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 <sys/mplock2.h> 64 #include <vm/vm_page2.h> 65 66 #include "opt_ddb.h" 67 #ifdef DDB 68 #include <ddb/ddb.h> 69 #endif 70 71 /* 72 * Buffer queues. 73 */ 74 enum bufq_type { 75 BQUEUE_NONE, /* not on any queue */ 76 BQUEUE_LOCKED, /* locked buffers */ 77 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 78 BQUEUE_DIRTY, /* B_DELWRI buffers */ 79 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */ 80 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 81 BQUEUE_EMPTY, /* empty buffer headers */ 82 83 BUFFER_QUEUES /* number of buffer queues */ 84 }; 85 86 typedef enum bufq_type bufq_type_t; 87 88 #define BD_WAKE_SIZE 16384 89 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1) 90 91 TAILQ_HEAD(bqueues, buf); 92 93 struct bufpcpu { 94 struct spinlock spin; 95 struct bqueues bufqueues[BUFFER_QUEUES]; 96 } __cachealign; 97 98 struct bufpcpu bufpcpu[MAXCPU]; 99 100 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 101 102 struct buf *buf; /* buffer header pool */ 103 104 static void vfs_clean_pages(struct buf *bp); 105 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m); 106 #if 0 107 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m); 108 #endif 109 static void vfs_vmio_release(struct buf *bp); 110 static int flushbufqueues(struct buf *marker, bufq_type_t q); 111 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj, 112 vm_pindex_t pg, int deficit); 113 114 static void bd_signal(long totalspace); 115 static void buf_daemon(void); 116 static void buf_daemon_hw(void); 117 118 /* 119 * bogus page -- for I/O to/from partially complete buffers 120 * this is a temporary solution to the problem, but it is not 121 * really that bad. it would be better to split the buffer 122 * for input in the case of buffers partially already in memory, 123 * but the code is intricate enough already. 124 */ 125 vm_page_t bogus_page; 126 127 /* 128 * These are all static, but make the ones we export globals so we do 129 * not need to use compiler magic. 130 */ 131 long bufspace; /* locked by buffer_map */ 132 long maxbufspace; 133 static long bufmallocspace; /* atomic ops */ 134 long maxbufmallocspace, lobufspace, hibufspace; 135 static long bufreusecnt, bufdefragcnt, buffreekvacnt; 136 static long lorunningspace; 137 static long hirunningspace; 138 static long dirtykvaspace; /* atomic */ 139 long dirtybufspace; /* atomic (global for systat) */ 140 static long dirtybufcount; /* atomic */ 141 static long dirtybufspacehw; /* atomic */ 142 static long dirtybufcounthw; /* atomic */ 143 static long runningbufspace; /* atomic */ 144 static long runningbufcount; /* atomic */ 145 long lodirtybufspace; 146 long hidirtybufspace; 147 static int getnewbufcalls; 148 static int getnewbufrestarts; 149 static int recoverbufcalls; 150 static int needsbuffer; /* atomic */ 151 static int runningbufreq; /* atomic */ 152 static int bd_request; /* atomic */ 153 static int bd_request_hw; /* atomic */ 154 static u_int bd_wake_ary[BD_WAKE_SIZE]; 155 static u_int bd_wake_index; 156 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 157 static int debug_commit; 158 static int debug_bufbio; 159 160 static struct thread *bufdaemon_td; 161 static struct thread *bufdaemonhw_td; 162 static u_int lowmempgallocs; 163 static u_int lowmempgfails; 164 static u_int flushperqueue = 1024; 165 166 /* 167 * Sysctls for operational control of the buffer cache. 168 */ 169 SYSCTL_UINT(_vfs, OID_AUTO, flushperqueue, CTLFLAG_RW, &flushperqueue, 0, 170 "Number of buffers to flush from each per-cpu queue"); 171 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 172 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 173 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 174 "High watermark used to trigger explicit flushing of dirty buffers"); 175 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 176 "Minimum amount of buffer space required for active I/O"); 177 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 178 "Maximum amount of buffer space to usable for active I/O"); 179 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 180 "Page allocations done during periods of very low free memory"); 181 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 182 "Page allocations which failed during periods of very low free memory"); 183 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 184 "Recycle pages to active or inactive queue transition pt 0-64"); 185 /* 186 * Sysctls determining current state of the buffer cache. 187 */ 188 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 189 "Total number of buffers in buffer cache"); 190 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0, 191 "KVA reserved by dirty buffers (all)"); 192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 193 "Pending bytes of dirty buffers (all)"); 194 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 195 "Pending bytes of dirty buffers (heavy weight)"); 196 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 197 "Pending number of dirty buffers"); 198 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 199 "Pending number of dirty buffers (heavy weight)"); 200 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 201 "I/O bytes currently in progress due to asynchronous writes"); 202 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 203 "I/O buffers currently in progress due to asynchronous writes"); 204 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 205 "Hard limit on maximum amount of memory usable for buffer space"); 206 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 207 "Soft limit on maximum amount of memory usable for buffer space"); 208 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 209 "Minimum amount of memory to reserve for system buffer space"); 210 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 211 "Amount of memory available for buffers"); 212 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 213 0, "Maximum amount of memory reserved for buffers using malloc"); 214 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 215 "Amount of memory left for buffers using malloc-scheme"); 216 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 217 "New buffer header acquisition requests"); 218 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 219 0, "New buffer header acquisition restarts"); 220 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 221 "Recover VM space in an emergency"); 222 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 223 "Buffer acquisition restarts due to fragmented buffer map"); 224 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 225 "Amount of time KVA space was deallocated in an arbitrary buffer"); 226 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 227 "Amount of time buffer re-use operations were successful"); 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 * bufspacewakeup: 242 * 243 * Called when buffer space is potentially available for recovery. 244 * getnewbuf() will block on this flag when it is unable to free 245 * sufficient buffer space. Buffer space becomes recoverable when 246 * bp's get placed back in the queues. 247 */ 248 static __inline void 249 bufspacewakeup(void) 250 { 251 /* 252 * If someone is waiting for BUF space, wake them up. Even 253 * though we haven't freed the kva space yet, the waiting 254 * process will be able to now. 255 */ 256 for (;;) { 257 int flags = needsbuffer; 258 cpu_ccfence(); 259 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0) 260 break; 261 if (atomic_cmpset_int(&needsbuffer, flags, 262 flags & ~VFS_BIO_NEED_BUFSPACE)) { 263 wakeup(&needsbuffer); 264 break; 265 } 266 /* retry */ 267 } 268 } 269 270 /* 271 * runningbufwakeup: 272 * 273 * Accounting for I/O in progress. 274 * 275 */ 276 static __inline void 277 runningbufwakeup(struct buf *bp) 278 { 279 long totalspace; 280 long flags; 281 282 if ((totalspace = bp->b_runningbufspace) != 0) { 283 atomic_add_long(&runningbufspace, -totalspace); 284 atomic_add_long(&runningbufcount, -1); 285 bp->b_runningbufspace = 0; 286 287 /* 288 * see waitrunningbufspace() for limit test. 289 */ 290 for (;;) { 291 flags = runningbufreq; 292 cpu_ccfence(); 293 if (flags == 0) 294 break; 295 if (atomic_cmpset_int(&runningbufreq, flags, 0)) { 296 wakeup(&runningbufreq); 297 break; 298 } 299 /* retry */ 300 } 301 bd_signal(totalspace); 302 } 303 } 304 305 /* 306 * bufcountwakeup: 307 * 308 * Called when a buffer has been added to one of the free queues to 309 * account for the buffer and to wakeup anyone waiting for free buffers. 310 * This typically occurs when large amounts of metadata are being handled 311 * by the buffer cache ( else buffer space runs out first, usually ). 312 */ 313 static __inline void 314 bufcountwakeup(void) 315 { 316 long flags; 317 318 for (;;) { 319 flags = needsbuffer; 320 if (flags == 0) 321 break; 322 if (atomic_cmpset_int(&needsbuffer, flags, 323 (flags & ~VFS_BIO_NEED_ANY))) { 324 wakeup(&needsbuffer); 325 break; 326 } 327 /* retry */ 328 } 329 } 330 331 /* 332 * waitrunningbufspace() 333 * 334 * If runningbufspace exceeds 4/6 hirunningspace we block until 335 * runningbufspace drops to 3/6 hirunningspace. We also block if another 336 * thread blocked here in order to be fair, even if runningbufspace 337 * is now lower than the limit. 338 * 339 * The caller may be using this function to block in a tight loop, we 340 * must block while runningbufspace is greater than at least 341 * hirunningspace * 3 / 6. 342 */ 343 void 344 waitrunningbufspace(void) 345 { 346 long limit = hirunningspace * 4 / 6; 347 long flags; 348 349 while (runningbufspace > limit || runningbufreq) { 350 tsleep_interlock(&runningbufreq, 0); 351 flags = atomic_fetchadd_int(&runningbufreq, 1); 352 if (runningbufspace > limit || flags) 353 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz); 354 } 355 } 356 357 /* 358 * buf_dirty_count_severe: 359 * 360 * Return true if we have too many dirty buffers. 361 */ 362 int 363 buf_dirty_count_severe(void) 364 { 365 return (runningbufspace + dirtykvaspace >= hidirtybufspace || 366 dirtybufcount >= nbuf / 2); 367 } 368 369 /* 370 * Return true if the amount of running I/O is severe and BIOQ should 371 * start bursting. 372 */ 373 int 374 buf_runningbufspace_severe(void) 375 { 376 return (runningbufspace >= hirunningspace * 4 / 6); 377 } 378 379 /* 380 * vfs_buf_test_cache: 381 * 382 * Called when a buffer is extended. This function clears the B_CACHE 383 * bit if the newly extended portion of the buffer does not contain 384 * valid data. 385 * 386 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 387 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 388 * them while a clean buffer was present. 389 */ 390 static __inline__ 391 void 392 vfs_buf_test_cache(struct buf *bp, 393 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 394 vm_page_t m) 395 { 396 if (bp->b_flags & B_CACHE) { 397 int base = (foff + off) & PAGE_MASK; 398 if (vm_page_is_valid(m, base, size) == 0) 399 bp->b_flags &= ~B_CACHE; 400 } 401 } 402 403 /* 404 * bd_speedup() 405 * 406 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 407 * low water mark. 408 */ 409 static __inline__ 410 void 411 bd_speedup(void) 412 { 413 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2) 414 return; 415 416 if (bd_request == 0 && 417 (dirtykvaspace > lodirtybufspace / 2 || 418 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 419 if (atomic_fetchadd_int(&bd_request, 1) == 0) 420 wakeup(&bd_request); 421 } 422 if (bd_request_hw == 0 && 423 (dirtykvaspace > lodirtybufspace / 2 || 424 dirtybufcounthw >= nbuf / 2)) { 425 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0) 426 wakeup(&bd_request_hw); 427 } 428 } 429 430 /* 431 * bd_heatup() 432 * 433 * Get the buf_daemon heated up when the number of running and dirty 434 * buffers exceeds the mid-point. 435 * 436 * Return the total number of dirty bytes past the second mid point 437 * as a measure of how much excess dirty data there is in the system. 438 */ 439 long 440 bd_heatup(void) 441 { 442 long mid1; 443 long mid2; 444 long totalspace; 445 446 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 447 448 totalspace = runningbufspace + dirtykvaspace; 449 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 450 bd_speedup(); 451 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 452 if (totalspace >= mid2) 453 return(totalspace - mid2); 454 } 455 return(0); 456 } 457 458 /* 459 * bd_wait() 460 * 461 * Wait for the buffer cache to flush (totalspace) bytes worth of 462 * buffers, then return. 463 * 464 * Regardless this function blocks while the number of dirty buffers 465 * exceeds hidirtybufspace. 466 */ 467 void 468 bd_wait(long totalspace) 469 { 470 u_int i; 471 u_int j; 472 u_int mi; 473 int count; 474 475 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 476 return; 477 478 while (totalspace > 0) { 479 bd_heatup(); 480 481 /* 482 * Order is important. Suppliers adjust bd_wake_index after 483 * updating runningbufspace/dirtykvaspace. We want to fetch 484 * bd_wake_index before accessing. Any error should thus 485 * be in our favor. 486 */ 487 i = atomic_fetchadd_int(&bd_wake_index, 0); 488 if (totalspace > runningbufspace + dirtykvaspace) 489 totalspace = runningbufspace + dirtykvaspace; 490 count = totalspace / BKVASIZE; 491 if (count >= BD_WAKE_SIZE / 2) 492 count = BD_WAKE_SIZE / 2; 493 i = i + count; 494 mi = i & BD_WAKE_MASK; 495 496 /* 497 * This is not a strict interlock, so we play a bit loose 498 * with locking access to dirtybufspace*. We have to re-check 499 * bd_wake_index to ensure that it hasn't passed us. 500 */ 501 tsleep_interlock(&bd_wake_ary[mi], 0); 502 atomic_add_int(&bd_wake_ary[mi], 1); 503 j = atomic_fetchadd_int(&bd_wake_index, 0); 504 if ((int)(i - j) >= 0) 505 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz); 506 507 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace; 508 } 509 } 510 511 /* 512 * bd_signal() 513 * 514 * This function is called whenever runningbufspace or dirtykvaspace 515 * is reduced. Track threads waiting for run+dirty buffer I/O 516 * complete. 517 */ 518 static void 519 bd_signal(long totalspace) 520 { 521 u_int i; 522 523 if (totalspace > 0) { 524 if (totalspace > BKVASIZE * BD_WAKE_SIZE) 525 totalspace = BKVASIZE * BD_WAKE_SIZE; 526 while (totalspace > 0) { 527 i = atomic_fetchadd_int(&bd_wake_index, 1); 528 i &= BD_WAKE_MASK; 529 if (atomic_readandclear_int(&bd_wake_ary[i])) 530 wakeup(&bd_wake_ary[i]); 531 totalspace -= BKVASIZE; 532 } 533 } 534 } 535 536 /* 537 * BIO tracking support routines. 538 * 539 * Release a ref on a bio_track. Wakeup requests are atomically released 540 * along with the last reference so bk_active will never wind up set to 541 * only 0x80000000. 542 */ 543 static 544 void 545 bio_track_rel(struct bio_track *track) 546 { 547 int active; 548 int desired; 549 550 /* 551 * Shortcut 552 */ 553 active = track->bk_active; 554 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 555 return; 556 557 /* 558 * Full-on. Note that the wait flag is only atomically released on 559 * the 1->0 count transition. 560 * 561 * We check for a negative count transition using bit 30 since bit 31 562 * has a different meaning. 563 */ 564 for (;;) { 565 desired = (active & 0x7FFFFFFF) - 1; 566 if (desired) 567 desired |= active & 0x80000000; 568 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 569 if (desired & 0x40000000) 570 panic("bio_track_rel: bad count: %p", track); 571 if (active & 0x80000000) 572 wakeup(track); 573 break; 574 } 575 active = track->bk_active; 576 } 577 } 578 579 /* 580 * Wait for the tracking count to reach 0. 581 * 582 * Use atomic ops such that the wait flag is only set atomically when 583 * bk_active is non-zero. 584 */ 585 int 586 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 587 { 588 int active; 589 int desired; 590 int error; 591 592 /* 593 * Shortcut 594 */ 595 if (track->bk_active == 0) 596 return(0); 597 598 /* 599 * Full-on. Note that the wait flag may only be atomically set if 600 * the active count is non-zero. 601 * 602 * NOTE: We cannot optimize active == desired since a wakeup could 603 * clear active prior to our tsleep_interlock(). 604 */ 605 error = 0; 606 while ((active = track->bk_active) != 0) { 607 cpu_ccfence(); 608 desired = active | 0x80000000; 609 tsleep_interlock(track, slp_flags); 610 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 611 error = tsleep(track, slp_flags | PINTERLOCKED, 612 "trwait", slp_timo); 613 if (error) 614 break; 615 } 616 } 617 return (error); 618 } 619 620 /* 621 * bufinit: 622 * 623 * Load time initialisation of the buffer cache, called from machine 624 * dependant initialization code. 625 */ 626 static 627 void 628 bufinit(void *dummy __unused) 629 { 630 struct bufpcpu *pcpu; 631 struct buf *bp; 632 vm_offset_t bogus_offset; 633 int i; 634 int j; 635 long n; 636 637 /* next, make a null set of free lists */ 638 for (i = 0; i < ncpus; ++i) { 639 pcpu = &bufpcpu[i]; 640 spin_init(&pcpu->spin, "bufinit"); 641 for (j = 0; j < BUFFER_QUEUES; j++) 642 TAILQ_INIT(&pcpu->bufqueues[j]); 643 } 644 645 /* finally, initialize each buffer header and stick on empty q */ 646 i = 0; 647 pcpu = &bufpcpu[i]; 648 649 for (n = 0; n < nbuf; n++) { 650 bp = &buf[n]; 651 bzero(bp, sizeof *bp); 652 bp->b_flags = B_INVAL; /* we're just an empty header */ 653 bp->b_cmd = BUF_CMD_DONE; 654 bp->b_qindex = BQUEUE_EMPTY; 655 bp->b_qcpu = i; 656 initbufbio(bp); 657 xio_init(&bp->b_xio); 658 buf_dep_init(bp); 659 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 660 bp, b_freelist); 661 662 i = (i + 1) % ncpus; 663 pcpu = &bufpcpu[i]; 664 } 665 666 /* 667 * maxbufspace is the absolute maximum amount of buffer space we are 668 * allowed to reserve in KVM and in real terms. The absolute maximum 669 * is nominally used by buf_daemon. hibufspace is the nominal maximum 670 * used by most other processes. The differential is required to 671 * ensure that buf_daemon is able to run when other processes might 672 * be blocked waiting for buffer space. 673 * 674 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 675 * this may result in KVM fragmentation which is not handled optimally 676 * by the system. 