition. 558 * 559 * We check for a negative count transition using bit 30 since bit 31 560 * has a different meaning. 561 */ 562 for (;;) { 563 desired = (active & 0x7FFFFFFF) - 1; 564 if (desired) 565 desired |= active & 0x80000000; 566 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 567 if (desired & 0x40000000) 568 panic("bio_track_rel: bad count: %p", track); 569 if (active & 0x80000000) 570 wakeup(track); 571 break; 572 } 573 active = track->bk_active; 574 } 575 } 576 577 /* 578 * Wait for the tracking count to reach 0. 579 * 580 * Use atomic ops such that the wait flag is only set atomically when 581 * bk_active is non-zero. 582 */ 583 int 584 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 585 { 586 int active; 587 int desired; 588 int error; 589 590 /* 591 * Shortcut 592 */ 593 if (track->bk_active == 0) 594 return(0); 595 596 /* 597 * Full-on. Note that the wait flag may only be atomically set if 598 * the active count is non-zero. 599 * 600 * NOTE: We cannot optimize active == desired since a wakeup could 601 * clear active prior to our tsleep_interlock(). 602 */ 603 error = 0; 604 while ((active = track->bk_active) != 0) { 605 cpu_ccfence(); 606 desired = active | 0x80000000; 607 tsleep_interlock(track, slp_flags); 608 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 609 error = tsleep(track, slp_flags | PINTERLOCKED, 610 "trwait", slp_timo); 611 if (error) 612 break; 613 } 614 } 615 return (error); 616 } 617 618 /* 619 * bufinit: 620 * 621 * Load time initialisation of the buffer cache, called from machine 622 * dependant initialization code. 623 */ 624 static 625 void 626 bufinit(void *dummy __unused) 627 { 628 struct bufpcpu *pcpu; 629 struct buf *bp; 630 vm_offset_t bogus_offset; 631 int i; 632 int j; 633 long n; 634 635 /* next, make a null set of free lists */ 636 for (i = 0; i < ncpus; ++i) { 637 pcpu = &bufpcpu[i]; 638 spin_init(&pcpu->spin, "bufinit"); 639 for (j = 0; j < BUFFER_QUEUES; j++) 640 TAILQ_INIT(&pcpu->bufqueues[j]); 641 } 642 643 /* 644 * Finally, initialize each buffer header and stick on empty q. 645 * Each buffer gets its own KVA reservation. 646 */ 647 i = 0; 648 pcpu = &bufpcpu[i]; 649 650 for (n = 0; n < nbuf; n++) { 651 bp = &buf[n]; 652 bzero(bp, sizeof *bp); 653 bp->b_flags = B_INVAL; /* we're just an empty header */ 654 bp->b_cmd = BUF_CMD_DONE; 655 bp->b_qindex = BQUEUE_EMPTY; 656 bp->b_qcpu = i; 657 bp->b_kvabase = (void *)(vm_map_min(&buffer_map) + 658 MAXBSIZE * n); 659 bp->b_kvasize = MAXBSIZE; 660 initbufbio(bp); 661 xio_init(&bp->b_xio); 662 buf_dep_init(bp); 663 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 664 bp, b_freelist); 665 666 i = (i + 1) % ncpus; 667 pcpu = &bufpcpu[i]; 668 } 669 670 /* 671 * maxbufspace is the absolute maximum amount of buffer space we are 672 * allowed to reserve in KVM and in real terms. The absolute maximum 673 * is nominally used by buf_daemon. hibufspace is the nominal maximum 674 * used by most other processes. The differential is required to 675 * ensure that buf_daemon is able to run when other processes might 676 * be blocked waiting for buffer space. 677 * 678 * Calculate hysteresis (lobufspace, hibufspace). Don't make it 679 * too large or we might lockup a cpu for too long a period of 680 * time in our tight loop. 681 */ 682 maxbufspace = nbuf * NBUFCALCSIZE; 683 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 684 lobufspace = hibufspace * 7 / 8; 685 if (hibufspace - lobufspace > 64 * 1024 * 1024) 686 lobufspace = hibufspace - 64 * 1024 * 1024; 687 if (lobufspace > hibufspace - MAXBSIZE) 688 lobufspace = hibufspace - MAXBSIZE; 689 690 lorunningspace = 512 * 1024; 691 /* hirunningspace -- see below */ 692 693 /* 694 * Limit the amount of malloc memory since it is wired permanently 695 * into the kernel space. Even though this is accounted for in 696 * the buffer allocation, we don't want the malloced region to grow 697 * uncontrolled. The malloc scheme improves memory utilization 698 * significantly on average (small) directories. 699 */ 700 maxbufmallocspace = hibufspace / 20; 701 702 /* 703 * Reduce the chance of a deadlock occuring by limiting the number 704 * of delayed-write dirty buffers we allow to stack up. 705 * 706 * We don't want too much actually queued to the device at once 707 * (XXX this needs to be per-mount!), because the buffers will 708 * wind up locked for a very long period of time while the I/O 709 * drains. 710 */ 711 hidirtybufspace = hibufspace / 2; /* dirty + running */ 712 hirunningspace = hibufspace / 16; /* locked & queued to device */ 713 if (hirunningspace < 1024 * 1024) 714 hirunningspace = 1024 * 1024; 715 716 dirtykvaspace = 0; 717 dirtybufspace = 0; 718 dirtybufspacehw = 0; 719 720 lodirtybufspace = hidirtybufspace / 2; 721 722 /* 723 * Maximum number of async ops initiated per buf_daemon loop. This is 724 * somewhat of a hack at the moment, we really need to limit ourselves 725 * based on the number of bytes of I/O in-transit that were initiated 726 * from buf_daemon. 727 */ 728 729 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE, 730 VM_SUBSYS_BOGUS); 731 vm_object_hold(&kernel_object); 732 bogus_page = vm_page_alloc(&kernel_object, 733 (bogus_offset >> PAGE_SHIFT), 734 VM_ALLOC_NORMAL); 735 vm_object_drop(&kernel_object); 736 vmstats.v_wire_count++; 737 738 } 739 740 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL); 741 742 /* 743 * Initialize the embedded bio structures, typically used by 744 * deprecated code which tries to allocate its own struct bufs. 745 */ 746 void 747 initbufbio(struct buf *bp) 748 { 749 bp->b_bio1.bio_buf = bp; 750 bp->b_bio1.bio_prev = NULL; 751 bp->b_bio1.bio_offset = NOOFFSET; 752 bp->b_bio1.bio_next = &bp->b_bio2; 753 bp->b_bio1.bio_done = NULL; 754 bp->b_bio1.bio_flags = 0; 755 756 bp->b_bio2.bio_buf = bp; 757 bp->b_bio2.bio_prev = &bp->b_bio1; 758 bp->b_bio2.bio_offset = NOOFFSET; 759 bp->b_bio2.bio_next = NULL; 760 bp->b_bio2.bio_done = NULL; 761 bp->b_bio2.bio_flags = 0; 762 763 BUF_LOCKINIT(bp); 764 } 765 766 /* 767 * Reinitialize the embedded bio structures as well as any additional 768 * translation cache layers. 769 */ 770 void 771 reinitbufbio(struct buf *bp) 772 { 773 struct bio *bio; 774 775 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 776 bio->bio_done = NULL; 777 bio->bio_offset = NOOFFSET; 778 } 779 } 780 781 /* 782 * Undo the effects of an initbufbio(). 783 */ 784 void 785 uninitbufbio(struct buf *bp) 786 { 787 dsched_buf_exit(bp); 788 BUF_LOCKFREE(bp); 789 } 790 791 /* 792 * Push another BIO layer onto an existing BIO and return it. The new 793 * BIO layer may already exist, holding cached translation data. 794 */ 795 struct bio * 796 push_bio(struct bio *bio) 797 { 798 struct bio *nbio; 799 800 if ((nbio = bio->bio_next) == NULL) { 801 int index = bio - &bio->bio_buf->b_bio_array[0]; 802 if (index >= NBUF_BIO - 1) { 803 panic("push_bio: too many layers %d for bp %p", 804 index, bio->bio_buf); 805 } 806 nbio = &bio->bio_buf->b_bio_array[index + 1]; 807 bio->bio_next = nbio; 808 nbio->bio_prev = bio; 809 nbio->bio_buf = bio->bio_buf; 810 nbio->bio_offset = NOOFFSET; 811 nbio->bio_done = NULL; 812 nbio->bio_next = NULL; 813 } 814 KKASSERT(nbio->bio_done == NULL); 815 return(nbio); 816 } 817 818 /* 819 * Pop a BIO translation layer, returning the previous layer. The 820 * must have been previously pushed. 821 */ 822 struct bio * 823 pop_bio(struct bio *bio) 824 { 825 return(bio->bio_prev); 826 } 827 828 void 829 clearbiocache(struct bio *bio) 830 { 831 while (bio) { 832 bio->bio_offset = NOOFFSET; 833 bio = bio->bio_next; 834 } 835 } 836 837 /* 838 * Remove the buffer from the appropriate free list. 839 * (caller must be locked) 840 */ 841 static __inline void 842 _bremfree(struct buf *bp) 843 { 844 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 845 846 if (bp->b_qindex != BQUEUE_NONE) { 847 KASSERT(BUF_REFCNTNB(bp) == 1, 848 ("bremfree: bp %p not locked",bp)); 849 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist); 850 bp->b_qindex = BQUEUE_NONE; 851 } else { 852 if (BUF_REFCNTNB(bp) <= 1) 853 panic("bremfree: removing a buffer not on a queue"); 854 } 855 } 856 857 /* 858 * bremfree() - must be called with a locked buffer 859 */ 860 void 861 bremfree(struct buf *bp) 862 { 863 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 864 865 spin_lock(&pcpu->spin); 866 _bremfree(bp); 867 spin_unlock(&pcpu->spin); 868 } 869 870 /* 871 * bremfree_locked - must be called with pcpu->spin locked 872 */ 873 static void 874 bremfree_locked(struct buf *bp) 875 { 876 _bremfree(bp); 877 } 878 879 /* 880 * This version of bread issues any required I/O asyncnronously and 881 * makes a callback on completion. 882 * 883 * The callback must check whether BIO_DONE is set in the bio and issue 884 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 885 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 886 */ 887 void 888 breadcb(struct vnode *vp, off_t loffset, int size, int bflags, 889 void (*func)(struct bio *), void *arg) 890 { 891 struct buf *bp; 892 893 bp = getblk(vp, loffset, size, 0, 0); 894 895 /* if not found in cache, do some I/O */ 896 if ((bp->b_flags & B_CACHE) == 0) { 897 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA); 898 bp->b_flags |= bflags; 899 bp->b_cmd = BUF_CMD_READ; 900 bp->b_bio1.bio_done = func; 901 bp->b_bio1.bio_caller_info1.ptr = arg; 902 vfs_busy_pages(vp, bp); 903 BUF_KERNPROC(bp); 904 vn_strategy(vp, &bp->b_bio1); 905 } else if (func) { 906 /* 907 * Since we are issuing the callback synchronously it cannot 908 * race the BIO_DONE, so no need for atomic ops here. 909 */ 910 /*bp->b_bio1.bio_done = func;*/ 911 bp->b_bio1.bio_caller_info1.ptr = arg; 912 bp->b_bio1.bio_flags |= BIO_DONE; 913 func(&bp->b_bio1); 914 } else { 915 bqrelse(bp); 916 } 917 } 918 919 /* 920 * breadnx() - Terminal function for bread() and breadn(). 921 * 922 * This function will start asynchronous I/O on read-ahead blocks as well 923 * as satisfy the primary request. 924 * 925 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is 926 * set, the buffer is valid and we do not have to do anything. 927 */ 928 int 929 breadnx(struct vnode *vp, off_t loffset, int size, int bflags, 930 off_t *raoffset, int *rabsize, 931 int cnt, struct buf **bpp) 932 { 933 struct buf *bp, *rabp; 934 int i; 935 int rv = 0, readwait = 0; 936 937 if (*bpp) 938 bp = *bpp; 939 else 940 *bpp = bp = getblk(vp, loffset, size, 0, 0); 941 942 /* if not found in cache, do some I/O */ 943 if ((bp->b_flags & B_CACHE) == 0) { 944 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA); 945 bp->b_flags |= bflags; 946 bp->b_cmd = BUF_CMD_READ; 947 bp->b_bio1.bio_done = biodone_sync; 948 bp->b_bio1.bio_flags |= BIO_SYNC; 949 vfs_busy_pages(vp, bp); 950 vn_strategy(vp, &bp->b_bio1); 951 ++readwait; 952 } 953 954 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 955 if (inmem(vp, *raoffset)) 956 continue; 957 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 958 959 if ((rabp->b_flags & B_CACHE) == 0) { 960 rabp->b_flags &= ~(B_ERROR | B_EINTR | 961 B_INVAL | B_NOTMETA); 962 rabp->b_flags |= bflags; 963 rabp->b_cmd = BUF_CMD_READ; 964 vfs_busy_pages(vp, rabp); 965 BUF_KERNPROC(rabp); 966 vn_strategy(vp, &rabp->b_bio1); 967 } else { 968 brelse(rabp); 969 } 970 } 971 if (readwait) 972 rv = biowait(&bp->b_bio1, "biord"); 973 return (rv); 974 } 975 976 /* 977 * bwrite: 978 * 979 * Synchronous write, waits for completion. 980 * 981 * Write, release buffer on completion. (Done by iodone 982 * if async). Do not bother writing anything if the buffer 983 * is invalid. 984 * 985 * Note that we set B_CACHE here, indicating that buffer is 986 * fully valid and thus cacheable. This is true even of NFS 987 * now so we set it generally. This could be set either here 988 * or in biodone() since the I/O is synchronous. We put it 989 * here. 990 */ 991 int 992 bwrite(struct buf *bp) 993 { 994 int error; 995 996 if (bp->b_flags & B_INVAL) { 997 brelse(bp); 998 return (0); 999 } 1000 if (BUF_REFCNTNB(bp) == 0) 1001 panic("bwrite: buffer is not busy???"); 1002 1003 /* 1004 * NOTE: We no longer mark the buffer clear prior to the vn_strategy() 1005 * call because it will remove the buffer from the vnode's 1006 * dirty buffer list prematurely and possibly cause filesystem 1007 * checks to race buffer flushes. This is now handled in 1008 * bpdone(). 1009 * 1010 * bundirty(bp); REMOVED 1011 */ 1012 1013 bp->b_flags &= ~(B_ERROR | B_EINTR); 1014 bp->b_flags |= B_CACHE; 1015 bp->b_cmd = BUF_CMD_WRITE; 1016 bp->b_bio1.bio_done = biodone_sync; 1017 bp->b_bio1.bio_flags |= BIO_SYNC; 1018 vfs_busy_pages(bp->b_vp, bp); 1019 1020 /* 1021 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1022 * valid for vnode-backed buffers. 1023 */ 1024 bsetrunningbufspace(bp, bp->b_bufsize); 1025 vn_strategy(bp->b_vp, &bp->b_bio1); 1026 error = biowait(&bp->b_bio1, "biows"); 1027 brelse(bp); 1028 1029 return (error); 1030 } 1031 1032 /* 1033 * bawrite: 1034 * 1035 * Asynchronous write. Start output on a buffer, but do not wait for 1036 * it to complete. The buffer is released when the output completes. 1037 * 1038 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1039 * B_INVAL buffers. Not us. 1040 */ 1041 void 1042 bawrite(struct buf *bp) 1043 { 1044 if (bp->b_flags & B_INVAL) { 1045 brelse(bp); 1046 return; 1047 } 1048 if (BUF_REFCNTNB(bp) == 0) 1049 panic("bawrite: buffer is not busy???"); 1050 1051 /* 1052 * NOTE: We no longer mark the buffer clear prior to the vn_strategy() 1053 * call because it will remove the buffer from the vnode's 1054 * dirty buffer list prematurely and possibly cause filesystem 1055 * checks to race buffer flushes. This is now handled in 1056 * bpdone(). 1057 * 1058 * bundirty(bp); REMOVED 1059 */ 1060 bp->b_flags &= ~(B_ERROR | B_EINTR); 1061 bp->b_flags |= B_CACHE; 1062 bp->b_cmd = BUF_CMD_WRITE; 1063 KKASSERT(bp->b_bio1.bio_done == NULL); 1064 vfs_busy_pages(bp->b_vp, bp); 1065 1066 /* 1067 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1068 * valid for vnode-backed buffers. 1069 */ 1070 bsetrunningbufspace(bp, bp->b_bufsize); 1071 BUF_KERNPROC(bp); 1072 vn_strategy(bp->b_vp, &bp->b_bio1); 1073 } 1074 1075 /* 1076 * bdwrite: 1077 * 1078 * Delayed write. (Buffer is marked dirty). Do not bother writing 1079 * anything if the buffer is marked invalid. 1080 * 1081 * Note that since the buffer must be completely valid, we can safely 1082 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1083 * biodone() in order to prevent getblk from writing the buffer 1084 * out synchronously. 1085 */ 1086 void 1087 bdwrite(struct buf *bp) 1088 { 1089 if (BUF_REFCNTNB(bp) == 0) 1090 panic("bdwrite: buffer is not busy"); 1091 1092 if (bp->b_flags & B_INVAL) { 1093 brelse(bp); 1094 return; 1095 } 1096 bdirty(bp); 1097 1098 dsched_buf_enter(bp); /* might stack */ 1099 1100 /* 1101 * Set B_CACHE, indicating that the buffer is fully valid. This is 1102 * true even of NFS now. 1103 */ 1104 bp->b_flags |= B_CACHE; 1105 1106 /* 1107 * This bmap keeps the system from needing to do the bmap later, 1108 * perhaps when the system is attempting to do a sync. Since it 1109 * is likely that the indirect block -- or whatever other datastructure 1110 * that the filesystem needs is still in memory now, it is a good 1111 * thing to do this. Note also, that if the pageout daemon is 1112 * requesting a sync -- there might not be enough memory to do 1113 * the bmap then... So, this is important to do. 1114 */ 1115 if (bp->b_bio2.bio_offset == NOOFFSET) { 1116 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1117 NULL, NULL, BUF_CMD_WRITE); 1118 } 1119 1120 /* 1121 * Because the underlying pages may still be mapped and 1122 * writable trying to set the dirty buffer (b_dirtyoff/end) 1123 * range here will be inaccurate. 1124 * 1125 * However, we must still clean the pages to satisfy the 1126 * vnode_pager and pageout daemon, so they think the pages 1127 * have been "cleaned". What has really occured is that 1128 * they've been earmarked for later writing by the buffer 1129 * cache. 1130 * 1131 * So we get the b_dirtyoff/end update but will not actually 1132 * depend on it (NFS that is) until the pages are busied for 1133 * writing later on. 1134 */ 1135 vfs_clean_pages(bp); 1136 bqrelse(bp); 1137 1138 /* 1139 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1140 * due to the softdep code. 1141 */ 1142 } 1143 1144 /* 1145 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1146 * This is used by tmpfs. 1147 * 1148 * It is important for any VFS using this routine to NOT use it for 1149 * IO_SYNC or IO_ASYNC operations which occur when the system really 1150 * wants to flush VM pages to backing store. 