1 /* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $ 15 */ 16 17 /* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30 #include <sys/param.h> 31 #include <sys/systm.h> 32 #include <sys/buf.h> 33 #include <sys/conf.h> 34 #include <sys/devicestat.h> 35 #include <sys/eventhandler.h> 36 #include <sys/lock.h> 37 #include <sys/malloc.h> 38 #include <sys/mount.h> 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/proc.h> 42 #include <sys/reboot.h> 43 #include <sys/resourcevar.h> 44 #include <sys/sysctl.h> 45 #include <sys/vmmeter.h> 46 #include <sys/vnode.h> 47 #include <sys/dsched.h> 48 #include <vm/vm.h> 49 #include <vm/vm_param.h> 50 #include <vm/vm_kern.h> 51 #include <vm/vm_pageout.h> 52 #include <vm/vm_page.h> 53 #include <vm/vm_object.h> 54 #include <vm/vm_extern.h> 55 #include <vm/vm_map.h> 56 #include <vm/vm_pager.h> 57 #include <vm/swap_pager.h> 58 59 #include <sys/buf2.h> 60 #include <sys/thread2.h> 61 #include <sys/spinlock2.h> 62 #include <sys/mplock2.h> 63 #include <vm/vm_page2.h> 64 65 #include "opt_ddb.h" 66 #ifdef DDB 67 #include <ddb/ddb.h> 68 #endif 69 70 /* 71 * Buffer queues. 72 */ 73 enum bufq_type { 74 BQUEUE_NONE, /* not on any queue */ 75 BQUEUE_LOCKED, /* locked buffers */ 76 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 77 BQUEUE_DIRTY, /* B_DELWRI buffers */ 78 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */ 79 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 80 BQUEUE_EMPTY, /* empty buffer headers */ 81 82 BUFFER_QUEUES /* number of buffer queues */ 83 }; 84 85 typedef enum bufq_type bufq_type_t; 86 87 #define BD_WAKE_SIZE 16384 88 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1) 89 90 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 91 static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin); 92 static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin); 93 94 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 95 96 struct buf *buf; /* buffer header pool */ 97 98 static void vfs_clean_pages(struct buf *bp); 99 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m); 100 #if 0 101 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m); 102 #endif 103 static void vfs_vmio_release(struct buf *bp); 104 static int flushbufqueues(struct buf *marker, bufq_type_t q); 105 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit); 106 107 static void bd_signal(long totalspace); 108 static void buf_daemon(void); 109 static void buf_daemon_hw(void); 110 111 /* 112 * bogus page -- for I/O to/from partially complete buffers 113 * this is a temporary solution to the problem, but it is not 114 * really that bad. it would be better to split the buffer 115 * for input in the case of buffers partially already in memory, 116 * but the code is intricate enough already. 117 */ 118 vm_page_t bogus_page; 119 120 /* 121 * These are all static, but make the ones we export globals so we do 122 * not need to use compiler magic. 123 */ 124 long bufspace; /* locked by buffer_map */ 125 long maxbufspace; 126 static long bufmallocspace; /* atomic ops */ 127 long maxbufmallocspace, lobufspace, hibufspace; 128 static long bufreusecnt, bufdefragcnt, buffreekvacnt; 129 static long lorunningspace; 130 static long hirunningspace; 131 static long runningbufreq; /* locked by bufcspin */ 132 static long dirtykvaspace; /* locked by bufcspin */ 133 static long dirtybufspace; /* locked by bufcspin */ 134 static long dirtybufcount; /* locked by bufcspin */ 135 static long dirtybufspacehw; /* locked by bufcspin */ 136 static long dirtybufcounthw; /* locked by bufcspin */ 137 static long runningbufspace; /* locked by bufcspin */ 138 static long runningbufcount; /* locked by bufcspin */ 139 long lodirtybufspace; 140 long hidirtybufspace; 141 static int getnewbufcalls; 142 static int getnewbufrestarts; 143 static int recoverbufcalls; 144 static int needsbuffer; /* locked by bufcspin */ 145 static int bd_request; /* locked by bufcspin */ 146 static int bd_request_hw; /* locked by bufcspin */ 147 static u_int bd_wake_ary[BD_WAKE_SIZE]; 148 static u_int bd_wake_index; 149 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 150 static int debug_commit; 151 152 static struct thread *bufdaemon_td; 153 static struct thread *bufdaemonhw_td; 154 static u_int lowmempgallocs; 155 static u_int lowmempgfails; 156 157 /* 158 * Sysctls for operational control of the buffer cache. 159 */ 160 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 161 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 162 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 163 "High watermark used to trigger explicit flushing of dirty buffers"); 164 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 165 "Minimum amount of buffer space required for active I/O"); 166 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 167 "Maximum amount of buffer space to usable for active I/O"); 168 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 169 "Page allocations done during periods of very low free memory"); 170 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 171 "Page allocations which failed during periods of very low free memory"); 172 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 173 "Recycle pages to active or inactive queue transition pt 0-64"); 174 /* 175 * Sysctls determining current state of the buffer cache. 176 */ 177 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 178 "Total number of buffers in buffer cache"); 179 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0, 180 "KVA reserved by dirty buffers (all)"); 181 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 182 "Pending bytes of dirty buffers (all)"); 183 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 184 "Pending bytes of dirty buffers (heavy weight)"); 185 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 186 "Pending number of dirty buffers"); 187 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 188 "Pending number of dirty buffers (heavy weight)"); 189 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 190 "I/O bytes currently in progress due to asynchronous writes"); 191 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 192 "I/O buffers currently in progress due to asynchronous writes"); 193 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 194 "Hard limit on maximum amount of memory usable for buffer space"); 195 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 196 "Soft limit on maximum amount of memory usable for buffer space"); 197 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 198 "Minimum amount of memory to reserve for system buffer space"); 199 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 200 "Amount of memory available for buffers"); 201 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 202 0, "Maximum amount of memory reserved for buffers using malloc"); 203 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 204 "Amount of memory left for buffers using malloc-scheme"); 205 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 206 "New buffer header acquisition requests"); 207 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 208 0, "New buffer header acquisition restarts"); 209 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 210 "Recover VM space in an emergency"); 211 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 212 "Buffer acquisition restarts due to fragmented buffer map"); 213 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 214 "Amount of time KVA space was deallocated in an arbitrary buffer"); 215 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 216 "Amount of time buffer re-use operations were successful"); 217 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, ""); 218 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 219 "sizeof(struct buf)"); 220 221 char *buf_wmesg = BUF_WMESG; 222 223 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 224 #define VFS_BIO_NEED_UNUSED02 0x02 225 #define VFS_BIO_NEED_UNUSED04 0x04 226 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 227 228 /* 229 * bufspacewakeup: 230 * 231 * Called when buffer space is potentially available for recovery. 