677 */ 678 maxbufspace = nbuf * BKVASIZE; 679 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 680 lobufspace = hibufspace - MAXBSIZE; 681 682 lorunningspace = 512 * 1024; 683 /* hirunningspace -- see below */ 684 685 /* 686 * Limit the amount of malloc memory since it is wired permanently 687 * into the kernel space. Even though this is accounted for in 688 * the buffer allocation, we don't want the malloced region to grow 689 * uncontrolled. The malloc scheme improves memory utilization 690 * significantly on average (small) directories. 691 */ 692 maxbufmallocspace = hibufspace / 20; 693 694 /* 695 * Reduce the chance of a deadlock occuring by limiting the number 696 * of delayed-write dirty buffers we allow to stack up. 697 * 698 * We don't want too much actually queued to the device at once 699 * (XXX this needs to be per-mount!), because the buffers will 700 * wind up locked for a very long period of time while the I/O 701 * drains. 702 */ 703 hidirtybufspace = hibufspace / 2; /* dirty + running */ 704 hirunningspace = hibufspace / 16; /* locked & queued to device */ 705 if (hirunningspace < 1024 * 1024) 706 hirunningspace = 1024 * 1024; 707 708 dirtykvaspace = 0; 709 dirtybufspace = 0; 710 dirtybufspacehw = 0; 711 712 lodirtybufspace = hidirtybufspace / 2; 713 714 /* 715 * Maximum number of async ops initiated per buf_daemon loop. This is 716 * somewhat of a hack at the moment, we really need to limit ourselves 717 * based on the number of bytes of I/O in-transit that were initiated 718 * from buf_daemon. 719 */ 720 721 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 722 vm_object_hold(&kernel_object); 723 bogus_page = vm_page_alloc(&kernel_object, 724 (bogus_offset >> PAGE_SHIFT), 725 VM_ALLOC_NORMAL); 726 vm_object_drop(&kernel_object); 727 vmstats.v_wire_count++; 728 729 } 730 731 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL); 732 733 /* 734 * Initialize the embedded bio structures, typically used by 735 * deprecated code which tries to allocate its own struct bufs. 736 */ 737 void 738 initbufbio(struct buf *bp) 739 { 740 bp->b_bio1.bio_buf = bp; 741 bp->b_bio1.bio_prev = NULL; 742 bp->b_bio1.bio_offset = NOOFFSET; 743 bp->b_bio1.bio_next = &bp->b_bio2; 744 bp->b_bio1.bio_done = NULL; 745 bp->b_bio1.bio_flags = 0; 746 747 bp->b_bio2.bio_buf = bp; 748 bp->b_bio2.bio_prev = &bp->b_bio1; 749 bp->b_bio2.bio_offset = NOOFFSET; 750 bp->b_bio2.bio_next = NULL; 751 bp->b_bio2.bio_done = NULL; 752 bp->b_bio2.bio_flags = 0; 753 754 BUF_LOCKINIT(bp); 755 } 756 757 /* 758 * Reinitialize the embedded bio structures as well as any additional 759 * translation cache layers. 760 */ 761 void 762 reinitbufbio(struct buf *bp) 763 { 764 struct bio *bio; 765 766 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 767 bio->bio_done = NULL; 768 bio->bio_offset = NOOFFSET; 769 } 770 } 771 772 /* 773 * Undo the effects of an initbufbio(). 774 */ 775 void 776 uninitbufbio(struct buf *bp) 777 { 778 dsched_buf_exit(bp); 779 BUF_LOCKFREE(bp); 780 } 781 782 /* 783 * Push another BIO layer onto an existing BIO and return it. The new 784 * BIO layer may already exist, holding cached translation data. 785 */ 786 struct bio * 787 push_bio(struct bio *bio) 788 { 789 struct bio *nbio; 790 791 if ((nbio = bio->bio_next) == NULL) { 792 int index = bio - &bio->bio_buf->b_bio_array[0]; 793 if (index >= NBUF_BIO - 1) { 794 panic("push_bio: too many layers %d for bp %p", 795 index, bio->bio_buf); 796 } 797 nbio = &bio->bio_buf->b_bio_array[index + 1]; 798 bio->bio_next = nbio; 799 nbio->bio_prev = bio; 800 nbio->bio_buf = bio->bio_buf; 801 nbio->bio_offset = NOOFFSET; 802 nbio->bio_done = NULL; 803 nbio->bio_next = NULL; 804 } 805 KKASSERT(nbio->bio_done == NULL); 806 return(nbio); 807 } 808 809 /* 810 * Pop a BIO translation layer, returning the previous layer. The 811 * must have been previously pushed. 812 */ 813 struct bio * 814 pop_bio(struct bio *bio) 815 { 816 return(bio->bio_prev); 817 } 818 819 void 820 clearbiocache(struct bio *bio) 821 { 822 while (bio) { 823 bio->bio_offset = NOOFFSET; 824 bio = bio->bio_next; 825 } 826 } 827 828 /* 829 * bfreekva: 830 * 831 * Free the KVA allocation for buffer 'bp'. 832 * 833 * Must be called from a critical section as this is the only locking for 834 * buffer_map. 835 * 836 * Since this call frees up buffer space, we call bufspacewakeup(). 837 */ 838 static void 839 bfreekva(struct buf *bp) 840 { 841 int count; 842 843 if (bp->b_kvasize) { 844 ++buffreekvacnt; 845 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 846 vm_map_lock(&buffer_map); 847 bufspace -= bp->b_kvasize; 848 vm_map_delete(&buffer_map, 849 (vm_offset_t) bp->b_kvabase, 850 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 851 &count 852 ); 853 vm_map_unlock(&buffer_map); 854 vm_map_entry_release(count); 855 bp->b_kvasize = 0; 856 bp->b_kvabase = NULL; 857 bufspacewakeup(); 858 } 859 } 860 861 /* 862 * Remove the buffer from the appropriate free list. 863 * (caller must be locked) 864 */ 865 static __inline void 866 _bremfree(struct buf *bp) 867 { 868 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 869 870 if (bp->b_qindex != BQUEUE_NONE) { 871 KASSERT(BUF_REFCNTNB(bp) == 1, 872 ("bremfree: bp %p not locked",bp)); 873 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist); 874 bp->b_qindex = BQUEUE_NONE; 875 } else { 876 if (BUF_REFCNTNB(bp) <= 1) 877 panic("bremfree: removing a buffer not on a queue"); 878 } 879 } 880 881 /* 882 * bremfree() - must be called with a locked buffer 883 */ 884 void 885 bremfree(struct buf *bp) 886 { 887 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 888 889 spin_lock(&pcpu->spin); 890 _bremfree(bp); 891 spin_unlock(&pcpu->spin); 892 } 893 894 /* 895 * bremfree_locked - must be called with pcpu->spin locked 896 */ 897 static void 898 bremfree_locked(struct buf *bp) 899 { 900 _bremfree(bp); 901 } 902 903 /* 904 * This version of bread issues any required I/O asyncnronously and 905 * makes a callback on completion. 906 * 907 * The callback must check whether BIO_DONE is set in the bio and issue 908 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 909 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 910 */ 911 void 912 breadcb(struct vnode *vp, off_t loffset, int size, 913 void (*func)(struct bio *), void *arg) 914 { 915 struct buf *bp; 916 917 bp = getblk(vp, loffset, size, 0, 0); 918 919 /* if not found in cache, do some I/O */ 920 if ((bp->b_flags & B_CACHE) == 0) { 921 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 922 bp->b_cmd = BUF_CMD_READ; 923 bp->b_bio1.bio_done = func; 924 bp->b_bio1.bio_caller_info1.ptr = arg; 925 vfs_busy_pages(vp, bp); 926 BUF_KERNPROC(bp); 927 vn_strategy(vp, &bp->b_bio1); 928 } else if (func) { 929 /* 930 * Since we are issuing the callback synchronously it cannot 931 * race the BIO_DONE, so no need for atomic ops here. 932 */ 933 /*bp->b_bio1.bio_done = func;*/ 934 bp->b_bio1.bio_caller_info1.ptr = arg; 935 bp->b_bio1.bio_flags |= BIO_DONE; 936 func(&bp->b_bio1); 937 } else { 938 bqrelse(bp); 939 } 940 } 941 942 /* 943 * breadnx() - Terminal function for bread() and breadn(). 944 * 945 * This function will start asynchronous I/O on read-ahead blocks as well 946 * as satisfy the primary request. 947 * 948 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is 949 * set, the buffer is valid and we do not have to do anything. 950 */ 951 int 952 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset, 953 int *rabsize, int cnt, struct buf **bpp) 954 { 955 struct buf *bp, *rabp; 956 int i; 957 int rv = 0, readwait = 0; 958 959 if (*bpp) 960 bp = *bpp; 961 else 962 *bpp = bp = getblk(vp, loffset, size, 0, 0); 963 964 /* if not found in cache, do some I/O */ 965 if ((bp->b_flags & B_CACHE) == 0) { 966 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 967 bp->b_cmd = BUF_CMD_READ; 968 bp->b_bio1.bio_done = biodone_sync; 969 bp->b_bio1.bio_flags |= BIO_SYNC; 970 vfs_busy_pages(vp, bp); 971 vn_strategy(vp, &bp->b_bio1); 972 ++readwait; 973 } 974 975 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 976 if (inmem(vp, *raoffset)) 977 continue; 978 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 979 980 if ((rabp->b_flags & B_CACHE) == 0) { 981 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 982 rabp->b_cmd = BUF_CMD_READ; 983 vfs_busy_pages(vp, rabp); 984 BUF_KERNPROC(rabp); 985 vn_strategy(vp, &rabp->b_bio1); 986 } else { 987 brelse(rabp); 988 } 989 } 990 if (readwait) 991 rv = biowait(&bp->b_bio1, "biord"); 992 return (rv); 993 } 994 995 /* 996 * bwrite: 997 * 998 * Synchronous write, waits for completion. 999 * 1000 * Write, release buffer on completion. (Done by iodone 1001 * if async). Do not bother writing anything if the buffer 1002 * is invalid. 1003 * 1004 * Note that we set B_CACHE here, indicating that buffer is 1005 * fully valid and thus cacheable. This is true even of NFS 1006 * now so we set it generally. This could be set either here 1007 * or in biodone() since the I/O is synchronous. We put it 1008 * here. 1009 */ 1010 int 1011 bwrite(struct buf *bp) 1012 { 1013 int error; 1014 1015 if (bp->b_flags & B_INVAL) { 1016 brelse(bp); 1017 return (0); 1018 } 1019 if (BUF_REFCNTNB(bp) == 0) 1020 panic("bwrite: buffer is not busy???"); 1021 1022 /* 1023 * NOTE: We no longer mark the buffer clear prior to the vn_strategy() 1024 * call because it will remove the buffer from the vnode's 1025 * dirty buffer list prematurely and possibly cause filesystem 1026 * checks to race buffer flushes. This is now handled in 1027 * bpdone(). 1028 * 1029 * bundirty(bp); REMOVED 1030 */ 1031 1032 bp->b_flags &= ~(B_ERROR | B_EINTR); 1033 bp->b_flags |= B_CACHE; 1034 bp->b_cmd = BUF_CMD_WRITE; 1035 bp->b_bio1.bio_done = biodone_sync; 1036 bp->b_bio1.bio_flags |= BIO_SYNC; 1037 vfs_busy_pages(bp->b_vp, bp); 1038 1039 /* 1040 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1041 * valid for vnode-backed buffers. 1042 */ 1043 bsetrunningbufspace(bp, bp->b_bufsize); 1044 vn_strategy(bp->b_vp, &bp->b_bio1); 1045 error = biowait(&bp->b_bio1, "biows"); 1046 brelse(bp); 1047 1048 return (error); 1049 } 1050 1051 /* 1052 * bawrite: 1053 * 1054 * Asynchronous write. Start output on a buffer, but do not wait for 1055 * it to complete. The buffer is released when the output completes. 1056 * 1057 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1058 * B_INVAL buffers. Not us. 1059 */ 1060 void 1061 bawrite(struct buf *bp) 1062 { 1063 if (bp->b_flags & B_INVAL) { 1064 brelse(bp); 1065 return; 1066 } 1067 if (BUF_REFCNTNB(bp) == 0) 1068 panic("bawrite: buffer is not busy???"); 1069 1070 /* 1071 * NOTE: We no longer mark the buffer clear prior to the vn_strategy() 1072 * call because it will remove the buffer from the vnode's 1073 * dirty buffer list prematurely and possibly cause filesystem 1074 * checks to race buffer flushes. This is now handled in 1075 * bpdone(). 1076 * 1077 * bundirty(bp); REMOVED 1078 */ 1079 bp->b_flags &= ~(B_ERROR | B_EINTR); 1080 bp->b_flags |= B_CACHE; 1081 bp->b_cmd = BUF_CMD_WRITE; 1082 KKASSERT(bp->b_bio1.bio_done == NULL); 1083 vfs_busy_pages(bp->b_vp, bp); 1084 1085 /* 1086 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1087 * valid for vnode-backed buffers. 1088 */ 1089 bsetrunningbufspace(bp, bp->b_bufsize); 1090 BUF_KERNPROC(bp); 1091 vn_strategy(bp->b_vp, &bp->b_bio1); 1092 } 1093 1094 /* 1095 * bowrite: 1096 * 1097 * Ordered write. Start output on a buffer, and flag it so that the 1098 * device will write it in the order it was queued. The buffer is 1099 * released when the output completes. bwrite() ( or the VOP routine 1100 * anyway ) is responsible for handling B_INVAL buffers. 1101 */ 1102 int 1103 bowrite(struct buf *bp) 1104 { 1105 bp->b_flags |= B_ORDERED; 1106 bawrite(bp); 1107 return (0); 1108 } 1109 1110 /* 1111 * bdwrite: 1112 * 1113 * Delayed write. (Buffer is marked dirty). Do not bother writing 1114 * anything if the buffer is marked invalid. 1115 * 1116 * Note that since the buffer must be completely valid, we can safely 1117 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1118 * biodone() in order to prevent getblk from writing the buffer 1119 * out synchronously. 1120 */ 1121 void 1122 bdwrite(struct buf *bp) 1123 { 1124 if (BUF_REFCNTNB(bp) == 0) 1125 panic("bdwrite: buffer is not busy"); 1126 1127 if (bp->b_flags & B_INVAL) { 1128 brelse(bp); 1129 return; 1130 } 1131 bdirty(bp); 1132 1133 dsched_buf_enter(bp); /* might stack */ 1134 1135 /* 1136 * Set B_CACHE, indicating that the buffer is fully valid. This is 1137 * true even of NFS now. 1138 */ 1139 bp->b_flags |= B_CACHE; 1140 1141 /* 1142 * This bmap keeps the system from needing to do the bmap later, 1143 * perhaps when the system is attempting to do a sync. Since it 1144 * is likely that the indirect block -- or whatever other datastructure 1145 * that the filesystem needs is still in memory now, it is a good 1146 * thing to do this. Note also, that if the pageout daemon is 1147 * requesting a sync -- there might not be enough memory to do 1148 * the bmap then... So, this is important to do. 1149 */ 1150 if (bp->b_bio2.bio_offset == NOOFFSET) { 1151 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1152 NULL, NULL, BUF_CMD_WRITE); 1153 } 1154 1155 /* 1156 * Because the underlying pages may still be mapped and 1157 * writable trying to set the dirty buffer (b_dirtyoff/end) 1158 * range here will be inaccurate. 1159 * 1160 * However, we must still clean the pages to satisfy the 1161 * vnode_pager and pageout daemon, so they think the pages 1162 * have been "cleaned". What has really occured is that 1163 * they've been earmarked for later writing by the buffer 1164 * cache. 1165 * 1166 * So we get the b_dirtyoff/end update but will not actually 1167 * depend on it (NFS that is) until the pages are busied for 1168 * writing later on. 1169 */ 1170 vfs_clean_pages(bp); 1171 bqrelse(bp); 1172 1173 /* 1174 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1175 * due to the softdep code. 1176 */ 1177 } 1178 1179 /* 1180 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1181 * This is used by tmpfs. 1182 * 1183 * It is important for any VFS using this routine to NOT use it for 1184 * IO_SYNC or IO_ASYNC operations which occur when the system really 1185 * wants to flush VM pages to backing store. 1186 */ 1187 void 1188 buwrite(struct buf *bp) 1189 { 1190 vm_page_t m; 1191 int i; 1192 1193 /* 1194 * Only works for VMIO buffers. If the buffer is already 1195 * marked for delayed-write we can't avoid the bdwrite(). 1196 */ 1197 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1198 bdwrite(bp); 1199 return; 1200 } 1201 1202 /* 1203 * Mark as needing a commit. 1204 */ 1205 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1206 m = bp->b_xio.xio_pages[i]; 1207 vm_page_need_commit(m); 1208 } 1209 bqrelse(bp); 1210 } 1211 1212 /* 1213 * bdirty: 1214 * 1215 * Turn buffer into delayed write request by marking it B_DELWRI. 1216 * B_RELBUF and B_NOCACHE must be cleared. 1217 * 1218 * We reassign the buffer to itself to properly update it in the 1219 * dirty/clean lists. 1220 * 1221 * Must be called from a critical section. 1222 * The buffer must be on BQUEUE_NONE. 1223 */ 1224 void 1225 bdirty(struct buf *bp) 1226 { 1227 KASSERT(bp->b_qindex == BQUEUE_NONE, 1228 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1229 if (bp->b_flags & B_NOCACHE) { 1230 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1231 bp->b_flags &= ~B_NOCACHE; 1232 } 1233 if (bp->b_flags & B_INVAL) { 1234 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1235 } 1236 bp->b_flags &= ~B_RELBUF; 1237 1238 if ((bp->b_flags & B_DELWRI) == 0) { 1239 lwkt_gettoken(&bp->b_vp->v_token); 1240 bp->b_flags |= B_DELWRI; 1241 reassignbuf(bp); 1242 lwkt_reltoken(&bp->b_vp->v_token); 1243 1244 atomic_add_long(&dirtybufcount, 1); 1245 atomic_add_long(&dirtykvaspace, bp->b_kvasize); 1246 atomic_add_long(&dirtybufspace, bp->b_bufsize); 1247 if (bp->b_flags & B_HEAVY) { 1248 atomic_add_long(&dirtybufcounthw, 1); 1249 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1250 } 1251 bd_heatup(); 1252 } 1253 } 1254 1255 /* 1256 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1257 * needs to be flushed with a different buf_daemon thread to avoid 1258 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1259 */ 1260 void 1261 bheavy(struct buf *bp) 1262 { 1263 if ((bp->b_flags & B_HEAVY) == 0) { 1264 bp->b_flags |= B_HEAVY; 1265 if (bp->b_flags & B_DELWRI) { 1266 atomic_add_long(&dirtybufcounthw, 1); 1267 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1268 } 1269 } 1270 } 1271 1272 /* 1273 * bundirty: 1274 * 1275 * Clear B_DELWRI for buffer. 1276 * 1277 * Must be called from a critical section. 1278 * 1279 * The buffer is typically on BQUEUE_NONE but there is one case in 1280 * brelse() that calls this function after placing the buffer on 1281 * a different queue. 1282 */ 1283 void 1284 bundirty(struct buf *bp) 1285 { 1286 if (bp->b_flags & B_DELWRI) { 1287 lwkt_gettoken(&bp->b_vp->v_token); 1288 bp->b_flags &= ~B_DELWRI; 1289 reassignbuf(bp); 1290 lwkt_reltoken(&bp->b_vp->v_token); 1291 1292 atomic_add_long(&dirtybufcount, -1); 1293 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1294 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1295 if (bp->b_flags & B_HEAVY) { 1296 atomic_add_long(&dirtybufcounthw, -1); 1297 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize); 1298 } 1299 bd_signal(bp->b_bufsize); 1300 } 1301 /* 1302 * Since it is now being written, we can clear its deferred write flag. 1303 */ 1304 bp->b_flags &= ~B_DEFERRED; 1305 } 1306 1307 /* 1308 * Set the b_runningbufspace field, used to track how much I/O is 1309 * in progress at any given moment. 1310 */ 1311 void 1312 bsetrunningbufspace(struct buf *bp, int bytes) 1313 { 1314 bp->b_runningbufspace = bytes; 1315 if (bytes) { 1316 atomic_add_long(&runningbufspace, bytes); 1317 atomic_add_long(&runningbufcount, 1); 1318 } 1319 } 1320 1321 /* 1322 * brelse: 1323 * 1324 * Release a busy buffer and, if requested, free its resources. The 1325 * buffer will be stashed in the appropriate bufqueue[] allowing it 1326 * to be accessed later as a cache entity or reused for other purposes. 