1151 */ 1152 void 1153 buwrite(struct buf *bp) 1154 { 1155 vm_page_t m; 1156 int i; 1157 1158 /* 1159 * Only works for VMIO buffers. If the buffer is already 1160 * marked for delayed-write we can't avoid the bdwrite(). 1161 */ 1162 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1163 bdwrite(bp); 1164 return; 1165 } 1166 1167 /* 1168 * Mark as needing a commit. 1169 */ 1170 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1171 m = bp->b_xio.xio_pages[i]; 1172 vm_page_need_commit(m); 1173 } 1174 bqrelse(bp); 1175 } 1176 1177 /* 1178 * bdirty: 1179 * 1180 * Turn buffer into delayed write request by marking it B_DELWRI. 1181 * B_RELBUF and B_NOCACHE must be cleared. 1182 * 1183 * We reassign the buffer to itself to properly update it in the 1184 * dirty/clean lists. 1185 * 1186 * Must be called from a critical section. 1187 * The buffer must be on BQUEUE_NONE. 1188 */ 1189 void 1190 bdirty(struct buf *bp) 1191 { 1192 KASSERT(bp->b_qindex == BQUEUE_NONE, 1193 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1194 if (bp->b_flags & B_NOCACHE) { 1195 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1196 bp->b_flags &= ~B_NOCACHE; 1197 } 1198 if (bp->b_flags & B_INVAL) { 1199 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1200 } 1201 bp->b_flags &= ~B_RELBUF; 1202 1203 if ((bp->b_flags & B_DELWRI) == 0) { 1204 lwkt_gettoken(&bp->b_vp->v_token); 1205 bp->b_flags |= B_DELWRI; 1206 reassignbuf(bp); 1207 lwkt_reltoken(&bp->b_vp->v_token); 1208 1209 atomic_add_long(&dirtybufcount, 1); 1210 atomic_add_long(&dirtykvaspace, bp->b_kvasize); 1211 atomic_add_long(&dirtybufspace, bp->b_bufsize); 1212 if (bp->b_flags & B_HEAVY) { 1213 atomic_add_long(&dirtybufcounthw, 1); 1214 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1215 } 1216 bd_heatup(); 1217 } 1218 } 1219 1220 /* 1221 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1222 * needs to be flushed with a different buf_daemon thread to avoid 1223 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1224 */ 1225 void 1226 bheavy(struct buf *bp) 1227 { 1228 if ((bp->b_flags & B_HEAVY) == 0) { 1229 bp->b_flags |= B_HEAVY; 1230 if (bp->b_flags & B_DELWRI) { 1231 atomic_add_long(&dirtybufcounthw, 1); 1232 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1233 } 1234 } 1235 } 1236 1237 /* 1238 * bundirty: 1239 * 1240 * Clear B_DELWRI for buffer. 1241 * 1242 * Must be called from a critical section. 1243 * 1244 * The buffer is typically on BQUEUE_NONE but there is one case in 1245 * brelse() that calls this function after placing the buffer on 1246 * a different queue. 1247 */ 1248 void 1249 bundirty(struct buf *bp) 1250 { 1251 if (bp->b_flags & B_DELWRI) { 1252 lwkt_gettoken(&bp->b_vp->v_token); 1253 bp->b_flags &= ~B_DELWRI; 1254 reassignbuf(bp); 1255 lwkt_reltoken(&bp->b_vp->v_token); 1256 1257 atomic_add_long(&dirtybufcount, -1); 1258 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1259 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1260 if (bp->b_flags & B_HEAVY) { 1261 atomic_add_long(&dirtybufcounthw, -1); 1262 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize); 1263 } 1264 bd_signal(bp->b_bufsize); 1265 } 1266 /* 1267 * Since it is now being written, we can clear its deferred write flag. 1268 */ 1269 bp->b_flags &= ~B_DEFERRED; 1270 } 1271 1272 /* 1273 * Set the b_runningbufspace field, used to track how much I/O is 1274 * in progress at any given moment. 1275 */ 1276 void 1277 bsetrunningbufspace(struct buf *bp, int bytes) 1278 { 1279 bp->b_runningbufspace = bytes; 1280 if (bytes) { 1281 atomic_add_long(&runningbufspace, bytes); 1282 atomic_add_long(&runningbufcount, 1); 1283 } 1284 } 1285 1286 /* 1287 * brelse: 1288 * 1289 * Release a busy buffer and, if requested, free its resources. The 1290 * buffer will be stashed in the appropriate bufqueue[] allowing it 1291 * to be accessed later as a cache entity or reused for other purposes. 1292 */ 1293 void 1294 brelse(struct buf *bp) 1295 { 1296 struct bufpcpu *pcpu; 1297 #ifdef INVARIANTS 1298 int saved_flags = bp->b_flags; 1299 #endif 1300 1301 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1302 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1303 1304 /* 1305 * If B_NOCACHE is set we are being asked to destroy the buffer and 1306 * its backing store. Clear B_DELWRI. 1307 * 1308 * B_NOCACHE is set in two cases: (1) when the caller really wants 1309 * to destroy the buffer and backing store and (2) when the caller 1310 * wants to destroy the buffer and backing store after a write 1311 * completes. 1312 */ 1313 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1314 bundirty(bp); 1315 } 1316 1317 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1318 /* 1319 * A re-dirtied buffer is only subject to destruction 1320 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1321 */ 1322 /* leave buffer intact */ 1323 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1324 (bp->b_bufsize <= 0)) { 1325 /* 1326 * Either a failed read or we were asked to free or not 1327 * cache the buffer. This path is reached with B_DELWRI 1328 * set only if B_INVAL is already set. B_NOCACHE governs 1329 * backing store destruction. 1330 * 1331 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1332 * buffer cannot be immediately freed. 1333 */ 1334 bp->b_flags |= B_INVAL; 1335 if (LIST_FIRST(&bp->b_dep) != NULL) 1336 buf_deallocate(bp); 1337 if (bp->b_flags & B_DELWRI) { 1338 atomic_add_long(&dirtybufcount, -1); 1339 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1340 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1341 if (bp->b_flags & B_HEAVY) { 1342 atomic_add_long(&dirtybufcounthw, -1); 1343 atomic_add_long(&dirtybufspacehw, 1344 -bp->b_bufsize); 1345 } 1346 bd_signal(bp->b_bufsize); 1347 } 1348 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1349 } 1350 1351 /* 1352 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set, 1353 * or if b_refs is non-zero. 1354 * 1355 * If vfs_vmio_release() is called with either bit set, the 1356 * underlying pages may wind up getting freed causing a previous 1357 * write (bdwrite()) to get 'lost' because pages associated with 1358 * a B_DELWRI bp are marked clean. Pages associated with a 1359 * B_LOCKED buffer may be mapped by the filesystem. 1360 * 1361 * If we want to release the buffer ourselves (rather then the 1362 * originator asking us to release it), give the originator a 1363 * chance to countermand the release by setting B_LOCKED. 1364 * 1365 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1366 * if B_DELWRI is set. 1367 * 1368 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1369 * on pages to return pages to the VM page queues. 1370 */ 1371 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) { 1372 bp->b_flags &= ~B_RELBUF; 1373 } else if (vm_page_count_min(0)) { 1374 if (LIST_FIRST(&bp->b_dep) != NULL) 1375 buf_deallocate(bp); /* can set B_LOCKED */ 1376 if (bp->b_flags & (B_DELWRI | B_LOCKED)) 1377 bp->b_flags &= ~B_RELBUF; 1378 else 1379 bp->b_flags |= B_RELBUF; 1380 } 1381 1382 /* 1383 * Make sure b_cmd is clear. It may have already been cleared by 1384 * biodone(). 1385 * 1386 * At this point destroying the buffer is governed by the B_INVAL 1387 * or B_RELBUF flags. 1388 */ 1389 bp->b_cmd = BUF_CMD_DONE; 1390 dsched_buf_exit(bp); 1391 1392 /* 1393 * VMIO buffer rundown. Make sure the VM page array is restored 1394 * after an I/O may have replaces some of the pages with bogus pages 1395 * in order to not destroy dirty pages in a fill-in read. 1396 * 1397 * Note that due to the code above, if a buffer is marked B_DELWRI 1398 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1399 * B_INVAL may still be set, however. 1400 * 1401 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1402 * but not the backing store. B_NOCACHE will destroy the backing 1403 * store. 1404 * 1405 * Note that dirty NFS buffers contain byte-granular write ranges 1406 * and should not be destroyed w/ B_INVAL even if the backing store 1407 * is left intact. 1408 */ 1409 if (bp->b_flags & B_VMIO) { 1410 /* 1411 * Rundown for VMIO buffers which are not dirty NFS buffers. 1412 */ 1413 int i, j, resid; 1414 vm_page_t m; 1415 off_t foff; 1416 vm_pindex_t poff; 1417 vm_object_t obj; 1418 struct vnode *vp; 1419 1420 vp = bp->b_vp; 1421 1422 /* 1423 * Get the base offset and length of the buffer. Note that 1424 * in the VMIO case if the buffer block size is not 1425 * page-aligned then b_data pointer may not be page-aligned. 1426 * But our b_xio.xio_pages array *IS* page aligned. 1427 * 1428 * block sizes less then DEV_BSIZE (usually 512) are not 1429 * supported due to the page granularity bits (m->valid, 1430 * m->dirty, etc...). 1431 * 1432 * See man buf(9) for more information 1433 */ 1434 1435 resid = bp->b_bufsize; 1436 foff = bp->b_loffset; 1437 1438 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1439 m = bp->b_xio.xio_pages[i]; 1440 1441 /* 1442 * If we hit a bogus page, fixup *all* of them 1443 * now. Note that we left these pages wired 1444 * when we removed them so they had better exist, 1445 * and they cannot be ripped out from under us so 1446 * no critical section protection is necessary. 1447 */ 1448 if (m == bogus_page) { 1449 obj = vp->v_object; 1450 poff = OFF_TO_IDX(bp->b_loffset); 1451 1452 vm_object_hold(obj); 1453 for (j = i; j < bp->b_xio.xio_npages; j++) { 1454 vm_page_t mtmp; 1455 1456 mtmp = bp->b_xio.xio_pages[j]; 1457 if (mtmp == bogus_page) { 1458 if ((bp->b_flags & B_HASBOGUS) == 0) 1459 panic("brelse: bp %p corrupt bogus", bp); 1460 mtmp = vm_page_lookup(obj, poff + j); 1461 if (!mtmp) 1462 panic("brelse: bp %p page %d missing", bp, j); 1463 bp->b_xio.xio_pages[j] = mtmp; 1464 } 1465 } 1466 vm_object_drop(obj); 1467 1468 if ((bp->b_flags & B_HASBOGUS) || (bp->b_flags & B_INVAL) == 0) { 1469 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1470 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1471 bp->b_flags &= ~B_HASBOGUS; 1472 } 1473 m = bp->b_xio.xio_pages[i]; 1474 } 1475 1476 /* 1477 * Invalidate the backing store if B_NOCACHE is set 1478 * (e.g. used with vinvalbuf()). If this is NFS 1479 * we impose a requirement that the block size be 1480 * a multiple of PAGE_SIZE and create a temporary 1481 * hack to basically invalidate the whole page. The 1482 * problem is that NFS uses really odd buffer sizes 1483 * especially when tracking piecemeal writes and 1484 * it also vinvalbuf()'s a lot, which would result 1485 * in only partial page validation and invalidation 1486 * here. If the file page is mmap()'d, however, 1487 * all the valid bits get set so after we invalidate 1488 * here we would end up with weird m->valid values 1489 * like 0xfc. nfs_getpages() can't handle this so 1490 * we clear all the valid bits for the NFS case 1491 * instead of just some of them. 1492 * 1493 * The real bug is the VM system having to set m->valid 1494 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1495 * itself is an artifact of the whole 512-byte 1496 * granular mess that exists to support odd block 1497 * sizes and UFS meta-data block sizes (e.g. 6144). 1498 * A complete rewrite is required. 1499 * 1500 * XXX 1501 */ 1502 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1503 int poffset = foff & PAGE_MASK; 1504 int presid; 1505 1506 presid = PAGE_SIZE - poffset; 1507 if (bp->b_vp->v_tag == VT_NFS && 1508 bp->b_vp->v_type == VREG) { 1509 ; /* entire page */ 1510 } else if (presid > resid) { 1511 presid = resid; 1512 } 1513 KASSERT(presid >= 0, ("brelse: extra page")); 1514 vm_page_set_invalid(m, poffset, presid); 1515 1516 /* 1517 * Also make sure any swap cache is removed 1518 * as it is now stale (HAMMER in particular 1519 * uses B_NOCACHE to deal with buffer 1520 * aliasing). 1521 */ 1522 swap_pager_unswapped(m); 1523 } 1524 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1525 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1526 } 1527 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1528 vfs_vmio_release(bp); 1529 } else { 1530 /* 1531 * Rundown for non-VMIO buffers. 1532 */ 1533 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1534 if (bp->b_bufsize) 1535 allocbuf(bp, 0); 1536 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1537 if (bp->b_vp) 1538 brelvp(bp); 1539 } 1540 } 1541 1542 if (bp->b_qindex != BQUEUE_NONE) 1543 panic("brelse: free buffer onto another queue???"); 1544 if (BUF_REFCNTNB(bp) > 1) { 1545 /* Temporary panic to verify exclusive locking */ 1546 /* This panic goes away when we allow shared refs */ 1547 panic("brelse: multiple refs"); 1548 /* NOT REACHED */ 1549 return; 1550 } 1551 1552 /* 1553 * Figure out the correct queue to place the cleaned up buffer on. 1554 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1555 * disassociated from their vnode. 1556 * 1557 * Return the buffer to its original pcpu area 1558 */ 1559 pcpu = &bufpcpu[bp->b_qcpu]; 1560 spin_lock(&pcpu->spin); 1561 1562 if (bp->b_flags & B_LOCKED) { 1563 /* 1564 * Buffers that are locked are placed in the locked queue 1565 * immediately, regardless of their state. 1566 */ 1567 bp->b_qindex = BQUEUE_LOCKED; 1568 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1569 bp, b_freelist); 1570 } else if (bp->b_bufsize == 0) { 1571 /* 1572 * Buffers with no memory. Due to conditionals near the top 1573 * of brelse() such buffers should probably already be 1574 * marked B_INVAL and disassociated from their vnode. 1575 */ 1576 bp->b_flags |= B_INVAL; 1577 KASSERT(bp->b_vp == NULL, 1578 ("bp1 %p flags %08x/%08x vnode %p " 1579 "unexpectededly still associated!", 1580 bp, saved_flags, bp->b_flags, bp->b_vp)); 1581 KKASSERT((bp->b_flags & B_HASHED) == 0); 1582 bp->b_qindex = BQUEUE_EMPTY; 1583 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1584 bp, b_freelist); 1585 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1586 /* 1587 * Buffers with junk contents. Again these buffers had better 1588 * already be disassociated from their vnode. 1589 */ 1590 KASSERT(bp->b_vp == NULL, 1591 ("bp2 %p flags %08x/%08x vnode %p unexpectededly " 1592 "still associated!", 1593 bp, saved_flags, bp->b_flags, bp->b_vp)); 1594 KKASSERT((bp->b_flags & B_HASHED) == 0); 1595 bp->b_flags |= B_INVAL; 1596 bp->b_qindex = BQUEUE_CLEAN; 1597 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1598 bp, b_freelist); 1599 } else { 1600 /* 1601 * Remaining buffers. These buffers are still associated with 1602 * their vnode. 1603 */ 1604 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1605 case B_DELWRI: 1606 bp->b_qindex = BQUEUE_DIRTY; 1607 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1608 bp, b_freelist); 1609 break; 1610 case B_DELWRI | B_HEAVY: 1611 bp->b_qindex = BQUEUE_DIRTY_HW; 1612 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1613 bp, b_freelist); 1614 break; 1615 default: 1616 /* 1617 * NOTE: Buffers are always placed at the end of the 1618 * queue. If B_AGE is not set the buffer will cycle 1619 * through the queue twice. 1620 */ 1621 bp->b_qindex = BQUEUE_CLEAN; 1622 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1623 bp, b_freelist); 1624 break; 1625 } 1626 } 1627 spin_unlock(&pcpu->spin); 1628 1629 /* 1630 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1631 * on the correct queue but we have not yet unlocked it. 1632 */ 1633 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1634 bundirty(bp); 1635 1636 /* 1637 * The bp is on an appropriate queue unless locked. If it is not 1638 * locked or dirty we can wakeup threads waiting for buffer space. 1639 * 1640 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1641 * if B_INVAL is set ). 1642 */ 1643 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1644 bufcountwakeup(); 1645 1646 /* 1647 * Something we can maybe free or reuse 1648 */ 1649 if (bp->b_bufsize || bp->b_kvasize) 1650 bufspacewakeup(); 1651 1652 /* 1653 * Clean up temporary flags and unlock the buffer. 1654 */ 1655 bp->b_flags &= ~(B_NOCACHE | B_RELBUF | B_DIRECT); 1656 BUF_UNLOCK(bp); 1657 } 1658 1659 /* 1660 * bqrelse: 1661 * 1662 * Release a buffer back to the appropriate queue but do not try to free 1663 * it. The buffer is expected to be used again soon. 1664 * 1665 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1666 * biodone() to requeue an async I/O on completion. It is also used when 1667 * known good buffers need to be requeued but we think we may need the data 1668 * again soon. 1669 * 1670 * XXX we should be able to leave the B_RELBUF hint set on completion. 