232 * getnewbuf() will block on this flag when it is unable to free 233 * sufficient buffer space. Buffer space becomes recoverable when 234 * bp's get placed back in the queues. 235 */ 236 static __inline void 237 bufspacewakeup(void) 238 { 239 /* 240 * If someone is waiting for BUF space, wake them up. Even 241 * though we haven't freed the kva space yet, the waiting 242 * process will be able to now. 243 */ 244 spin_lock(&bufcspin); 245 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 246 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 247 spin_unlock(&bufcspin); 248 wakeup(&needsbuffer); 249 } else { 250 spin_unlock(&bufcspin); 251 } 252 } 253 254 /* 255 * runningbufwakeup: 256 * 257 * Accounting for I/O in progress. 258 * 259 */ 260 static __inline void 261 runningbufwakeup(struct buf *bp) 262 { 263 long totalspace; 264 long limit; 265 266 if ((totalspace = bp->b_runningbufspace) != 0) { 267 spin_lock(&bufcspin); 268 runningbufspace -= totalspace; 269 --runningbufcount; 270 bp->b_runningbufspace = 0; 271 272 /* 273 * see waitrunningbufspace() for limit test. 274 */ 275 limit = hirunningspace * 3 / 6; 276 if (runningbufreq && runningbufspace <= limit) { 277 runningbufreq = 0; 278 spin_unlock(&bufcspin); 279 wakeup(&runningbufreq); 280 } else { 281 spin_unlock(&bufcspin); 282 } 283 bd_signal(totalspace); 284 } 285 } 286 287 /* 288 * bufcountwakeup: 289 * 290 * Called when a buffer has been added to one of the free queues to 291 * account for the buffer and to wakeup anyone waiting for free buffers. 292 * This typically occurs when large amounts of metadata are being handled 293 * by the buffer cache ( else buffer space runs out first, usually ). 294 * 295 * MPSAFE 296 */ 297 static __inline void 298 bufcountwakeup(void) 299 { 300 spin_lock(&bufcspin); 301 if (needsbuffer) { 302 needsbuffer &= ~VFS_BIO_NEED_ANY; 303 spin_unlock(&bufcspin); 304 wakeup(&needsbuffer); 305 } else { 306 spin_unlock(&bufcspin); 307 } 308 } 309 310 /* 311 * waitrunningbufspace() 312 * 313 * If runningbufspace exceeds 4/6 hirunningspace we block until 314 * runningbufspace drops to 3/6 hirunningspace. We also block if another 315 * thread blocked here in order to be fair, even if runningbufspace 316 * is now lower than the limit. 317 * 318 * The caller may be using this function to block in a tight loop, we 319 * must block while runningbufspace is greater than at least 320 * hirunningspace * 3 / 6. 321 */ 322 void 323 waitrunningbufspace(void) 324 { 325 long limit = hirunningspace * 4 / 6; 326 327 if (runningbufspace > limit || runningbufreq) { 328 spin_lock(&bufcspin); 329 while (runningbufspace > limit || runningbufreq) { 330 runningbufreq = 1; 331 ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0); 332 } 333 spin_unlock(&bufcspin); 334 } 335 } 336 337 /* 338 * buf_dirty_count_severe: 339 * 340 * Return true if we have too many dirty buffers. 341 */ 342 int 343 buf_dirty_count_severe(void) 344 { 345 return (runningbufspace + dirtykvaspace >= hidirtybufspace || 346 dirtybufcount >= nbuf / 2); 347 } 348 349 /* 350 * Return true if the amount of running I/O is severe and BIOQ should 351 * start bursting. 352 */ 353 int 354 buf_runningbufspace_severe(void) 355 { 356 return (runningbufspace >= hirunningspace * 4 / 6); 357 } 358 359 /* 360 * vfs_buf_test_cache: 361 * 362 * Called when a buffer is extended. This function clears the B_CACHE 363 * bit if the newly extended portion of the buffer does not contain 364 * valid data. 365 * 366 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 367 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 368 * them while a clean buffer was present. 369 */ 370 static __inline__ 371 void 372 vfs_buf_test_cache(struct buf *bp, 373 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 374 vm_page_t m) 375 { 376 if (bp->b_flags & B_CACHE) { 377 int base = (foff + off) & PAGE_MASK; 378 if (vm_page_is_valid(m, base, size) == 0) 379 bp->b_flags &= ~B_CACHE; 380 } 381 } 382 383 /* 384 * bd_speedup() 385 * 386 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 387 * low water mark. 388 * 389 * MPSAFE 390 */ 391 static __inline__ 392 void 393 bd_speedup(void) 394 { 395 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2) 396 return; 397 398 if (bd_request == 0 && 399 (dirtykvaspace > lodirtybufspace / 2 || 400 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 401 spin_lock(&bufcspin); 402 bd_request = 1; 403 spin_unlock(&bufcspin); 404 wakeup(&bd_request); 405 } 406 if (bd_request_hw == 0 && 407 (dirtykvaspace > lodirtybufspace / 2 || 408 dirtybufcounthw >= nbuf / 2)) { 409 spin_lock(&bufcspin); 410 bd_request_hw = 1; 411 spin_unlock(&bufcspin); 412 wakeup(&bd_request_hw); 413 } 414 } 415 416 /* 417 * bd_heatup() 418 * 419 * Get the buf_daemon heated up when the number of running and dirty 420 * buffers exceeds the mid-point. 421 * 422 * Return the total number of dirty bytes past the second mid point 423 * as a measure of how much excess dirty data there is in the system. 424 * 425 * MPSAFE 426 */ 427 long 428 bd_heatup(void) 429 { 430 long mid1; 431 long mid2; 432 long totalspace; 433 434 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 435 436 totalspace = runningbufspace + dirtykvaspace; 437 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 438 bd_speedup(); 439 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 440 if (totalspace >= mid2) 441 return(totalspace - mid2); 442 } 443 return(0); 444 } 445 446 /* 447 * bd_wait() 448 * 449 * Wait for the buffer cache to flush (totalspace) bytes worth of 450 * buffers, then return. 451 * 452 * Regardless this function blocks while the number of dirty buffers 453 * exceeds hidirtybufspace. 454 * 455 * MPSAFE 456 */ 457 void 458 bd_wait(long totalspace) 459 { 460 u_int i; 461 int count; 462 463 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 464 return; 465 466 while (totalspace > 0) { 467 bd_heatup(); 468 if (totalspace > runningbufspace + dirtykvaspace) 469 totalspace = runningbufspace + dirtykvaspace; 470 count = totalspace / BKVASIZE; 471 if (count >= BD_WAKE_SIZE) 472 count = BD_WAKE_SIZE - 1; 473 474 spin_lock(&bufcspin); 475 i = (bd_wake_index + count) & BD_WAKE_MASK; 476 ++bd_wake_ary[i]; 477 478 /* 479 * This is not a strict interlock, so we play a bit loose 480 * with locking access to dirtybufspace* 481 */ 482 tsleep_interlock(&bd_wake_ary[i], 0); 483 spin_unlock(&bufcspin); 484 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz); 485 486 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace; 487 } 488 } 489 490 /* 491 * bd_signal() 492 * 493 * This function is called whenever runningbufspace or dirtykvaspace 494 * is reduced. Track threads waiting for run+dirty buffer I/O 495 * complete. 