1327 */ 1328 void 1329 brelse(struct buf *bp) 1330 { 1331 struct bufpcpu *pcpu; 1332 #ifdef INVARIANTS 1333 int saved_flags = bp->b_flags; 1334 #endif 1335 1336 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1337 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1338 1339 /* 1340 * If B_NOCACHE is set we are being asked to destroy the buffer and 1341 * its backing store. Clear B_DELWRI. 1342 * 1343 * B_NOCACHE is set in two cases: (1) when the caller really wants 1344 * to destroy the buffer and backing store and (2) when the caller 1345 * wants to destroy the buffer and backing store after a write 1346 * completes. 1347 */ 1348 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1349 bundirty(bp); 1350 } 1351 1352 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1353 /* 1354 * A re-dirtied buffer is only subject to destruction 1355 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1356 */ 1357 /* leave buffer intact */ 1358 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1359 (bp->b_bufsize <= 0)) { 1360 /* 1361 * Either a failed read or we were asked to free or not 1362 * cache the buffer. This path is reached with B_DELWRI 1363 * set only if B_INVAL is already set. B_NOCACHE governs 1364 * backing store destruction. 1365 * 1366 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1367 * buffer cannot be immediately freed. 1368 */ 1369 bp->b_flags |= B_INVAL; 1370 if (LIST_FIRST(&bp->b_dep) != NULL) 1371 buf_deallocate(bp); 1372 if (bp->b_flags & B_DELWRI) { 1373 atomic_add_long(&dirtybufcount, -1); 1374 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1375 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1376 if (bp->b_flags & B_HEAVY) { 1377 atomic_add_long(&dirtybufcounthw, -1); 1378 atomic_add_long(&dirtybufspacehw, 1379 -bp->b_bufsize); 1380 } 1381 bd_signal(bp->b_bufsize); 1382 } 1383 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1384 } 1385 1386 /* 1387 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set, 1388 * or if b_refs is non-zero. 1389 * 1390 * If vfs_vmio_release() is called with either bit set, the 1391 * underlying pages may wind up getting freed causing a previous 1392 * write (bdwrite()) to get 'lost' because pages associated with 1393 * a B_DELWRI bp are marked clean. Pages associated with a 1394 * B_LOCKED buffer may be mapped by the filesystem. 1395 * 1396 * If we want to release the buffer ourselves (rather then the 1397 * originator asking us to release it), give the originator a 1398 * chance to countermand the release by setting B_LOCKED. 1399 * 1400 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1401 * if B_DELWRI is set. 1402 * 1403 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1404 * on pages to return pages to the VM page queues. 1405 */ 1406 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) { 1407 bp->b_flags &= ~B_RELBUF; 1408 } else if (vm_page_count_min(0)) { 1409 if (LIST_FIRST(&bp->b_dep) != NULL) 1410 buf_deallocate(bp); /* can set B_LOCKED */ 1411 if (bp->b_flags & (B_DELWRI | B_LOCKED)) 1412 bp->b_flags &= ~B_RELBUF; 1413 else 1414 bp->b_flags |= B_RELBUF; 1415 } 1416 1417 /* 1418 * Make sure b_cmd is clear. It may have already been cleared by 1419 * biodone(). 1420 * 1421 * At this point destroying the buffer is governed by the B_INVAL 1422 * or B_RELBUF flags. 1423 */ 1424 bp->b_cmd = BUF_CMD_DONE; 1425 dsched_buf_exit(bp); 1426 1427 /* 1428 * VMIO buffer rundown. Make sure the VM page array is restored 1429 * after an I/O may have replaces some of the pages with bogus pages 1430 * in order to not destroy dirty pages in a fill-in read. 1431 * 1432 * Note that due to the code above, if a buffer is marked B_DELWRI 1433 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1434 * B_INVAL may still be set, however. 1435 * 1436 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1437 * but not the backing store. B_NOCACHE will destroy the backing 1438 * store. 1439 * 1440 * Note that dirty NFS buffers contain byte-granular write ranges 1441 * and should not be destroyed w/ B_INVAL even if the backing store 1442 * is left intact. 1443 */ 1444 if (bp->b_flags & B_VMIO) { 1445 /* 1446 * Rundown for VMIO buffers which are not dirty NFS buffers. 1447 */ 1448 int i, j, resid; 1449 vm_page_t m; 1450 off_t foff; 1451 vm_pindex_t poff; 1452 vm_object_t obj; 1453 struct vnode *vp; 1454 1455 vp = bp->b_vp; 1456 1457 /* 1458 * Get the base offset and length of the buffer. Note that 1459 * in the VMIO case if the buffer block size is not 1460 * page-aligned then b_data pointer may not be page-aligned. 1461 * But our b_xio.xio_pages array *IS* page aligned. 1462 * 1463 * block sizes less then DEV_BSIZE (usually 512) are not 1464 * supported due to the page granularity bits (m->valid, 1465 * m->dirty, etc...). 1466 * 1467 * See man buf(9) for more information 1468 */ 1469 1470 resid = bp->b_bufsize; 1471 foff = bp->b_loffset; 1472 1473 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1474 m = bp->b_xio.xio_pages[i]; 1475 vm_page_flag_clear(m, PG_ZERO); 1476 /* 1477 * If we hit a bogus page, fixup *all* of them 1478 * now. Note that we left these pages wired 1479 * when we removed them so they had better exist, 1480 * and they cannot be ripped out from under us so 1481 * no critical section protection is necessary. 1482 */ 1483 if (m == bogus_page) { 1484 obj = vp->v_object; 1485 poff = OFF_TO_IDX(bp->b_loffset); 1486 1487 vm_object_hold(obj); 1488 for (j = i; j < bp->b_xio.xio_npages; j++) { 1489 vm_page_t mtmp; 1490 1491 mtmp = bp->b_xio.xio_pages[j]; 1492 if (mtmp == bogus_page) { 1493 mtmp = vm_page_lookup(obj, poff + j); 1494 if (!mtmp) { 1495 panic("brelse: page missing"); 1496 } 1497 bp->b_xio.xio_pages[j] = mtmp; 1498 } 1499 } 1500 bp->b_flags &= ~B_HASBOGUS; 1501 vm_object_drop(obj); 1502 1503 if ((bp->b_flags & B_INVAL) == 0) { 1504 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1505 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1506 } 1507 m = bp->b_xio.xio_pages[i]; 1508 } 1509 1510 /* 1511 * Invalidate the backing store if B_NOCACHE is set 1512 * (e.g. used with vinvalbuf()). If this is NFS 1513 * we impose a requirement that the block size be 1514 * a multiple of PAGE_SIZE and create a temporary 1515 * hack to basically invalidate the whole page. The 1516 * problem is that NFS uses really odd buffer sizes 1517 * especially when tracking piecemeal writes and 1518 * it also vinvalbuf()'s a lot, which would result 1519 * in only partial page validation and invalidation 1520 * here. If the file page is mmap()'d, however, 1521 * all the valid bits get set so after we invalidate 1522 * here we would end up with weird m->valid values 1523 * like 0xfc. nfs_getpages() can't handle this so 1524 * we clear all the valid bits for the NFS case 1525 * instead of just some of them. 1526 * 1527 * The real bug is the VM system having to set m->valid 1528 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1529 * itself is an artifact of the whole 512-byte 1530 * granular mess that exists to support odd block 1531 * sizes and UFS meta-data block sizes (e.g. 6144). 1532 * A complete rewrite is required. 1533 * 1534 * XXX 1535 */ 1536 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1537 int poffset = foff & PAGE_MASK; 1538 int presid; 1539 1540 presid = PAGE_SIZE - poffset; 1541 if (bp->b_vp->v_tag == VT_NFS && 1542 bp->b_vp->v_type == VREG) { 1543 ; /* entire page */ 1544 } else if (presid > resid) { 1545 presid = resid; 1546 } 1547 KASSERT(presid >= 0, ("brelse: extra page")); 1548 vm_page_set_invalid(m, poffset, presid); 1549 1550 /* 1551 * Also make sure any swap cache is removed 1552 * as it is now stale (HAMMER in particular 1553 * uses B_NOCACHE to deal with buffer 1554 * aliasing). 1555 */ 1556 swap_pager_unswapped(m); 1557 } 1558 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1559 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1560 } 1561 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1562 vfs_vmio_release(bp); 1563 } else { 1564 /* 1565 * Rundown for non-VMIO buffers. 1566 */ 1567 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1568 if (bp->b_bufsize) 1569 allocbuf(bp, 0); 1570 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1571 if (bp->b_vp) 1572 brelvp(bp); 1573 } 1574 } 1575 1576 if (bp->b_qindex != BQUEUE_NONE) 1577 panic("brelse: free buffer onto another queue???"); 1578 if (BUF_REFCNTNB(bp) > 1) { 1579 /* Temporary panic to verify exclusive locking */ 1580 /* This panic goes away when we allow shared refs */ 1581 panic("brelse: multiple refs"); 1582 /* NOT REACHED */ 1583 return; 1584 } 1585 1586 /* 1587 * Figure out the correct queue to place the cleaned up buffer on. 1588 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1589 * disassociated from their vnode. 1590 * 1591 * Return the buffer to its original pcpu area 1592 */ 1593 pcpu = &bufpcpu[bp->b_qcpu]; 1594 spin_lock(&pcpu->spin); 1595 1596 if (bp->b_flags & B_LOCKED) { 1597 /* 1598 * Buffers that are locked are placed in the locked queue 1599 * immediately, regardless of their state. 1600 */ 1601 bp->b_qindex = BQUEUE_LOCKED; 1602 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1603 bp, b_freelist); 1604 } else if (bp->b_bufsize == 0) { 1605 /* 1606 * Buffers with no memory. Due to conditionals near the top 1607 * of brelse() such buffers should probably already be 1608 * marked B_INVAL and disassociated from their vnode. 1609 */ 1610 bp->b_flags |= B_INVAL; 1611 KASSERT(bp->b_vp == NULL, 1612 ("bp1 %p flags %08x/%08x vnode %p " 1613 "unexpectededly still associated!", 1614 bp, saved_flags, bp->b_flags, bp->b_vp)); 1615 KKASSERT((bp->b_flags & B_HASHED) == 0); 1616 if (bp->b_kvasize) { 1617 bp->b_qindex = BQUEUE_EMPTYKVA; 1618 } else { 1619 bp->b_qindex = BQUEUE_EMPTY; 1620 } 1621 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1622 bp, b_freelist); 1623 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1624 /* 1625 * Buffers with junk contents. Again these buffers had better 1626 * already be disassociated from their vnode. 1627 */ 1628 KASSERT(bp->b_vp == NULL, 1629 ("bp2 %p flags %08x/%08x vnode %p unexpectededly " 1630 "still associated!", 1631 bp, saved_flags, bp->b_flags, bp->b_vp)); 1632 KKASSERT((bp->b_flags & B_HASHED) == 0); 1633 bp->b_flags |= B_INVAL; 1634 bp->b_qindex = BQUEUE_CLEAN; 1635 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1636 bp, b_freelist); 1637 } else { 1638 /* 1639 * Remaining buffers. These buffers are still associated with 1640 * their vnode. 1641 */ 1642 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1643 case B_DELWRI: 1644 bp->b_qindex = BQUEUE_DIRTY; 1645 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1646 bp, b_freelist); 1647 break; 1648 case B_DELWRI | B_HEAVY: 1649 bp->b_qindex = BQUEUE_DIRTY_HW; 1650 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1651 bp, b_freelist); 1652 break; 1653 default: 1654 /* 1655 * NOTE: Buffers are always placed at the end of the 1656 * queue. If B_AGE is not set the buffer will cycle 1657 * through the queue twice. 1658 */ 1659 bp->b_qindex = BQUEUE_CLEAN; 1660 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1661 bp, b_freelist); 1662 break; 1663 } 1664 } 1665 spin_unlock(&pcpu->spin); 1666 1667 /* 1668 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1669 * on the correct queue but we have not yet unlocked it. 1670 */ 1671 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1672 bundirty(bp); 1673 1674 /* 1675 * The bp is on an appropriate queue unless locked. If it is not 1676 * locked or dirty we can wakeup threads waiting for buffer space. 1677 * 1678 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1679 * if B_INVAL is set ). 1680 */ 1681 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1682 bufcountwakeup(); 1683 1684 /* 1685 * Something we can maybe free or reuse 1686 */ 1687 if (bp->b_bufsize || bp->b_kvasize) 1688 bufspacewakeup(); 1689 1690 /* 1691 * Clean up temporary flags and unlock the buffer. 1692 */ 1693 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT); 1694 BUF_UNLOCK(bp); 1695 } 1696 1697 /* 1698 * bqrelse: 1699 * 1700 * Release a buffer back to the appropriate queue but do not try to free 1701 * it. The buffer is expected to be used again soon. 1702 * 1703 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1704 * biodone() to requeue an async I/O on completion. It is also used when 1705 * known good buffers need to be requeued but we think we may need the data 1706 * again soon. 1707 * 1708 * XXX we should be able to leave the B_RELBUF hint set on completion. 1709 */ 1710 void 1711 bqrelse(struct buf *bp) 1712 { 1713 struct bufpcpu *pcpu; 1714 1715 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1716 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1717 1718 if (bp->b_qindex != BQUEUE_NONE) 1719 panic("bqrelse: free buffer onto another queue???"); 1720 if (BUF_REFCNTNB(bp) > 1) { 1721 /* do not release to free list */ 1722 panic("bqrelse: multiple refs"); 1723 return; 1724 } 1725 1726 buf_act_advance(bp); 1727 1728 pcpu = &bufpcpu[bp->b_qcpu]; 1729 spin_lock(&pcpu->spin); 1730 1731 if (bp->b_flags & B_LOCKED) { 1732 /* 1733 * Locked buffers are released to the locked queue. However, 1734 * if the buffer is dirty it will first go into the dirty 1735 * queue and later on after the I/O completes successfully it 1736 * will be released to the locked queue. 1737 */ 1738 bp->b_qindex = BQUEUE_LOCKED; 1739 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1740 bp, b_freelist); 1741 } else if (bp->b_flags & B_DELWRI) { 1742 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1743 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1744 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1745 bp, b_freelist); 1746 } else if (vm_page_count_min(0)) { 1747 /* 1748 * We are too low on memory, we have to try to free the 1749 * buffer (most importantly: the wired pages making up its 1750 * backing store) *now*. 1751 */ 1752 spin_unlock(&pcpu->spin); 1753 brelse(bp); 1754 return; 1755 } else { 1756 bp->b_qindex = BQUEUE_CLEAN; 1757 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1758 bp, b_freelist); 1759 } 1760 spin_unlock(&pcpu->spin); 1761 1762 /* 1763 * We have now placed the buffer on the proper queue, but have yet 1764 * to unlock it. 1765 */ 1766 if ((bp->b_flags & B_LOCKED) == 0 && 1767 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1768 bufcountwakeup(); 1769 } 1770 1771 /* 1772 * Something we can maybe free or reuse. 1773 */ 1774 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1775 bufspacewakeup(); 1776 1777 /* 1778 * Final cleanup and unlock. Clear bits that are only used while a 1779 * buffer is actively locked. 1780 */ 1781 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF); 1782 dsched_buf_exit(bp); 1783 BUF_UNLOCK(bp); 1784 } 1785 1786 /* 1787 * Hold a buffer, preventing it from being reused. This will prevent 1788 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL 1789 * operations. If a B_INVAL operation occurs the buffer will remain held 1790 * but the underlying pages may get ripped out. 1791 * 1792 * These functions are typically used in VOP_READ/VOP_WRITE functions 1793 * to hold a buffer during a copyin or copyout, preventing deadlocks 1794 * or recursive lock panics when read()/write() is used over mmap()'d 1795 * space. 1796 * 1797 * NOTE: bqhold() requires that the buffer be locked at the time of the 1798 * hold. bqdrop() has no requirements other than the buffer having 1799 * previously been held. 1800 */ 1801 void 1802 bqhold(struct buf *bp) 1803 { 1804 atomic_add_int(&bp->b_refs, 1); 1805 } 1806 1807 void 1808 bqdrop(struct buf *bp) 1809 { 1810 KKASSERT(bp->b_refs > 0); 1811 atomic_add_int(&bp->b_refs, -1); 1812 } 1813 1814 /* 1815 * Return backing pages held by the buffer 'bp' back to the VM system. 1816 * This routine is called when the bp is invalidated, released, or 1817 * reused. 1818 * 1819 * The KVA mapping (b_data) for the underlying pages is removed by 1820 * this function. 1821 * 1822 * WARNING! This routine is integral to the low memory critical path 1823 * when a buffer is B_RELBUF'd. If the system has a severe page 1824 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE 1825 * queues so they can be reused in the current pageout daemon 1826 * pass. 1827 */ 1828 static void 1829 vfs_vmio_release(struct buf *bp) 1830 { 1831 int i; 1832 vm_page_t m; 1833 1834 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1835 m = bp->b_xio.xio_pages[i]; 1836 bp->b_xio.xio_pages[i] = NULL; 1837 1838 /* 1839 * We need to own the page in order to safely unwire it. 1840 */ 1841 vm_page_busy_wait(m, FALSE, "vmiopg"); 1842 1843 /* 1844 * The VFS is telling us this is not a meta-data buffer 1845 * even if it is backed by a block device. 1846 */ 1847 if (bp->b_flags & B_NOTMETA) 1848 vm_page_flag_set(m, PG_NOTMETA); 1849 1850 /* 1851 * This is a very important bit of code. We try to track 1852 * VM page use whether the pages are wired into the buffer 1853 * cache or not. While wired into the buffer cache the 1854 * bp tracks the act_count. 1855 * 1856 * We can choose to place unwired pages on the inactive 1857 * queue (0) or active queue (1). If we place too many 1858 * on the active queue the queue will cycle the act_count 1859 * on pages we'd like to keep, just from single-use pages 1860 * (such as when doing a tar-up or file scan). 1861 */ 1862 if (bp->b_act_count < vm_cycle_point) 1863 vm_page_unwire(m, 0); 1864 else 1865 vm_page_unwire(m, 1); 1866 1867 /* 1868 * If the wire_count has dropped to 0 we may need to take 1869 * further action before unbusying the page. 1870 * 1871 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us. 1872 */ 1873 if (m->wire_count == 0) { 1874 vm_page_flag_clear(m, PG_ZERO); 1875 1876 if (bp->b_flags & B_DIRECT) { 1877 /* 1878 * Attempt to free the page if B_DIRECT is 1879 * set, the caller does not desire the page 1880 * to be cached. 1881 */ 1882 vm_page_wakeup(m); 1883 vm_page_try_to_free(m); 1884 } else if ((bp->b_flags & B_NOTMETA) || 1885 vm_page_count_min(0)) { 1886 /* 1887 * Attempt to move the page to PQ_CACHE 1888 * if B_NOTMETA is set. This flag is set 1889 * by HAMMER to remove one of the two pages 1890 * present when double buffering is enabled. 