1671 */ 1672 void 1673 bqrelse(struct buf *bp) 1674 { 1675 struct bufpcpu *pcpu; 1676 1677 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1678 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1679 1680 if (bp->b_qindex != BQUEUE_NONE) 1681 panic("bqrelse: free buffer onto another queue???"); 1682 if (BUF_REFCNTNB(bp) > 1) { 1683 /* do not release to free list */ 1684 panic("bqrelse: multiple refs"); 1685 return; 1686 } 1687 1688 buf_act_advance(bp); 1689 1690 pcpu = &bufpcpu[bp->b_qcpu]; 1691 spin_lock(&pcpu->spin); 1692 1693 if (bp->b_flags & B_LOCKED) { 1694 /* 1695 * Locked buffers are released to the locked queue. However, 1696 * if the buffer is dirty it will first go into the dirty 1697 * queue and later on after the I/O completes successfully it 1698 * will be released to the locked queue. 1699 */ 1700 bp->b_qindex = BQUEUE_LOCKED; 1701 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1702 bp, b_freelist); 1703 } else if (bp->b_flags & B_DELWRI) { 1704 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1705 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1706 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1707 bp, b_freelist); 1708 } else if (vm_page_count_min(0)) { 1709 /* 1710 * We are too low on memory, we have to try to free the 1711 * buffer (most importantly: the wired pages making up its 1712 * backing store) *now*. 1713 */ 1714 spin_unlock(&pcpu->spin); 1715 brelse(bp); 1716 return; 1717 } else { 1718 bp->b_qindex = BQUEUE_CLEAN; 1719 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1720 bp, b_freelist); 1721 } 1722 spin_unlock(&pcpu->spin); 1723 1724 /* 1725 * We have now placed the buffer on the proper queue, but have yet 1726 * to unlock it. 1727 */ 1728 if ((bp->b_flags & B_LOCKED) == 0 && 1729 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1730 bufcountwakeup(); 1731 } 1732 1733 /* 1734 * Something we can maybe free or reuse. 1735 */ 1736 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1737 bufspacewakeup(); 1738 1739 /* 1740 * Final cleanup and unlock. Clear bits that are only used while a 1741 * buffer is actively locked. 1742 */ 1743 bp->b_flags &= ~(B_NOCACHE | B_RELBUF); 1744 dsched_buf_exit(bp); 1745 BUF_UNLOCK(bp); 1746 } 1747 1748 /* 1749 * Hold a buffer, preventing it from being reused. This will prevent 1750 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL 1751 * operations. If a B_INVAL operation occurs the buffer will remain held 1752 * but the underlying pages may get ripped out. 1753 * 1754 * These functions are typically used in VOP_READ/VOP_WRITE functions 1755 * to hold a buffer during a copyin or copyout, preventing deadlocks 1756 * or recursive lock panics when read()/write() is used over mmap()'d 1757 * space. 1758 * 1759 * NOTE: bqhold() requires that the buffer be locked at the time of the 1760 * hold. bqdrop() has no requirements other than the buffer having 1761 * previously been held. 1762 */ 1763 void 1764 bqhold(struct buf *bp) 1765 { 1766 atomic_add_int(&bp->b_refs, 1); 1767 } 1768 1769 void 1770 bqdrop(struct buf *bp) 1771 { 1772 KKASSERT(bp->b_refs > 0); 1773 atomic_add_int(&bp->b_refs, -1); 1774 } 1775 1776 /* 1777 * Return backing pages held by the buffer 'bp' back to the VM system. 1778 * This routine is called when the bp is invalidated, released, or 1779 * reused. 1780 * 1781 * The KVA mapping (b_data) for the underlying pages is removed by 1782 * this function. 1783 * 1784 * WARNING! This routine is integral to the low memory critical path 1785 * when a buffer is B_RELBUF'd. If the system has a severe page 1786 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE 1787 * queues so they can be reused in the current pageout daemon 1788 * pass. 1789 */ 1790 static void 1791 vfs_vmio_release(struct buf *bp) 1792 { 1793 int i; 1794 vm_page_t m; 1795 1796 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1797 m = bp->b_xio.xio_pages[i]; 1798 bp->b_xio.xio_pages[i] = NULL; 1799 1800 /* 1801 * We need to own the page in order to safely unwire it. 1802 */ 1803 vm_page_busy_wait(m, FALSE, "vmiopg"); 1804 1805 /* 1806 * The VFS is telling us this is not a meta-data buffer 1807 * even if it is backed by a block device. 1808 */ 1809 if (bp->b_flags & B_NOTMETA) 1810 vm_page_flag_set(m, PG_NOTMETA); 1811 1812 /* 1813 * This is a very important bit of code. We try to track 1814 * VM page use whether the pages are wired into the buffer 1815 * cache or not. While wired into the buffer cache the 1816 * bp tracks the act_count. 1817 * 1818 * We can choose to place unwired pages on the inactive 1819 * queue (0) or active queue (1). If we place too many 1820 * on the active queue the queue will cycle the act_count 1821 * on pages we'd like to keep, just from single-use pages 1822 * (such as when doing a tar-up or file scan). 1823 */ 1824 if (bp->b_act_count < vm_cycle_point) 1825 vm_page_unwire(m, 0); 1826 else 1827 vm_page_unwire(m, 1); 1828 1829 /* 1830 * If the wire_count has dropped to 0 we may need to take 1831 * further action before unbusying the page. 1832 * 1833 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us. 1834 */ 1835 if (m->wire_count == 0) { 1836 if (bp->b_flags & B_DIRECT) { 1837 /* 1838 * Attempt to free the page if B_DIRECT is 1839 * set, the caller does not desire the page 1840 * to be cached. 1841 */ 1842 vm_page_wakeup(m); 1843 vm_page_try_to_free(m); 1844 } else if ((bp->b_flags & B_NOTMETA) || 1845 vm_page_count_min(0)) { 1846 /* 1847 * Attempt to move the page to PQ_CACHE 1848 * if B_NOTMETA is set. This flag is set 1849 * by HAMMER to remove one of the two pages 1850 * present when double buffering is enabled. 1851 * 1852 * Attempt to move the page to PQ_CACHE 1853 * If we have a severe page deficit. This 1854 * will cause buffer cache operations related 1855 * to pageouts to recycle the related pages 1856 * in order to avoid a low memory deadlock. 1857 */ 1858 m->act_count = bp->b_act_count; 1859 vm_page_try_to_cache(m); 1860 } else { 1861 /* 1862 * Nominal case, leave the page on the 1863 * queue the original unwiring placed it on 1864 * (active or inactive). 1865 */ 1866 m->act_count = bp->b_act_count; 1867 vm_page_wakeup(m); 1868 } 1869 } else { 1870 vm_page_wakeup(m); 1871 } 1872 } 1873 1874 /* 1875 * Zero out the pmap pte's for the mapping, but don't bother 1876 * invalidating the TLB. The range will be properly invalidating 1877 * when new pages are entered into the mapping. 1878 * 1879 * This in particular reduces tmpfs tear-down overhead and reduces 1880 * buffer cache re-use overhead (one invalidation sequence instead 1881 * of two per re-use). 1882 */ 1883 pmap_qremove_noinval(trunc_page((vm_offset_t) bp->b_data), 1884 bp->b_xio.xio_npages); 1885 if (bp->b_bufsize) { 1886 atomic_add_long(&bufspace, -bp->b_bufsize); 1887 bp->b_bufsize = 0; 1888 bufspacewakeup(); 1889 } 1890 bp->b_xio.xio_npages = 0; 1891 bp->b_flags &= ~B_VMIO; 1892 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1893 if (bp->b_vp) 1894 brelvp(bp); 1895 } 1896 1897 /* 1898 * Find and initialize a new buffer header, freeing up existing buffers 1899 * in the bufqueues as necessary. The new buffer is returned locked. 1900 * 1901 * If repurpose is non-NULL getnewbuf() is allowed to re-purpose an existing 1902 * buffer. The buffer will be disassociated, its page and page mappings 1903 * left intact, and returned with *repurpose set to 1. Else *repurpose is set 1904 * to 0. If 1, the caller must repurpose the underlying VM pages. 1905 * 1906 * If repurpose is NULL getnewbuf() is not allowed to re-purpose an 1907 * existing buffer. That is, it must completely initialize the returned 1908 * buffer. 1909 * 1910 * Important: B_INVAL is not set. If the caller wishes to throw the 1911 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1912 * 1913 * We block if: 1914 * We have insufficient buffer headers 1915 * We have insufficient buffer space 1916 * 1917 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1918 * Instead we ask the buf daemon to do it for us. We attempt to 1919 * avoid piecemeal wakeups of the pageout daemon. 1920 */ 1921 struct buf * 1922 getnewbuf(int blkflags, int slptimeo, int size, int maxsize, 1923 struct vm_object **repurposep) 1924 { 1925 struct bufpcpu *pcpu; 1926 struct buf *bp; 1927 struct buf *nbp; 1928 int nqindex; 1929 int nqcpu; 1930 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1931 int maxloops = 200000; 1932 int restart_reason = 0; 1933 struct buf *restart_bp = NULL; 1934 static char flushingbufs[MAXCPU]; 1935 char *flushingp; 1936 1937 /* 1938 * We can't afford to block since we might be holding a vnode lock, 1939 * which may prevent system daemons from running. We deal with 1940 * low-memory situations by proactively returning memory and running 1941 * async I/O rather then sync I/O. 1942 */ 1943 1944 ++getnewbufcalls; 1945 nqcpu = mycpu->gd_cpuid; 1946 flushingp = &flushingbufs[nqcpu]; 1947 restart: 1948 if (bufspace < lobufspace) 1949 *flushingp = 0; 1950 1951 if (debug_bufbio && --maxloops == 0) 1952 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)", 1953 mycpu->gd_cpuid, restart_reason, restart_bp); 1954 1955 /* 1956 * Setup for scan. If we do not have enough free buffers, 1957 * we setup a degenerate case that immediately fails. Note 1958 * that if we are specially marked process, we are allowed to 1959 * dip into our reserves. 1960 * 1961 * The scanning sequence is nominally: EMPTY->CLEAN 1962 */ 1963 pcpu = &bufpcpu[nqcpu]; 1964 spin_lock(&pcpu->spin); 1965 1966 /* 1967 * Determine if repurposing should be disallowed. Generally speaking 1968 * do not repurpose buffers if the buffer cache hasn't capped. Also 1969 * control repurposing based on buffer-cache -> main-memory bandwidth. 1970 * That is, we want to recycle buffers normally up until the buffer 1971 * cache bandwidth (new-buffer bw) exceeds bufcache_bw. 1972 * 1973 * (This is heuristical, SMP collisions are ok) 1974 */ 1975 if (repurposep) { 1976 int delta = ticks - bufcache_bw_ticks; 1977 if (delta < 0 || delta >= hz) { 1978 atomic_swap_long(&bufcache_bw_accum, 0); 1979 atomic_swap_int(&bufcache_bw_ticks, ticks); 1980 } 1981 atomic_add_long(&bufcache_bw_accum, size); 1982 if (bufspace < lobufspace) { 1983 repurposep = NULL; 1984 } else if (bufcache_bw_accum < bufcache_bw) { 1985 repurposep = NULL; 1986 } 1987 } 1988 1989 /* 1990 * Prime the scan for this cpu. Locate the first buffer to 1991 * check. If we are flushing buffers we must skip the 1992 * EMPTY queue. 1993 */ 1994 nqindex = BQUEUE_EMPTY; 1995 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]); 1996 if (nbp == NULL || *flushingp || repurposep) { 1997 nqindex = BQUEUE_CLEAN; 1998 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]); 1999 } 2000 2001 /* 2002 * Run scan, possibly freeing data and/or kva mappings on the fly, 2003 * depending. 2004 * 2005 * WARNING! spin is held! 2006 */ 2007 while ((bp = nbp) != NULL) { 2008 int qindex = nqindex; 2009 2010 nbp = TAILQ_NEXT(bp, b_freelist); 2011 2012 /* 2013 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2014 * cycles through the queue twice before being selected. 2015 */ 2016 if (qindex == BQUEUE_CLEAN && 2017 (bp->b_flags & B_AGE) == 0 && nbp) { 2018 bp->b_flags |= B_AGE; 2019 TAILQ_REMOVE(&pcpu->bufqueues[qindex], 2020 bp, b_freelist); 2021 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex], 2022 bp, b_freelist); 2023 continue; 2024 } 2025 2026 /* 2027 * Calculate next bp ( we can only use it if we do not block 2028 * or do other fancy things ). 2029 */ 2030 if (nbp == NULL) { 2031 switch(qindex) { 2032 case BQUEUE_EMPTY: 2033 nqindex = BQUEUE_CLEAN; 2034 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]))) 2035 break; 2036 /* fall through */ 2037 case BQUEUE_CLEAN: 2038 /* 2039 * nbp is NULL. 2040 */ 2041 break; 2042 } 2043 } 2044 2045 /* 2046 * Sanity Checks 2047 */ 2048 KASSERT(bp->b_qindex == qindex, 2049 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2050 2051 /* 2052 * Note: we no longer distinguish between VMIO and non-VMIO 2053 * buffers. 2054 */ 2055 KASSERT((bp->b_flags & B_DELWRI) == 0, 2056 ("delwri buffer %p found in queue %d", bp, qindex)); 2057 2058 /* 2059 * Do not try to reuse a buffer with a non-zero b_refs. 2060 * This is an unsynchronized test. A synchronized test 2061 * is also performed after we lock the buffer. 2062 */ 2063 if (bp->b_refs) 2064 continue; 2065 2066 /* 2067 * Start freeing the bp. This is somewhat involved. nbp 2068 * remains valid only for BQUEUE_EMPTY bp's. Buffers 2069 * on the clean list must be disassociated from their 2070 * current vnode. Buffers on the empty lists have 2071 * already been disassociated. 2072 * 2073 * b_refs is checked after locking along with queue changes. 2074 * We must check here to deal with zero->nonzero transitions 2075 * made by the owner of the buffer lock, which is used by 2076 * VFS's to hold the buffer while issuing an unlocked 2077 * uiomove()s. We cannot invalidate the buffer's pages 2078 * for this case. Once we successfully lock a buffer the 2079 * only 0->1 transitions of b_refs will occur via findblk(). 2080 * 2081 * We must also check for queue changes after successful 2082 * locking as the current lock holder may dispose of the 2083 * buffer and change its queue. 2084 */ 2085 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2086 spin_unlock(&pcpu->spin); 2087 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2088 restart_reason = 1; 2089 restart_bp = bp; 2090 goto restart; 2091 } 2092 if (bp->b_qindex != qindex || bp->b_refs) { 2093 spin_unlock(&pcpu->spin); 2094 BUF_UNLOCK(bp); 2095 restart_reason = 2; 2096 restart_bp = bp; 2097 goto restart; 2098 } 2099 bremfree_locked(bp); 2100 spin_unlock(&pcpu->spin); 2101 2102 /* 2103 * Dependancies must be handled before we disassociate the 2104 * vnode. 2105 * 2106 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2107 * be immediately disassociated. HAMMER then becomes 2108 * responsible for releasing the buffer. 2109 * 2110 * NOTE: spin is UNLOCKED now. 2111 */ 2112 if (LIST_FIRST(&bp->b_dep) != NULL) { 2113 buf_deallocate(bp); 2114 if (bp->b_flags & B_LOCKED) { 2115 bqrelse(bp); 2116 restart_reason = 3; 2117 restart_bp = bp; 2118 goto restart; 2119 } 2120 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2121 } 2122 2123 /* 2124 * CLEAN buffers have content or associations that must be 2125 * cleaned out if not repurposing. 2126 */ 2127 if (qindex == BQUEUE_CLEAN) { 2128 if (bp->b_flags & B_VMIO) { 2129 if (repurpose_enable && 2130 repurposep && bp->b_bufsize && 2131 (bp->b_flags & (B_DELWRI | B_MALLOC)) == 0) { 2132 *repurposep = bp->b_vp->v_object; 2133 vm_object_hold(*repurposep); 2134 } else { 2135 vfs_vmio_release(bp); 2136 } 2137 } 2138 if (bp->b_vp) 2139 brelvp(bp); 2140 } 2141 2142 /* 2143 * NOTE: nbp is now entirely invalid. We can only restart 2144 * the scan from this point on. 2145 * 2146 * Get the rest of the buffer freed up. b_kva* is still 2147 * valid after this operation. 2148 */ 2149 KASSERT(bp->b_vp == NULL, 2150 ("bp3 %p flags %08x vnode %p qindex %d " 2151 "unexpectededly still associated!", 2152 bp, bp->b_flags, bp->b_vp, qindex)); 2153 KKASSERT((bp->b_flags & B_HASHED) == 0); 2154 2155 if (repurposep == NULL || *repurposep == NULL) { 2156 if (bp->b_bufsize) 2157 allocbuf(bp, 0); 2158 } 2159 2160 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) { 2161 kprintf("getnewbuf: caught bug vp queue " 2162 "%p/%08x qidx %d\n", 2163 bp, bp->b_flags, qindex); 2164 brelvp(bp); 2165 } 2166 bp->b_flags = B_BNOCLIP; 2167 bp->b_cmd = BUF_CMD_DONE; 2168 bp->b_vp = NULL; 2169 bp->b_error = 0; 2170 bp->b_resid = 0; 2171 bp->b_bcount = 0; 2172 if (repurposep == NULL || *repurposep == NULL) 2173 bp->b_xio.xio_npages = 0; 2174 bp->b_dirtyoff = bp->b_dirtyend = 0; 2175 bp->b_act_count = ACT_INIT; 2176 reinitbufbio(bp); 2177 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2178 buf_dep_init(bp); 2179 if (blkflags & GETBLK_BHEAVY) 2180 bp->b_flags |= B_HEAVY; 2181 2182 if (bufspace >= hibufspace) 2183 *flushingp = 1; 2184 if (bufspace < lobufspace) 2185 *flushingp = 0; 2186 if (*flushingp) { 2187 if (repurposep && *repurposep != NULL) { 2188 bp->b_flags |= B_VMIO; 2189 vfs_vmio_release(bp); 2190 if (bp->b_bufsize) 2191 allocbuf(bp, 0); 2192 vm_object_drop(*repurposep); 2193 *repurposep = NULL; 2194 } 2195 bp->b_flags |= B_INVAL; 2196 brelse(bp); 2197 restart_reason = 5; 2198 restart_bp = bp; 2199 goto restart; 2200 } 2201 2202 /* 2203 * b_refs can transition to a non-zero value while we hold 2204 * the buffer locked due to a findblk(). Our brelvp() above 2205 * interlocked any future possible transitions due to 2206 * findblk()s. 2207 * 2208 * If we find b_refs to be non-zero we can destroy the 2209 * buffer's contents but we cannot yet reuse the buffer. 