496 * 497 * MPSAFE 498 */ 499 static void 500 bd_signal(long totalspace) 501 { 502 u_int i; 503 504 if (totalspace > 0) { 505 if (totalspace > BKVASIZE * BD_WAKE_SIZE) 506 totalspace = BKVASIZE * BD_WAKE_SIZE; 507 spin_lock(&bufcspin); 508 while (totalspace > 0) { 509 i = bd_wake_index++; 510 i &= BD_WAKE_MASK; 511 if (bd_wake_ary[i]) { 512 bd_wake_ary[i] = 0; 513 spin_unlock(&bufcspin); 514 wakeup(&bd_wake_ary[i]); 515 spin_lock(&bufcspin); 516 } 517 totalspace -= BKVASIZE; 518 } 519 spin_unlock(&bufcspin); 520 } 521 } 522 523 /* 524 * BIO tracking support routines. 525 * 526 * Release a ref on a bio_track. Wakeup requests are atomically released 527 * along with the last reference so bk_active will never wind up set to 528 * only 0x80000000. 529 * 530 * MPSAFE 531 */ 532 static 533 void 534 bio_track_rel(struct bio_track *track) 535 { 536 int active; 537 int desired; 538 539 /* 540 * Shortcut 541 */ 542 active = track->bk_active; 543 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 544 return; 545 546 /* 547 * Full-on. Note that the wait flag is only atomically released on 548 * the 1->0 count transition. 549 * 550 * We check for a negative count transition using bit 30 since bit 31 551 * has a different meaning. 552 */ 553 for (;;) { 554 desired = (active & 0x7FFFFFFF) - 1; 555 if (desired) 556 desired |= active & 0x80000000; 557 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 558 if (desired & 0x40000000) 559 panic("bio_track_rel: bad count: %p", track); 560 if (active & 0x80000000) 561 wakeup(track); 562 break; 563 } 564 active = track->bk_active; 565 } 566 } 567 568 /* 569 * Wait for the tracking count to reach 0. 570 * 571 * Use atomic ops such that the wait flag is only set atomically when 572 * bk_active is non-zero. 573 * 574 * MPSAFE 575 */ 576 int 577 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 578 { 579 int active; 580 int desired; 581 int error; 582 583 /* 584 * Shortcut 585 */ 586 if (track->bk_active == 0) 587 return(0); 588 589 /* 590 * Full-on. Note that the wait flag may only be atomically set if 591 * the active count is non-zero. 592 * 593 * NOTE: We cannot optimize active == desired since a wakeup could 594 * clear active prior to our tsleep_interlock(). 595 */ 596 error = 0; 597 while ((active = track->bk_active) != 0) { 598 cpu_ccfence(); 599 desired = active | 0x80000000; 600 tsleep_interlock(track, slp_flags); 601 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 602 error = tsleep(track, slp_flags | PINTERLOCKED, 603 "trwait", slp_timo); 604 if (error) 605 break; 606 } 607 } 608 return (error); 609 } 610 611 /* 612 * bufinit: 613 * 614 * Load time initialisation of the buffer cache, called from machine 615 * dependant initialization code. 616 */ 617 void 618 bufinit(void) 619 { 620 struct buf *bp; 621 vm_offset_t bogus_offset; 622 long i; 623 624 /* next, make a null set of free lists */ 625 for (i = 0; i < BUFFER_QUEUES; i++) 626 TAILQ_INIT(&bufqueues[i]); 627 628 /* finally, initialize each buffer header and stick on empty q */ 629 for (i = 0; i < nbuf; i++) { 630 bp = &buf[i]; 631 bzero(bp, sizeof *bp); 632 bp->b_flags = B_INVAL; /* we're just an empty header */ 633 bp->b_cmd = BUF_CMD_DONE; 634 bp->b_qindex = BQUEUE_EMPTY; 635 initbufbio(bp); 636 xio_init(&bp->b_xio); 637 buf_dep_init(bp); 638 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist); 639 } 640 641 /* 642 * maxbufspace is the absolute maximum amount of buffer space we are 643 * allowed to reserve in KVM and in real terms. The absolute maximum 644 * is nominally used by buf_daemon. hibufspace is the nominal maximum 645 * used by most other processes. The differential is required to 646 * ensure that buf_daemon is able to run when other processes might 647 * be blocked waiting for buffer space. 648 * 649 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 650 * this may result in KVM fragmentation which is not handled optimally 651 * by the system. 652 */ 653 maxbufspace = nbuf * BKVASIZE; 654 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 655 lobufspace = hibufspace - MAXBSIZE; 656 657 lorunningspace = 512 * 1024; 658 /* hirunningspace -- see below */ 659 660 /* 661 * Limit the amount of malloc memory since it is wired permanently 662 * into the kernel space. Even though this is accounted for in 663 * the buffer allocation, we don't want the malloced region to grow 664 * uncontrolled. The malloc scheme improves memory utilization 665 * significantly on average (small) directories. 666 */ 667 maxbufmallocspace = hibufspace / 20; 668 669 /* 670 * Reduce the chance of a deadlock occuring by limiting the number 671 * of delayed-write dirty buffers we allow to stack up. 672 * 673 * We don't want too much actually queued to the device at once 674 * (XXX this needs to be per-mount!), because the buffers will 675 * wind up locked for a very long period of time while the I/O 676 * drains. 677 */ 678 hidirtybufspace = hibufspace / 2; /* dirty + running */ 679 hirunningspace = hibufspace / 16; /* locked & queued to device */ 680 if (hirunningspace < 1024 * 1024) 681 hirunningspace = 1024 * 1024; 682 683 dirtykvaspace = 0; 684 dirtybufspace = 0; 685 dirtybufspacehw = 0; 686 687 lodirtybufspace = hidirtybufspace / 2; 688 689 /* 690 * Maximum number of async ops initiated per buf_daemon loop. This is 691 * somewhat of a hack at the moment, we really need to limit ourselves 692 * based on the number of bytes of I/O in-transit that were initiated 693 * from buf_daemon. 694 */ 695 696 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 697 vm_object_hold(&kernel_object); 698 bogus_page = vm_page_alloc(&kernel_object, 699 (bogus_offset >> PAGE_SHIFT), 700 VM_ALLOC_NORMAL); 701 vm_object_drop(&kernel_object); 702 vmstats.v_wire_count++; 703 704 } 705 706 /* 707 * Initialize the embedded bio structures, typically used by 708 * deprecated code which tries to allocate its own struct bufs. 709 */ 710 void 711 initbufbio(struct buf *bp) 712 { 713 bp->b_bio1.bio_buf = bp; 714 bp->b_bio1.bio_prev = NULL; 715 bp->b_bio1.bio_offset = NOOFFSET; 716 bp->b_bio1.bio_next = &bp->b_bio2; 717 bp->b_bio1.bio_done = NULL; 718 bp->b_bio1.bio_flags = 0; 719 720 bp->b_bio2.bio_buf = bp; 721 bp->b_bio2.bio_prev = &bp->b_bio1; 722 bp->b_bio2.bio_offset = NOOFFSET; 723 bp->b_bio2.bio_next = NULL; 724 bp->b_bio2.bio_done = NULL; 725 bp->b_bio2.bio_flags = 0; 726 727 BUF_LOCKINIT(bp); 728 } 729 730 /* 731 * Reinitialize the embedded bio structures as well as any additional 732 * translation cache layers. 733 */ 734 void 735 reinitbufbio(struct buf *bp) 736 { 737 struct bio *bio; 738 739 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 740 bio->bio_done = NULL; 741 bio->bio_offset = NOOFFSET; 742 } 743 } 744 745 /* 746 * Undo the effects of an initbufbio(). 