1891 * 1892 * Attempt to move the page to PQ_CACHE 1893 * If we have a severe page deficit. This 1894 * will cause buffer cache operations related 1895 * to pageouts to recycle the related pages 1896 * in order to avoid a low memory deadlock. 1897 */ 1898 m->act_count = bp->b_act_count; 1899 vm_page_wakeup(m); 1900 vm_page_try_to_cache(m); 1901 } else { 1902 /* 1903 * Nominal case, leave the page on the 1904 * queue the original unwiring placed it on 1905 * (active or inactive). 1906 */ 1907 m->act_count = bp->b_act_count; 1908 vm_page_wakeup(m); 1909 } 1910 } else { 1911 vm_page_wakeup(m); 1912 } 1913 } 1914 1915 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1916 bp->b_xio.xio_npages); 1917 if (bp->b_bufsize) { 1918 bufspacewakeup(); 1919 bp->b_bufsize = 0; 1920 } 1921 bp->b_xio.xio_npages = 0; 1922 bp->b_flags &= ~B_VMIO; 1923 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1924 if (bp->b_vp) 1925 brelvp(bp); 1926 } 1927 1928 /* 1929 * Find and initialize a new buffer header, freeing up existing buffers 1930 * in the bufqueues as necessary. The new buffer is returned locked. 1931 * 1932 * Important: B_INVAL is not set. If the caller wishes to throw the 1933 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1934 * 1935 * We block if: 1936 * We have insufficient buffer headers 1937 * We have insufficient buffer space 1938 * buffer_map is too fragmented ( space reservation fails ) 1939 * If we have to flush dirty buffers ( but we try to avoid this ) 1940 * 1941 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1942 * Instead we ask the buf daemon to do it for us. We attempt to 1943 * avoid piecemeal wakeups of the pageout daemon. 1944 */ 1945 struct buf * 1946 getnewbuf(int blkflags, int slptimeo, int size, int maxsize) 1947 { 1948 struct bufpcpu *pcpu; 1949 struct buf *bp; 1950 struct buf *nbp; 1951 int defrag = 0; 1952 int nqindex; 1953 int nqcpu; 1954 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1955 int maxloops = 200000; 1956 int restart_reason = 0; 1957 struct buf *restart_bp = NULL; 1958 static int flushingbufs; 1959 1960 /* 1961 * We can't afford to block since we might be holding a vnode lock, 1962 * which may prevent system daemons from running. We deal with 1963 * low-memory situations by proactively returning memory and running 1964 * async I/O rather then sync I/O. 1965 */ 1966 1967 ++getnewbufcalls; 1968 --getnewbufrestarts; 1969 nqcpu = mycpu->gd_cpuid; 1970 restart: 1971 ++getnewbufrestarts; 1972 1973 if (debug_bufbio && --maxloops == 0) 1974 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)", 1975 mycpu->gd_cpuid, restart_reason, restart_bp); 1976 1977 /* 1978 * Setup for scan. If we do not have enough free buffers, 1979 * we setup a degenerate case that immediately fails. Note 1980 * that if we are specially marked process, we are allowed to 1981 * dip into our reserves. 1982 * 1983 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1984 * 1985 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1986 * However, there are a number of cases (defragging, reusing, ...) 1987 * where we cannot backup. 1988 */ 1989 pcpu = &bufpcpu[nqcpu]; 1990 nqindex = BQUEUE_EMPTYKVA; 1991 spin_lock(&pcpu->spin); 1992 1993 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]); 1994 1995 if (nbp == NULL) { 1996 /* 1997 * If no EMPTYKVA buffers and we are either 1998 * defragging or reusing, locate a CLEAN buffer 1999 * to free or reuse. If bufspace useage is low 2000 * skip this step so we can allocate a new buffer. 2001 */ 2002 if (defrag || bufspace >= lobufspace) { 2003 nqindex = BQUEUE_CLEAN; 2004 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]); 2005 } 2006 2007 /* 2008 * If we could not find or were not allowed to reuse a 2009 * CLEAN buffer, check to see if it is ok to use an EMPTY 2010 * buffer. We can only use an EMPTY buffer if allocating 2011 * its KVA would not otherwise run us out of buffer space. 2012 */ 2013 if (nbp == NULL && defrag == 0 && 2014 bufspace + maxsize < hibufspace) { 2015 nqindex = BQUEUE_EMPTY; 2016 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]); 2017 } 2018 } 2019 2020 /* 2021 * Run scan, possibly freeing data and/or kva mappings on the fly 2022 * depending. 2023 * 2024 * WARNING! spin is held! 2025 */ 2026 while ((bp = nbp) != NULL) { 2027 int qindex = nqindex; 2028 2029 nbp = TAILQ_NEXT(bp, b_freelist); 2030 2031 /* 2032 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2033 * cycles through the queue twice before being selected. 2034 */ 2035 if (qindex == BQUEUE_CLEAN && 2036 (bp->b_flags & B_AGE) == 0 && nbp) { 2037 bp->b_flags |= B_AGE; 2038 TAILQ_REMOVE(&pcpu->bufqueues[qindex], 2039 bp, b_freelist); 2040 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex], 2041 bp, b_freelist); 2042 continue; 2043 } 2044 2045 /* 2046 * Calculate next bp ( we can only use it if we do not block 2047 * or do other fancy things ). 2048 */ 2049 if (nbp == NULL) { 2050 switch(qindex) { 2051 case BQUEUE_EMPTY: 2052 nqindex = BQUEUE_EMPTYKVA; 2053 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]))) 2054 break; 2055 /* fall through */ 2056 case BQUEUE_EMPTYKVA: 2057 nqindex = BQUEUE_CLEAN; 2058 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]))) 2059 break; 2060 /* fall through */ 2061 case BQUEUE_CLEAN: 2062 /* 2063 * nbp is NULL. 2064 */ 2065 break; 2066 } 2067 } 2068 2069 /* 2070 * Sanity Checks 2071 */ 2072 KASSERT(bp->b_qindex == qindex, 2073 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2074 2075 /* 2076 * Note: we no longer distinguish between VMIO and non-VMIO 2077 * buffers. 2078 */ 2079 KASSERT((bp->b_flags & B_DELWRI) == 0, 2080 ("delwri buffer %p found in queue %d", bp, qindex)); 2081 2082 /* 2083 * Do not try to reuse a buffer with a non-zero b_refs. 2084 * This is an unsynchronized test. A synchronized test 2085 * is also performed after we lock the buffer. 2086 */ 2087 if (bp->b_refs) 2088 continue; 2089 2090 /* 2091 * If we are defragging then we need a buffer with 2092 * b_kvasize != 0. XXX this situation should no longer 2093 * occur, if defrag is non-zero the buffer's b_kvasize 2094 * should also be non-zero at this point. XXX 2095 */ 2096 if (defrag && bp->b_kvasize == 0) { 2097 kprintf("Warning: defrag empty buffer %p\n", bp); 2098 continue; 2099 } 2100 2101 /* 2102 * Start freeing the bp. This is somewhat involved. nbp 2103 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 2104 * on the clean list must be disassociated from their 2105 * current vnode. Buffers on the empty[kva] lists have 2106 * already been disassociated. 2107 * 2108 * b_refs is checked after locking along with queue changes. 2109 * We must check here to deal with zero->nonzero transitions 2110 * made by the owner of the buffer lock, which is used by 2111 * VFS's to hold the buffer while issuing an unlocked 2112 * uiomove()s. We cannot invalidate the buffer's pages 2113 * for this case. Once we successfully lock a buffer the 2114 * only 0->1 transitions of b_refs will occur via findblk(). 2115 * 2116 * We must also check for queue changes after successful 2117 * locking as the current lock holder may dispose of the 2118 * buffer and change its queue. 2119 */ 2120 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2121 spin_unlock(&pcpu->spin); 2122 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2123 restart_reason = 1; 2124 restart_bp = bp; 2125 goto restart; 2126 } 2127 if (bp->b_qindex != qindex || bp->b_refs) { 2128 spin_unlock(&pcpu->spin); 2129 BUF_UNLOCK(bp); 2130 restart_reason = 2; 2131 restart_bp = bp; 2132 goto restart; 2133 } 2134 bremfree_locked(bp); 2135 spin_unlock(&pcpu->spin); 2136 2137 /* 2138 * Dependancies must be handled before we disassociate the 2139 * vnode. 2140 * 2141 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2142 * be immediately disassociated. HAMMER then becomes 2143 * responsible for releasing the buffer. 2144 * 2145 * NOTE: spin is UNLOCKED now. 2146 */ 2147 if (LIST_FIRST(&bp->b_dep) != NULL) { 2148 buf_deallocate(bp); 2149 if (bp->b_flags & B_LOCKED) { 2150 bqrelse(bp); 2151 restart_reason = 3; 2152 restart_bp = bp; 2153 goto restart; 2154 } 2155 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2156 } 2157 2158 if (qindex == BQUEUE_CLEAN) { 2159 if (bp->b_flags & B_VMIO) 2160 vfs_vmio_release(bp); 2161 if (bp->b_vp) 2162 brelvp(bp); 2163 } 2164 2165 /* 2166 * NOTE: nbp is now entirely invalid. We can only restart 2167 * the scan from this point on. 2168 * 2169 * Get the rest of the buffer freed up. b_kva* is still 2170 * valid after this operation. 2171 */ 2172 KASSERT(bp->b_vp == NULL, 2173 ("bp3 %p flags %08x vnode %p qindex %d " 2174 "unexpectededly still associated!", 2175 bp, bp->b_flags, bp->b_vp, qindex)); 2176 KKASSERT((bp->b_flags & B_HASHED) == 0); 2177 2178 /* 2179 * critical section protection is not required when 2180 * scrapping a buffer's contents because it is already 2181 * wired. 2182 */ 2183 if (bp->b_bufsize) 2184 allocbuf(bp, 0); 2185 2186 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) { 2187 kprintf("getnewbuf: caught bug vp queue " 2188 "%p/%08x qidx %d\n", 2189 bp, bp->b_flags, qindex); 2190 brelvp(bp); 2191 } 2192 bp->b_flags = B_BNOCLIP; 2193 bp->b_cmd = BUF_CMD_DONE; 2194 bp->b_vp = NULL; 2195 bp->b_error = 0; 2196 bp->b_resid = 0; 2197 bp->b_bcount = 0; 2198 bp->b_xio.xio_npages = 0; 2199 bp->b_dirtyoff = bp->b_dirtyend = 0; 2200 bp->b_act_count = ACT_INIT; 2201 reinitbufbio(bp); 2202 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2203 buf_dep_init(bp); 2204 if (blkflags & GETBLK_BHEAVY) 2205 bp->b_flags |= B_HEAVY; 2206 2207 /* 2208 * If we are defragging then free the buffer. 2209 */ 2210 if (defrag) { 2211 bp->b_flags |= B_INVAL; 2212 bfreekva(bp); 2213 brelse(bp); 2214 defrag = 0; 2215 restart_reason = 4; 2216 restart_bp = bp; 2217 goto restart; 2218 } 2219 2220 /* 2221 * If we are overcomitted then recover the buffer and its 2222 * KVM space. This occurs in rare situations when multiple 2223 * processes are blocked in getnewbuf() or allocbuf(). 2224 * 2225 * On 64-bit systems BKVASIZE == MAXBSIZE and overcommit 2226 * should not be possible. 2227 */ 2228 if (bufspace >= hibufspace) 2229 flushingbufs = 1; 2230 if (BKVASIZE != MAXBSIZE) { 2231 if (flushingbufs && bp->b_kvasize != 0) { 2232 bp->b_flags |= B_INVAL; 2233 bfreekva(bp); 2234 brelse(bp); 2235 restart_reason = 5; 2236 restart_bp = bp; 2237 goto restart; 2238 } 2239 } 2240 if (bufspace < lobufspace) 2241 flushingbufs = 0; 2242 2243 /* 2244 * b_refs can transition to a non-zero value while we hold 2245 * the buffer locked due to a findblk(). Our brelvp() above 2246 * interlocked any future possible transitions due to 2247 * findblk()s. 2248 * 2249 * If we find b_refs to be non-zero we can destroy the 2250 * buffer's contents but we cannot yet reuse the buffer. 2251 */ 2252 if (bp->b_refs) { 2253 bp->b_flags |= B_INVAL; 2254 if (BKVASIZE != MAXBSIZE) 2255 bfreekva(bp); 2256 brelse(bp); 2257 restart_reason = 6; 2258 restart_bp = bp; 2259 goto restart; 2260 } 2261 break; 2262 /* NOT REACHED, spin not held */ 2263 } 2264 2265 /* 2266 * If we exhausted our list, iterate other cpus. If that fails, 2267 * sleep as appropriate. We may have to wakeup various daemons 2268 * and write out some dirty buffers. 2269 * 2270 * Generally we are sleeping due to insufficient buffer space. 2271 * 2272 * NOTE: spin is held if bp is NULL, else it is not held. 2273 */ 2274 if (bp == NULL) { 2275 int flags; 2276 char *waitmsg; 2277 2278 spin_unlock(&pcpu->spin); 2279 2280 nqcpu = (nqcpu + 1) % ncpus; 2281 if (nqcpu != mycpu->gd_cpuid) { 2282 restart_reason = 7; 2283 restart_bp = bp; 2284 goto restart; 2285 } 2286 2287 if (defrag) { 2288 flags = VFS_BIO_NEED_BUFSPACE; 2289 waitmsg = "nbufkv"; 2290 } else if (bufspace >= hibufspace) { 2291 waitmsg = "nbufbs"; 2292 flags = VFS_BIO_NEED_BUFSPACE; 2293 } else { 2294 waitmsg = "newbuf"; 2295 flags = VFS_BIO_NEED_ANY; 2296 } 2297 2298 bd_speedup(); /* heeeelp */ 2299 atomic_set_int(&needsbuffer, flags); 2300 while (needsbuffer & flags) { 2301 int value; 2302 2303 tsleep_interlock(&needsbuffer, 0); 2304 value = atomic_fetchadd_int(&needsbuffer, 0); 2305 if (value & flags) { 2306 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags, 2307 waitmsg, slptimeo)) { 2308 return (NULL); 2309 } 2310 } 2311 } 2312 } else { 2313 /* 2314 * We finally have a valid bp. We aren't quite out of the 2315 * woods, we still have to reserve kva space. In order 2316 * to keep fragmentation sane we only allocate kva in 2317 * BKVASIZE chunks. 2318 * 2319 * (spin is not held) 2320 */ 2321 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2322 2323 if (maxsize != bp->b_kvasize) { 2324 vm_offset_t addr = 0; 2325 int count; 2326 2327 bfreekva(bp); 2328 2329 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2330 vm_map_lock(&buffer_map); 2331 2332 if (vm_map_findspace(&buffer_map, 2333 vm_map_min(&buffer_map), maxsize, 2334 maxsize, 0, &addr)) { 2335 /* 2336 * Uh oh. Buffer map is too fragmented. We 2337 * must defragment the map. 2338 */ 2339 vm_map_unlock(&buffer_map); 2340 vm_map_entry_release(count); 2341 ++bufdefragcnt; 2342 defrag = 1; 2343 bp->b_flags |= B_INVAL; 2344 brelse(bp); 2345 restart_reason = 8; 2346 restart_bp = bp; 2347 goto restart; 2348 } 2349 if (addr) { 2350 vm_map_insert(&buffer_map, &count, 2351 NULL, NULL, 2352 0, addr, addr + maxsize, 2353 VM_MAPTYPE_NORMAL, 2354 VM_PROT_ALL, VM_PROT_ALL, 2355 MAP_NOFAULT); 2356 2357 bp->b_kvabase = (caddr_t) addr; 2358 bp->b_kvasize = maxsize; 2359 bufspace += bp->b_kvasize; 2360 ++bufreusecnt; 2361 } 2362 vm_map_unlock(&buffer_map); 2363 vm_map_entry_release(count); 2364 } 2365 bp->b_data = bp->b_kvabase; 2366 } 2367 return(bp); 2368 } 2369 2370 /* 2371 * buf_daemon: 2372 * 2373 * Buffer flushing daemon. Buffers are normally flushed by the 2374 * update daemon but if it cannot keep up this process starts to 2375 * take the load in an attempt to prevent getnewbuf() from blocking. 2376 * 2377 * Once a flush is initiated it does not stop until the number 2378 * of buffers falls below lodirtybuffers, but we will wake up anyone 2379 * waiting at the mid-point. 2380 */ 2381 static struct kproc_desc buf_kp = { 2382 "bufdaemon", 2383 buf_daemon, 2384 &bufdaemon_td 2385 }; 2386 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2387 kproc_start, &buf_kp); 2388 2389 static struct kproc_desc bufhw_kp = { 2390 "bufdaemon_hw", 2391 buf_daemon_hw, 2392 &bufdaemonhw_td 2393 }; 2394 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2395 kproc_start, &bufhw_kp); 2396 2397 static void 2398 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long), 2399 int *bd_req) 2400 { 2401 long limit; 2402 struct buf *marker; 2403 2404 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO); 2405 marker->b_flags |= B_MARKER; 2406 marker->b_qindex = BQUEUE_NONE; 2407 marker->b_qcpu = 0; 2408 2409 /* 2410 * This process needs to be suspended prior to shutdown sync. 2411 */ 2412 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2413 td, SHUTDOWN_PRI_LAST); 2414 curthread->td_flags |= TDF_SYSTHREAD; 2415 2416 /* 2417 * This process is allowed to take the buffer cache to the limit 2418 */ 2419 for (;;) { 2420 kproc_suspend_loop(); 2421 2422 /* 2423 * Do the flush as long as the number of dirty buffers 2424 * (including those running) exceeds lodirtybufspace. 2425 * 2426 * When flushing limit running I/O to hirunningspace 2427 * Do the flush. Limit the amount of in-transit I/O we 2428 * allow to build up, otherwise we would completely saturate 2429 * the I/O system. Wakeup any waiting processes before we 2430 * normally would so they can run in parallel with our drain. 2431 * 2432 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2433 * but because we split the operation into two threads we 2434 * have to cut it in half for each thread. 2435 */ 2436 waitrunningbufspace(); 2437 limit = lodirtybufspace / 2; 2438 while (buf_limit_fn(limit)) { 2439 if (flushbufqueues(marker, queue) == 0) 2440 break; 2441 if (runningbufspace < hirunningspace) 2442 continue; 2443 waitrunningbufspace(); 2444 } 2445 2446 /* 2447 * We reached our low water mark, reset the 2448 * request and sleep until we are needed again. 2449 * The sleep is just so the suspend code works. 2450 */ 2451 tsleep_interlock(bd_req, 0); 2452 if (atomic_swap_int(bd_req, 0) == 0) 2453 tsleep(bd_req, PINTERLOCKED, "psleep", hz); 2454 } 2455 /* NOT REACHED */ 2456 /*kfree(marker, M_BIOBUF);*/ 2457 } 2458 2459 static int 2460 buf_daemon_limit(long limit) 2461 { 2462 return (runningbufspace + dirtykvaspace > limit || 2463 dirtybufcount - dirtybufcounthw >= nbuf / 2); 2464 } 2465 2466 static int 2467 buf_daemon_hw_limit(long limit) 2468 { 2469 return (runningbufspace + dirtykvaspace > limit || 2470 dirtybufcounthw >= nbuf / 2); 2471 } 2472 2473 static void 2474 buf_daemon(void) 2475 { 2476 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit, 2477 &bd_request); 2478 } 2479 2480 static void 2481 buf_daemon_hw(void) 2482 { 2483 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit, 2484 &bd_request_hw); 2485 } 2486 2487 /* 2488 * Flush up to (flushperqueue) buffers in the dirty queue. Each cpu has a 2489 * localized version of the queue. Each call made to this function iterates 2490 * to another cpu. It is desireable to flush several buffers from the same 2491 * cpu's queue at once, as these are likely going to be linear. 2492 * 2493 * We must be careful to free up B_INVAL buffers instead of write them, which 2494 * NFS is particularly sensitive to. 2495 * 2496 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate that we 2497 * really want to try to get the buffer out and reuse it due to the write 2498 * load on the machine. 2499 * 2500 * We must lock the buffer in order to check its validity before we can mess 2501 * with its contents. spin isn't enough. 