2210 */ 2211 if (bp->b_refs) { 2212 if (repurposep && *repurposep != NULL) { 2213 bp->b_flags |= B_VMIO; 2214 vfs_vmio_release(bp); 2215 if (bp->b_bufsize) 2216 allocbuf(bp, 0); 2217 vm_object_drop(*repurposep); 2218 *repurposep = NULL; 2219 } 2220 bp->b_flags |= B_INVAL; 2221 brelse(bp); 2222 restart_reason = 6; 2223 restart_bp = bp; 2224 2225 goto restart; 2226 } 2227 2228 /* 2229 * We found our buffer! 2230 */ 2231 break; 2232 } 2233 2234 /* 2235 * If we exhausted our list, iterate other cpus. If that fails, 2236 * sleep as appropriate. We may have to wakeup various daemons 2237 * and write out some dirty buffers. 2238 * 2239 * Generally we are sleeping due to insufficient buffer space. 2240 * 2241 * NOTE: spin is held if bp is NULL, else it is not held. 2242 */ 2243 if (bp == NULL) { 2244 int flags; 2245 char *waitmsg; 2246 2247 spin_unlock(&pcpu->spin); 2248 2249 nqcpu = (nqcpu + 1) % ncpus; 2250 if (nqcpu != mycpu->gd_cpuid) { 2251 restart_reason = 7; 2252 restart_bp = bp; 2253 goto restart; 2254 } 2255 2256 if (bufspace >= hibufspace) { 2257 waitmsg = "bufspc"; 2258 flags = VFS_BIO_NEED_BUFSPACE; 2259 } else { 2260 waitmsg = "newbuf"; 2261 flags = VFS_BIO_NEED_ANY; 2262 } 2263 2264 bd_speedup(); /* heeeelp */ 2265 atomic_set_int(&needsbuffer, flags); 2266 while (needsbuffer & flags) { 2267 int value; 2268 2269 tsleep_interlock(&needsbuffer, 0); 2270 value = atomic_fetchadd_int(&needsbuffer, 0); 2271 if (value & flags) { 2272 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags, 2273 waitmsg, slptimeo)) { 2274 return (NULL); 2275 } 2276 } 2277 } 2278 } else { 2279 /* 2280 * We finally have a valid bp. Reset b_data. 2281 * 2282 * (spin is not held) 2283 */ 2284 bp->b_data = bp->b_kvabase; 2285 } 2286 return(bp); 2287 } 2288 2289 /* 2290 * buf_daemon: 2291 * 2292 * Buffer flushing daemon. Buffers are normally flushed by the 2293 * update daemon but if it cannot keep up this process starts to 2294 * take the load in an attempt to prevent getnewbuf() from blocking. 2295 * 2296 * Once a flush is initiated it does not stop until the number 2297 * of buffers falls below lodirtybuffers, but we will wake up anyone 2298 * waiting at the mid-point. 2299 */ 2300 static struct kproc_desc buf_kp = { 2301 "bufdaemon", 2302 buf_daemon, 2303 &bufdaemon_td 2304 }; 2305 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2306 kproc_start, &buf_kp); 2307 2308 static struct kproc_desc bufhw_kp = { 2309 "bufdaemon_hw", 2310 buf_daemon_hw, 2311 &bufdaemonhw_td 2312 }; 2313 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2314 kproc_start, &bufhw_kp); 2315 2316 static void 2317 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long), 2318 int *bd_req) 2319 { 2320 long limit; 2321 struct buf *marker; 2322 2323 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO); 2324 marker->b_flags |= B_MARKER; 2325 marker->b_qindex = BQUEUE_NONE; 2326 marker->b_qcpu = 0; 2327 2328 /* 2329 * This process needs to be suspended prior to shutdown sync. 2330 */ 2331 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2332 td, SHUTDOWN_PRI_LAST); 2333 curthread->td_flags |= TDF_SYSTHREAD; 2334 2335 /* 2336 * This process is allowed to take the buffer cache to the limit 2337 */ 2338 for (;;) { 2339 kproc_suspend_loop(); 2340 2341 /* 2342 * Do the flush as long as the number of dirty buffers 2343 * (including those running) exceeds lodirtybufspace. 2344 * 2345 * When flushing limit running I/O to hirunningspace 2346 * Do the flush. Limit the amount of in-transit I/O we 2347 * allow to build up, otherwise we would completely saturate 2348 * the I/O system. Wakeup any waiting processes before we 2349 * normally would so they can run in parallel with our drain. 2350 * 2351 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2352 * but because we split the operation into two threads we 2353 * have to cut it in half for each thread. 2354 */ 2355 waitrunningbufspace(); 2356 limit = lodirtybufspace / 2; 2357 while (buf_limit_fn(limit)) { 2358 if (flushbufqueues(marker, queue) == 0) 2359 break; 2360 if (runningbufspace < hirunningspace) 2361 continue; 2362 waitrunningbufspace(); 2363 } 2364 2365 /* 2366 * We reached our low water mark, reset the 2367 * request and sleep until we are needed again. 2368 * The sleep is just so the suspend code works. 2369 */ 2370 tsleep_interlock(bd_req, 0); 2371 if (atomic_swap_int(bd_req, 0) == 0) 2372 tsleep(bd_req, PINTERLOCKED, "psleep", hz); 2373 } 2374 /* NOT REACHED */ 2375 /*kfree(marker, M_BIOBUF);*/ 2376 } 2377 2378 static int 2379 buf_daemon_limit(long limit) 2380 { 2381 return (runningbufspace + dirtykvaspace > limit || 2382 dirtybufcount - dirtybufcounthw >= nbuf / 2); 2383 } 2384 2385 static int 2386 buf_daemon_hw_limit(long limit) 2387 { 2388 return (runningbufspace + dirtykvaspace > limit || 2389 dirtybufcounthw >= nbuf / 2); 2390 } 2391 2392 static void 2393 buf_daemon(void) 2394 { 2395 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit, 2396 &bd_request); 2397 } 2398 2399 static void 2400 buf_daemon_hw(void) 2401 { 2402 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit, 2403 &bd_request_hw); 2404 } 2405 2406 /* 2407 * Flush up to (flushperqueue) buffers in the dirty queue. Each cpu has a 2408 * localized version of the queue. Each call made to this function iterates 2409 * to another cpu. It is desireable to flush several buffers from the same 2410 * cpu's queue at once, as these are likely going to be linear. 2411 * 2412 * We must be careful to free up B_INVAL buffers instead of write them, which 2413 * NFS is particularly sensitive to. 2414 * 2415 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate that we 2416 * really want to try to get the buffer out and reuse it due to the write 2417 * load on the machine. 2418 * 2419 * We must lock the buffer in order to check its validity before we can mess 2420 * with its contents. spin isn't enough. 2421 */ 2422 static int 2423 flushbufqueues(struct buf *marker, bufq_type_t q) 2424 { 2425 struct bufpcpu *pcpu; 2426 struct buf *bp; 2427 int r = 0; 2428 u_int loops = flushperqueue; 2429 int lcpu = marker->b_qcpu; 2430 2431 KKASSERT(marker->b_qindex == BQUEUE_NONE); 2432 KKASSERT(marker->b_flags & B_MARKER); 2433 2434 again: 2435 /* 2436 * Spinlock needed to perform operations on the queue and may be 2437 * held through a non-blocking BUF_LOCK(), but cannot be held when 2438 * BUF_UNLOCK()ing or through any other major operation. 2439 */ 2440 pcpu = &bufpcpu[marker->b_qcpu]; 2441 spin_lock(&pcpu->spin); 2442 marker->b_qindex = q; 2443 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist); 2444 bp = marker; 2445 2446 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) { 2447 /* 2448 * NOTE: spinlock is always held at the top of the loop 2449 */ 2450 if (bp->b_flags & B_MARKER) 2451 continue; 2452 if ((bp->b_flags & B_DELWRI) == 0) { 2453 kprintf("Unexpected clean buffer %p\n", bp); 2454 continue; 2455 } 2456 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 2457 continue; 2458 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q); 2459 2460 /* 2461 * Once the buffer is locked we will have no choice but to 2462 * unlock the spinlock around a later BUF_UNLOCK and re-set 2463 * bp = marker when looping. Move the marker now to make 2464 * things easier. 2465 */ 2466 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2467 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist); 2468 2469 /* 2470 * Must recheck B_DELWRI after successfully locking 2471 * the buffer. 2472 */ 2473 if ((bp->b_flags & B_DELWRI) == 0) { 2474 spin_unlock(&pcpu->spin); 2475 BUF_UNLOCK(bp); 2476 spin_lock(&pcpu->spin); 2477 bp = marker; 2478 continue; 2479 } 2480 2481 /* 2482 * Remove the buffer from its queue. We still own the 2483 * spinlock here. 2484 */ 2485 _bremfree(bp); 2486 2487 /* 2488 * Disposing of an invalid buffer counts as a flush op 2489 */ 2490 if (bp->b_flags & B_INVAL) { 2491 spin_unlock(&pcpu->spin); 2492 brelse(bp); 2493 goto doloop; 2494 } 2495 2496 /* 2497 * Release the spinlock for the more complex ops we 2498 * are now going to do. 2499 */ 2500 spin_unlock(&pcpu->spin); 2501 lwkt_yield(); 2502 2503 /* 2504 * This is a bit messy 2505 */ 2506 if (LIST_FIRST(&bp->b_dep) != NULL && 2507 (bp->b_flags & B_DEFERRED) == 0 && 2508 buf_countdeps(bp, 0)) { 2509 spin_lock(&pcpu->spin); 2510 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist); 2511 bp->b_qindex = q; 2512 bp->b_flags |= B_DEFERRED; 2513 spin_unlock(&pcpu->spin); 2514 BUF_UNLOCK(bp); 2515 spin_lock(&pcpu->spin); 2516 bp = marker; 2517 continue; 2518 } 2519 2520 /* 2521 * spinlock not held here. 2522 * 2523 * If the buffer has a dependancy, buf_checkwrite() must 2524 * also return 0 for us to be able to initate the write. 2525 * 2526 * If the buffer is flagged B_ERROR it may be requeued 2527 * over and over again, we try to avoid a live lock. 2528 */ 2529 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2530 brelse(bp); 2531 } else if (bp->b_flags & B_ERROR) { 2532 tsleep(bp, 0, "bioer", 1); 2533 bp->b_flags &= ~B_AGE; 2534 cluster_awrite(bp); 2535 } else { 2536 bp->b_flags |= B_AGE; 2537 cluster_awrite(bp); 2538 } 2539 /* bp invalid but needs to be NULL-tested if we break out */ 2540 doloop: 2541 spin_lock(&pcpu->spin); 2542 ++r; 2543 if (--loops == 0) 2544 break; 2545 bp = marker; 2546 } 2547 /* bp is invalid here but can be NULL-tested to advance */ 2548 2549 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2550 marker->b_qindex = BQUEUE_NONE; 2551 spin_unlock(&pcpu->spin); 2552 2553 /* 2554 * Advance the marker to be fair. 2555 */ 2556 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus; 2557 if (bp == NULL) { 2558 if (marker->b_qcpu != lcpu) 2559 goto again; 2560 } 2561 2562 return (r); 2563 } 2564 2565 /* 2566 * inmem: 2567 * 2568 * Returns true if no I/O is needed to access the associated VM object. 2569 * This is like findblk except it also hunts around in the VM system for 2570 * the data. 2571 * 2572 * Note that we ignore vm_page_free() races from interrupts against our 2573 * lookup, since if the caller is not protected our return value will not 2574 * be any more valid then otherwise once we exit the critical section. 2575 */ 2576 int 2577 inmem(struct vnode *vp, off_t loffset) 2578 { 2579 vm_object_t obj; 2580 vm_offset_t toff, tinc, size; 2581 vm_page_t m; 2582 int res = 1; 2583 2584 if (findblk(vp, loffset, FINDBLK_TEST)) 2585 return 1; 2586 if (vp->v_mount == NULL) 2587 return 0; 2588 if ((obj = vp->v_object) == NULL) 2589 return 0; 2590 2591 size = PAGE_SIZE; 2592 if (size > vp->v_mount->mnt_stat.f_iosize) 2593 size = vp->v_mount->mnt_stat.f_iosize; 2594 2595 vm_object_hold(obj); 2596 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2597 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2598 if (m == NULL) { 2599 res = 0; 2600 break; 2601 } 2602 tinc = size; 2603 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2604 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2605 if (vm_page_is_valid(m, 2606 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2607 res = 0; 2608 break; 2609 } 2610 } 2611 vm_object_drop(obj); 2612 return (res); 2613 } 2614 2615 /* 2616 * findblk: 2617 * 2618 * Locate and return the specified buffer. Unless flagged otherwise, 2619 * a locked buffer will be returned if it exists or NULL if it does not. 2620 * 2621 * findblk()'d buffers are still on the bufqueues and if you intend 2622 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2623 * and possibly do other stuff to it. 2624 * 2625 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2626 * for locking the buffer and ensuring that it remains 2627 * the desired buffer after locking. 2628 * 2629 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2630 * to acquire the lock we return NULL, even if the 2631 * buffer exists. 2632 * 2633 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2634 * reuse by getnewbuf() but does not prevent 2635 * disassociation (B_INVAL). Used to avoid deadlocks 2636 * against random (vp,loffset)s due to reassignment. 2637 * 2638 * (0) - Lock the buffer blocking. 2639 */ 2640 struct buf * 2641 findblk(struct vnode *vp, off_t loffset, int flags) 2642 { 2643 struct buf *bp; 2644 int lkflags; 2645 2646 lkflags = LK_EXCLUSIVE; 2647 if (flags & FINDBLK_NBLOCK) 2648 lkflags |= LK_NOWAIT; 2649 2650 for (;;) { 2651 /* 2652 * Lookup. Ref the buf while holding v_token to prevent 2653 * reuse (but does not prevent diassociation). 2654 */ 2655 lwkt_gettoken_shared(&vp->v_token); 2656 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2657 if (bp == NULL) { 2658 lwkt_reltoken(&vp->v_token); 2659 return(NULL); 2660 } 2661 bqhold(bp); 2662 lwkt_reltoken(&vp->v_token); 2663 2664 /* 2665 * If testing only break and return bp, do not lock. 2666 */ 2667 if (flags & FINDBLK_TEST) 2668 break; 2669 2670 /* 2671 * Lock the buffer, return an error if the lock fails. 2672 * (only FINDBLK_NBLOCK can cause the lock to fail). 2673 */ 2674 if (BUF_LOCK(bp, lkflags)) { 2675 atomic_subtract_int(&bp->b_refs, 1); 2676 /* bp = NULL; not needed */ 2677 return(NULL); 2678 } 2679 2680 /* 2681 * Revalidate the locked buf before allowing it to be 2682 * returned. 2683 */ 2684 if (bp->b_vp == vp && bp->b_loffset == loffset) 2685 break; 2686 atomic_subtract_int(&bp->b_refs, 1); 2687 BUF_UNLOCK(bp); 2688 } 2689 2690 /* 2691 * Success 2692 */ 2693 if ((flags & FINDBLK_REF) == 0) 2694 atomic_subtract_int(&bp->b_refs, 1); 2695 return(bp); 2696 } 2697 2698 /* 2699 * getcacheblk: 2700 * 2701 * Similar to getblk() except only returns the buffer if it is 2702 * B_CACHE and requires no other manipulation. Otherwise NULL 2703 * is returned. NULL is also returned if GETBLK_NOWAIT is set 2704 * and the getblk() would block. 2705 * 2706 * If B_RAM is set the buffer might be just fine, but we return 2707 * NULL anyway because we want the code to fall through to the 2708 * cluster read to issue more read-aheads. Otherwise read-ahead breaks. 2709 * 2710 * If blksize is 0 the buffer cache buffer must already be fully 2711 * cached. 2712 * 2713 * If blksize is non-zero getblk() will be used, allowing a buffer 2714 * to be reinstantiated from its VM backing store. The buffer must 2715 * still be fully cached after reinstantiation to be returned. 2716 */ 2717 struct buf * 2718 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags) 2719 { 2720 struct buf *bp; 2721 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0; 2722 2723 if (blksize) { 2724 bp = getblk(vp, loffset, blksize, blkflags, 0); 2725 if (bp) { 2726 if ((bp->b_flags & (B_INVAL | B_CACHE)) == B_CACHE) { 2727 bp->b_flags &= ~B_AGE; 2728 if (bp->b_flags & B_RAM) { 2729 bqrelse(bp); 2730 bp = NULL; 2731 } 2732 } else { 2733 brelse(bp); 2734 bp = NULL; 2735 } 2736 } 2737 } else { 2738 bp = findblk(vp, loffset, fndflags); 2739 if (bp) { 2740 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2741 B_CACHE) { 2742 bp->b_flags &= ~B_AGE; 2743 bremfree(bp); 2744 } else { 2745 BUF_UNLOCK(bp); 2746 bp = NULL; 2747 } 2748 } 2749 } 2750 return (bp); 2751 } 2752 2753 /* 2754 * getblk: 2755 * 2756 * Get a block given a specified block and offset into a file/device. 2757 * B_INVAL may or may not be set on return. The caller should clear 2758 * B_INVAL prior to initiating a READ. 2759 * 2760 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2761 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2762 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2763 * without doing any of those things the system will likely believe 2764 * the buffer to be valid (especially if it is not B_VMIO), and the 2765 * next getblk() will return the buffer with B_CACHE set. 2766 * 2767 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2768 * an existing buffer. 2769 * 2770 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2771 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2772 * and then cleared based on the backing VM. If the previous buffer is 2773 * non-0-sized but invalid, B_CACHE will be cleared. 2774 * 2775 * If getblk() must create a new buffer, the new buffer is returned with 2776 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2777 * case it is returned with B_INVAL clear and B_CACHE set based on the 2778 * backing VM. 2779 * 2780 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2781 * B_CACHE bit is clear. 