747 */ 748 void 749 uninitbufbio(struct buf *bp) 750 { 751 dsched_exit_buf(bp); 752 BUF_LOCKFREE(bp); 753 } 754 755 /* 756 * Push another BIO layer onto an existing BIO and return it. The new 757 * BIO layer may already exist, holding cached translation data. 758 */ 759 struct bio * 760 push_bio(struct bio *bio) 761 { 762 struct bio *nbio; 763 764 if ((nbio = bio->bio_next) == NULL) { 765 int index = bio - &bio->bio_buf->b_bio_array[0]; 766 if (index >= NBUF_BIO - 1) { 767 panic("push_bio: too many layers bp %p", 768 bio->bio_buf); 769 } 770 nbio = &bio->bio_buf->b_bio_array[index + 1]; 771 bio->bio_next = nbio; 772 nbio->bio_prev = bio; 773 nbio->bio_buf = bio->bio_buf; 774 nbio->bio_offset = NOOFFSET; 775 nbio->bio_done = NULL; 776 nbio->bio_next = NULL; 777 } 778 KKASSERT(nbio->bio_done == NULL); 779 return(nbio); 780 } 781 782 /* 783 * Pop a BIO translation layer, returning the previous layer. The 784 * must have been previously pushed. 785 */ 786 struct bio * 787 pop_bio(struct bio *bio) 788 { 789 return(bio->bio_prev); 790 } 791 792 void 793 clearbiocache(struct bio *bio) 794 { 795 while (bio) { 796 bio->bio_offset = NOOFFSET; 797 bio = bio->bio_next; 798 } 799 } 800 801 /* 802 * bfreekva: 803 * 804 * Free the KVA allocation for buffer 'bp'. 805 * 806 * Must be called from a critical section as this is the only locking for 807 * buffer_map. 808 * 809 * Since this call frees up buffer space, we call bufspacewakeup(). 810 * 811 * MPALMOSTSAFE 812 */ 813 static void 814 bfreekva(struct buf *bp) 815 { 816 int count; 817 818 if (bp->b_kvasize) { 819 ++buffreekvacnt; 820 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 821 vm_map_lock(&buffer_map); 822 bufspace -= bp->b_kvasize; 823 vm_map_delete(&buffer_map, 824 (vm_offset_t) bp->b_kvabase, 825 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 826 &count 827 ); 828 vm_map_unlock(&buffer_map); 829 vm_map_entry_release(count); 830 bp->b_kvasize = 0; 831 bp->b_kvabase = NULL; 832 bufspacewakeup(); 833 } 834 } 835 836 /* 837 * bremfree: 838 * 839 * Remove the buffer from the appropriate free list. 840 */ 841 static __inline void 842 _bremfree(struct buf *bp) 843 { 844 if (bp->b_qindex != BQUEUE_NONE) { 845 KASSERT(BUF_REFCNTNB(bp) == 1, 846 ("bremfree: bp %p not locked",bp)); 847 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 848 bp->b_qindex = BQUEUE_NONE; 849 } else { 850 if (BUF_REFCNTNB(bp) <= 1) 851 panic("bremfree: removing a buffer not on a queue"); 852 } 853 } 854 855 void 856 bremfree(struct buf *bp) 857 { 858 spin_lock(&bufqspin); 859 _bremfree(bp); 860 spin_unlock(&bufqspin); 861 } 862 863 static void 864 bremfree_locked(struct buf *bp) 865 { 866 _bremfree(bp); 867 } 868 869 /* 870 * This version of bread issues any required I/O asyncnronously and 871 * makes a callback on completion. 872 * 873 * The callback must check whether BIO_DONE is set in the bio and issue 874 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 875 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 876 */ 877 void 878 breadcb(struct vnode *vp, off_t loffset, int size, 879 void (*func)(struct bio *), void *arg) 880 { 881 struct buf *bp; 882 883 bp = getblk(vp, loffset, size, 0, 0); 884 885 /* if not found in cache, do some I/O */ 886 if ((bp->b_flags & B_CACHE) == 0) { 887 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 888 bp->b_cmd = BUF_CMD_READ; 889 bp->b_bio1.bio_done = func; 890 bp->b_bio1.bio_caller_info1.ptr = arg; 891 vfs_busy_pages(vp, bp); 892 BUF_KERNPROC(bp); 893 vn_strategy(vp, &bp->b_bio1); 894 } else if (func) { 895 /* 896 * Since we are issuing the callback synchronously it cannot 897 * race the BIO_DONE, so no need for atomic ops here. 898 */ 899 /*bp->b_bio1.bio_done = func;*/ 900 bp->b_bio1.bio_caller_info1.ptr = arg; 901 bp->b_bio1.bio_flags |= BIO_DONE; 902 func(&bp->b_bio1); 903 } else { 904 bqrelse(bp); 905 } 906 } 907 908 /* 909 * breadnx() - Terminal function for bread() and breadn(). 910 * 911 * This function will start asynchronous I/O on read-ahead blocks as well 912 * as satisfy the primary request. 913 * 914 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is 915 * set, the buffer is valid and we do not have to do anything. 916 */ 917 int 918 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset, 919 int *rabsize, int cnt, struct buf **bpp) 920 { 921 struct buf *bp, *rabp; 922 int i; 923 int rv = 0, readwait = 0; 924 925 if (*bpp) 926 bp = *bpp; 927 else 928 *bpp = bp = getblk(vp, loffset, size, 0, 0); 929 930 /* if not found in cache, do some I/O */ 931 if ((bp->b_flags & B_CACHE) == 0) { 932 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 933 bp->b_cmd = BUF_CMD_READ; 934 bp->b_bio1.bio_done = biodone_sync; 935 bp->b_bio1.bio_flags |= BIO_SYNC; 936 vfs_busy_pages(vp, bp); 937 vn_strategy(vp, &bp->b_bio1); 938 ++readwait; 939 } 940 941 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 942 if (inmem(vp, *raoffset)) 943 continue; 944 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 945 946 if ((rabp->b_flags & B_CACHE) == 0) { 947 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 948 rabp->b_cmd = BUF_CMD_READ; 949 vfs_busy_pages(vp, rabp); 950 BUF_KERNPROC(rabp); 951 vn_strategy(vp, &rabp->b_bio1); 952 } else { 953 brelse(rabp); 954 } 955 } 956 if (readwait) 957 rv = biowait(&bp->b_bio1, "biord"); 958 return (rv); 959 } 960 961 /* 962 * bwrite: 963 * 964 * Synchronous write, waits for completion. 965 * 966 * Write, release buffer on completion. (Done by iodone 967 * if async). Do not bother writing anything if the buffer 968 * is invalid. 969 * 970 * Note that we set B_CACHE here, indicating that buffer is 971 * fully valid and thus cacheable. This is true even of NFS 972 * now so we set it generally. This could be set either here 973 * or in biodone() since the I/O is synchronous. We put it 974 * here. 975 */ 976 int 977 bwrite(struct buf *bp) 978 { 979 int error; 980 981 if (bp->b_flags & B_INVAL) { 982 brelse(bp); 983 return (0); 984 } 985 if (BUF_REFCNTNB(bp) == 0) 986 panic("bwrite: buffer is not busy???"); 987 988 /* Mark the buffer clean */ 989 bundirty(bp); 990 991 bp->b_flags &= ~(B_ERROR | B_EINTR); 992 bp->b_flags |= B_CACHE; 993 bp->b_cmd = BUF_CMD_WRITE; 994 bp->b_bio1.bio_done = biodone_sync; 995 bp->b_bio1.bio_flags |= BIO_SYNC; 996 vfs_busy_pages(bp->b_vp, bp); 997 998 /* 999 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1000 * valid for vnode-backed buffers. 1001 */ 1002 bsetrunningbufspace(bp, bp->b_bufsize); 1003 vn_strategy(bp->b_vp, &bp->b_bio1); 1004 error = biowait(&bp->b_bio1, "biows"); 1005 brelse(bp); 1006 1007 return (error); 1008 } 1009 1010 /* 1011 * bawrite: 1012 * 1013 * Asynchronous write. Start output on a buffer, but do not wait for 1014 * it to complete. The buffer is released when the output completes. 1015 * 1016 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1017 * B_INVAL buffers. Not us. 1018 */ 1019 void 1020 bawrite(struct buf *bp) 1021 { 1022 if (bp->b_flags & B_INVAL) { 1023 brelse(bp); 1024 return; 1025 } 1026 if (BUF_REFCNTNB(bp) == 0) 1027 panic("bwrite: buffer is not busy???"); 1028 1029 /* Mark the buffer clean */ 1030 bundirty(bp); 1031 1032 bp->b_flags &= ~(B_ERROR | B_EINTR); 1033 bp->b_flags |= B_CACHE; 1034 bp->b_cmd = BUF_CMD_WRITE; 1035 KKASSERT(bp->b_bio1.bio_done == NULL); 1036 vfs_busy_pages(bp->b_vp, bp); 1037 1038 /* 1039 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1040 * valid for vnode-backed buffers. 