2502 */ 2503 static int 2504 flushbufqueues(struct buf *marker, bufq_type_t q) 2505 { 2506 struct bufpcpu *pcpu; 2507 struct buf *bp; 2508 int r = 0; 2509 u_int loops = flushperqueue; 2510 int lcpu = marker->b_qcpu; 2511 2512 KKASSERT(marker->b_qindex == BQUEUE_NONE); 2513 KKASSERT(marker->b_flags & B_MARKER); 2514 2515 again: 2516 /* 2517 * Spinlock needed to perform operations on the queue and may be 2518 * held through a non-blocking BUF_LOCK(), but cannot be held when 2519 * BUF_UNLOCK()ing or through any other major operation. 2520 */ 2521 pcpu = &bufpcpu[marker->b_qcpu]; 2522 spin_lock(&pcpu->spin); 2523 marker->b_qindex = q; 2524 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist); 2525 bp = marker; 2526 2527 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) { 2528 /* 2529 * NOTE: spinlock is always held at the top of the loop 2530 */ 2531 if (bp->b_flags & B_MARKER) 2532 continue; 2533 if ((bp->b_flags & B_DELWRI) == 0) { 2534 kprintf("Unexpected clean buffer %p\n", bp); 2535 continue; 2536 } 2537 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 2538 continue; 2539 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q); 2540 2541 /* 2542 * Once the buffer is locked we will have no choice but to 2543 * unlock the spinlock around a later BUF_UNLOCK and re-set 2544 * bp = marker when looping. Move the marker now to make 2545 * things easier. 2546 */ 2547 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2548 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist); 2549 2550 /* 2551 * Must recheck B_DELWRI after successfully locking 2552 * the buffer. 2553 */ 2554 if ((bp->b_flags & B_DELWRI) == 0) { 2555 spin_unlock(&pcpu->spin); 2556 BUF_UNLOCK(bp); 2557 spin_lock(&pcpu->spin); 2558 bp = marker; 2559 continue; 2560 } 2561 2562 /* 2563 * Remove the buffer from its queue. We still own the 2564 * spinlock here. 2565 */ 2566 _bremfree(bp); 2567 2568 /* 2569 * Disposing of an invalid buffer counts as a flush op 2570 */ 2571 if (bp->b_flags & B_INVAL) { 2572 spin_unlock(&pcpu->spin); 2573 brelse(bp); 2574 goto doloop; 2575 } 2576 2577 /* 2578 * Release the spinlock for the more complex ops we 2579 * are now going to do. 2580 */ 2581 spin_unlock(&pcpu->spin); 2582 lwkt_yield(); 2583 2584 /* 2585 * This is a bit messy 2586 */ 2587 if (LIST_FIRST(&bp->b_dep) != NULL && 2588 (bp->b_flags & B_DEFERRED) == 0 && 2589 buf_countdeps(bp, 0)) { 2590 spin_lock(&pcpu->spin); 2591 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist); 2592 bp->b_qindex = q; 2593 bp->b_flags |= B_DEFERRED; 2594 spin_unlock(&pcpu->spin); 2595 BUF_UNLOCK(bp); 2596 spin_lock(&pcpu->spin); 2597 bp = marker; 2598 continue; 2599 } 2600 2601 /* 2602 * spinlock not held here. 2603 * 2604 * If the buffer has a dependancy, buf_checkwrite() must 2605 * also return 0 for us to be able to initate the write. 2606 * 2607 * If the buffer is flagged B_ERROR it may be requeued 2608 * over and over again, we try to avoid a live lock. 2609 */ 2610 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2611 brelse(bp); 2612 } else if (bp->b_flags & B_ERROR) { 2613 tsleep(bp, 0, "bioer", 1); 2614 bp->b_flags &= ~B_AGE; 2615 cluster_awrite(bp); 2616 } else { 2617 bp->b_flags |= B_AGE; 2618 cluster_awrite(bp); 2619 } 2620 /* bp invalid but needs to be NULL-tested if we break out */ 2621 doloop: 2622 spin_lock(&pcpu->spin); 2623 ++r; 2624 if (--loops == 0) 2625 break; 2626 bp = marker; 2627 } 2628 /* bp is invalid here but can be NULL-tested to advance */ 2629 2630 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2631 marker->b_qindex = BQUEUE_NONE; 2632 spin_unlock(&pcpu->spin); 2633 2634 /* 2635 * Advance the marker to be fair. 2636 */ 2637 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus; 2638 if (bp == NULL) { 2639 if (marker->b_qcpu != lcpu) 2640 goto again; 2641 } 2642 2643 return (r); 2644 } 2645 2646 /* 2647 * inmem: 2648 * 2649 * Returns true if no I/O is needed to access the associated VM object. 2650 * This is like findblk except it also hunts around in the VM system for 2651 * the data. 2652 * 2653 * Note that we ignore vm_page_free() races from interrupts against our 2654 * lookup, since if the caller is not protected our return value will not 2655 * be any more valid then otherwise once we exit the critical section. 2656 */ 2657 int 2658 inmem(struct vnode *vp, off_t loffset) 2659 { 2660 vm_object_t obj; 2661 vm_offset_t toff, tinc, size; 2662 vm_page_t m; 2663 int res = 1; 2664 2665 if (findblk(vp, loffset, FINDBLK_TEST)) 2666 return 1; 2667 if (vp->v_mount == NULL) 2668 return 0; 2669 if ((obj = vp->v_object) == NULL) 2670 return 0; 2671 2672 size = PAGE_SIZE; 2673 if (size > vp->v_mount->mnt_stat.f_iosize) 2674 size = vp->v_mount->mnt_stat.f_iosize; 2675 2676 vm_object_hold(obj); 2677 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2678 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2679 if (m == NULL) { 2680 res = 0; 2681 break; 2682 } 2683 tinc = size; 2684 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2685 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2686 if (vm_page_is_valid(m, 2687 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2688 res = 0; 2689 break; 2690 } 2691 } 2692 vm_object_drop(obj); 2693 return (res); 2694 } 2695 2696 /* 2697 * findblk: 2698 * 2699 * Locate and return the specified buffer. Unless flagged otherwise, 2700 * a locked buffer will be returned if it exists or NULL if it does not. 2701 * 2702 * findblk()'d buffers are still on the bufqueues and if you intend 2703 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2704 * and possibly do other stuff to it. 2705 * 2706 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2707 * for locking the buffer and ensuring that it remains 2708 * the desired buffer after locking. 2709 * 2710 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2711 * to acquire the lock we return NULL, even if the 2712 * buffer exists. 2713 * 2714 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2715 * reuse by getnewbuf() but does not prevent 2716 * disassociation (B_INVAL). Used to avoid deadlocks 2717 * against random (vp,loffset)s due to reassignment. 2718 * 2719 * (0) - Lock the buffer blocking. 2720 */ 2721 struct buf * 2722 findblk(struct vnode *vp, off_t loffset, int flags) 2723 { 2724 struct buf *bp; 2725 int lkflags; 2726 2727 lkflags = LK_EXCLUSIVE; 2728 if (flags & FINDBLK_NBLOCK) 2729 lkflags |= LK_NOWAIT; 2730 2731 for (;;) { 2732 /* 2733 * Lookup. Ref the buf while holding v_token to prevent 2734 * reuse (but does not prevent diassociation). 2735 */ 2736 lwkt_gettoken_shared(&vp->v_token); 2737 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2738 if (bp == NULL) { 2739 lwkt_reltoken(&vp->v_token); 2740 return(NULL); 2741 } 2742 bqhold(bp); 2743 lwkt_reltoken(&vp->v_token); 2744 2745 /* 2746 * If testing only break and return bp, do not lock. 2747 */ 2748 if (flags & FINDBLK_TEST) 2749 break; 2750 2751 /* 2752 * Lock the buffer, return an error if the lock fails. 2753 * (only FINDBLK_NBLOCK can cause the lock to fail). 2754 */ 2755 if (BUF_LOCK(bp, lkflags)) { 2756 atomic_subtract_int(&bp->b_refs, 1); 2757 /* bp = NULL; not needed */ 2758 return(NULL); 2759 } 2760 2761 /* 2762 * Revalidate the locked buf before allowing it to be 2763 * returned. 2764 */ 2765 if (bp->b_vp == vp && bp->b_loffset == loffset) 2766 break; 2767 atomic_subtract_int(&bp->b_refs, 1); 2768 BUF_UNLOCK(bp); 2769 } 2770 2771 /* 2772 * Success 2773 */ 2774 if ((flags & FINDBLK_REF) == 0) 2775 atomic_subtract_int(&bp->b_refs, 1); 2776 return(bp); 2777 } 2778 2779 /* 2780 * getcacheblk: 2781 * 2782 * Similar to getblk() except only returns the buffer if it is 2783 * B_CACHE and requires no other manipulation. Otherwise NULL 2784 * is returned. NULL is also returned if GETBLK_NOWAIT is set 2785 * and the getblk() would block. 2786 * 2787 * If B_RAM is set the buffer might be just fine, but we return 2788 * NULL anyway because we want the code to fall through to the 2789 * cluster read to issue more read-aheads. Otherwise read-ahead breaks. 2790 * 2791 * If blksize is 0 the buffer cache buffer must already be fully 2792 * cached. 2793 * 2794 * If blksize is non-zero getblk() will be used, allowing a buffer 2795 * to be reinstantiated from its VM backing store. The buffer must 2796 * still be fully cached after reinstantiation to be returned. 2797 */ 2798 struct buf * 2799 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags) 2800 { 2801 struct buf *bp; 2802 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0; 2803 2804 if (blksize) { 2805 bp = getblk(vp, loffset, blksize, blkflags, 0); 2806 if (bp) { 2807 if ((bp->b_flags & (B_INVAL | B_CACHE)) == B_CACHE) { 2808 bp->b_flags &= ~B_AGE; 2809 if (bp->b_flags & B_RAM) { 2810 bqrelse(bp); 2811 bp = NULL; 2812 } 2813 } else { 2814 brelse(bp); 2815 bp = NULL; 2816 } 2817 } 2818 } else { 2819 bp = findblk(vp, loffset, fndflags); 2820 if (bp) { 2821 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2822 B_CACHE) { 2823 bp->b_flags &= ~B_AGE; 2824 bremfree(bp); 2825 } else { 2826 BUF_UNLOCK(bp); 2827 bp = NULL; 2828 } 2829 } 2830 } 2831 return (bp); 2832 } 2833 2834 /* 2835 * getblk: 2836 * 2837 * Get a block given a specified block and offset into a file/device. 2838 * B_INVAL may or may not be set on return. The caller should clear 2839 * B_INVAL prior to initiating a READ. 2840 * 2841 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2842 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2843 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2844 * without doing any of those things the system will likely believe 2845 * the buffer to be valid (especially if it is not B_VMIO), and the 2846 * next getblk() will return the buffer with B_CACHE set. 2847 * 2848 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2849 * an existing buffer. 2850 * 2851 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2852 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2853 * and then cleared based on the backing VM. If the previous buffer is 2854 * non-0-sized but invalid, B_CACHE will be cleared. 2855 * 2856 * If getblk() must create a new buffer, the new buffer is returned with 2857 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2858 * case it is returned with B_INVAL clear and B_CACHE set based on the 2859 * backing VM. 2860 * 2861 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2862 * B_CACHE bit is clear. 2863 * 2864 * What this means, basically, is that the caller should use B_CACHE to 2865 * determine whether the buffer is fully valid or not and should clear 2866 * B_INVAL prior to issuing a read. If the caller intends to validate 2867 * the buffer by loading its data area with something, the caller needs 2868 * to clear B_INVAL. If the caller does this without issuing an I/O, 2869 * the caller should set B_CACHE ( as an optimization ), else the caller 2870 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2871 * a write attempt or if it was a successfull read. If the caller 2872 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2873 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2874 * 2875 * getblk flags: 2876 * 2877 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2878 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2879 */ 2880 struct buf * 2881 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2882 { 2883 struct buf *bp; 2884 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2885 int error; 2886 int lkflags; 2887 2888 if (size > MAXBSIZE) 2889 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2890 if (vp->v_object == NULL) 2891 panic("getblk: vnode %p has no object!", vp); 2892 2893 loop: 2894 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2895 /* 2896 * The buffer was found in the cache, but we need to lock it. 2897 * We must acquire a ref on the bp to prevent reuse, but 2898 * this will not prevent disassociation (brelvp()) so we 2899 * must recheck (vp,loffset) after acquiring the lock. 2900 * 2901 * Without the ref the buffer could potentially be reused 2902 * before we acquire the lock and create a deadlock 2903 * situation between the thread trying to reuse the buffer 2904 * and us due to the fact that we would wind up blocking 2905 * on a random (vp,loffset). 2906 */ 2907 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2908 if (blkflags & GETBLK_NOWAIT) { 2909 bqdrop(bp); 2910 return(NULL); 2911 } 2912 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2913 if (blkflags & GETBLK_PCATCH) 2914 lkflags |= LK_PCATCH; 2915 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2916 if (error) { 2917 bqdrop(bp); 2918 if (error == ENOLCK) 2919 goto loop; 2920 return (NULL); 2921 } 2922 /* buffer may have changed on us */ 2923 } 2924 bqdrop(bp); 2925 2926 /* 2927 * Once the buffer has been locked, make sure we didn't race 2928 * a buffer recyclement. Buffers that are no longer hashed 2929 * will have b_vp == NULL, so this takes care of that check 2930 * as well. 2931 */ 2932 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2933 #if 0 2934 kprintf("Warning buffer %p (vp %p loffset %lld) " 2935 "was recycled\n", 2936 bp, vp, (long long)loffset); 2937 #endif 2938 BUF_UNLOCK(bp); 2939 goto loop; 2940 } 2941 2942 /* 2943 * If SZMATCH any pre-existing buffer must be of the requested 2944 * size or NULL is returned. The caller absolutely does not 2945 * want getblk() to bwrite() the buffer on a size mismatch. 2946 */ 2947 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2948 BUF_UNLOCK(bp); 2949 return(NULL); 2950 } 2951 2952 /* 2953 * All vnode-based buffers must be backed by a VM object. 2954 */ 2955 KKASSERT(bp->b_flags & B_VMIO); 2956 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2957 bp->b_flags &= ~B_AGE; 2958 2959 /* 2960 * Make sure that B_INVAL buffers do not have a cached 2961 * block number translation. 2962 */ 2963 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2964 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2965 " did not have cleared bio_offset cache\n", 2966 bp, vp, (long long)loffset); 2967 clearbiocache(&bp->b_bio2); 2968 } 2969 2970 /* 2971 * The buffer is locked. B_CACHE is cleared if the buffer is 2972 * invalid. 2973 */ 2974 if (bp->b_flags & B_INVAL) 2975 bp->b_flags &= ~B_CACHE; 2976 bremfree(bp); 2977 2978 /* 2979 * Any size inconsistancy with a dirty buffer or a buffer 2980 * with a softupdates dependancy must be resolved. Resizing 2981 * the buffer in such circumstances can lead to problems. 2982 * 2983 * Dirty or dependant buffers are written synchronously. 2984 * Other types of buffers are simply released and 2985 * reconstituted as they may be backed by valid, dirty VM 2986 * pages (but not marked B_DELWRI). 2987 * 2988 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2989 * and may be left over from a prior truncation (and thus 2990 * no longer represent the actual EOF point), so we 2991 * definitely do not want to B_NOCACHE the backing store. 2992 */ 2993 if (size != bp->b_bcount) { 2994 if (bp->b_flags & B_DELWRI) { 2995 bp->b_flags |= B_RELBUF; 2996 bwrite(bp); 2997 } else if (LIST_FIRST(&bp->b_dep)) { 2998 bp->b_flags |= B_RELBUF; 2999 bwrite(bp); 3000 } else { 3001 bp->b_flags |= B_RELBUF; 3002 brelse(bp); 3003 } 3004 goto loop; 3005 } 3006 KKASSERT(size <= bp->b_kvasize); 3007 KASSERT(bp->b_loffset != NOOFFSET, 3008 ("getblk: no buffer offset")); 3009 3010 /* 3011 * A buffer with B_DELWRI set and B_CACHE clear must 3012 * be committed before we can return the buffer in 3013 * order to prevent the caller from issuing a read 3014 * ( due to B_CACHE not being set ) and overwriting 3015 * it. 3016 * 3017 * Most callers, including NFS and FFS, need this to 3018 * operate properly either because they assume they 3019 * can issue a read if B_CACHE is not set, or because 3020 * ( for example ) an uncached B_DELWRI might loop due 3021 * to softupdates re-dirtying the buffer. In the latter 3022 * case, B_CACHE is set after the first write completes, 3023 * preventing further loops. 3024 * 3025 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3026 * above while extending the buffer, we cannot allow the 3027 * buffer to remain with B_CACHE set after the write 3028 * completes or it will represent a corrupt state. To 3029 * deal with this we set B_NOCACHE to scrap the buffer 3030 * after the write. 3031 * 3032 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3033 * I'm not even sure this state is still possible 3034 * now that getblk() writes out any dirty buffers 3035 * on size changes. 3036 * 3037 * We might be able to do something fancy, like setting 3038 * B_CACHE in bwrite() except if B_DELWRI is already set, 3039 * so the below call doesn't set B_CACHE, but that gets real 3040 * confusing. This is much easier. 3041 */ 3042 3043 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3044 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3045 "and CACHE clear, b_flags %08x\n", 3046 bp, (uintmax_t)bp->b_loffset, bp->b_flags); 3047 bp->b_flags |= B_NOCACHE; 3048 bwrite(bp); 3049 goto loop; 3050 } 3051 } else { 3052 /* 3053 * Buffer is not in-core, create new buffer. The buffer 3054 * returned by getnewbuf() is locked. Note that the returned 3055 * buffer is also considered valid (not marked B_INVAL). 3056 * 3057 * Calculating the offset for the I/O requires figuring out 3058 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3059 * the mount's f_iosize otherwise. If the vnode does not 3060 * have an associated mount we assume that the passed size is 3061 * the block size. 3062 * 3063 * Note that vn_isdisk() cannot be used here since it may 3064 * return a failure for numerous reasons. Note that the 3065 * buffer size may be larger then the block size (the caller 3066 * will use block numbers with the proper multiple). Beware 3067 * of using any v_* fields which are part of unions. In 3068 * particular, in DragonFly the mount point overloading 3069 * mechanism uses the namecache only and the underlying 3070 * directory vnode is not a special case. 3071 */ 3072 int bsize, maxsize; 3073 3074 if (vp->v_type == VBLK || vp->v_type == VCHR) 3075 bsize = DEV_BSIZE; 3076 else if (vp->v_mount) 3077 bsize = vp->v_mount->mnt_stat.f_iosize; 3078 else 3079 bsize = size; 3080 3081 maxsize = size + (loffset & PAGE_MASK); 3082 maxsize = imax(maxsize, bsize); 3083 3084 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3085 if (bp == NULL) { 3086 if (slpflags || slptimeo) 3087 return NULL; 3088 goto loop; 3089 } 3090 3091 /* 3092 * Atomically insert the buffer into the hash, so that it can 3093 * be found by findblk(). 