2782 * 2783 * What this means, basically, is that the caller should use B_CACHE to 2784 * determine whether the buffer is fully valid or not and should clear 2785 * B_INVAL prior to issuing a read. If the caller intends to validate 2786 * the buffer by loading its data area with something, the caller needs 2787 * to clear B_INVAL. If the caller does this without issuing an I/O, 2788 * the caller should set B_CACHE ( as an optimization ), else the caller 2789 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2790 * a write attempt or if it was a successfull read. If the caller 2791 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2792 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2793 * 2794 * getblk flags: 2795 * 2796 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2797 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2798 */ 2799 struct buf * 2800 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2801 { 2802 struct buf *bp; 2803 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2804 int error; 2805 int lkflags; 2806 2807 if (size > MAXBSIZE) 2808 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2809 if (vp->v_object == NULL) 2810 panic("getblk: vnode %p has no object!", vp); 2811 2812 loop: 2813 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2814 /* 2815 * The buffer was found in the cache, but we need to lock it. 2816 * We must acquire a ref on the bp to prevent reuse, but 2817 * this will not prevent disassociation (brelvp()) so we 2818 * must recheck (vp,loffset) after acquiring the lock. 2819 * 2820 * Without the ref the buffer could potentially be reused 2821 * before we acquire the lock and create a deadlock 2822 * situation between the thread trying to reuse the buffer 2823 * and us due to the fact that we would wind up blocking 2824 * on a random (vp,loffset). 2825 */ 2826 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2827 if (blkflags & GETBLK_NOWAIT) { 2828 bqdrop(bp); 2829 return(NULL); 2830 } 2831 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2832 if (blkflags & GETBLK_PCATCH) 2833 lkflags |= LK_PCATCH; 2834 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2835 if (error) { 2836 bqdrop(bp); 2837 if (error == ENOLCK) 2838 goto loop; 2839 return (NULL); 2840 } 2841 /* buffer may have changed on us */ 2842 } 2843 bqdrop(bp); 2844 2845 /* 2846 * Once the buffer has been locked, make sure we didn't race 2847 * a buffer recyclement. Buffers that are no longer hashed 2848 * will have b_vp == NULL, so this takes care of that check 2849 * as well. 2850 */ 2851 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2852 #if 0 2853 kprintf("Warning buffer %p (vp %p loffset %lld) " 2854 "was recycled\n", 2855 bp, vp, (long long)loffset); 2856 #endif 2857 BUF_UNLOCK(bp); 2858 goto loop; 2859 } 2860 2861 /* 2862 * If SZMATCH any pre-existing buffer must be of the requested 2863 * size or NULL is returned. The caller absolutely does not 2864 * want getblk() to bwrite() the buffer on a size mismatch. 2865 */ 2866 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2867 BUF_UNLOCK(bp); 2868 return(NULL); 2869 } 2870 2871 /* 2872 * All vnode-based buffers must be backed by a VM object. 2873 */ 2874 KKASSERT(bp->b_flags & B_VMIO); 2875 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2876 bp->b_flags &= ~B_AGE; 2877 2878 /* 2879 * Make sure that B_INVAL buffers do not have a cached 2880 * block number translation. 2881 */ 2882 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2883 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2884 " did not have cleared bio_offset cache\n", 2885 bp, vp, (long long)loffset); 2886 clearbiocache(&bp->b_bio2); 2887 } 2888 2889 /* 2890 * The buffer is locked. B_CACHE is cleared if the buffer is 2891 * invalid. 2892 */ 2893 if (bp->b_flags & B_INVAL) 2894 bp->b_flags &= ~B_CACHE; 2895 bremfree(bp); 2896 2897 /* 2898 * Any size inconsistancy with a dirty buffer or a buffer 2899 * with a softupdates dependancy must be resolved. Resizing 2900 * the buffer in such circumstances can lead to problems. 2901 * 2902 * Dirty or dependant buffers are written synchronously. 2903 * Other types of buffers are simply released and 2904 * reconstituted as they may be backed by valid, dirty VM 2905 * pages (but not marked B_DELWRI). 2906 * 2907 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2908 * and may be left over from a prior truncation (and thus 2909 * no longer represent the actual EOF point), so we 2910 * definitely do not want to B_NOCACHE the backing store. 2911 */ 2912 if (size != bp->b_bcount) { 2913 if (bp->b_flags & B_DELWRI) { 2914 bp->b_flags |= B_RELBUF; 2915 bwrite(bp); 2916 } else if (LIST_FIRST(&bp->b_dep)) { 2917 bp->b_flags |= B_RELBUF; 2918 bwrite(bp); 2919 } else { 2920 bp->b_flags |= B_RELBUF; 2921 brelse(bp); 2922 } 2923 goto loop; 2924 } 2925 KKASSERT(size <= bp->b_kvasize); 2926 KASSERT(bp->b_loffset != NOOFFSET, 2927 ("getblk: no buffer offset")); 2928 2929 /* 2930 * A buffer with B_DELWRI set and B_CACHE clear must 2931 * be committed before we can return the buffer in 2932 * order to prevent the caller from issuing a read 2933 * ( due to B_CACHE not being set ) and overwriting 2934 * it. 2935 * 2936 * Most callers, including NFS and FFS, need this to 2937 * operate properly either because they assume they 2938 * can issue a read if B_CACHE is not set, or because 2939 * ( for example ) an uncached B_DELWRI might loop due 2940 * to softupdates re-dirtying the buffer. In the latter 2941 * case, B_CACHE is set after the first write completes, 2942 * preventing further loops. 2943 * 2944 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2945 * above while extending the buffer, we cannot allow the 2946 * buffer to remain with B_CACHE set after the write 2947 * completes or it will represent a corrupt state. To 2948 * deal with this we set B_NOCACHE to scrap the buffer 2949 * after the write. 2950 * 2951 * XXX Should this be B_RELBUF instead of B_NOCACHE? 2952 * I'm not even sure this state is still possible 2953 * now that getblk() writes out any dirty buffers 2954 * on size changes. 2955 * 2956 * We might be able to do something fancy, like setting 2957 * B_CACHE in bwrite() except if B_DELWRI is already set, 2958 * so the below call doesn't set B_CACHE, but that gets real 2959 * confusing. This is much easier. 2960 */ 2961 2962 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2963 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 2964 "and CACHE clear, b_flags %08x\n", 2965 bp, (uintmax_t)bp->b_loffset, bp->b_flags); 2966 bp->b_flags |= B_NOCACHE; 2967 bwrite(bp); 2968 goto loop; 2969 } 2970 } else { 2971 /* 2972 * Buffer is not in-core, create new buffer. The buffer 2973 * returned by getnewbuf() is locked. Note that the returned 2974 * buffer is also considered valid (not marked B_INVAL). 2975 * 2976 * Calculating the offset for the I/O requires figuring out 2977 * the block size. We use DEV_BSIZE for VBLK or VCHR and 2978 * the mount's f_iosize otherwise. If the vnode does not 2979 * have an associated mount we assume that the passed size is 2980 * the block size. 2981 * 2982 * Note that vn_isdisk() cannot be used here since it may 2983 * return a failure for numerous reasons. Note that the 2984 * buffer size may be larger then the block size (the caller 2985 * will use block numbers with the proper multiple). Beware 2986 * of using any v_* fields which are part of unions. In 2987 * particular, in DragonFly the mount point overloading 2988 * mechanism uses the namecache only and the underlying 2989 * directory vnode is not a special case. 2990 */ 2991 int bsize, maxsize; 2992 vm_object_t repurpose; 2993 2994 if (vp->v_type == VBLK || vp->v_type == VCHR) 2995 bsize = DEV_BSIZE; 2996 else if (vp->v_mount) 2997 bsize = vp->v_mount->mnt_stat.f_iosize; 2998 else 2999 bsize = size; 3000 3001 maxsize = size + (loffset & PAGE_MASK); 3002 maxsize = imax(maxsize, bsize); 3003 repurpose = NULL; 3004 3005 /* 3006 * Allow repurposing. The returned buffer may contain VM 3007 * pages associated with its previous incarnation. These 3008 * pages must be repurposed for the new buffer (hopefully 3009 * without disturbing the KVM mapping). 3010 * 3011 * WARNING! If repurpose != NULL on return, the buffer will 3012 * still contain some data from its prior 3013 * incarnation. We MUST properly dispose of this 3014 * data. 3015 */ 3016 bp = getnewbuf(blkflags, slptimeo, size, maxsize, &repurpose); 3017 if (bp == NULL) { 3018 if (slpflags || slptimeo) 3019 return NULL; 3020 goto loop; 3021 } 3022 3023 /* 3024 * Atomically insert the buffer into the hash, so that it can 3025 * be found by findblk(). 3026 * 3027 * If bgetvp() returns non-zero a collision occured, and the 3028 * bp will not be associated with the vnode. 3029 * 3030 * Make sure the translation layer has been cleared. 3031 */ 3032 bp->b_loffset = loffset; 3033 bp->b_bio2.bio_offset = NOOFFSET; 3034 /* bp->b_bio2.bio_next = NULL; */ 3035 3036 if (bgetvp(vp, bp, size)) { 3037 if (repurpose) { 3038 bp->b_flags |= B_VMIO; 3039 repurposebuf(bp, 0); 3040 vm_object_drop(repurpose); 3041 } 3042 bp->b_flags |= B_INVAL; 3043 brelse(bp); 3044 goto loop; 3045 } 3046 3047 /* 3048 * All vnode-based buffers must be backed by a VM object. 3049 */ 3050 KKASSERT(vp->v_object != NULL); 3051 bp->b_flags |= B_VMIO; 3052 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3053 3054 /* 3055 * If we allowed repurposing of the buffer it will contain 3056 * free-but-held vm_page's, already kmapped, that can be 3057 * repurposed. The repurposebuf() code handles reassigning 3058 * those pages to the new (object, offsets) and dealing with 3059 * the case where the pages already exist. 3060 */ 3061 if (repurpose) { 3062 repurposebuf(bp, size); 3063 vm_object_drop(repurpose); 3064 } else { 3065 allocbuf(bp, size); 3066 } 3067 } 3068 return (bp); 3069 } 3070 3071 /* 3072 * regetblk(bp) 3073 * 3074 * Reacquire a buffer that was previously released to the locked queue, 3075 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3076 * set B_LOCKED (which handles the acquisition race). 3077 * 3078 * To this end, either B_LOCKED must be set or the dependancy list must be 3079 * non-empty. 3080 */ 3081 void 3082 regetblk(struct buf *bp) 3083 { 3084 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3085 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3086 bremfree(bp); 3087 } 3088 3089 /* 3090 * geteblk: 3091 * 3092 * Get an empty, disassociated buffer of given size. The buffer is 3093 * initially set to B_INVAL. 3094 * 3095 * critical section protection is not required for the allocbuf() 3096 * call because races are impossible here. 3097 */ 3098 struct buf * 3099 geteblk(int size) 3100 { 3101 struct buf *bp; 3102 3103 while ((bp = getnewbuf(0, 0, size, MAXBSIZE, NULL)) == NULL) 3104 ; 3105 allocbuf(bp, size); 3106 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3107 3108 return (bp); 3109 } 3110 3111 /* 3112 * allocbuf: 3113 * 3114 * This code constitutes the buffer memory from either anonymous system 3115 * memory (in the case of non-VMIO operations) or from an associated 3116 * VM object (in the case of VMIO operations). This code is able to 3117 * resize a buffer up or down. 3118 * 3119 * Note that this code is tricky, and has many complications to resolve 3120 * deadlock or inconsistant data situations. Tread lightly!!! 3121 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3122 * the caller. Calling this code willy nilly can result in the loss of 3123 * data. 3124 * 3125 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3126 * B_CACHE for the non-VMIO case. 3127 * 3128 * This routine does not need to be called from a critical section but you 3129 * must own the buffer. 3130 */ 3131 void 3132 allocbuf(struct buf *bp, int size) 3133 { 3134 int newbsize, mbsize; 3135 int i; 3136 3137 if (BUF_REFCNT(bp) == 0) 3138 panic("allocbuf: buffer not busy"); 3139 3140 if (bp->b_kvasize < size) 3141 panic("allocbuf: buffer too small"); 3142 3143 if ((bp->b_flags & B_VMIO) == 0) { 3144 caddr_t origbuf; 3145 int origbufsize; 3146 /* 3147 * Just get anonymous memory from the kernel. Don't 3148 * mess with B_CACHE. 3149 */ 3150 mbsize = roundup2(size, DEV_BSIZE); 3151 if (bp->b_flags & B_MALLOC) 3152 newbsize = mbsize; 3153 else 3154 newbsize = round_page(size); 3155 3156 if (newbsize < bp->b_bufsize) { 3157 /* 3158 * Malloced buffers are not shrunk 3159 */ 3160 if (bp->b_flags & B_MALLOC) { 3161 if (newbsize) { 3162 bp->b_bcount = size; 3163 } else { 3164 kfree(bp->b_data, M_BIOBUF); 3165 if (bp->b_bufsize) { 3166 atomic_subtract_long(&bufmallocspace, bp->b_bufsize); 3167 bp->b_bufsize = 0; 3168 bufspacewakeup(); 3169 } 3170 bp->b_data = bp->b_kvabase; 3171 bp->b_bcount = 0; 3172 bp->b_flags &= ~B_MALLOC; 3173 } 3174 return; 3175 } 3176 vm_hold_free_pages( 3177 bp, 3178 (vm_offset_t) bp->b_data + newbsize, 3179 (vm_offset_t) bp->b_data + bp->b_bufsize); 3180 } else if (newbsize > bp->b_bufsize) { 3181 /* 3182 * We only use malloced memory on the first allocation. 3183 * and revert to page-allocated memory when the buffer 3184 * grows. 3185 */ 3186 if ((bufmallocspace < maxbufmallocspace) && 3187 (bp->b_bufsize == 0) && 3188 (mbsize <= PAGE_SIZE/2)) { 3189 3190 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3191 bp->b_bufsize = mbsize; 3192 bp->b_bcount = size; 3193 bp->b_flags |= B_MALLOC; 3194 atomic_add_long(&bufmallocspace, mbsize); 3195 return; 3196 } 3197 origbuf = NULL; 3198 origbufsize = 0; 3199 /* 3200 * If the buffer is growing on its other-than-first 3201 * allocation, then we revert to the page-allocation 3202 * scheme. 3203 */ 3204 if (bp->b_flags & B_MALLOC) { 3205 origbuf = bp->b_data; 3206 origbufsize = bp->b_bufsize; 3207 bp->b_data = bp->b_kvabase; 3208 if (bp->b_bufsize) { 3209 atomic_subtract_long(&bufmallocspace, 3210 bp->b_bufsize); 3211 bp->b_bufsize = 0; 3212 bufspacewakeup(); 3213 } 3214 bp->b_flags &= ~B_MALLOC; 3215 newbsize = round_page(newbsize); 3216 } 3217 vm_hold_load_pages( 3218 bp, 3219 (vm_offset_t) bp->b_data + bp->b_bufsize, 3220 (vm_offset_t) bp->b_data + newbsize); 3221 if (origbuf) { 3222 bcopy(origbuf, bp->b_data, origbufsize); 3223 kfree(origbuf, M_BIOBUF); 3224 } 3225 } 3226 } else { 3227 vm_page_t m; 3228 int desiredpages; 3229 3230 newbsize = roundup2(size, DEV_BSIZE); 3231 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3232 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3233 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3234 3235 if (bp->b_flags & B_MALLOC) 3236 panic("allocbuf: VMIO buffer can't be malloced"); 3237 /* 3238 * Set B_CACHE initially if buffer is 0 length or will become 3239 * 0-length. 3240 */ 3241 if (size == 0 || bp->b_bufsize == 0) 3242 bp->b_flags |= B_CACHE; 3243 3244 if (newbsize < bp->b_bufsize) { 3245 /* 3246 * DEV_BSIZE aligned new buffer size is less then the 3247 * DEV_BSIZE aligned existing buffer size. Figure out 3248 * if we have to remove any pages. 3249 */ 3250 if (desiredpages < bp->b_xio.xio_npages) { 3251 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3252 /* 3253 * the page is not freed here -- it 3254 * is the responsibility of 3255 * vnode_pager_setsize 3256 */ 3257 m = bp->b_xio.xio_pages[i]; 3258 KASSERT(m != bogus_page, 3259 ("allocbuf: bogus page found")); 3260 vm_page_busy_wait(m, TRUE, "biodep"); 3261 bp->b_xio.xio_pages[i] = NULL; 3262 vm_page_unwire(m, 0); 3263 vm_page_wakeup(m); 3264 } 3265 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3266 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3267 bp->b_xio.xio_npages = desiredpages; 3268 } 3269 } else if (size > bp->b_bcount) { 3270 /* 3271 * We are growing the buffer, possibly in a 3272 * byte-granular fashion. 3273 */ 3274 struct vnode *vp; 3275 vm_object_t obj; 3276 vm_offset_t toff; 3277 vm_offset_t tinc; 3278 3279 /* 3280 * Step 1, bring in the VM pages from the object, 3281 * allocating them if necessary. We must clear 3282 * B_CACHE if these pages are not valid for the 3283 * range covered by the buffer. 