1041 */ 1042 bsetrunningbufspace(bp, bp->b_bufsize); 1043 BUF_KERNPROC(bp); 1044 vn_strategy(bp->b_vp, &bp->b_bio1); 1045 } 1046 1047 /* 1048 * bowrite: 1049 * 1050 * Ordered write. Start output on a buffer, and flag it so that the 1051 * device will write it in the order it was queued. The buffer is 1052 * released when the output completes. bwrite() ( or the VOP routine 1053 * anyway ) is responsible for handling B_INVAL buffers. 1054 */ 1055 int 1056 bowrite(struct buf *bp) 1057 { 1058 bp->b_flags |= B_ORDERED; 1059 bawrite(bp); 1060 return (0); 1061 } 1062 1063 /* 1064 * bdwrite: 1065 * 1066 * Delayed write. (Buffer is marked dirty). Do not bother writing 1067 * anything if the buffer is marked invalid. 1068 * 1069 * Note that since the buffer must be completely valid, we can safely 1070 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1071 * biodone() in order to prevent getblk from writing the buffer 1072 * out synchronously. 1073 */ 1074 void 1075 bdwrite(struct buf *bp) 1076 { 1077 if (BUF_REFCNTNB(bp) == 0) 1078 panic("bdwrite: buffer is not busy"); 1079 1080 if (bp->b_flags & B_INVAL) { 1081 brelse(bp); 1082 return; 1083 } 1084 bdirty(bp); 1085 1086 if (dsched_is_clear_buf_priv(bp)) 1087 dsched_new_buf(bp); 1088 1089 /* 1090 * Set B_CACHE, indicating that the buffer is fully valid. This is 1091 * true even of NFS now. 1092 */ 1093 bp->b_flags |= B_CACHE; 1094 1095 /* 1096 * This bmap keeps the system from needing to do the bmap later, 1097 * perhaps when the system is attempting to do a sync. Since it 1098 * is likely that the indirect block -- or whatever other datastructure 1099 * that the filesystem needs is still in memory now, it is a good 1100 * thing to do this. Note also, that if the pageout daemon is 1101 * requesting a sync -- there might not be enough memory to do 1102 * the bmap then... So, this is important to do. 1103 */ 1104 if (bp->b_bio2.bio_offset == NOOFFSET) { 1105 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1106 NULL, NULL, BUF_CMD_WRITE); 1107 } 1108 1109 /* 1110 * Because the underlying pages may still be mapped and 1111 * writable trying to set the dirty buffer (b_dirtyoff/end) 1112 * range here will be inaccurate. 1113 * 1114 * However, we must still clean the pages to satisfy the 1115 * vnode_pager and pageout daemon, so theythink the pages 1116 * have been "cleaned". What has really occured is that 1117 * they've been earmarked for later writing by the buffer 1118 * cache. 1119 * 1120 * So we get the b_dirtyoff/end update but will not actually 1121 * depend on it (NFS that is) until the pages are busied for 1122 * writing later on. 1123 */ 1124 vfs_clean_pages(bp); 1125 bqrelse(bp); 1126 1127 /* 1128 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1129 * due to the softdep code. 1130 */ 1131 } 1132 1133 /* 1134 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1135 * This is used by tmpfs. 1136 * 1137 * It is important for any VFS using this routine to NOT use it for 1138 * IO_SYNC or IO_ASYNC operations which occur when the system really 1139 * wants to flush VM pages to backing store. 1140 */ 1141 void 1142 buwrite(struct buf *bp) 1143 { 1144 vm_page_t m; 1145 int i; 1146 1147 /* 1148 * Only works for VMIO buffers. If the buffer is already 1149 * marked for delayed-write we can't avoid the bdwrite(). 1150 */ 1151 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1152 bdwrite(bp); 1153 return; 1154 } 1155 1156 /* 1157 * Mark as needing a commit. 1158 */ 1159 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1160 m = bp->b_xio.xio_pages[i]; 1161 vm_page_need_commit(m); 1162 } 1163 bqrelse(bp); 1164 } 1165 1166 /* 1167 * bdirty: 1168 * 1169 * Turn buffer into delayed write request by marking it B_DELWRI. 1170 * B_RELBUF and B_NOCACHE must be cleared. 1171 * 1172 * We reassign the buffer to itself to properly update it in the 1173 * dirty/clean lists. 1174 * 1175 * Must be called from a critical section. 1176 * The buffer must be on BQUEUE_NONE. 1177 */ 1178 void 1179 bdirty(struct buf *bp) 1180 { 1181 KASSERT(bp->b_qindex == BQUEUE_NONE, 1182 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1183 if (bp->b_flags & B_NOCACHE) { 1184 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1185 bp->b_flags &= ~B_NOCACHE; 1186 } 1187 if (bp->b_flags & B_INVAL) { 1188 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1189 } 1190 bp->b_flags &= ~B_RELBUF; 1191 1192 if ((bp->b_flags & B_DELWRI) == 0) { 1193 lwkt_gettoken(&bp->b_vp->v_token); 1194 bp->b_flags |= B_DELWRI; 1195 reassignbuf(bp); 1196 lwkt_reltoken(&bp->b_vp->v_token); 1197 1198 spin_lock(&bufcspin); 1199 ++dirtybufcount; 1200 dirtykvaspace += bp->b_kvasize; 1201 dirtybufspace += bp->b_bufsize; 1202 if (bp->b_flags & B_HEAVY) { 1203 ++dirtybufcounthw; 1204 dirtybufspacehw += bp->b_bufsize; 1205 } 1206 spin_unlock(&bufcspin); 1207 1208 bd_heatup(); 1209 } 1210 } 1211 1212 /* 1213 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1214 * needs to be flushed with a different buf_daemon thread to avoid 1215 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1216 */ 1217 void 1218 bheavy(struct buf *bp) 1219 { 1220 if ((bp->b_flags & B_HEAVY) == 0) { 1221 bp->b_flags |= B_HEAVY; 1222 if (bp->b_flags & B_DELWRI) { 1223 spin_lock(&bufcspin); 1224 ++dirtybufcounthw; 1225 dirtybufspacehw += bp->b_bufsize; 1226 spin_unlock(&bufcspin); 1227 } 1228 } 1229 } 1230 1231 /* 1232 * bundirty: 1233 * 1234 * Clear B_DELWRI for buffer. 1235 * 1236 * Must be called from a critical section. 1237 * 1238 * The buffer is typically on BQUEUE_NONE but there is one case in 1239 * brelse() that calls this function after placing the buffer on 1240 * a different queue. 1241 * 1242 * MPSAFE 1243 */ 1244 void 1245 bundirty(struct buf *bp) 1246 { 1247 if (bp->b_flags & B_DELWRI) { 1248 lwkt_gettoken(&bp->b_vp->v_token); 1249 bp->b_flags &= ~B_DELWRI; 1250 reassignbuf(bp); 1251 lwkt_reltoken(&bp->b_vp->v_token); 1252 1253 spin_lock(&bufcspin); 1254 --dirtybufcount; 1255 dirtykvaspace -= bp->b_kvasize; 1256 dirtybufspace -= bp->b_bufsize; 1257 if (bp->b_flags & B_HEAVY) { 1258 --dirtybufcounthw; 1259 dirtybufspacehw -= bp->b_bufsize; 1260 } 1261 spin_unlock(&bufcspin); 1262 1263 bd_signal(bp->b_bufsize); 1264 } 1265 /* 1266 * Since it is now being written, we can clear its deferred write flag. 1267 */ 1268 bp->b_flags &= ~B_DEFERRED; 1269 } 1270 1271 /* 1272 * Set the b_runningbufspace field, used to track how much I/O is 1273 * in progress at any given moment. 1274 */ 1275 void 1276 bsetrunningbufspace(struct buf *bp, int bytes) 1277 { 1278 bp->b_runningbufspace = bytes; 1279 if (bytes) { 1280 spin_lock(&bufcspin); 1281 runningbufspace += bytes; 1282 ++runningbufcount; 1283 spin_unlock(&bufcspin); 1284 } 1285 } 1286 1287 /* 1288 * brelse: 1289 * 1290 * Release a busy buffer and, if requested, free its resources. The 1291 * buffer will be stashed in the appropriate bufqueue[] allowing it 1292 * to be accessed later as a cache entity or reused for other purposes. 1293 * 1294 * MPALMOSTSAFE 1295 */ 1296 void 1297 brelse(struct buf *bp) 1298 { 1299 #ifdef INVARIANTS 1300 int saved_flags = bp->b_flags; 1301 #endif 1302 1303 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1304 1305 /* 1306 * If B_NOCACHE is set we are being asked to destroy the buffer and 1307 * its backing store. Clear B_DELWRI. 