3094 * 3095 * If bgetvp() returns non-zero a collision occured, and the 3096 * bp will not be associated with the vnode. 3097 * 3098 * Make sure the translation layer has been cleared. 3099 */ 3100 bp->b_loffset = loffset; 3101 bp->b_bio2.bio_offset = NOOFFSET; 3102 /* bp->b_bio2.bio_next = NULL; */ 3103 3104 if (bgetvp(vp, bp, size)) { 3105 bp->b_flags |= B_INVAL; 3106 brelse(bp); 3107 goto loop; 3108 } 3109 3110 /* 3111 * All vnode-based buffers must be backed by a VM object. 3112 */ 3113 KKASSERT(vp->v_object != NULL); 3114 bp->b_flags |= B_VMIO; 3115 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3116 3117 allocbuf(bp, size); 3118 } 3119 return (bp); 3120 } 3121 3122 /* 3123 * regetblk(bp) 3124 * 3125 * Reacquire a buffer that was previously released to the locked queue, 3126 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3127 * set B_LOCKED (which handles the acquisition race). 3128 * 3129 * To this end, either B_LOCKED must be set or the dependancy list must be 3130 * non-empty. 3131 */ 3132 void 3133 regetblk(struct buf *bp) 3134 { 3135 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3136 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3137 bremfree(bp); 3138 } 3139 3140 /* 3141 * geteblk: 3142 * 3143 * Get an empty, disassociated buffer of given size. The buffer is 3144 * initially set to B_INVAL. 3145 * 3146 * critical section protection is not required for the allocbuf() 3147 * call because races are impossible here. 3148 */ 3149 struct buf * 3150 geteblk(int size) 3151 { 3152 struct buf *bp; 3153 int maxsize; 3154 3155 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3156 3157 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL) 3158 ; 3159 allocbuf(bp, size); 3160 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3161 return (bp); 3162 } 3163 3164 3165 /* 3166 * allocbuf: 3167 * 3168 * This code constitutes the buffer memory from either anonymous system 3169 * memory (in the case of non-VMIO operations) or from an associated 3170 * VM object (in the case of VMIO operations). This code is able to 3171 * resize a buffer up or down. 3172 * 3173 * Note that this code is tricky, and has many complications to resolve 3174 * deadlock or inconsistant data situations. Tread lightly!!! 3175 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3176 * the caller. Calling this code willy nilly can result in the loss of 3177 * data. 3178 * 3179 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3180 * B_CACHE for the non-VMIO case. 3181 * 3182 * This routine does not need to be called from a critical section but you 3183 * must own the buffer. 3184 */ 3185 int 3186 allocbuf(struct buf *bp, int size) 3187 { 3188 int newbsize, mbsize; 3189 int i; 3190 3191 if (BUF_REFCNT(bp) == 0) 3192 panic("allocbuf: buffer not busy"); 3193 3194 if (bp->b_kvasize < size) 3195 panic("allocbuf: buffer too small"); 3196 3197 if ((bp->b_flags & B_VMIO) == 0) { 3198 caddr_t origbuf; 3199 int origbufsize; 3200 /* 3201 * Just get anonymous memory from the kernel. Don't 3202 * mess with B_CACHE. 3203 */ 3204 mbsize = roundup2(size, DEV_BSIZE); 3205 if (bp->b_flags & B_MALLOC) 3206 newbsize = mbsize; 3207 else 3208 newbsize = round_page(size); 3209 3210 if (newbsize < bp->b_bufsize) { 3211 /* 3212 * Malloced buffers are not shrunk 3213 */ 3214 if (bp->b_flags & B_MALLOC) { 3215 if (newbsize) { 3216 bp->b_bcount = size; 3217 } else { 3218 kfree(bp->b_data, M_BIOBUF); 3219 if (bp->b_bufsize) { 3220 atomic_subtract_long(&bufmallocspace, bp->b_bufsize); 3221 bufspacewakeup(); 3222 bp->b_bufsize = 0; 3223 } 3224 bp->b_data = bp->b_kvabase; 3225 bp->b_bcount = 0; 3226 bp->b_flags &= ~B_MALLOC; 3227 } 3228 return 1; 3229 } 3230 vm_hold_free_pages( 3231 bp, 3232 (vm_offset_t) bp->b_data + newbsize, 3233 (vm_offset_t) bp->b_data + bp->b_bufsize); 3234 } else if (newbsize > bp->b_bufsize) { 3235 /* 3236 * We only use malloced memory on the first allocation. 3237 * and revert to page-allocated memory when the buffer 3238 * grows. 3239 */ 3240 if ((bufmallocspace < maxbufmallocspace) && 3241 (bp->b_bufsize == 0) && 3242 (mbsize <= PAGE_SIZE/2)) { 3243 3244 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3245 bp->b_bufsize = mbsize; 3246 bp->b_bcount = size; 3247 bp->b_flags |= B_MALLOC; 3248 atomic_add_long(&bufmallocspace, mbsize); 3249 return 1; 3250 } 3251 origbuf = NULL; 3252 origbufsize = 0; 3253 /* 3254 * If the buffer is growing on its other-than-first 3255 * allocation, then we revert to the page-allocation 3256 * scheme. 3257 */ 3258 if (bp->b_flags & B_MALLOC) { 3259 origbuf = bp->b_data; 3260 origbufsize = bp->b_bufsize; 3261 bp->b_data = bp->b_kvabase; 3262 if (bp->b_bufsize) { 3263 atomic_subtract_long(&bufmallocspace, 3264 bp->b_bufsize); 3265 bufspacewakeup(); 3266 bp->b_bufsize = 0; 3267 } 3268 bp->b_flags &= ~B_MALLOC; 3269 newbsize = round_page(newbsize); 3270 } 3271 vm_hold_load_pages( 3272 bp, 3273 (vm_offset_t) bp->b_data + bp->b_bufsize, 3274 (vm_offset_t) bp->b_data + newbsize); 3275 if (origbuf) { 3276 bcopy(origbuf, bp->b_data, origbufsize); 3277 kfree(origbuf, M_BIOBUF); 3278 } 3279 } 3280 } else { 3281 vm_page_t m; 3282 int desiredpages; 3283 3284 newbsize = roundup2(size, DEV_BSIZE); 3285 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3286 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3287 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3288 3289 if (bp->b_flags & B_MALLOC) 3290 panic("allocbuf: VMIO buffer can't be malloced"); 3291 /* 3292 * Set B_CACHE initially if buffer is 0 length or will become 3293 * 0-length. 3294 */ 3295 if (size == 0 || bp->b_bufsize == 0) 3296 bp->b_flags |= B_CACHE; 3297 3298 if (newbsize < bp->b_bufsize) { 3299 /* 3300 * DEV_BSIZE aligned new buffer size is less then the 3301 * DEV_BSIZE aligned existing buffer size. Figure out 3302 * if we have to remove any pages. 3303 */ 3304 if (desiredpages < bp->b_xio.xio_npages) { 3305 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3306 /* 3307 * the page is not freed here -- it 3308 * is the responsibility of 3309 * vnode_pager_setsize 3310 */ 3311 m = bp->b_xio.xio_pages[i]; 3312 KASSERT(m != bogus_page, 3313 ("allocbuf: bogus page found")); 3314 vm_page_busy_wait(m, TRUE, "biodep"); 3315 bp->b_xio.xio_pages[i] = NULL; 3316 vm_page_unwire(m, 0); 3317 vm_page_wakeup(m); 3318 } 3319 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3320 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3321 bp->b_xio.xio_npages = desiredpages; 3322 } 3323 } else if (size > bp->b_bcount) { 3324 /* 3325 * We are growing the buffer, possibly in a 3326 * byte-granular fashion. 3327 */ 3328 struct vnode *vp; 3329 vm_object_t obj; 3330 vm_offset_t toff; 3331 vm_offset_t tinc; 3332 3333 /* 3334 * Step 1, bring in the VM pages from the object, 3335 * allocating them if necessary. We must clear 3336 * B_CACHE if these pages are not valid for the 3337 * range covered by the buffer. 3338 * 3339 * critical section protection is required to protect 3340 * against interrupts unbusying and freeing pages 3341 * between our vm_page_lookup() and our 3342 * busycheck/wiring call. 3343 */ 3344 vp = bp->b_vp; 3345 obj = vp->v_object; 3346 3347 vm_object_hold(obj); 3348 while (bp->b_xio.xio_npages < desiredpages) { 3349 vm_page_t m; 3350 vm_pindex_t pi; 3351 int error; 3352 3353 pi = OFF_TO_IDX(bp->b_loffset) + 3354 bp->b_xio.xio_npages; 3355 3356 /* 3357 * Blocking on m->busy might lead to a 3358 * deadlock: 3359 * 3360 * vm_fault->getpages->cluster_read->allocbuf 3361 */ 3362 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3363 &error); 3364 if (error) { 3365 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3366 continue; 3367 } 3368 if (m == NULL) { 3369 /* 3370 * note: must allocate system pages 3371 * since blocking here could intefere 3372 * with paging I/O, no matter which 3373 * process we are. 3374 */ 3375 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages); 3376 if (m) { 3377 vm_page_wire(m); 3378 vm_page_flag_clear(m, PG_ZERO); 3379 vm_page_wakeup(m); 3380 bp->b_flags &= ~B_CACHE; 3381 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3382 ++bp->b_xio.xio_npages; 3383 } 3384 continue; 3385 } 3386 3387 /* 3388 * We found a page and were able to busy it. 3389 */ 3390 vm_page_flag_clear(m, PG_ZERO); 3391 vm_page_wire(m); 3392 vm_page_wakeup(m); 3393 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3394 ++bp->b_xio.xio_npages; 3395 if (bp->b_act_count < m->act_count) 3396 bp->b_act_count = m->act_count; 3397 } 3398 vm_object_drop(obj); 3399 3400 /* 3401 * Step 2. We've loaded the pages into the buffer, 3402 * we have to figure out if we can still have B_CACHE 3403 * set. Note that B_CACHE is set according to the 3404 * byte-granular range ( bcount and size ), not the 3405 * aligned range ( newbsize ). 3406 * 3407 * The VM test is against m->valid, which is DEV_BSIZE 3408 * aligned. Needless to say, the validity of the data 3409 * needs to also be DEV_BSIZE aligned. Note that this 3410 * fails with NFS if the server or some other client 3411 * extends the file's EOF. If our buffer is resized, 3412 * B_CACHE may remain set! XXX 3413 */ 3414 3415 toff = bp->b_bcount; 3416 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3417 3418 while ((bp->b_flags & B_CACHE) && toff < size) { 3419 vm_pindex_t pi; 3420 3421 if (tinc > (size - toff)) 3422 tinc = size - toff; 3423 3424 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3425 PAGE_SHIFT; 3426 3427 vfs_buf_test_cache( 3428 bp, 3429 bp->b_loffset, 3430 toff, 3431 tinc, 3432 bp->b_xio.xio_pages[pi] 3433 ); 3434 toff += tinc; 3435 tinc = PAGE_SIZE; 3436 } 3437 3438 /* 3439 * Step 3, fixup the KVM pmap. Remember that 3440 * bp->b_data is relative to bp->b_loffset, but 3441 * bp->b_loffset may be offset into the first page. 3442 */ 3443 3444 bp->b_data = (caddr_t) 3445 trunc_page((vm_offset_t)bp->b_data); 3446 pmap_qenter( 3447 (vm_offset_t)bp->b_data, 3448 bp->b_xio.xio_pages, 3449 bp->b_xio.xio_npages 3450 ); 3451 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3452 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3453 } 3454 } 3455 3456 /* adjust space use on already-dirty buffer */ 3457 if (bp->b_flags & B_DELWRI) { 3458 /* dirtykvaspace unchanged */ 3459 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize); 3460 if (bp->b_flags & B_HEAVY) { 3461 atomic_add_long(&dirtybufspacehw, 3462 newbsize - bp->b_bufsize); 3463 } 3464 } 3465 if (newbsize < bp->b_bufsize) 3466 bufspacewakeup(); 3467 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3468 bp->b_bcount = size; /* requested buffer size */ 3469 return 1; 3470 } 3471 3472 /* 3473 * biowait: 3474 * 3475 * Wait for buffer I/O completion, returning error status. B_EINTR 3476 * is converted into an EINTR error but not cleared (since a chain 3477 * of biowait() calls may occur). 3478 * 3479 * On return bpdone() will have been called but the buffer will remain 3480 * locked and will not have been brelse()'d. 3481 * 3482 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3483 * likely still in progress on return. 3484 * 3485 * NOTE! This operation is on a BIO, not a BUF. 3486 * 3487 * NOTE! BIO_DONE is cleared by vn_strategy() 3488 */ 3489 static __inline int 3490 _biowait(struct bio *bio, const char *wmesg, int to) 3491 { 3492 struct buf *bp = bio->bio_buf; 3493 u_int32_t flags; 3494 u_int32_t nflags; 3495 int error; 3496 3497 KKASSERT(bio == &bp->b_bio1); 3498 for (;;) { 3499 flags = bio->bio_flags; 3500 if (flags & BIO_DONE) 3501 break; 3502 nflags = flags | BIO_WANT; 3503 tsleep_interlock(bio, 0); 3504 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3505 if (wmesg) 3506 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3507 else if (bp->b_cmd == BUF_CMD_READ) 3508 error = tsleep(bio, PINTERLOCKED, "biord", to); 3509 else 3510 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3511 if (error) { 3512 kprintf("tsleep error biowait %d\n", error); 3513 return (error); 3514 } 3515 } 3516 } 3517 3518 /* 3519 * Finish up. 3520 */ 3521 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3522 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3523 if (bp->b_flags & B_EINTR) 3524 return (EINTR); 3525 if (bp->b_flags & B_ERROR) 3526 return (bp->b_error ? bp->b_error : EIO); 3527 return (0); 3528 } 3529 3530 int 3531 biowait(struct bio *bio, const char *wmesg) 3532 { 3533 return(_biowait(bio, wmesg, 0)); 3534 } 3535 3536 int 3537 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3538 { 3539 return(_biowait(bio, wmesg, to)); 3540 } 3541 3542 /* 3543 * This associates a tracking count with an I/O. vn_strategy() and 3544 * dev_dstrategy() do this automatically but there are a few cases 3545 * where a vnode or device layer is bypassed when a block translation 3546 * is cached. In such cases bio_start_transaction() may be called on 3547 * the bypassed layers so the system gets an I/O in progress indication 3548 * for those higher layers. 3549 */ 3550 void 3551 bio_start_transaction(struct bio *bio, struct bio_track *track) 3552 { 3553 bio->bio_track = track; 3554 bio_track_ref(track); 3555 dsched_buf_enter(bio->bio_buf); /* might stack */ 3556 } 3557 3558 /* 3559 * Initiate I/O on a vnode. 3560 * 3561 * SWAPCACHE OPERATION: 3562 * 3563 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3564 * devfs also uses b_vp for fake buffers so we also have to check 3565 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3566 * underlying block device. The swap assignments are related to the 3567 * buffer cache buffer's b_vp, not the passed vp. 3568 * 3569 * The passed vp == bp->b_vp only in the case where the strategy call 3570 * is made on the vp itself for its own buffers (a regular file or 3571 * block device vp). The filesystem usually then re-calls vn_strategy() 3572 * after translating the request to an underlying device. 3573 * 3574 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3575 * underlying buffer cache buffers. 3576 * 3577 * We can only deal with page-aligned buffers at the moment, because 3578 * we can't tell what the real dirty state for pages straddling a buffer 3579 * are. 3580 * 3581 * In order to call swap_pager_strategy() we must provide the VM object 3582 * and base offset for the underlying buffer cache pages so it can find 3583 * the swap blocks. 3584 */ 3585 void 3586 vn_strategy(struct vnode *vp, struct bio *bio) 3587 { 3588 struct bio_track *track; 3589 struct buf *bp = bio->bio_buf; 3590 3591 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3592 3593 /* 3594 * Set when an I/O is issued on the bp. Cleared by consumers 3595 * (aka HAMMER), allowing the consumer to determine if I/O had 3596 * actually occurred. 3597 */ 3598 bp->b_flags |= B_IOISSUED; 3599 3600 /* 3601 * Handle the swap cache intercept. 3602 */ 3603 if (vn_cache_strategy(vp, bio)) 3604 return; 3605 3606 /* 3607 * Otherwise do the operation through the filesystem 3608 */ 3609 if (bp->b_cmd == BUF_CMD_READ) 3610 track = &vp->v_track_read; 3611 else 3612 track = &vp->v_track_write; 3613 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3614 bio->bio_track = track; 3615 bio_track_ref(track); 3616 dsched_buf_enter(bp); /* might stack */ 3617 vop_strategy(*vp->v_ops, vp, bio); 3618 } 3619 3620 static void vn_cache_strategy_callback(struct bio *bio); 3621 3622 int 3623 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3624 { 3625 struct buf *bp = bio->bio_buf; 3626 struct bio *nbio; 3627 vm_object_t object; 3628 vm_page_t m; 3629 int i; 3630 3631 /* 3632 * Stop using swapcache if paniced, dumping, or dumped 3633 */ 3634 if (panicstr || dumping) 3635 return(0); 3636 3637 /* 3638 * Is this buffer cache buffer suitable for reading from 3639 * the swap cache? 3640 */ 3641 if (vm_swapcache_read_enable == 0 || 3642 bp->b_cmd != BUF_CMD_READ || 3643 ((bp->b_flags & B_CLUSTER) == 0 && 3644 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3645 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3646 (bp->b_bcount & PAGE_MASK) != 0) { 3647 return(0); 3648 } 3649 3650 /* 3651 * Figure out the original VM object (it will match the underlying 3652 * VM pages). Note that swap cached data uses page indices relative 3653 * to that object, not relative to bio->bio_offset. 3654 */ 3655 if (bp->b_flags & B_CLUSTER) 3656 object = vp->v_object; 3657 else 3658 object = bp->b_vp->v_object; 3659 3660 /* 3661 * In order to be able to use the swap cache all underlying VM 3662 * pages must be marked as such, and we can't have any bogus pages. 3663 */ 3664 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3665 m = bp->b_xio.xio_pages[i]; 3666 if ((m->flags & PG_SWAPPED) == 0) 3667 break; 3668 if (m == bogus_page) 3669 break; 3670 } 3671 3672 /* 3673 * If we are good then issue the I/O using swap_pager_strategy(). 3674 * 3675 * We can only do this if the buffer actually supports object-backed 3676 * I/O. If it doesn't npages will be 0. 3677 */ 3678 if (i && i == bp->b_xio.xio_npages) { 3679 m = bp->b_xio.xio_pages[0]; 3680 nbio = push_bio(bio); 3681 nbio->bio_done = vn_cache_strategy_callback; 3682 nbio->bio_offset = ptoa(m->pindex); 3683 KKASSERT(m->object == object); 3684 swap_pager_strategy(object, nbio); 3685 return(1); 3686 } 3687 return(0); 3688 } 3689 3690 /* 3691 * This is a bit of a hack but since the vn_cache_strategy() function can 3692 * override a VFS's strategy function we must make sure that the bio, which 3693 * is probably bio2, doesn't leak an unexpected offset value back to the 3694 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3695 * bio went through its own file strategy function and the the bio2 offset 3696 * is a cached disk offset when, in fact, it isn't. 3697 */ 3698 static void 3699 vn_cache_strategy_callback(struct bio *bio) 3700 { 3701 bio->bio_offset = NOOFFSET; 3702 biodone(pop_bio(bio)); 3703 } 3704 3705 /* 3706 * bpdone: 3707 * 3708 * Finish I/O on a buffer after all BIOs have been processed. 3709 * Called when the bio chain is exhausted or by biowait. If called 3710 * by biowait, elseit is typically 0. 