3284 */ 3285 vp = bp->b_vp; 3286 obj = vp->v_object; 3287 3288 vm_object_hold(obj); 3289 while (bp->b_xio.xio_npages < desiredpages) { 3290 vm_page_t m; 3291 vm_pindex_t pi; 3292 int error; 3293 3294 pi = OFF_TO_IDX(bp->b_loffset) + 3295 bp->b_xio.xio_npages; 3296 3297 /* 3298 * Blocking on m->busy might lead to a 3299 * deadlock: 3300 * 3301 * vm_fault->getpages->cluster_read->allocbuf 3302 */ 3303 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3304 &error); 3305 if (error) { 3306 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3307 continue; 3308 } 3309 if (m == NULL) { 3310 /* 3311 * note: must allocate system pages 3312 * since blocking here could intefere 3313 * with paging I/O, no matter which 3314 * process we are. 3315 */ 3316 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages); 3317 if (m) { 3318 vm_page_wire(m); 3319 vm_page_wakeup(m); 3320 bp->b_flags &= ~B_CACHE; 3321 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3322 ++bp->b_xio.xio_npages; 3323 } 3324 continue; 3325 } 3326 3327 /* 3328 * We found a page and were able to busy it. 3329 */ 3330 vm_page_wire(m); 3331 vm_page_wakeup(m); 3332 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3333 ++bp->b_xio.xio_npages; 3334 if (bp->b_act_count < m->act_count) 3335 bp->b_act_count = m->act_count; 3336 } 3337 vm_object_drop(obj); 3338 3339 /* 3340 * Step 2. We've loaded the pages into the buffer, 3341 * we have to figure out if we can still have B_CACHE 3342 * set. Note that B_CACHE is set according to the 3343 * byte-granular range ( bcount and size ), not the 3344 * aligned range ( newbsize ). 3345 * 3346 * The VM test is against m->valid, which is DEV_BSIZE 3347 * aligned. Needless to say, the validity of the data 3348 * needs to also be DEV_BSIZE aligned. Note that this 3349 * fails with NFS if the server or some other client 3350 * extends the file's EOF. If our buffer is resized, 3351 * B_CACHE may remain set! XXX 3352 */ 3353 3354 toff = bp->b_bcount; 3355 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3356 3357 while ((bp->b_flags & B_CACHE) && toff < size) { 3358 vm_pindex_t pi; 3359 3360 if (tinc > (size - toff)) 3361 tinc = size - toff; 3362 3363 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3364 PAGE_SHIFT; 3365 3366 vfs_buf_test_cache( 3367 bp, 3368 bp->b_loffset, 3369 toff, 3370 tinc, 3371 bp->b_xio.xio_pages[pi] 3372 ); 3373 toff += tinc; 3374 tinc = PAGE_SIZE; 3375 } 3376 3377 /* 3378 * Step 3, fixup the KVM pmap. Remember that 3379 * bp->b_data is relative to bp->b_loffset, but 3380 * bp->b_loffset may be offset into the first page. 3381 */ 3382 bp->b_data = (caddr_t) 3383 trunc_page((vm_offset_t)bp->b_data); 3384 pmap_qenter((vm_offset_t)bp->b_data, 3385 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3386 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3387 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3388 } 3389 atomic_add_long(&bufspace, newbsize - bp->b_bufsize); 3390 } 3391 3392 /* adjust space use on already-dirty buffer */ 3393 if (bp->b_flags & B_DELWRI) { 3394 /* dirtykvaspace unchanged */ 3395 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize); 3396 if (bp->b_flags & B_HEAVY) { 3397 atomic_add_long(&dirtybufspacehw, 3398 newbsize - bp->b_bufsize); 3399 } 3400 } 3401 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3402 bp->b_bcount = size; /* requested buffer size */ 3403 bufspacewakeup(); 3404 } 3405 3406 /* 3407 * repurposebuf() (VMIO only) 3408 * 3409 * This performs a function similar to allocbuf() but the passed-in buffer 3410 * may contain some detrius from its previous incarnation in the form of 3411 * the page array. We try to repurpose the underlying pages. 3412 * 3413 * This code is nominally called to recycle buffer cache buffers AND (if 3414 * they are clean) to also recycle their underlying pages. We currently 3415 * can only recycle unmapped, clean pages. The code is called when buffer 3416 * cache 'newbuf' bandwidth exceeds (bufrate_cache) bytes per second. 3417 */ 3418 static 3419 void 3420 repurposebuf(struct buf *bp, int size) 3421 { 3422 int newbsize; 3423 int desiredpages; 3424 vm_offset_t toff; 3425 vm_offset_t tinc; 3426 vm_object_t obj; 3427 vm_page_t m; 3428 int i; 3429 int must_reenter = 0; 3430 long deaccumulate = 0; 3431 3432 3433 KKASSERT((bp->b_flags & (B_VMIO | B_DELWRI | B_MALLOC)) == B_VMIO); 3434 if (BUF_REFCNT(bp) == 0) 3435 panic("repurposebuf: buffer not busy"); 3436 3437 if (bp->b_kvasize < size) 3438 panic("repurposebuf: buffer too small"); 3439 3440 newbsize = roundup2(size, DEV_BSIZE); 3441 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3442 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3443 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3444 3445 /* 3446 * Buffer starts out 0-length with B_CACHE set. We will clear 3447 * As we check the backing store we will clear B_CACHE if necessary. 3448 */ 3449 atomic_add_long(&bufspace, newbsize - bp->b_bufsize); 3450 bp->b_bufsize = 0; 3451 bp->b_bcount = 0; 3452 bp->b_flags |= B_CACHE; 3453 3454 if (desiredpages) { 3455 obj = bp->b_vp->v_object; 3456 vm_object_hold(obj); 3457 } else { 3458 obj = NULL; 3459 } 3460 3461 /* 3462 * Step 1, bring in the VM pages from the object, repurposing or 3463 * allocating them if necessary. We must clear B_CACHE if these 3464 * pages are not valid for the range covered by the buffer. 3465 * 3466 * We are growing the buffer, possibly in a byte-granular fashion. 3467 */ 3468 for (i = 0; i < desiredpages; ++i) { 3469 vm_pindex_t pi; 3470 int error; 3471 int iswired; 3472 3473 pi = OFF_TO_IDX(bp->b_loffset) + i; 3474 3475 /* 3476 * Blocking on m->busy might lead to a 3477 * deadlock: 3478 * 3479 * vm_fault->getpages->cluster_read->allocbuf 3480 */ 3481 m = (i < bp->b_xio.xio_npages) ? bp->b_xio.xio_pages[i] : NULL; 3482 bp->b_xio.xio_pages[i] = NULL; 3483 KASSERT(m != bogus_page, ("repurposebuf: bogus page found")); 3484 m = vm_page_repurpose(obj, pi, FALSE, &error, m, 3485 &must_reenter, &iswired); 3486 3487 if (error) { 3488 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3489 --i; /* retry */ 3490 continue; 3491 } 3492 if (m == NULL) { 3493 /* 3494 * note: must allocate system pages 3495 * since blocking here could intefere 3496 * with paging I/O, no matter which 3497 * process we are. 3498 */ 3499 must_reenter = 1; 3500 m = bio_page_alloc(bp, obj, pi, desiredpages - i); 3501 if (m) { 3502 vm_page_wire(m); 3503 vm_page_wakeup(m); 3504 bp->b_flags &= ~B_CACHE; 3505 bp->b_xio.xio_pages[i] = m; 3506 if (m->valid) 3507 deaccumulate += PAGE_SIZE; 3508 } else { 3509 --i; /* retry */ 3510 } 3511 continue; 3512 } 3513 if (m->valid) 3514 deaccumulate += PAGE_SIZE; 3515 3516 /* 3517 * We found a page and were able to busy it. 3518 */ 3519 if (!iswired) 3520 vm_page_wire(m); 3521 vm_page_wakeup(m); 3522 bp->b_xio.xio_pages[i] = m; 3523 if (bp->b_act_count < m->act_count) 3524 bp->b_act_count = m->act_count; 3525 } 3526 if (desiredpages) 3527 vm_object_drop(obj); 3528 3529 /* 3530 * Even though its a new buffer, any pages already in the VM 3531 * page cache should not count towards I/O bandwidth. 3532 */ 3533 if (deaccumulate) 3534 atomic_add_long(&bufcache_bw_accum, -deaccumulate); 3535 3536 /* 3537 * Clean-up any loose pages. 3538 */ 3539 while (i < bp->b_xio.xio_npages) { 3540 m = bp->b_xio.xio_pages[i]; 3541 KASSERT(m != bogus_page, ("repurposebuf: bogus page found")); 3542 vm_page_busy_wait(m, TRUE, "biodep"); 3543 bp->b_xio.xio_pages[i] = NULL; 3544 vm_page_unwire(m, 0); 3545 vm_page_wakeup(m); 3546 ++i; 3547 } 3548 if (desiredpages < bp->b_xio.xio_npages) { 3549 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 3550 (desiredpages << PAGE_SHIFT), 3551 (bp->b_xio.xio_npages - desiredpages)); 3552 } 3553 bp->b_xio.xio_npages = desiredpages; 3554 3555 /* 3556 * Step 2. We've loaded the pages into the buffer, 3557 * we have to figure out if we can still have B_CACHE 3558 * set. Note that B_CACHE is set according to the 3559 * byte-granular range ( bcount and size ), not the 3560 * aligned range ( newbsize ). 3561 * 3562 * The VM test is against m->valid, which is DEV_BSIZE 3563 * aligned. Needless to say, the validity of the data 3564 * needs to also be DEV_BSIZE aligned. Note that this 3565 * fails with NFS if the server or some other client 3566 * extends the file's EOF. If our buffer is resized, 3567 * B_CACHE may remain set! XXX 3568 */ 3569 toff = bp->b_bcount; 3570 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3571 3572 while ((bp->b_flags & B_CACHE) && toff < size) { 3573 vm_pindex_t pi; 3574 3575 if (tinc > (size - toff)) 3576 tinc = size - toff; 3577 3578 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3579 3580 vfs_buf_test_cache(bp, bp->b_loffset, toff, 3581 tinc, bp->b_xio.xio_pages[pi]); 3582 toff += tinc; 3583 tinc = PAGE_SIZE; 3584 } 3585 3586 /* 3587 * Step 3, fixup the KVM pmap. Remember that 3588 * bp->b_data is relative to bp->b_loffset, but 3589 * bp->b_loffset may be offset into the first page. 3590 */ 3591 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 3592 if (must_reenter) { 3593 pmap_qenter((vm_offset_t)bp->b_data, 3594 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3595 } else { 3596 atomic_add_long(&repurposedspace, newbsize); 3597 } 3598 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3599 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3600 3601 if (newbsize < bp->b_bufsize) 3602 bufspacewakeup(); 3603 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3604 bp->b_bcount = size; /* requested buffer size */ 3605 } 3606 3607 /* 3608 * biowait: 3609 * 3610 * Wait for buffer I/O completion, returning error status. B_EINTR 3611 * is converted into an EINTR error but not cleared (since a chain 3612 * of biowait() calls may occur). 3613 * 3614 * On return bpdone() will have been called but the buffer will remain 3615 * locked and will not have been brelse()'d. 3616 * 3617 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3618 * likely still in progress on return. 3619 * 3620 * NOTE! This operation is on a BIO, not a BUF. 3621 * 3622 * NOTE! BIO_DONE is cleared by vn_strategy() 3623 */ 3624 static __inline int 3625 _biowait(struct bio *bio, const char *wmesg, int to) 3626 { 3627 struct buf *bp = bio->bio_buf; 3628 u_int32_t flags; 3629 u_int32_t nflags; 3630 int error; 3631 3632 KKASSERT(bio == &bp->b_bio1); 3633 for (;;) { 3634 flags = bio->bio_flags; 3635 if (flags & BIO_DONE) 3636 break; 3637 nflags = flags | BIO_WANT; 3638 tsleep_interlock(bio, 0); 3639 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3640 if (wmesg) 3641 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3642 else if (bp->b_cmd == BUF_CMD_READ) 3643 error = tsleep(bio, PINTERLOCKED, "biord", to); 3644 else 3645 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3646 if (error) { 3647 kprintf("tsleep error biowait %d\n", error); 3648 return (error); 3649 } 3650 } 3651 } 3652 3653 /* 3654 * Finish up. 3655 */ 3656 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3657 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3658 if (bp->b_flags & B_EINTR) 3659 return (EINTR); 3660 if (bp->b_flags & B_ERROR) 3661 return (bp->b_error ? bp->b_error : EIO); 3662 return (0); 3663 } 3664 3665 int 3666 biowait(struct bio *bio, const char *wmesg) 3667 { 3668 return(_biowait(bio, wmesg, 0)); 3669 } 3670 3671 int 3672 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3673 { 3674 return(_biowait(bio, wmesg, to)); 3675 } 3676 3677 /* 3678 * This associates a tracking count with an I/O. vn_strategy() and 3679 * dev_dstrategy() do this automatically but there are a few cases 3680 * where a vnode or device layer is bypassed when a block translation 3681 * is cached. In such cases bio_start_transaction() may be called on 3682 * the bypassed layers so the system gets an I/O in progress indication 3683 * for those higher layers. 3684 */ 3685 void 3686 bio_start_transaction(struct bio *bio, struct bio_track *track) 3687 { 3688 bio->bio_track = track; 3689 bio_track_ref(track); 3690 dsched_buf_enter(bio->bio_buf); /* might stack */ 3691 } 3692 3693 /* 3694 * Initiate I/O on a vnode. 3695 * 3696 * SWAPCACHE OPERATION: 3697 * 3698 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3699 * devfs also uses b_vp for fake buffers so we also have to check 3700 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3701 * underlying block device. The swap assignments are related to the 3702 * buffer cache buffer's b_vp, not the passed vp. 3703 * 3704 * The passed vp == bp->b_vp only in the case where the strategy call 3705 * is made on the vp itself for its own buffers (a regular file or 3706 * block device vp). The filesystem usually then re-calls vn_strategy() 3707 * after translating the request to an underlying device. 3708 * 3709 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3710 * underlying buffer cache buffers. 3711 * 3712 * We can only deal with page-aligned buffers at the moment, because 3713 * we can't tell what the real dirty state for pages straddling a buffer 3714 * are. 3715 * 3716 * In order to call swap_pager_strategy() we must provide the VM object 3717 * and base offset for the underlying buffer cache pages so it can find 3718 * the swap blocks. 3719 */ 3720 void 3721 vn_strategy(struct vnode *vp, struct bio *bio) 3722 { 3723 struct bio_track *track; 3724 struct buf *bp = bio->bio_buf; 3725 3726 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3727 3728 /* 3729 * Set when an I/O is issued on the bp. Cleared by consumers 3730 * (aka HAMMER), allowing the consumer to determine if I/O had 3731 * actually occurred. 3732 */ 3733 bp->b_flags |= B_IOISSUED; 3734 3735 /* 3736 * Handle the swap cache intercept. 3737 */ 3738 if (vn_cache_strategy(vp, bio)) 3739 return; 3740 3741 /* 3742 * Otherwise do the operation through the filesystem 3743 */ 3744 if (bp->b_cmd == BUF_CMD_READ) 3745 track = &vp->v_track_read; 3746 else 3747 track = &vp->v_track_write; 3748 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3749 bio->bio_track = track; 3750 bio_track_ref(track); 3751 dsched_buf_enter(bp); /* might stack */ 3752 vop_strategy(*vp->v_ops, vp, bio); 3753 } 3754 3755 static void vn_cache_strategy_callback(struct bio *bio); 3756 3757 int 3758 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3759 { 3760 struct buf *bp = bio->bio_buf; 3761 struct bio *nbio; 3762 vm_object_t object; 3763 vm_page_t m; 3764 int i; 3765 3766 /* 3767 * Stop using swapcache if paniced, dumping, or dumped 3768 */ 3769 if (panicstr || dumping) 3770 return(0); 3771 3772 /* 3773 * Is this buffer cache buffer suitable for reading from 3774 * the swap cache? 3775 */ 3776 if (vm_swapcache_read_enable == 0 || 3777 bp->b_cmd != BUF_CMD_READ || 3778 ((bp->b_flags & B_CLUSTER) == 0 && 3779 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3780 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3781 (bp->b_bcount & PAGE_MASK) != 0) { 3782 return(0); 3783 } 3784 3785 /* 3786 * Figure out the original VM object (it will match the underlying 3787 * VM pages). Note that swap cached data uses page indices relative 3788 * to that object, not relative to bio->bio_offset. 3789 */ 3790 if (bp->b_flags & B_CLUSTER) 3791 object = vp->v_object; 3792 else 3793 object = bp->b_vp->v_object; 3794 3795 /* 3796 * In order to be able to use the swap cache all underlying VM 3797 * pages must be marked as such, and we can't have any bogus pages. 3798 */ 3799 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3800 m = bp->b_xio.xio_pages[i]; 3801 if ((m->flags & PG_SWAPPED) == 0) 3802 break; 3803 if (m == bogus_page) 3804 break; 3805 } 3806 3807 /* 3808 * If we are good then issue the I/O using swap_pager_strategy(). 3809 * 3810 * We can only do this if the buffer actually supports object-backed 3811 * I/O. If it doesn't npages will be 0. 3812 */ 3813 if (i && i == bp->b_xio.xio_npages) { 3814 m = bp->b_xio.xio_pages[0]; 3815 nbio = push_bio(bio); 3816 nbio->bio_done = vn_cache_strategy_callback; 3817 nbio->bio_offset = ptoa(m->pindex); 3818 KKASSERT(m->object == object); 3819 swap_pager_strategy(object, nbio); 3820 return(1); 3821 } 3822 return(0); 3823 } 3824 3825 /* 3826 * This is a bit of a hack but since the vn_cache_strategy() function can 3827 * override a VFS's strategy function we must make sure that the bio, which 3828 * is probably bio2, doesn't leak an unexpected offset value back to the 3829 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3830 * bio went through its own file strategy function and the the bio2 offset 3831 * is a cached disk offset when, in fact, it isn't. 3832 */ 3833 static void 3834 vn_cache_strategy_callback(struct bio *bio) 3835 { 3836 bio->bio_offset = NOOFFSET; 3837 biodone(pop_bio(bio)); 3838 } 3839 3840 /* 3841 * bpdone: 3842 * 3843 * Finish I/O on a buffer after all BIOs have been processed. 3844 * Called when the bio chain is exhausted or by biowait. If called 3845 * by biowait, elseit is typically 0. 3846 * 3847 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3848 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3849 * assuming B_INVAL is clear. 3850 * 3851 * For the VMIO case, we set B_CACHE if the op was a read and no 3852 * read error occured, or if the op was a write. B_CACHE is never 3853 * set if the buffer is invalid or otherwise uncacheable. 