1308 * 1309 * B_NOCACHE is set in two cases: (1) when the caller really wants 1310 * to destroy the buffer and backing store and (2) when the caller 1311 * wants to destroy the buffer and backing store after a write 1312 * completes. 1313 */ 1314 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1315 bundirty(bp); 1316 } 1317 1318 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1319 /* 1320 * A re-dirtied buffer is only subject to destruction 1321 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1322 */ 1323 /* leave buffer intact */ 1324 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1325 (bp->b_bufsize <= 0)) { 1326 /* 1327 * Either a failed read or we were asked to free or not 1328 * cache the buffer. This path is reached with B_DELWRI 1329 * set only if B_INVAL is already set. B_NOCACHE governs 1330 * backing store destruction. 1331 * 1332 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1333 * buffer cannot be immediately freed. 1334 */ 1335 bp->b_flags |= B_INVAL; 1336 if (LIST_FIRST(&bp->b_dep) != NULL) 1337 buf_deallocate(bp); 1338 if (bp->b_flags & B_DELWRI) { 1339 spin_lock(&bufcspin); 1340 --dirtybufcount; 1341 dirtykvaspace -= bp->b_kvasize; 1342 dirtybufspace -= bp->b_bufsize; 1343 if (bp->b_flags & B_HEAVY) { 1344 --dirtybufcounthw; 1345 dirtybufspacehw -= bp->b_bufsize; 1346 } 1347 spin_unlock(&bufcspin); 1348 1349 bd_signal(bp->b_bufsize); 1350 } 1351 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1352 } 1353 1354 /* 1355 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set, 1356 * or if b_refs is non-zero. 1357 * 1358 * If vfs_vmio_release() is called with either bit set, the 1359 * underlying pages may wind up getting freed causing a previous 1360 * write (bdwrite()) to get 'lost' because pages associated with 1361 * a B_DELWRI bp are marked clean. Pages associated with a 1362 * B_LOCKED buffer may be mapped by the filesystem. 1363 * 1364 * If we want to release the buffer ourselves (rather then the 1365 * originator asking us to release it), give the originator a 1366 * chance to countermand the release by setting B_LOCKED. 1367 * 1368 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1369 * if B_DELWRI is set. 1370 * 1371 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1372 * on pages to return pages to the VM page queues. 1373 */ 1374 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) { 1375 bp->b_flags &= ~B_RELBUF; 1376 } else if (vm_page_count_min(0)) { 1377 if (LIST_FIRST(&bp->b_dep) != NULL) 1378 buf_deallocate(bp); /* can set B_LOCKED */ 1379 if (bp->b_flags & (B_DELWRI | B_LOCKED)) 1380 bp->b_flags &= ~B_RELBUF; 1381 else 1382 bp->b_flags |= B_RELBUF; 1383 } 1384 1385 /* 1386 * Make sure b_cmd is clear. It may have already been cleared by 1387 * biodone(). 1388 * 1389 * At this point destroying the buffer is governed by the B_INVAL 1390 * or B_RELBUF flags. 1391 */ 1392 bp->b_cmd = BUF_CMD_DONE; 1393 dsched_exit_buf(bp); 1394 1395 /* 1396 * VMIO buffer rundown. Make sure the VM page array is restored 1397 * after an I/O may have replaces some of the pages with bogus pages 1398 * in order to not destroy dirty pages in a fill-in read. 1399 * 1400 * Note that due to the code above, if a buffer is marked B_DELWRI 1401 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1402 * B_INVAL may still be set, however. 1403 * 1404 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1405 * but not the backing store. B_NOCACHE will destroy the backing 1406 * store. 1407 * 1408 * Note that dirty NFS buffers contain byte-granular write ranges 1409 * and should not be destroyed w/ B_INVAL even if the backing store 1410 * is left intact. 1411 */ 1412 if (bp->b_flags & B_VMIO) { 1413 /* 1414 * Rundown for VMIO buffers which are not dirty NFS buffers. 1415 */ 1416 int i, j, resid; 1417 vm_page_t m; 1418 off_t foff; 1419 vm_pindex_t poff; 1420 vm_object_t obj; 1421 struct vnode *vp; 1422 1423 vp = bp->b_vp; 1424 1425 /* 1426 * Get the base offset and length of the buffer. Note that 1427 * in the VMIO case if the buffer block size is not 1428 * page-aligned then b_data pointer may not be page-aligned. 1429 * But our b_xio.xio_pages array *IS* page aligned. 1430 * 1431 * block sizes less then DEV_BSIZE (usually 512) are not 1432 * supported due to the page granularity bits (m->valid, 1433 * m->dirty, etc...). 1434 * 1435 * See man buf(9) for more information 1436 */ 1437 1438 resid = bp->b_bufsize; 1439 foff = bp->b_loffset; 1440 1441 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1442 m = bp->b_xio.xio_pages[i]; 1443 vm_page_flag_clear(m, PG_ZERO); 1444 /* 1445 * If we hit a bogus page, fixup *all* of them 1446 * now. Note that we left these pages wired 1447 * when we removed them so they had better exist, 1448 * and they cannot be ripped out from under us so 1449 * no critical section protection is necessary. 1450 */ 1451 if (m == bogus_page) { 1452 obj = vp->v_object; 1453 poff = OFF_TO_IDX(bp->b_loffset); 1454 1455 vm_object_hold(obj); 1456 for (j = i; j < bp->b_xio.xio_npages; j++) { 1457 vm_page_t mtmp; 1458 1459 mtmp = bp->b_xio.xio_pages[j]; 1460 if (mtmp == bogus_page) { 1461 mtmp = vm_page_lookup(obj, poff + j); 1462 if (!mtmp) { 1463 panic("brelse: page missing"); 1464 } 1465 bp->b_xio.xio_pages[j] = mtmp; 1466 } 1467 } 1468 bp->b_flags &= ~B_HASBOGUS; 1469 vm_object_drop(obj); 1470 1471 if ((bp->b_flags & B_INVAL) == 0) { 1472 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1473 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1474 } 1475 m = bp->b_xio.xio_pages[i]; 1476 } 1477 1478 /* 1479 * Invalidate the backing store if B_NOCACHE is set 1480 * (e.g. used with vinvalbuf()). If this is NFS 1481 * we impose a requirement that the block size be 1482 * a multiple of PAGE_SIZE and create a temporary 1483 * hack to basically invalidate the whole page. The 1484 * problem is that NFS uses really odd buffer sizes 1485 * especially when tracking piecemeal writes and 1486 * it also vinvalbuf()'s a lot, which would result 1487 * in only partial page validation and invalidation 1488 * here. If the file page is mmap()'d, however, 1489 * all the valid bits get set so after we invalidate 1490 * here we would end up with weird m->valid values 1491 * like 0xfc. nfs_getpages() can't handle this so 1492 * we clear all the valid bits for the NFS case 1493 * instead of just some of them. 1494 * 1495 * The real bug is the VM system having to set m->valid 1496 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1497 * itself is an artifact of the whole 512-byte 1498 * granular mess that exists to support odd block 1499 * sizes and UFS meta-data block sizes (e.g. 6144). 1500 * A complete rewrite is required. 1501 * 1502 * XXX 1503 */ 1504 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1505 int poffset = foff & PAGE_MASK; 1506 int presid; 1507 1508 presid = PAGE_SIZE - poffset; 1509 if (bp->b_vp->v_tag == VT_NFS && 1510 bp->b_vp->v_type == VREG) { 1511 ; /* entire page */ 1512 } else if (presid > resid) { 1513 presid = resid; 1514 } 1515 KASSERT(presid >= 0, ("brelse: extra page")); 1516 vm_page_set_invalid(m, poffset, presid); 1517 1518 /* 1519 * Also make sure any swap cache is removed 1520 * as it is now stale (HAMMER in particular 1521 * uses B_NOCACHE to deal with buffer 1522 * aliasing). 