3711 * 3712 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3713 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3714 * assuming B_INVAL is clear. 3715 * 3716 * For the VMIO case, we set B_CACHE if the op was a read and no 3717 * read error occured, or if the op was a write. B_CACHE is never 3718 * set if the buffer is invalid or otherwise uncacheable. 3719 * 3720 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3721 * initiator to leave B_INVAL set to brelse the buffer out of existance 3722 * in the biodone routine. 3723 * 3724 * bpdone is responsible for calling bundirty() on the buffer after a 3725 * successful write. We previously did this prior to initiating the 3726 * write under the assumption that the buffer might be dirtied again 3727 * while the write was in progress, however doing it before-hand creates 3728 * a race condition prior to the call to vn_strategy() where the 3729 * filesystem may not be aware that a dirty buffer is present. 3730 * It should not be possible for the buffer or its underlying pages to 3731 * be redirtied prior to bpdone()'s unbusying of the underlying VM 3732 * pages. 3733 */ 3734 void 3735 bpdone(struct buf *bp, int elseit) 3736 { 3737 buf_cmd_t cmd; 3738 3739 KASSERT(BUF_REFCNTNB(bp) > 0, 3740 ("bpdone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3741 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3742 ("bpdone: bp %p already done!", bp)); 3743 3744 /* 3745 * No more BIOs are left. All completion functions have been dealt 3746 * with, now we clean up the buffer. 3747 */ 3748 cmd = bp->b_cmd; 3749 bp->b_cmd = BUF_CMD_DONE; 3750 3751 /* 3752 * Only reads and writes are processed past this point. 3753 */ 3754 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3755 if (cmd == BUF_CMD_FREEBLKS) 3756 bp->b_flags |= B_NOCACHE; 3757 if (elseit) 3758 brelse(bp); 3759 return; 3760 } 3761 3762 /* 3763 * A failed write must re-dirty the buffer unless B_INVAL 3764 * was set. 3765 * 3766 * A successful write must clear the dirty flag. This is done after 3767 * the write to ensure that the buffer remains on the vnode's dirty 3768 * list for filesystem interlocks / checks until the write is actually 3769 * complete. HAMMER2 is sensitive to this issue. 3770 * 3771 * Only applicable to normal buffers (with VPs). vinum buffers may 3772 * not have a vp. 3773 * 3774 * Must be done prior to calling buf_complete() as the callback might 3775 * re-dirty the buffer. 3776 */ 3777 if (cmd == BUF_CMD_WRITE) { 3778 if ((bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3779 bp->b_flags &= ~B_NOCACHE; 3780 if (bp->b_vp) 3781 bdirty(bp); 3782 } else { 3783 if (bp->b_vp) 3784 bundirty(bp); 3785 } 3786 } 3787 3788 /* 3789 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3790 * a lot worse. XXX - move this above the clearing of b_cmd 3791 */ 3792 if (LIST_FIRST(&bp->b_dep) != NULL) 3793 buf_complete(bp); 3794 3795 if (bp->b_flags & B_VMIO) { 3796 int i; 3797 vm_ooffset_t foff; 3798 vm_page_t m; 3799 vm_object_t obj; 3800 int iosize; 3801 struct vnode *vp = bp->b_vp; 3802 3803 obj = vp->v_object; 3804 3805 #if defined(VFS_BIO_DEBUG) 3806 if (vp->v_auxrefs == 0) 3807 panic("bpdone: zero vnode hold count"); 3808 if ((vp->v_flag & VOBJBUF) == 0) 3809 panic("bpdone: vnode is not setup for merged cache"); 3810 #endif 3811 3812 foff = bp->b_loffset; 3813 KASSERT(foff != NOOFFSET, ("bpdone: no buffer offset")); 3814 KASSERT(obj != NULL, ("bpdone: missing VM object")); 3815 3816 #if defined(VFS_BIO_DEBUG) 3817 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3818 kprintf("bpdone: paging in progress(%d) < " 3819 "bp->b_xio.xio_npages(%d)\n", 3820 obj->paging_in_progress, 3821 bp->b_xio.xio_npages); 3822 } 3823 #endif 3824 3825 /* 3826 * Set B_CACHE if the op was a normal read and no error 3827 * occured. B_CACHE is set for writes in the b*write() 3828 * routines. 3829 */ 3830 iosize = bp->b_bcount - bp->b_resid; 3831 if (cmd == BUF_CMD_READ && 3832 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3833 bp->b_flags |= B_CACHE; 3834 } 3835 3836 vm_object_hold(obj); 3837 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3838 int bogusflag = 0; 3839 int resid; 3840 3841 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3842 if (resid > iosize) 3843 resid = iosize; 3844 3845 /* 3846 * cleanup bogus pages, restoring the originals. Since 3847 * the originals should still be wired, we don't have 3848 * to worry about interrupt/freeing races destroying 3849 * the VM object association. 3850 */ 3851 m = bp->b_xio.xio_pages[i]; 3852 if (m == bogus_page) { 3853 bogusflag = 1; 3854 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3855 if (m == NULL) 3856 panic("bpdone: page disappeared"); 3857 bp->b_xio.xio_pages[i] = m; 3858 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3859 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3860 } 3861 #if defined(VFS_BIO_DEBUG) 3862 if (OFF_TO_IDX(foff) != m->pindex) { 3863 kprintf("bpdone: foff(%lu)/m->pindex(%ld) " 3864 "mismatch\n", 3865 (unsigned long)foff, (long)m->pindex); 3866 } 3867 #endif 3868 3869 /* 3870 * In the write case, the valid and clean bits are 3871 * already changed correctly (see bdwrite()), so we 3872 * only need to do this here in the read case. 3873 */ 3874 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3875 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3876 vfs_clean_one_page(bp, i, m); 3877 } 3878 vm_page_flag_clear(m, PG_ZERO); 3879 3880 /* 3881 * when debugging new filesystems or buffer I/O 3882 * methods, this is the most common error that pops 3883 * up. if you see this, you have not set the page 3884 * busy flag correctly!!! 3885 */ 3886 if (m->busy == 0) { 3887 kprintf("bpdone: page busy < 0, " 3888 "pindex: %d, foff: 0x(%x,%x), " 3889 "resid: %d, index: %d\n", 3890 (int) m->pindex, (int)(foff >> 32), 3891 (int) foff & 0xffffffff, resid, i); 3892 if (!vn_isdisk(vp, NULL)) 3893 kprintf(" iosize: %ld, loffset: %lld, " 3894 "flags: 0x%08x, npages: %d\n", 3895 bp->b_vp->v_mount->mnt_stat.f_iosize, 3896 (long long)bp->b_loffset, 3897 bp->b_flags, bp->b_xio.xio_npages); 3898 else 3899 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3900 (long long)bp->b_loffset, 3901 bp->b_flags, bp->b_xio.xio_npages); 3902 kprintf(" valid: 0x%x, dirty: 0x%x, " 3903 "wired: %d\n", 3904 m->valid, m->dirty, 3905 m->wire_count); 3906 panic("bpdone: page busy < 0"); 3907 } 3908 vm_page_io_finish(m); 3909 vm_page_wakeup(m); 3910 vm_object_pip_wakeup(obj); 3911 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3912 iosize -= resid; 3913 } 3914 bp->b_flags &= ~B_HASBOGUS; 3915 vm_object_drop(obj); 3916 } 3917 3918 /* 3919 * Finish up by releasing the buffer. There are no more synchronous 3920 * or asynchronous completions, those were handled by bio_done 3921 * callbacks. 3922 */ 3923 if (elseit) { 3924 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3925 brelse(bp); 3926 else 3927 bqrelse(bp); 3928 } 3929 } 3930 3931 /* 3932 * Normal biodone. 3933 */ 3934 void 3935 biodone(struct bio *bio) 3936 { 3937 struct buf *bp = bio->bio_buf; 3938 3939 runningbufwakeup(bp); 3940 3941 /* 3942 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3943 */ 3944 while (bio) { 3945 biodone_t *done_func; 3946 struct bio_track *track; 3947 3948 /* 3949 * BIO tracking. Most but not all BIOs are tracked. 3950 */ 3951 if ((track = bio->bio_track) != NULL) { 3952 bio_track_rel(track); 3953 bio->bio_track = NULL; 3954 } 3955 3956 /* 3957 * A bio_done function terminates the loop. The function 3958 * will be responsible for any further chaining and/or 3959 * buffer management. 3960 * 3961 * WARNING! The done function can deallocate the buffer! 3962 */ 3963 if ((done_func = bio->bio_done) != NULL) { 3964 bio->bio_done = NULL; 3965 done_func(bio); 3966 return; 3967 } 3968 bio = bio->bio_prev; 3969 } 3970 3971 /* 3972 * If we've run out of bio's do normal [a]synchronous completion. 3973 */ 3974 bpdone(bp, 1); 3975 } 3976 3977 /* 3978 * Synchronous biodone - this terminates a synchronous BIO. 3979 * 3980 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3981 * but still locked. The caller must brelse() the buffer after waiting 3982 * for completion. 3983 */ 3984 void 3985 biodone_sync(struct bio *bio) 3986 { 3987 struct buf *bp = bio->bio_buf; 3988 int flags; 3989 int nflags; 3990 3991 KKASSERT(bio == &bp->b_bio1); 3992 bpdone(bp, 0); 3993 3994 for (;;) { 3995 flags = bio->bio_flags; 3996 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3997 3998 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3999 if (flags & BIO_WANT) 4000 wakeup(bio); 4001 break; 4002 } 4003 } 4004 } 4005 4006 /* 4007 * vfs_unbusy_pages: 4008 * 4009 * This routine is called in lieu of iodone in the case of 4010 * incomplete I/O. This keeps the busy status for pages 4011 * consistant. 4012 */ 4013 void 4014 vfs_unbusy_pages(struct buf *bp) 4015 { 4016 int i; 4017 4018 runningbufwakeup(bp); 4019 4020 if (bp->b_flags & B_VMIO) { 4021 struct vnode *vp = bp->b_vp; 4022 vm_object_t obj; 4023 4024 obj = vp->v_object; 4025 vm_object_hold(obj); 4026 4027 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4028 vm_page_t m = bp->b_xio.xio_pages[i]; 4029 4030 /* 4031 * When restoring bogus changes the original pages 4032 * should still be wired, so we are in no danger of 4033 * losing the object association and do not need 4034 * critical section protection particularly. 4035 */ 4036 if (m == bogus_page) { 4037 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 4038 if (!m) { 4039 panic("vfs_unbusy_pages: page missing"); 4040 } 4041 bp->b_xio.xio_pages[i] = m; 4042 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4043 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4044 } 4045 vm_page_busy_wait(m, FALSE, "bpdpgw"); 4046 vm_page_flag_clear(m, PG_ZERO); 4047 vm_page_io_finish(m); 4048 vm_page_wakeup(m); 4049 vm_object_pip_wakeup(obj); 4050 } 4051 bp->b_flags &= ~B_HASBOGUS; 4052 vm_object_drop(obj); 4053 } 4054 } 4055 4056 /* 4057 * vfs_busy_pages: 4058 * 4059 * This routine is called before a device strategy routine. 4060 * It is used to tell the VM system that paging I/O is in 4061 * progress, and treat the pages associated with the buffer 4062 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4063 * flag is handled to make sure that the object doesn't become 4064 * inconsistant. 4065 * 4066 * Since I/O has not been initiated yet, certain buffer flags 4067 * such as B_ERROR or B_INVAL may be in an inconsistant state 4068 * and should be ignored. 4069 */ 4070 void 4071 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4072 { 4073 int i, bogus; 4074 struct lwp *lp = curthread->td_lwp; 4075 4076 /* 4077 * The buffer's I/O command must already be set. If reading, 4078 * B_CACHE must be 0 (double check against callers only doing 4079 * I/O when B_CACHE is 0). 4080 */ 4081 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4082 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4083 4084 if (bp->b_flags & B_VMIO) { 4085 vm_object_t obj; 4086 4087 obj = vp->v_object; 4088 KASSERT(bp->b_loffset != NOOFFSET, 4089 ("vfs_busy_pages: no buffer offset")); 4090 4091 /* 4092 * Busy all the pages. We have to busy them all at once 4093 * to avoid deadlocks. 4094 */ 4095 retry: 4096 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4097 vm_page_t m = bp->b_xio.xio_pages[i]; 4098 4099 if (vm_page_busy_try(m, FALSE)) { 4100 vm_page_sleep_busy(m, FALSE, "vbpage"); 4101 while (--i >= 0) 4102 vm_page_wakeup(bp->b_xio.xio_pages[i]); 4103 goto retry; 4104 } 4105 } 4106 4107 /* 4108 * Setup for I/O, soft-busy the page right now because 4109 * the next loop may block. 4110 */ 4111 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4112 vm_page_t m = bp->b_xio.xio_pages[i]; 4113 4114 vm_page_flag_clear(m, PG_ZERO); 4115 if ((bp->b_flags & B_CLUSTER) == 0) { 4116 vm_object_pip_add(obj, 1); 4117 vm_page_io_start(m); 4118 } 4119 } 4120 4121 /* 4122 * Adjust protections for I/O and do bogus-page mapping. 4123 * Assume that vm_page_protect() can block (it can block 4124 * if VM_PROT_NONE, don't take any chances regardless). 4125 * 4126 * In particular note that for writes we must incorporate 4127 * page dirtyness from the VM system into the buffer's 4128 * dirty range. 4129 * 4130 * For reads we theoretically must incorporate page dirtyness 4131 * from the VM system to determine if the page needs bogus 4132 * replacement, but we shortcut the test by simply checking 4133 * that all m->valid bits are set, indicating that the page 4134 * is fully valid and does not need to be re-read. For any 4135 * VM system dirtyness the page will also be fully valid 4136 * since it was mapped at one point. 4137 */ 4138 bogus = 0; 4139 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4140 vm_page_t m = bp->b_xio.xio_pages[i]; 4141 4142 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4143 if (bp->b_cmd == BUF_CMD_WRITE) { 4144 /* 4145 * When readying a vnode-backed buffer for 4146 * a write we must zero-fill any invalid 4147 * portions of the backing VM pages, mark 4148 * it valid and clear related dirty bits. 4149 * 4150 * vfs_clean_one_page() incorporates any 4151 * VM dirtyness and updates the b_dirtyoff 4152 * range (after we've made the page RO). 4153 * 4154 * It is also expected that the pmap modified 4155 * bit has already been cleared by the 4156 * vm_page_protect(). We may not be able 4157 * to clear all dirty bits for a page if it 4158 * was also memory mapped (NFS). 4159 * 4160 * Finally be sure to unassign any swap-cache 4161 * backing store as it is now stale. 4162 */ 4163 vm_page_protect(m, VM_PROT_READ); 4164 vfs_clean_one_page(bp, i, m); 4165 swap_pager_unswapped(m); 4166 } else if (m->valid == VM_PAGE_BITS_ALL) { 4167 /* 4168 * When readying a vnode-backed buffer for 4169 * read we must replace any dirty pages with 4170 * a bogus page so dirty data is not destroyed 4171 * when filling gaps. 4172 * 4173 * To avoid testing whether the page is 4174 * dirty we instead test that the page was 4175 * at some point mapped (m->valid fully 4176 * valid) with the understanding that 4177 * this also covers the dirty case. 4178 */ 4179 bp->b_xio.xio_pages[i] = bogus_page; 4180 bp->b_flags |= B_HASBOGUS; 4181 bogus++; 4182 } else if (m->valid & m->dirty) { 4183 /* 4184 * This case should not occur as partial 4185 * dirtyment can only happen if the buffer 4186 * is B_CACHE, and this code is not entered 4187 * if the buffer is B_CACHE. 4188 */ 4189 kprintf("Warning: vfs_busy_pages - page not " 4190 "fully valid! loff=%jx bpf=%08x " 4191 "idx=%d val=%02x dir=%02x\n", 4192 (uintmax_t)bp->b_loffset, bp->b_flags, 4193 i, m->valid, m->dirty); 4194 vm_page_protect(m, VM_PROT_NONE); 4195 } else { 4196 /* 4197 * The page is not valid and can be made 4198 * part of the read. 4199 */ 4200 vm_page_protect(m, VM_PROT_NONE); 4201 } 4202 vm_page_wakeup(m); 4203 } 4204 if (bogus) { 4205 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4206 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4207 } 4208 } 4209 4210 /* 4211 * This is the easiest place to put the process accounting for the I/O 4212 * for now. 4213 */ 4214 if (lp != NULL) { 4215 if (bp->b_cmd == BUF_CMD_READ) 4216 lp->lwp_ru.ru_inblock++; 4217 else 4218 lp->lwp_ru.ru_oublock++; 4219 } 4220 } 4221 4222 /* 4223 * Tell the VM system that the pages associated with this buffer 4224 * are clean. This is used for delayed writes where the data is 4225 * going to go to disk eventually without additional VM intevention. 4226 * 4227 * NOTE: While we only really need to clean through to b_bcount, we 4228 * just go ahead and clean through to b_bufsize. 4229 */ 4230 static void 4231 vfs_clean_pages(struct buf *bp) 4232 { 4233 vm_page_t m; 4234 int i; 4235 4236 if ((bp->b_flags & B_VMIO) == 0) 4237 return; 4238 4239 KASSERT(bp->b_loffset != NOOFFSET, 4240 ("vfs_clean_pages: no buffer offset")); 4241 4242 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4243 m = bp->b_xio.xio_pages[i]; 4244 vfs_clean_one_page(bp, i, m); 4245 } 4246 } 4247 4248 /* 4249 * vfs_clean_one_page: 4250 * 4251 * Set the valid bits and clear the dirty bits in a page within a 4252 * buffer. The range is restricted to the buffer's size and the 4253 * buffer's logical offset might index into the first page. 4254 * 4255 * The caller has busied or soft-busied the page and it is not mapped, 4256 * test and incorporate the dirty bits into b_dirtyoff/end before 4257 * clearing them. Note that we need to clear the pmap modified bits 4258 * after determining the the page was dirty, vm_page_set_validclean() 4259 * does not do it for us. 4260 * 4261 * This routine is typically called after a read completes (dirty should 4262 * be zero in that case as we are not called on bogus-replace pages), 4263 * or before a write is initiated. 4264 */ 4265 static void 4266 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4267 { 4268 int bcount; 4269 int xoff; 4270 int soff; 4271 int eoff; 4272 4273 /* 4274 * Calculate offset range within the page but relative to buffer's 4275 * loffset. loffset might be offset into the first page. 4276 */ 4277 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4278 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4279 4280 if (pageno == 0) { 4281 soff = xoff; 4282 eoff = PAGE_SIZE; 4283 } else { 4284 soff = (pageno << PAGE_SHIFT); 4285 eoff = soff + PAGE_SIZE; 4286 } 4287 if (eoff > bcount) 4288 eoff = bcount; 4289 if (soff >= eoff) 4290 return; 4291 4292 /* 4293 * Test dirty bits and adjust b_dirtyoff/end. 4294 * 4295 * If dirty pages are incorporated into the bp any prior 4296 * B_NEEDCOMMIT state (NFS) must be cleared because the 4297 * caller has not taken into account the new dirty data. 4298 * 4299 * If the page was memory mapped the dirty bits might go beyond the 4300 * end of the buffer, but we can't really make the assumption that 4301 * a file EOF straddles the buffer (even though this is the case for 4302 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4303 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4304 * This also saves some console spam. 4305 * 4306 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4307 * NFS can handle huge commits but not huge writes. 