3854 * 3855 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3856 * initiator to leave B_INVAL set to brelse the buffer out of existance 3857 * in the biodone routine. 3858 * 3859 * bpdone is responsible for calling bundirty() on the buffer after a 3860 * successful write. We previously did this prior to initiating the 3861 * write under the assumption that the buffer might be dirtied again 3862 * while the write was in progress, however doing it before-hand creates 3863 * a race condition prior to the call to vn_strategy() where the 3864 * filesystem may not be aware that a dirty buffer is present. 3865 * It should not be possible for the buffer or its underlying pages to 3866 * be redirtied prior to bpdone()'s unbusying of the underlying VM 3867 * pages. 3868 */ 3869 void 3870 bpdone(struct buf *bp, int elseit) 3871 { 3872 buf_cmd_t cmd; 3873 3874 KASSERT(BUF_REFCNTNB(bp) > 0, 3875 ("bpdone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3876 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3877 ("bpdone: bp %p already done!", bp)); 3878 3879 /* 3880 * No more BIOs are left. All completion functions have been dealt 3881 * with, now we clean up the buffer. 3882 */ 3883 cmd = bp->b_cmd; 3884 bp->b_cmd = BUF_CMD_DONE; 3885 3886 /* 3887 * Only reads and writes are processed past this point. 3888 */ 3889 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3890 if (cmd == BUF_CMD_FREEBLKS) 3891 bp->b_flags |= B_NOCACHE; 3892 if (elseit) 3893 brelse(bp); 3894 return; 3895 } 3896 3897 /* 3898 * A failed write must re-dirty the buffer unless B_INVAL 3899 * was set. 3900 * 3901 * A successful write must clear the dirty flag. This is done after 3902 * the write to ensure that the buffer remains on the vnode's dirty 3903 * list for filesystem interlocks / checks until the write is actually 3904 * complete. HAMMER2 is sensitive to this issue. 3905 * 3906 * Only applicable to normal buffers (with VPs). vinum buffers may 3907 * not have a vp. 3908 * 3909 * Must be done prior to calling buf_complete() as the callback might 3910 * re-dirty the buffer. 3911 */ 3912 if (cmd == BUF_CMD_WRITE) { 3913 if ((bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3914 bp->b_flags &= ~B_NOCACHE; 3915 if (bp->b_vp) 3916 bdirty(bp); 3917 } else { 3918 if (bp->b_vp) 3919 bundirty(bp); 3920 } 3921 } 3922 3923 /* 3924 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3925 * a lot worse. XXX - move this above the clearing of b_cmd 3926 */ 3927 if (LIST_FIRST(&bp->b_dep) != NULL) 3928 buf_complete(bp); 3929 3930 if (bp->b_flags & B_VMIO) { 3931 int i; 3932 vm_ooffset_t foff; 3933 vm_page_t m; 3934 vm_object_t obj; 3935 int iosize; 3936 struct vnode *vp = bp->b_vp; 3937 3938 obj = vp->v_object; 3939 3940 #if defined(VFS_BIO_DEBUG) 3941 if (vp->v_auxrefs == 0) 3942 panic("bpdone: zero vnode hold count"); 3943 if ((vp->v_flag & VOBJBUF) == 0) 3944 panic("bpdone: vnode is not setup for merged cache"); 3945 #endif 3946 3947 foff = bp->b_loffset; 3948 KASSERT(foff != NOOFFSET, ("bpdone: no buffer offset")); 3949 KASSERT(obj != NULL, ("bpdone: missing VM object")); 3950 3951 #if defined(VFS_BIO_DEBUG) 3952 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3953 kprintf("bpdone: paging in progress(%d) < " 3954 "bp->b_xio.xio_npages(%d)\n", 3955 obj->paging_in_progress, 3956 bp->b_xio.xio_npages); 3957 } 3958 #endif 3959 3960 /* 3961 * Set B_CACHE if the op was a normal read and no error 3962 * occured. B_CACHE is set for writes in the b*write() 3963 * routines. 3964 */ 3965 iosize = bp->b_bcount - bp->b_resid; 3966 if (cmd == BUF_CMD_READ && 3967 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3968 bp->b_flags |= B_CACHE; 3969 } 3970 3971 vm_object_hold(obj); 3972 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3973 int resid; 3974 int isbogus; 3975 3976 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3977 if (resid > iosize) 3978 resid = iosize; 3979 3980 /* 3981 * cleanup bogus pages, restoring the originals. Since 3982 * the originals should still be wired, we don't have 3983 * to worry about interrupt/freeing races destroying 3984 * the VM object association. 3985 */ 3986 m = bp->b_xio.xio_pages[i]; 3987 if (m == bogus_page) { 3988 if ((bp->b_flags & B_HASBOGUS) == 0) 3989 panic("bpdone: bp %p corrupt bogus", bp); 3990 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3991 if (m == NULL) 3992 panic("bpdone: page disappeared"); 3993 bp->b_xio.xio_pages[i] = m; 3994 isbogus = 1; 3995 } else { 3996 isbogus = 0; 3997 } 3998 #if defined(VFS_BIO_DEBUG) 3999 if (OFF_TO_IDX(foff) != m->pindex) { 4000 kprintf("bpdone: foff(%lu)/m->pindex(%ld) " 4001 "mismatch\n", 4002 (unsigned long)foff, (long)m->pindex); 4003 } 4004 #endif 4005 4006 /* 4007 * In the write case, the valid and clean bits are 4008 * already changed correctly (see bdwrite()), so we 4009 * only need to do this here in the read case. 4010 */ 4011 vm_page_busy_wait(m, FALSE, "bpdpgw"); 4012 if (cmd == BUF_CMD_READ && isbogus == 0 && resid > 0) 4013 vfs_clean_one_page(bp, i, m); 4014 4015 /* 4016 * when debugging new filesystems or buffer I/O 4017 * methods, this is the most common error that pops 4018 * up. if you see this, you have not set the page 4019 * busy flag correctly!!! 4020 */ 4021 if (m->busy == 0) { 4022 kprintf("bpdone: page busy < 0, " 4023 "pindex: %d, foff: 0x(%x,%x), " 4024 "resid: %d, index: %d\n", 4025 (int) m->pindex, (int)(foff >> 32), 4026 (int) foff & 0xffffffff, resid, i); 4027 if (!vn_isdisk(vp, NULL)) 4028 kprintf(" iosize: %ld, loffset: %lld, " 4029 "flags: 0x%08x, npages: %d\n", 4030 bp->b_vp->v_mount->mnt_stat.f_iosize, 4031 (long long)bp->b_loffset, 4032 bp->b_flags, bp->b_xio.xio_npages); 4033 else 4034 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 4035 (long long)bp->b_loffset, 4036 bp->b_flags, bp->b_xio.xio_npages); 4037 kprintf(" valid: 0x%x, dirty: 0x%x, " 4038 "wired: %d\n", 4039 m->valid, m->dirty, 4040 m->wire_count); 4041 panic("bpdone: page busy < 0"); 4042 } 4043 vm_page_io_finish(m); 4044 vm_page_wakeup(m); 4045 vm_object_pip_wakeup(obj); 4046 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4047 iosize -= resid; 4048 } 4049 if (bp->b_flags & B_HASBOGUS) { 4050 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4051 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4052 bp->b_flags &= ~B_HASBOGUS; 4053 } 4054 vm_object_drop(obj); 4055 } 4056 4057 /* 4058 * Finish up by releasing the buffer. There are no more synchronous 4059 * or asynchronous completions, those were handled by bio_done 4060 * callbacks. 4061 */ 4062 if (elseit) { 4063 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 4064 brelse(bp); 4065 else 4066 bqrelse(bp); 4067 } 4068 } 4069 4070 /* 4071 * Normal biodone. 4072 */ 4073 void 4074 biodone(struct bio *bio) 4075 { 4076 struct buf *bp = bio->bio_buf; 4077 4078 runningbufwakeup(bp); 4079 4080 /* 4081 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 4082 */ 4083 while (bio) { 4084 biodone_t *done_func; 4085 struct bio_track *track; 4086 4087 /* 4088 * BIO tracking. Most but not all BIOs are tracked. 4089 */ 4090 if ((track = bio->bio_track) != NULL) { 4091 bio_track_rel(track); 4092 bio->bio_track = NULL; 4093 } 4094 4095 /* 4096 * A bio_done function terminates the loop. The function 4097 * will be responsible for any further chaining and/or 4098 * buffer management. 4099 * 4100 * WARNING! The done function can deallocate the buffer! 4101 */ 4102 if ((done_func = bio->bio_done) != NULL) { 4103 bio->bio_done = NULL; 4104 done_func(bio); 4105 return; 4106 } 4107 bio = bio->bio_prev; 4108 } 4109 4110 /* 4111 * If we've run out of bio's do normal [a]synchronous completion. 4112 */ 4113 bpdone(bp, 1); 4114 } 4115 4116 /* 4117 * Synchronous biodone - this terminates a synchronous BIO. 4118 * 4119 * bpdone() is called with elseit=FALSE, leaving the buffer completed 4120 * but still locked. The caller must brelse() the buffer after waiting 4121 * for completion. 4122 */ 4123 void 4124 biodone_sync(struct bio *bio) 4125 { 4126 struct buf *bp = bio->bio_buf; 4127 int flags; 4128 int nflags; 4129 4130 KKASSERT(bio == &bp->b_bio1); 4131 bpdone(bp, 0); 4132 4133 for (;;) { 4134 flags = bio->bio_flags; 4135 nflags = (flags | BIO_DONE) & ~BIO_WANT; 4136 4137 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 4138 if (flags & BIO_WANT) 4139 wakeup(bio); 4140 break; 4141 } 4142 } 4143 } 4144 4145 /* 4146 * vfs_unbusy_pages: 4147 * 4148 * This routine is called in lieu of iodone in the case of 4149 * incomplete I/O. This keeps the busy status for pages 4150 * consistant. 4151 */ 4152 void 4153 vfs_unbusy_pages(struct buf *bp) 4154 { 4155 int i; 4156 4157 runningbufwakeup(bp); 4158 4159 if (bp->b_flags & B_VMIO) { 4160 struct vnode *vp = bp->b_vp; 4161 vm_object_t obj; 4162 4163 obj = vp->v_object; 4164 vm_object_hold(obj); 4165 4166 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4167 vm_page_t m = bp->b_xio.xio_pages[i]; 4168 4169 /* 4170 * When restoring bogus changes the original pages 4171 * should still be wired, so we are in no danger of 4172 * losing the object association and do not need 4173 * critical section protection particularly. 4174 */ 4175 if (m == bogus_page) { 4176 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 4177 if (!m) { 4178 panic("vfs_unbusy_pages: page missing"); 4179 } 4180 bp->b_xio.xio_pages[i] = m; 4181 } 4182 vm_page_busy_wait(m, FALSE, "bpdpgw"); 4183 vm_page_io_finish(m); 4184 vm_page_wakeup(m); 4185 vm_object_pip_wakeup(obj); 4186 } 4187 if (bp->b_flags & B_HASBOGUS) { 4188 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4189 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4190 bp->b_flags &= ~B_HASBOGUS; 4191 } 4192 vm_object_drop(obj); 4193 } 4194 } 4195 4196 /* 4197 * vfs_busy_pages: 4198 * 4199 * This routine is called before a device strategy routine. 4200 * It is used to tell the VM system that paging I/O is in 4201 * progress, and treat the pages associated with the buffer 4202 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4203 * flag is handled to make sure that the object doesn't become 4204 * inconsistant. 4205 * 4206 * Since I/O has not been initiated yet, certain buffer flags 4207 * such as B_ERROR or B_INVAL may be in an inconsistant state 4208 * and should be ignored. 4209 */ 4210 void 4211 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4212 { 4213 int i, bogus; 4214 struct lwp *lp = curthread->td_lwp; 4215 4216 /* 4217 * The buffer's I/O command must already be set. If reading, 4218 * B_CACHE must be 0 (double check against callers only doing 4219 * I/O when B_CACHE is 0). 4220 */ 4221 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4222 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4223 4224 if (bp->b_flags & B_VMIO) { 4225 vm_object_t obj; 4226 4227 obj = vp->v_object; 4228 KASSERT(bp->b_loffset != NOOFFSET, 4229 ("vfs_busy_pages: no buffer offset")); 4230 4231 /* 4232 * Busy all the pages. We have to busy them all at once 4233 * to avoid deadlocks. 4234 */ 4235 retry: 4236 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4237 vm_page_t m = bp->b_xio.xio_pages[i]; 4238 4239 if (vm_page_busy_try(m, FALSE)) { 4240 vm_page_sleep_busy(m, FALSE, "vbpage"); 4241 while (--i >= 0) 4242 vm_page_wakeup(bp->b_xio.xio_pages[i]); 4243 goto retry; 4244 } 4245 } 4246 4247 /* 4248 * Setup for I/O, soft-busy the page right now because 4249 * the next loop may block. 4250 */ 4251 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4252 vm_page_t m = bp->b_xio.xio_pages[i]; 4253 4254 if ((bp->b_flags & B_CLUSTER) == 0) { 4255 vm_object_pip_add(obj, 1); 4256 vm_page_io_start(m); 4257 } 4258 } 4259 4260 /* 4261 * Adjust protections for I/O and do bogus-page mapping. 4262 * Assume that vm_page_protect() can block (it can block 4263 * if VM_PROT_NONE, don't take any chances regardless). 4264 * 4265 * In particular note that for writes we must incorporate 4266 * page dirtyness from the VM system into the buffer's 4267 * dirty range. 4268 * 4269 * For reads we theoretically must incorporate page dirtyness 4270 * from the VM system to determine if the page needs bogus 4271 * replacement, but we shortcut the test by simply checking 4272 * that all m->valid bits are set, indicating that the page 4273 * is fully valid and does not need to be re-read. For any 4274 * VM system dirtyness the page will also be fully valid 4275 * since it was mapped at one point. 4276 */ 4277 bogus = 0; 4278 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4279 vm_page_t m = bp->b_xio.xio_pages[i]; 4280 4281 if (bp->b_cmd == BUF_CMD_WRITE) { 4282 /* 4283 * When readying a vnode-backed buffer for 4284 * a write we must zero-fill any invalid 4285 * portions of the backing VM pages, mark 4286 * it valid and clear related dirty bits. 4287 * 4288 * vfs_clean_one_page() incorporates any 4289 * VM dirtyness and updates the b_dirtyoff 4290 * range (after we've made the page RO). 4291 * 4292 * It is also expected that the pmap modified 4293 * bit has already been cleared by the 4294 * vm_page_protect(). We may not be able 4295 * to clear all dirty bits for a page if it 4296 * was also memory mapped (NFS). 4297 * 4298 * Finally be sure to unassign any swap-cache 4299 * backing store as it is now stale. 4300 */ 4301 vm_page_protect(m, VM_PROT_READ); 4302 vfs_clean_one_page(bp, i, m); 4303 swap_pager_unswapped(m); 4304 } else if (m->valid == VM_PAGE_BITS_ALL) { 4305 /* 4306 * When readying a vnode-backed buffer for 4307 * read we must replace any dirty pages with 4308 * a bogus page so dirty data is not destroyed 4309 * when filling gaps. 4310 * 4311 * To avoid testing whether the page is 4312 * dirty we instead test that the page was 4313 * at some point mapped (m->valid fully 4314 * valid) with the understanding that 4315 * this also covers the dirty case. 4316 */ 4317 bp->b_xio.xio_pages[i] = bogus_page; 4318 bp->b_flags |= B_HASBOGUS; 4319 bogus++; 4320 } else if (m->valid & m->dirty) { 4321 /* 4322 * This case should not occur as partial 4323 * dirtyment can only happen if the buffer 4324 * is B_CACHE, and this code is not entered 4325 * if the buffer is B_CACHE. 4326 */ 4327 kprintf("Warning: vfs_busy_pages - page not " 4328 "fully valid! loff=%jx bpf=%08x " 4329 "idx=%d val=%02x dir=%02x\n", 4330 (uintmax_t)bp->b_loffset, bp->b_flags, 4331 i, m->valid, m->dirty); 4332 vm_page_protect(m, VM_PROT_NONE); 4333 } else { 4334 /* 4335 * The page is not valid and can be made 4336 * part of the read. 4337 */ 4338 vm_page_protect(m, VM_PROT_NONE); 4339 } 4340 vm_page_wakeup(m); 4341 } 4342 if (bogus) { 4343 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4344 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4345 } 4346 } 4347 4348 /* 4349 * This is the easiest place to put the process accounting for the I/O 4350 * for now. 4351 */ 4352 if (lp != NULL) { 4353 if (bp->b_cmd == BUF_CMD_READ) 4354 lp->lwp_ru.ru_inblock++; 4355 else 4356 lp->lwp_ru.ru_oublock++; 4357 } 4358 } 4359 4360 /* 4361 * Tell the VM system that the pages associated with this buffer 4362 * are clean. This is used for delayed writes where the data is 4363 * going to go to disk eventually without additional VM intevention. 4364 * 4365 * NOTE: While we only really need to clean through to b_bcount, we 4366 * just go ahead and clean through to b_bufsize. 4367 */ 4368 static void 4369 vfs_clean_pages(struct buf *bp) 4370 { 4371 vm_page_t m; 4372 int i; 4373 4374 if ((bp->b_flags & B_VMIO) == 0) 4375 return; 4376 4377 KASSERT(bp->b_loffset != NOOFFSET, 4378 ("vfs_clean_pages: no buffer offset")); 4379 4380 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4381 m = bp->b_xio.xio_pages[i]; 4382 vfs_clean_one_page(bp, i, m); 4383 } 4384 } 4385 4386 /* 4387 * vfs_clean_one_page: 4388 * 4389 * Set the valid bits and clear the dirty bits in a page within a 4390 * buffer. The range is restricted to the buffer's size and the 4391 * buffer's logical offset might index into the first page. 4392 * 4393 * The caller has busied or soft-busied the page and it is not mapped, 4394 * test and incorporate the dirty bits into b_dirtyoff/end before 4395 * clearing them. Note that we need to clear the pmap modified bits 4396 * after determining the the page was dirty, vm_page_set_validclean() 4397 * does not do it for us. 4398 * 4399 * This routine is typically called after a read completes (dirty should 4400 * be zero in that case as we are not called on bogus-replace pages), 4401 * or before a write is initiated. 