1523 */ 1524 swap_pager_unswapped(m); 1525 } 1526 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1527 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1528 } 1529 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1530 vfs_vmio_release(bp); 1531 } else { 1532 /* 1533 * Rundown for non-VMIO buffers. 1534 */ 1535 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1536 if (bp->b_bufsize) 1537 allocbuf(bp, 0); 1538 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1539 if (bp->b_vp) 1540 brelvp(bp); 1541 } 1542 } 1543 1544 if (bp->b_qindex != BQUEUE_NONE) 1545 panic("brelse: free buffer onto another queue???"); 1546 if (BUF_REFCNTNB(bp) > 1) { 1547 /* Temporary panic to verify exclusive locking */ 1548 /* This panic goes away when we allow shared refs */ 1549 panic("brelse: multiple refs"); 1550 /* NOT REACHED */ 1551 return; 1552 } 1553 1554 /* 1555 * Figure out the correct queue to place the cleaned up buffer on. 1556 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1557 * disassociated from their vnode. 1558 */ 1559 spin_lock(&bufqspin); 1560 if (bp->b_flags & B_LOCKED) { 1561 /* 1562 * Buffers that are locked are placed in the locked queue 1563 * immediately, regardless of their state. 1564 */ 1565 bp->b_qindex = BQUEUE_LOCKED; 1566 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1567 } else if (bp->b_bufsize == 0) { 1568 /* 1569 * Buffers with no memory. Due to conditionals near the top 1570 * of brelse() such buffers should probably already be 1571 * marked B_INVAL and disassociated from their vnode. 1572 */ 1573 bp->b_flags |= B_INVAL; 1574 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1575 KKASSERT((bp->b_flags & B_HASHED) == 0); 1576 if (bp->b_kvasize) { 1577 bp->b_qindex = BQUEUE_EMPTYKVA; 1578 } else { 1579 bp->b_qindex = BQUEUE_EMPTY; 1580 } 1581 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1582 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1583 /* 1584 * Buffers with junk contents. Again these buffers had better 1585 * already be disassociated from their vnode. 1586 */ 1587 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1588 KKASSERT((bp->b_flags & B_HASHED) == 0); 1589 bp->b_flags |= B_INVAL; 1590 bp->b_qindex = BQUEUE_CLEAN; 1591 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1592 } else { 1593 /* 1594 * Remaining buffers. These buffers are still associated with 1595 * their vnode. 1596 */ 1597 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1598 case B_DELWRI: 1599 bp->b_qindex = BQUEUE_DIRTY; 1600 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1601 break; 1602 case B_DELWRI | B_HEAVY: 1603 bp->b_qindex = BQUEUE_DIRTY_HW; 1604 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp, 1605 b_freelist); 1606 break; 1607 default: 1608 /* 1609 * NOTE: Buffers are always placed at the end of the 1610 * queue. If B_AGE is not set the buffer will cycle 1611 * through the queue twice. 1612 */ 1613 bp->b_qindex = BQUEUE_CLEAN; 1614 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1615 break; 1616 } 1617 } 1618 spin_unlock(&bufqspin); 1619 1620 /* 1621 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1622 * on the correct queue. 1623 */ 1624 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1625 bundirty(bp); 1626 1627 /* 1628 * The bp is on an appropriate queue unless locked. If it is not 1629 * locked or dirty we can wakeup threads waiting for buffer space. 1630 * 1631 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1632 * if B_INVAL is set ). 1633 */ 1634 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1635 bufcountwakeup(); 1636 1637 /* 1638 * Something we can maybe free or reuse 1639 */ 1640 if (bp->b_bufsize || bp->b_kvasize) 1641 bufspacewakeup(); 1642 1643 /* 1644 * Clean up temporary flags and unlock the buffer. 1645 */ 1646 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT); 1647 BUF_UNLOCK(bp); 1648 } 1649 1650 /* 1651 * bqrelse: 1652 * 1653 * Release a buffer back to the appropriate queue but do not try to free 1654 * it. The buffer is expected to be used again soon. 1655 * 1656 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1657 * biodone() to requeue an async I/O on completion. It is also used when 1658 * known good buffers need to be requeued but we think we may need the data 1659 * again soon. 1660 * 1661 * XXX we should be able to leave the B_RELBUF hint set on completion. 1662 * 1663 * MPSAFE 1664 */ 1665 void 1666 bqrelse(struct buf *bp) 1667 { 1668 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1669 1670 if (bp->b_qindex != BQUEUE_NONE) 1671 panic("bqrelse: free buffer onto another queue???"); 1672 if (BUF_REFCNTNB(bp) > 1) { 1673 /* do not release to free list */ 1674 panic("bqrelse: multiple refs"); 1675 return; 1676 } 1677 1678 buf_act_advance(bp); 1679 1680 spin_lock(&bufqspin); 1681 if (bp->b_flags & B_LOCKED) { 1682 /* 1683 * Locked buffers are released to the locked queue. However, 1684 * if the buffer is dirty it will first go into the dirty 1685 * queue and later on after the I/O completes successfully it 1686 * will be released to the locked queue. 1687 */ 1688 bp->b_qindex = BQUEUE_LOCKED; 1689 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1690 } else if (bp->b_flags & B_DELWRI) { 1691 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1692 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1693 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1694 } else if (vm_page_count_min(0)) { 1695 /* 1696 * We are too low on memory, we have to try to free the 1697 * buffer (most importantly: the wired pages making up its 1698 * backing store) *now*. 1699 */ 1700 spin_unlock(&bufqspin); 1701 brelse(bp); 1702 return; 1703 } else { 1704 bp->b_qindex = BQUEUE_CLEAN; 1705 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1706 } 1707 spin_unlock(&bufqspin); 1708 1709 if ((bp->b_flags & B_LOCKED) == 0 && 1710 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1711 bufcountwakeup(); 1712 } 1713 1714 /* 1715 * Something we can maybe free or reuse. 1716 */ 1717 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1718 bufspacewakeup(); 1719 1720 /* 1721 * Final cleanup and unlock. Clear bits that are only used while a 1722 * buffer is actively locked. 1723 */ 1724 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF); 1725 dsched_exit_buf(bp); 1726 BUF_UNLOCK(bp); 1727 } 1728 1729 /* 1730 * Hold a buffer, preventing it from being reused. This will prevent 1731 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL 1732 * operations. If a B_INVAL operation occurs the buffer will remain held 1733 * but the underlying pages may get ripped out. 1734 * 1735 * These functions are typically used in VOP_READ/VOP_WRITE functions 1736 * to hold a buffer during a copyin or copyout, preventing deadlocks 1737 * or recursive lock panics when read()/write() is used over mmap()'d 1738 * space. 1739 * 1740 * NOTE: bqhold() requires that the buffer be locked at the time of the 1741 * hold. bqdrop() has no requirements other than the buffer having 1742 * previously been held. 1743 */ 1744 void 1745 bqhold(struct buf *bp) 1746 { 1747 atomic_add_int(&bp->b_refs, 1); 1748 } 1749 1750 void 1751 bqdrop(struct buf *bp) 1752 { 1753 KKASSERT(bp->b_refs > 0); 1754 atomic_add_int(&bp->b_refs, -1); 1755 } 1756 1757 /* 1758 * Return backing pages held by the buffer 'bp' back to the VM system. 