4308 */ 4309 vm_page_test_dirty(m); 4310 if (m->dirty) { 4311 if ((bp->b_flags & B_NEEDCOMMIT) && 4312 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4313 if (debug_commit) 4314 kprintf("Warning: vfs_clean_one_page: bp %p " 4315 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4316 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4317 "doff/end %d %d\n", 4318 bp, (uintmax_t)bp->b_loffset, bp->b_bcount, 4319 bp->b_flags, bp->b_cmd, 4320 m->valid, m->dirty, xoff, soff, eoff, 4321 bp->b_dirtyoff, bp->b_dirtyend); 4322 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4323 if (debug_commit) 4324 print_backtrace(-1); 4325 } 4326 /* 4327 * Only clear the pmap modified bits if ALL the dirty bits 4328 * are set, otherwise the system might mis-clear portions 4329 * of a page. 4330 */ 4331 if (m->dirty == VM_PAGE_BITS_ALL && 4332 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4333 pmap_clear_modify(m); 4334 } 4335 if (bp->b_dirtyoff > soff - xoff) 4336 bp->b_dirtyoff = soff - xoff; 4337 if (bp->b_dirtyend < eoff - xoff) 4338 bp->b_dirtyend = eoff - xoff; 4339 } 4340 4341 /* 4342 * Set related valid bits, clear related dirty bits. 4343 * Does not mess with the pmap modified bit. 4344 * 4345 * WARNING! We cannot just clear all of m->dirty here as the 4346 * buffer cache buffers may use a DEV_BSIZE'd aligned 4347 * block size, or have an odd size (e.g. NFS at file EOF). 4348 * The putpages code can clear m->dirty to 0. 4349 * 4350 * If a VOP_WRITE generates a buffer cache buffer which 4351 * covers the same space as mapped writable pages the 4352 * buffer flush might not be able to clear all the dirty 4353 * bits and still require a putpages from the VM system 4354 * to finish it off. 4355 * 4356 * WARNING! vm_page_set_validclean() currently assumes vm_token 4357 * is held. The page might not be busied (bdwrite() case). 4358 * XXX remove this comment once we've validated that this 4359 * is no longer an issue. 4360 */ 4361 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4362 } 4363 4364 #if 0 4365 /* 4366 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4367 * The page data is assumed to be valid (there is no zeroing here). 4368 */ 4369 static void 4370 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4371 { 4372 int bcount; 4373 int xoff; 4374 int soff; 4375 int eoff; 4376 4377 /* 4378 * Calculate offset range within the page but relative to buffer's 4379 * loffset. loffset might be offset into the first page. 4380 */ 4381 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4382 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4383 4384 if (pageno == 0) { 4385 soff = xoff; 4386 eoff = PAGE_SIZE; 4387 } else { 4388 soff = (pageno << PAGE_SHIFT); 4389 eoff = soff + PAGE_SIZE; 4390 } 4391 if (eoff > bcount) 4392 eoff = bcount; 4393 if (soff >= eoff) 4394 return; 4395 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4396 } 4397 #endif 4398 4399 /* 4400 * vfs_bio_clrbuf: 4401 * 4402 * Clear a buffer. This routine essentially fakes an I/O, so we need 4403 * to clear B_ERROR and B_INVAL. 4404 * 4405 * Note that while we only theoretically need to clear through b_bcount, 4406 * we go ahead and clear through b_bufsize. 4407 */ 4408 4409 void 4410 vfs_bio_clrbuf(struct buf *bp) 4411 { 4412 int i, mask = 0; 4413 caddr_t sa, ea; 4414 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4415 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4416 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4417 (bp->b_loffset & PAGE_MASK) == 0) { 4418 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4419 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4420 bp->b_resid = 0; 4421 return; 4422 } 4423 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4424 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4425 bzero(bp->b_data, bp->b_bufsize); 4426 bp->b_xio.xio_pages[0]->valid |= mask; 4427 bp->b_resid = 0; 4428 return; 4429 } 4430 } 4431 sa = bp->b_data; 4432 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4433 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4434 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4435 ea = (caddr_t)(vm_offset_t)ulmin( 4436 (u_long)(vm_offset_t)ea, 4437 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4438 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4439 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4440 continue; 4441 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4442 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4443 bzero(sa, ea - sa); 4444 } 4445 } else { 4446 for (; sa < ea; sa += DEV_BSIZE, j++) { 4447 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4448 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4449 bzero(sa, DEV_BSIZE); 4450 } 4451 } 4452 bp->b_xio.xio_pages[i]->valid |= mask; 4453 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4454 } 4455 bp->b_resid = 0; 4456 } else { 4457 clrbuf(bp); 4458 } 4459 } 4460 4461 /* 4462 * vm_hold_load_pages: 4463 * 4464 * Load pages into the buffer's address space. The pages are 4465 * allocated from the kernel object in order to reduce interference 4466 * with the any VM paging I/O activity. The range of loaded 4467 * pages will be wired. 4468 * 4469 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4470 * retrieve the full range (to - from) of pages. 4471 */ 4472 void 4473 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4474 { 4475 vm_offset_t pg; 4476 vm_page_t p; 4477 int index; 4478 4479 to = round_page(to); 4480 from = round_page(from); 4481 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4482 4483 pg = from; 4484 while (pg < to) { 4485 /* 4486 * Note: must allocate system pages since blocking here 4487 * could intefere with paging I/O, no matter which 4488 * process we are. 4489 */ 4490 vm_object_hold(&kernel_object); 4491 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT, 4492 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4493 vm_object_drop(&kernel_object); 4494 if (p) { 4495 vm_page_wire(p); 4496 p->valid = VM_PAGE_BITS_ALL; 4497 vm_page_flag_clear(p, PG_ZERO); 4498 pmap_kenter_noinval(pg, VM_PAGE_TO_PHYS(p)); 4499 bp->b_xio.xio_pages[index] = p; 4500 vm_page_wakeup(p); 4501 4502 pg += PAGE_SIZE; 4503 ++index; 4504 } 4505 } 4506 pmap_invalidate_range(&kernel_pmap, from, to); 4507 bp->b_xio.xio_npages = index; 4508 } 4509 4510 /* 4511 * Allocate a page for a buffer cache buffer. 4512 * 4513 * If NULL is returned the caller is expected to retry (typically check if 4514 * the page already exists on retry before trying to allocate one). 4515 * 4516 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This 4517 * function will use the system reserve with the hope that the page 4518 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller 4519 * is done with the buffer. 4520 * 4521 * NOTE! However, TMPFS is a special case because flushing a dirty buffer 4522 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon 4523 * is capable of retiring pages (to swap). For TMPFS we don't dig 4524 * into the system reserve because doing so could stall out pretty 4525 * much every process running on the system. 4526 */ 4527 static 4528 vm_page_t 4529 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit) 4530 { 4531 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK; 4532 vm_page_t p; 4533 4534 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4535 4536 /* 4537 * Try a normal allocation first. 4538 */ 4539 p = vm_page_alloc(obj, pg, vmflags); 4540 if (p) 4541 return(p); 4542 if (vm_page_lookup(obj, pg)) 4543 return(NULL); 4544 vm_pageout_deficit += deficit; 4545 4546 /* 4547 * Try again, digging into the system reserve. 4548 * 4549 * Trying to recover pages from the buffer cache here can deadlock 4550 * against other threads trying to busy underlying pages so we 4551 * depend on the code in brelse() and bqrelse() to free/cache the 4552 * underlying buffer cache pages when memory is low. 4553 */ 4554 if (curthread->td_flags & TDF_SYSTHREAD) 4555 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT; 4556 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS) 4557 vmflags |= 0; 4558 else 4559 vmflags |= VM_ALLOC_SYSTEM; 4560 4561 /*recoverbufpages();*/ 4562 p = vm_page_alloc(obj, pg, vmflags); 4563 if (p) 4564 return(p); 4565 if (vm_page_lookup(obj, pg)) 4566 return(NULL); 4567 4568 /* 4569 * Wait for memory to free up and try again 4570 */ 4571 if (vm_page_count_severe()) 4572 ++lowmempgallocs; 4573 vm_wait(hz / 20 + 1); 4574 4575 p = vm_page_alloc(obj, pg, vmflags); 4576 if (p) 4577 return(p); 4578 if (vm_page_lookup(obj, pg)) 4579 return(NULL); 4580 4581 /* 4582 * Ok, now we are really in trouble. 4583 */ 4584 { 4585 static struct krate biokrate = { .freq = 1 }; 4586 krateprintf(&biokrate, 4587 "Warning: bio_page_alloc: memory exhausted " 4588 "during buffer cache page allocation from %s\n", 4589 curthread->td_comm); 4590 } 4591 if (curthread->td_flags & TDF_SYSTHREAD) 4592 vm_wait(hz / 20 + 1); 4593 else 4594 vm_wait(hz / 2 + 1); 4595 return (NULL); 4596 } 4597 4598 /* 4599 * vm_hold_free_pages: 4600 * 4601 * Return pages associated with the buffer back to the VM system. 4602 * 4603 * The range of pages underlying the buffer's address space will 4604 * be unmapped and un-wired. 4605 */ 4606 void 4607 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4608 { 4609 vm_offset_t pg; 4610 vm_page_t p; 4611 int index, newnpages; 4612 4613 from = round_page(from); 4614 to = round_page(to); 4615 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4616 newnpages = index; 4617 4618 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4619 p = bp->b_xio.xio_pages[index]; 4620 if (p && (index < bp->b_xio.xio_npages)) { 4621 if (p->busy) { 4622 kprintf("vm_hold_free_pages: doffset: %lld, " 4623 "loffset: %lld\n", 4624 (long long)bp->b_bio2.bio_offset, 4625 (long long)bp->b_loffset); 4626 } 4627 bp->b_xio.xio_pages[index] = NULL; 4628 pmap_kremove_noinval(pg); 4629 vm_page_busy_wait(p, FALSE, "vmhldpg"); 4630 vm_page_unwire(p, 0); 4631 vm_page_free(p); 4632 } 4633 } 4634 pmap_invalidate_range(&kernel_pmap, from, to); 4635 bp->b_xio.xio_npages = newnpages; 4636 } 4637 4638 /* 4639 * vmapbuf: 4640 * 4641 * Map a user buffer into KVM via a pbuf. On return the buffer's 4642 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4643 * initialized. 4644 */ 4645 int 4646 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4647 { 4648 caddr_t addr; 4649 vm_offset_t va; 4650 vm_page_t m; 4651 int vmprot; 4652 int error; 4653 int pidx; 4654 int i; 4655 4656 /* 4657 * bp had better have a command and it better be a pbuf. 4658 */ 4659 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4660 KKASSERT(bp->b_flags & B_PAGING); 4661 KKASSERT(bp->b_kvabase); 4662 4663 if (bytes < 0) 4664 return (-1); 4665 4666 /* 4667 * Map the user data into KVM. Mappings have to be page-aligned. 4668 */ 4669 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4670 pidx = 0; 4671 4672 vmprot = VM_PROT_READ; 4673 if (bp->b_cmd == BUF_CMD_READ) 4674 vmprot |= VM_PROT_WRITE; 4675 4676 while (addr < udata + bytes) { 4677 /* 4678 * Do the vm_fault if needed; do the copy-on-write thing 4679 * when reading stuff off device into memory. 4680 * 4681 * vm_fault_page*() returns a held VM page. 4682 */ 4683 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4684 va = trunc_page(va); 4685 4686 m = vm_fault_page_quick(va, vmprot, &error); 4687 if (m == NULL) { 4688 for (i = 0; i < pidx; ++i) { 4689 vm_page_unhold(bp->b_xio.xio_pages[i]); 4690 bp->b_xio.xio_pages[i] = NULL; 4691 } 4692 return(-1); 4693 } 4694 bp->b_xio.xio_pages[pidx] = m; 4695 addr += PAGE_SIZE; 4696 ++pidx; 4697 } 4698 4699 /* 4700 * Map the page array and set the buffer fields to point to 4701 * the mapped data buffer. 4702 */ 4703 if (pidx > btoc(MAXPHYS)) 4704 panic("vmapbuf: mapped more than MAXPHYS"); 4705 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4706 4707 bp->b_xio.xio_npages = pidx; 4708 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4709 bp->b_bcount = bytes; 4710 bp->b_bufsize = bytes; 4711 return(0); 4712 } 4713 4714 /* 4715 * vunmapbuf: 4716 * 4717 * Free the io map PTEs associated with this IO operation. 4718 * We also invalidate the TLB entries and restore the original b_addr. 4719 */ 4720 void 4721 vunmapbuf(struct buf *bp) 4722 { 4723 int pidx; 4724 int npages; 4725 4726 KKASSERT(bp->b_flags & B_PAGING); 4727 4728 npages = bp->b_xio.xio_npages; 4729 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4730 for (pidx = 0; pidx < npages; ++pidx) { 4731 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4732 bp->b_xio.xio_pages[pidx] = NULL; 4733 } 4734 bp->b_xio.xio_npages = 0; 4735 bp->b_data = bp->b_kvabase; 4736 } 4737 4738 /* 4739 * Scan all buffers in the system and issue the callback. 4740 */ 4741 int 4742 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4743 { 4744 int count = 0; 4745 int error; 4746 long n; 4747 4748 for (n = 0; n < nbuf; ++n) { 4749 if ((error = callback(&buf[n], info)) < 0) { 4750 count = error; 4751 break; 4752 } 4753 count += error; 4754 } 4755 return (count); 4756 } 4757 4758 /* 4759 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4760 * completion to the master buffer. 4761 */ 4762 static void 4763 nestiobuf_iodone(struct bio *bio) 4764 { 4765 struct bio *mbio; 4766 struct buf *mbp, *bp; 4767 struct devstat *stats; 4768 int error; 4769 int donebytes; 4770 4771 bp = bio->bio_buf; 4772 mbio = bio->bio_caller_info1.ptr; 4773 stats = bio->bio_caller_info2.ptr; 4774 mbp = mbio->bio_buf; 4775 4776 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4777 KKASSERT(mbp != bp); 4778 4779 error = bp->b_error; 4780 if (bp->b_error == 0 && 4781 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4782 /* 4783 * Not all got transfered, raise an error. We have no way to 4784 * propagate these conditions to mbp. 4785 */ 4786 error = EIO; 4787 } 4788 4789 donebytes = bp->b_bufsize; 4790 4791 relpbuf(bp, NULL); 4792 4793 nestiobuf_done(mbio, donebytes, error, stats); 4794 } 4795 4796 void 4797 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4798 { 4799 struct buf *mbp; 4800 4801 mbp = mbio->bio_buf; 4802 4803 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4804 4805 /* 4806 * If an error occured, propagate it to the master buffer. 4807 * 4808 * Several biodone()s may wind up running concurrently so 4809 * use an atomic op to adjust b_flags. 4810 */ 4811 if (error) { 4812 mbp->b_error = error; 4813 atomic_set_int(&mbp->b_flags, B_ERROR); 4814 } 4815 4816 /* 4817 * Decrement the operations in progress counter and terminate the 4818 * I/O if this was the last bit. 4819 */ 4820 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4821 mbp->b_resid = 0; 4822 if (stats) 4823 devstat_end_transaction_buf(stats, mbp); 4824 biodone(mbio); 4825 } 4826 } 4827 4828 /* 4829 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4830 * the mbio from being biodone()'d while we are still adding sub-bios to 4831 * it. 4832 */ 4833 void 4834 nestiobuf_init(struct bio *bio) 4835 { 4836 bio->bio_driver_info = (void *)1; 4837 } 4838 4839 /* 4840 * The BIOs added to the nestedio have already been started, remove the 4841 * count that placeheld our mbio and biodone() it if the count would 4842 * transition to 0. 4843 */ 4844 void 4845 nestiobuf_start(struct bio *mbio) 4846 { 4847 struct buf *mbp = mbio->bio_buf; 4848 4849 /* 4850 * Decrement the operations in progress counter and terminate the 4851 * I/O if this was the last bit. 4852 */ 4853 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4854 if (mbp->b_flags & B_ERROR) 4855 mbp->b_resid = mbp->b_bcount; 4856 else 4857 mbp->b_resid = 0; 4858 biodone(mbio); 4859 } 4860 } 4861 4862 /* 4863 * Set an intermediate error prior to calling nestiobuf_start() 4864 */ 4865 void 4866 nestiobuf_error(struct bio *mbio, int error) 4867 { 4868 struct buf *mbp = mbio->bio_buf; 4869 4870 if (error) { 4871 mbp->b_error = error; 4872 atomic_set_int(&mbp->b_flags, B_ERROR); 4873 } 4874 } 4875 4876 /* 4877 * nestiobuf_add: setup a "nested" buffer. 4878 * 4879 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4880 * => 'bp' should be a buffer allocated by getiobuf. 4881 * => 'offset' is a byte offset in the master buffer. 4882 * => 'size' is a size in bytes of this nested buffer. 4883 */ 4884 void 4885 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4886 { 4887 struct buf *mbp = mbio->bio_buf; 4888 struct vnode *vp = mbp->b_vp; 4889 4890 KKASSERT(mbp->b_bcount >= offset + size); 4891 4892 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4893 4894 /* kernel needs to own the lock for it to be released in biodone */ 4895 BUF_KERNPROC(bp); 4896 bp->b_vp = vp; 4897 bp->b_cmd = mbp->b_cmd; 4898 bp->b_bio1.bio_done = nestiobuf_iodone; 4899 bp->b_data = (char *)mbp->b_data + offset; 4900 bp->b_resid = bp->b_bcount = size; 4901 bp->b_bufsize = bp->b_bcount; 4902 4903 bp->b_bio1.bio_track = NULL; 4904 bp->b_bio1.bio_caller_info1.ptr = mbio; 4905 bp->b_bio1.bio_caller_info2.ptr = stats; 4906 } 4907 4908 #ifdef DDB 4909 4910 DB_SHOW_COMMAND(buffer, db_show_buffer) 4911 { 4912 /* get args */ 4913 struct buf *bp = (struct buf *)addr; 4914 4915 if (!have_addr) { 4916 db_printf("usage: show buffer <addr>\n"); 4917 return; 4918 } 4919 4920 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4921 db_printf("b_cmd = %d\n", bp->b_cmd); 4922 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4923 "b_resid = %d\n, b_data = %p, " 4924 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4925 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4926 bp->b_data, 4927 (long long)bp->b_bio2.bio_offset, 4928 (long long)(bp->b_bio2.bio_next ? 4929 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4930 if (bp->b_xio.xio_npages) { 4931 int i; 4932 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4933 bp->b_xio.xio_npages); 4934 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4935 vm_page_t m; 4936 m = bp->b_xio.xio_pages[i]; 4937 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4938 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4939 if ((i + 1) < bp->b_xio.xio_npages) 4940 db_printf(","); 4941 } 4942 db_printf("\n"); 4943 } 4944 } 4945 #endif /* DDB */ 4946