4402 */ 4403 static void 4404 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4405 { 4406 int bcount; 4407 int xoff; 4408 int soff; 4409 int eoff; 4410 4411 /* 4412 * Calculate offset range within the page but relative to buffer's 4413 * loffset. loffset might be offset into the first page. 4414 */ 4415 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4416 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4417 4418 if (pageno == 0) { 4419 soff = xoff; 4420 eoff = PAGE_SIZE; 4421 } else { 4422 soff = (pageno << PAGE_SHIFT); 4423 eoff = soff + PAGE_SIZE; 4424 } 4425 if (eoff > bcount) 4426 eoff = bcount; 4427 if (soff >= eoff) 4428 return; 4429 4430 /* 4431 * Test dirty bits and adjust b_dirtyoff/end. 4432 * 4433 * If dirty pages are incorporated into the bp any prior 4434 * B_NEEDCOMMIT state (NFS) must be cleared because the 4435 * caller has not taken into account the new dirty data. 4436 * 4437 * If the page was memory mapped the dirty bits might go beyond the 4438 * end of the buffer, but we can't really make the assumption that 4439 * a file EOF straddles the buffer (even though this is the case for 4440 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4441 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4442 * This also saves some console spam. 4443 * 4444 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4445 * NFS can handle huge commits but not huge writes. 4446 */ 4447 vm_page_test_dirty(m); 4448 if (m->dirty) { 4449 if ((bp->b_flags & B_NEEDCOMMIT) && 4450 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4451 if (debug_commit) 4452 kprintf("Warning: vfs_clean_one_page: bp %p " 4453 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4454 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4455 "doff/end %d %d\n", 4456 bp, (uintmax_t)bp->b_loffset, bp->b_bcount, 4457 bp->b_flags, bp->b_cmd, 4458 m->valid, m->dirty, xoff, soff, eoff, 4459 bp->b_dirtyoff, bp->b_dirtyend); 4460 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4461 if (debug_commit) 4462 print_backtrace(-1); 4463 } 4464 /* 4465 * Only clear the pmap modified bits if ALL the dirty bits 4466 * are set, otherwise the system might mis-clear portions 4467 * of a page. 4468 */ 4469 if (m->dirty == VM_PAGE_BITS_ALL && 4470 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4471 pmap_clear_modify(m); 4472 } 4473 if (bp->b_dirtyoff > soff - xoff) 4474 bp->b_dirtyoff = soff - xoff; 4475 if (bp->b_dirtyend < eoff - xoff) 4476 bp->b_dirtyend = eoff - xoff; 4477 } 4478 4479 /* 4480 * Set related valid bits, clear related dirty bits. 4481 * Does not mess with the pmap modified bit. 4482 * 4483 * WARNING! We cannot just clear all of m->dirty here as the 4484 * buffer cache buffers may use a DEV_BSIZE'd aligned 4485 * block size, or have an odd size (e.g. NFS at file EOF). 4486 * The putpages code can clear m->dirty to 0. 4487 * 4488 * If a VOP_WRITE generates a buffer cache buffer which 4489 * covers the same space as mapped writable pages the 4490 * buffer flush might not be able to clear all the dirty 4491 * bits and still require a putpages from the VM system 4492 * to finish it off. 4493 * 4494 * WARNING! vm_page_set_validclean() currently assumes vm_token 4495 * is held. The page might not be busied (bdwrite() case). 4496 * XXX remove this comment once we've validated that this 4497 * is no longer an issue. 4498 */ 4499 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4500 } 4501 4502 #if 0 4503 /* 4504 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4505 * The page data is assumed to be valid (there is no zeroing here). 4506 */ 4507 static void 4508 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4509 { 4510 int bcount; 4511 int xoff; 4512 int soff; 4513 int eoff; 4514 4515 /* 4516 * Calculate offset range within the page but relative to buffer's 4517 * loffset. loffset might be offset into the first page. 4518 */ 4519 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4520 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4521 4522 if (pageno == 0) { 4523 soff = xoff; 4524 eoff = PAGE_SIZE; 4525 } else { 4526 soff = (pageno << PAGE_SHIFT); 4527 eoff = soff + PAGE_SIZE; 4528 } 4529 if (eoff > bcount) 4530 eoff = bcount; 4531 if (soff >= eoff) 4532 return; 4533 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4534 } 4535 #endif 4536 4537 /* 4538 * vfs_bio_clrbuf: 4539 * 4540 * Clear a buffer. This routine essentially fakes an I/O, so we need 4541 * to clear B_ERROR and B_INVAL. 4542 * 4543 * Note that while we only theoretically need to clear through b_bcount, 4544 * we go ahead and clear through b_bufsize. 4545 */ 4546 4547 void 4548 vfs_bio_clrbuf(struct buf *bp) 4549 { 4550 int i, mask = 0; 4551 caddr_t sa, ea; 4552 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4553 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4554 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4555 (bp->b_loffset & PAGE_MASK) == 0) { 4556 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4557 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4558 bp->b_resid = 0; 4559 return; 4560 } 4561 if ((bp->b_xio.xio_pages[0]->valid & mask) == 0) { 4562 bzero(bp->b_data, bp->b_bufsize); 4563 bp->b_xio.xio_pages[0]->valid |= mask; 4564 bp->b_resid = 0; 4565 return; 4566 } 4567 } 4568 sa = bp->b_data; 4569 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4570 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4571 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4572 ea = (caddr_t)(vm_offset_t)ulmin( 4573 (u_long)(vm_offset_t)ea, 4574 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4575 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4576 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4577 continue; 4578 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4579 bzero(sa, ea - sa); 4580 } else { 4581 for (; sa < ea; sa += DEV_BSIZE, j++) { 4582 if ((bp->b_xio.xio_pages[i]->valid & 4583 (1<<j)) == 0) { 4584 bzero(sa, DEV_BSIZE); 4585 } 4586 } 4587 } 4588 bp->b_xio.xio_pages[i]->valid |= mask; 4589 } 4590 bp->b_resid = 0; 4591 } else { 4592 clrbuf(bp); 4593 } 4594 } 4595 4596 /* 4597 * vm_hold_load_pages: 4598 * 4599 * Load pages into the buffer's address space. The pages are 4600 * allocated from the kernel object in order to reduce interference 4601 * with the any VM paging I/O activity. The range of loaded 4602 * pages will be wired. 4603 * 4604 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4605 * retrieve the full range (to - from) of pages. 4606 */ 4607 void 4608 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4609 { 4610 vm_offset_t pg; 4611 vm_page_t p; 4612 int index; 4613 4614 to = round_page(to); 4615 from = round_page(from); 4616 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4617 4618 pg = from; 4619 while (pg < to) { 4620 /* 4621 * Note: must allocate system pages since blocking here 4622 * could intefere with paging I/O, no matter which 4623 * process we are. 4624 */ 4625 vm_object_hold(&kernel_object); 4626 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT, 4627 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4628 vm_object_drop(&kernel_object); 4629 if (p) { 4630 vm_page_wire(p); 4631 p->valid = VM_PAGE_BITS_ALL; 4632 pmap_kenter_noinval(pg, VM_PAGE_TO_PHYS(p)); 4633 bp->b_xio.xio_pages[index] = p; 4634 vm_page_wakeup(p); 4635 4636 pg += PAGE_SIZE; 4637 ++index; 4638 } 4639 } 4640 pmap_invalidate_range(&kernel_pmap, from, to); 4641 bp->b_xio.xio_npages = index; 4642 } 4643 4644 /* 4645 * Allocate a page for a buffer cache buffer. 4646 * 4647 * If NULL is returned the caller is expected to retry (typically check if 4648 * the page already exists on retry before trying to allocate one). 4649 * 4650 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This 4651 * function will use the system reserve with the hope that the page 4652 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller 4653 * is done with the buffer. 4654 * 4655 * NOTE! However, TMPFS is a special case because flushing a dirty buffer 4656 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon 4657 * is capable of retiring pages (to swap). For TMPFS we don't dig 4658 * into the system reserve because doing so could stall out pretty 4659 * much every process running on the system. 4660 */ 4661 static 4662 vm_page_t 4663 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit) 4664 { 4665 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK; 4666 vm_page_t p; 4667 4668 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4669 4670 /* 4671 * Try a normal allocation first. 4672 */ 4673 p = vm_page_alloc(obj, pg, vmflags); 4674 if (p) 4675 return(p); 4676 if (vm_page_lookup(obj, pg)) 4677 return(NULL); 4678 vm_pageout_deficit += deficit; 4679 4680 /* 4681 * Try again, digging into the system reserve. 4682 * 4683 * Trying to recover pages from the buffer cache here can deadlock 4684 * against other threads trying to busy underlying pages so we 4685 * depend on the code in brelse() and bqrelse() to free/cache the 4686 * underlying buffer cache pages when memory is low. 4687 */ 4688 if (curthread->td_flags & TDF_SYSTHREAD) 4689 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT; 4690 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS) 4691 vmflags |= 0; 4692 else 4693 vmflags |= VM_ALLOC_SYSTEM; 4694 4695 /*recoverbufpages();*/ 4696 p = vm_page_alloc(obj, pg, vmflags); 4697 if (p) 4698 return(p); 4699 if (vm_page_lookup(obj, pg)) 4700 return(NULL); 4701 4702 /* 4703 * Wait for memory to free up and try again 4704 */ 4705 if (vm_page_count_severe()) 4706 ++lowmempgallocs; 4707 vm_wait(hz / 20 + 1); 4708 4709 p = vm_page_alloc(obj, pg, vmflags); 4710 if (p) 4711 return(p); 4712 if (vm_page_lookup(obj, pg)) 4713 return(NULL); 4714 4715 /* 4716 * Ok, now we are really in trouble. 4717 */ 4718 if (bootverbose) { 4719 static struct krate biokrate = { .freq = 1 }; 4720 krateprintf(&biokrate, 4721 "Warning: bio_page_alloc: memory exhausted " 4722 "during buffer cache page allocation from %s\n", 4723 curthread->td_comm); 4724 } 4725 if (curthread->td_flags & TDF_SYSTHREAD) 4726 vm_wait(hz / 20 + 1); 4727 else 4728 vm_wait(hz / 2 + 1); 4729 return (NULL); 4730 } 4731 4732 /* 4733 * vm_hold_free_pages: 4734 * 4735 * Return pages associated with the buffer back to the VM system. 4736 * 4737 * The range of pages underlying the buffer's address space will 4738 * be unmapped and un-wired. 4739 */ 4740 void 4741 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4742 { 4743 vm_offset_t pg; 4744 vm_page_t p; 4745 int index, newnpages; 4746 4747 from = round_page(from); 4748 to = round_page(to); 4749 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4750 newnpages = index; 4751 4752 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4753 p = bp->b_xio.xio_pages[index]; 4754 if (p && (index < bp->b_xio.xio_npages)) { 4755 if (p->busy) { 4756 kprintf("vm_hold_free_pages: doffset: %lld, " 4757 "loffset: %lld\n", 4758 (long long)bp->b_bio2.bio_offset, 4759 (long long)bp->b_loffset); 4760 } 4761 bp->b_xio.xio_pages[index] = NULL; 4762 pmap_kremove_noinval(pg); 4763 vm_page_busy_wait(p, FALSE, "vmhldpg"); 4764 vm_page_unwire(p, 0); 4765 vm_page_free(p); 4766 } 4767 } 4768 pmap_invalidate_range(&kernel_pmap, from, to); 4769 bp->b_xio.xio_npages = newnpages; 4770 } 4771 4772 /* 4773 * Scan all buffers in the system and issue the callback. 4774 */ 4775 int 4776 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4777 { 4778 int count = 0; 4779 int error; 4780 long n; 4781 4782 for (n = 0; n < nbuf; ++n) { 4783 if ((error = callback(&buf[n], info)) < 0) { 4784 count = error; 4785 break; 4786 } 4787 count += error; 4788 } 4789 return (count); 4790 } 4791 4792 /* 4793 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4794 * completion to the master buffer. 4795 */ 4796 static void 4797 nestiobuf_iodone(struct bio *bio) 4798 { 4799 struct bio *mbio; 4800 struct buf *mbp, *bp; 4801 struct devstat *stats; 4802 int error; 4803 int donebytes; 4804 4805 bp = bio->bio_buf; 4806 mbio = bio->bio_caller_info1.ptr; 4807 stats = bio->bio_caller_info2.ptr; 4808 mbp = mbio->bio_buf; 4809 4810 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4811 KKASSERT(mbp != bp); 4812 4813 error = bp->b_error; 4814 if (bp->b_error == 0 && 4815 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4816 /* 4817 * Not all got transfered, raise an error. We have no way to 4818 * propagate these conditions to mbp. 4819 */ 4820 error = EIO; 4821 } 4822 4823 donebytes = bp->b_bufsize; 4824 4825 relpbuf(bp, NULL); 4826 4827 nestiobuf_done(mbio, donebytes, error, stats); 4828 } 4829 4830 void 4831 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4832 { 4833 struct buf *mbp; 4834 4835 mbp = mbio->bio_buf; 4836 4837 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4838 4839 /* 4840 * If an error occured, propagate it to the master buffer. 4841 * 4842 * Several biodone()s may wind up running concurrently so 4843 * use an atomic op to adjust b_flags. 4844 */ 4845 if (error) { 4846 mbp->b_error = error; 4847 atomic_set_int(&mbp->b_flags, B_ERROR); 4848 } 4849 4850 /* 4851 * Decrement the operations in progress counter and terminate the 4852 * I/O if this was the last bit. 4853 */ 4854 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4855 mbp->b_resid = 0; 4856 if (stats) 4857 devstat_end_transaction_buf(stats, mbp); 4858 biodone(mbio); 4859 } 4860 } 4861 4862 /* 4863 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4864 * the mbio from being biodone()'d while we are still adding sub-bios to 4865 * it. 4866 */ 4867 void 4868 nestiobuf_init(struct bio *bio) 4869 { 4870 bio->bio_driver_info = (void *)1; 4871 } 4872 4873 /* 4874 * The BIOs added to the nestedio have already been started, remove the 4875 * count that placeheld our mbio and biodone() it if the count would 4876 * transition to 0. 4877 */ 4878 void 4879 nestiobuf_start(struct bio *mbio) 4880 { 4881 struct buf *mbp = mbio->bio_buf; 4882 4883 /* 4884 * Decrement the operations in progress counter and terminate the 4885 * I/O if this was the last bit. 4886 */ 4887 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4888 if (mbp->b_flags & B_ERROR) 4889 mbp->b_resid = mbp->b_bcount; 4890 else 4891 mbp->b_resid = 0; 4892 biodone(mbio); 4893 } 4894 } 4895 4896 /* 4897 * Set an intermediate error prior to calling nestiobuf_start() 4898 */ 4899 void 4900 nestiobuf_error(struct bio *mbio, int error) 4901 { 4902 struct buf *mbp = mbio->bio_buf; 4903 4904 if (error) { 4905 mbp->b_error = error; 4906 atomic_set_int(&mbp->b_flags, B_ERROR); 4907 } 4908 } 4909 4910 /* 4911 * nestiobuf_add: setup a "nested" buffer. 4912 * 4913 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4914 * => 'bp' should be a buffer allocated by getiobuf. 4915 * => 'offset' is a byte offset in the master buffer. 4916 * => 'size' is a size in bytes of this nested buffer. 4917 */ 4918 void 4919 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4920 { 4921 struct buf *mbp = mbio->bio_buf; 4922 struct vnode *vp = mbp->b_vp; 4923 4924 KKASSERT(mbp->b_bcount >= offset + size); 4925 4926 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4927 4928 /* kernel needs to own the lock for it to be released in biodone */ 4929 BUF_KERNPROC(bp); 4930 bp->b_vp = vp; 4931 bp->b_cmd = mbp->b_cmd; 4932 bp->b_bio1.bio_done = nestiobuf_iodone; 4933 bp->b_data = (char *)mbp->b_data + offset; 4934 bp->b_resid = bp->b_bcount = size; 4935 bp->b_bufsize = bp->b_bcount; 4936 4937 bp->b_bio1.bio_track = NULL; 4938 bp->b_bio1.bio_caller_info1.ptr = mbio; 4939 bp->b_bio1.bio_caller_info2.ptr = stats; 4940 } 4941 4942 #ifdef DDB 4943 4944 DB_SHOW_COMMAND(buffer, db_show_buffer) 4945 { 4946 /* get args */ 4947 struct buf *bp = (struct buf *)addr; 4948 4949 if (!have_addr) { 4950 db_printf("usage: show buffer <addr>\n"); 4951 return; 4952 } 4953 4954 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4955 db_printf("b_cmd = %d\n", bp->b_cmd); 4956 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4957 "b_resid = %d\n, b_data = %p, " 4958 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4959 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4960 bp->b_data, 4961 (long long)bp->b_bio2.bio_offset, 4962 (long long)(bp->b_bio2.bio_next ? 4963 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4964 if (bp->b_xio.xio_npages) { 4965 int i; 4966 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4967 bp->b_xio.xio_npages); 4968 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4969 vm_page_t m; 4970 m = bp->b_xio.xio_pages[i]; 4971 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4972 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4973 if ((i + 1) < bp->b_xio.xio_npages) 4974 db_printf(","); 4975 } 4976 db_printf("\n"); 4977 } 4978 } 4979 #endif /* DDB */ 4980