1759 * This routine is called when the bp is invalidated, released, or 1760 * reused. 1761 * 1762 * The KVA mapping (b_data) for the underlying pages is removed by 1763 * this function. 1764 * 1765 * WARNING! This routine is integral to the low memory critical path 1766 * when a buffer is B_RELBUF'd. If the system has a severe page 1767 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE 1768 * queues so they can be reused in the current pageout daemon 1769 * pass. 1770 */ 1771 static void 1772 vfs_vmio_release(struct buf *bp) 1773 { 1774 int i; 1775 vm_page_t m; 1776 1777 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1778 m = bp->b_xio.xio_pages[i]; 1779 bp->b_xio.xio_pages[i] = NULL; 1780 1781 /* 1782 * We need to own the page in order to safely unwire it. 1783 */ 1784 vm_page_busy_wait(m, FALSE, "vmiopg"); 1785 1786 /* 1787 * The VFS is telling us this is not a meta-data buffer 1788 * even if it is backed by a block device. 1789 */ 1790 if (bp->b_flags & B_NOTMETA) 1791 vm_page_flag_set(m, PG_NOTMETA); 1792 1793 /* 1794 * This is a very important bit of code. We try to track 1795 * VM page use whether the pages are wired into the buffer 1796 * cache or not. While wired into the buffer cache the 1797 * bp tracks the act_count. 1798 * 1799 * We can choose to place unwired pages on the inactive 1800 * queue (0) or active queue (1). If we place too many 1801 * on the active queue the queue will cycle the act_count 1802 * on pages we'd like to keep, just from single-use pages 1803 * (such as when doing a tar-up or file scan). 1804 */ 1805 if (bp->b_act_count < vm_cycle_point) 1806 vm_page_unwire(m, 0); 1807 else 1808 vm_page_unwire(m, 1); 1809 1810 /* 1811 * If the wire_count has dropped to 0 we may need to take 1812 * further action before unbusying the page. 1813 * 1814 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us. 1815 */ 1816 if (m->wire_count == 0) { 1817 vm_page_flag_clear(m, PG_ZERO); 1818 1819 if (bp->b_flags & B_DIRECT) { 1820 /* 1821 * Attempt to free the page if B_DIRECT is 1822 * set, the caller does not desire the page 1823 * to be cached. 1824 */ 1825 vm_page_wakeup(m); 1826 vm_page_try_to_free(m); 1827 } else if ((bp->b_flags & B_NOTMETA) || 1828 vm_page_count_min(0)) { 1829 /* 1830 * Attempt to move the page to PQ_CACHE 1831 * if B_NOTMETA is set. This flag is set 1832 * by HAMMER to remove one of the two pages 1833 * present when double buffering is enabled. 1834 * 1835 * Attempt to move the page to PQ_CACHE 1836 * If we have a severe page deficit. This 1837 * will cause buffer cache operations related 1838 * to pageouts to recycle the related pages 1839 * in order to avoid a low memory deadlock. 1840 */ 1841 m->act_count = bp->b_act_count; 1842 vm_page_wakeup(m); 1843 vm_page_try_to_cache(m); 1844 } else { 1845 /* 1846 * Nominal case, leave the page on the 1847 * queue the original unwiring placed it on 1848 * (active or inactive). 1849 */ 1850 m->act_count = bp->b_act_count; 1851 vm_page_wakeup(m); 1852 } 1853 } else { 1854 vm_page_wakeup(m); 1855 } 1856 } 1857 1858 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1859 bp->b_xio.xio_npages); 1860 if (bp->b_bufsize) { 1861 bufspacewakeup(); 1862 bp->b_bufsize = 0; 1863 } 1864 bp->b_xio.xio_npages = 0; 1865 bp->b_flags &= ~B_VMIO; 1866 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1867 if (bp->b_vp) 1868 brelvp(bp); 1869 } 1870 1871 /* 1872 * getnewbuf: 1873 * 1874 * Find and initialize a new buffer header, freeing up existing buffers 1875 * in the bufqueues as necessary. The new buffer is returned locked. 1876 * 1877 * Important: B_INVAL is not set. If the caller wishes to throw the 1878 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1879 * 1880 * We block if: 1881 * We have insufficient buffer headers 1882 * We have insufficient buffer space 1883 * buffer_map is too fragmented ( space reservation fails ) 1884 * If we have to flush dirty buffers ( but we try to avoid this ) 1885 * 1886 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1887 * Instead we ask the buf daemon to do it for us. We attempt to 1888 * avoid piecemeal wakeups of the pageout daemon. 1889 * 1890 * MPALMOSTSAFE 1891 */ 1892 struct buf * 1893 getnewbuf(int blkflags, int slptimeo, int size, int maxsize) 1894 { 1895 struct buf *bp; 1896 struct buf *nbp; 1897 int defrag = 0; 1898 int nqindex; 1899 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1900 static int flushingbufs; 1901 1902 /* 1903 * We can't afford to block since we might be holding a vnode lock, 1904 * which may prevent system daemons from running. We deal with 1905 * low-memory situations by proactively returning memory and running 1906 * async I/O rather then sync I/O. 1907 */ 1908 1909 ++getnewbufcalls; 1910 --getnewbufrestarts; 1911 restart: 1912 ++getnewbufrestarts; 1913 1914 /* 1915 * Setup for scan. If we do not have enough free buffers, 1916 * we setup a degenerate case that immediately fails. Note 1917 * that if we are specially marked process, we are allowed to 1918 * dip into our reserves. 1919 * 1920 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1921 * 1922 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1923 * However, there are a number of cases (defragging, reusing, ...) 1924 * where we cannot backup. 1925 */ 1926 nqindex = BQUEUE_EMPTYKVA; 1927 spin_lock(&bufqspin); 1928 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]); 1929 1930 if (nbp == NULL) { 1931 /* 1932 * If no EMPTYKVA buffers and we are either 1933 * defragging or reusing, locate a CLEAN buffer 1934 * to free or reuse. If bufspace useage is low 1935 * skip this step so we can allocate a new buffer. 1936 */ 1937 if (defrag || bufspace >= lobufspace) { 1938 nqindex = BQUEUE_CLEAN; 1939 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 1940 } 1941 1942 /* 1943 * If we could not find or were not allowed to reuse a 1944 * CLEAN buffer, check to see if it is ok to use an EMPTY 1945 * buffer. We can only use an EMPTY buffer if allocating 1946 * its KVA would not otherwise run us out of buffer space. 1947 */ 1948 if (nbp == NULL && defrag == 0 && 1949 bufspace + maxsize < hibufspace) { 1950 nqindex = BQUEUE_EMPTY; 1951 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]); 1952 } 1953 } 1954 1955 /* 1956 * Run scan, possibly freeing data and/or kva mappings on the fly 1957 * depending. 1958 * 1959 * WARNING! bufqspin is held! 1960 */ 1961 while ((bp = nbp) != NULL) { 1962 int qindex = nqindex; 1963 1964 nbp = TAILQ_NEXT(bp, b_freelist); 1965 1966 /* 1967 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 1968 * cycles through the queue twice before being selected. 1969 */ 1970 if (qindex == BQUEUE_CLEAN && 1971 (bp->b_flags & B_AGE) == 0 && nbp) { 1972 bp->b_flags |= B_AGE; 1973 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist); 1974 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist); 1975 continue; 1976 } 1977 1978 /* 1979 * Calculate next bp ( we can only use it if we do not block 1980 * or do other fancy things ). 1981 */ 1982 if (nbp == NULL) { 1983 switch(qindex) { 1984 case BQUEUE_EMPTY: 1985 nqindex = BQUEUE_EMPTYKVA; 1986 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]))) 1987 break; 1988 /* fall through */ 1989 case BQUEUE_EMPTYKVA: 1990 nqindex = BQUEUE_CLEAN; 1991 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]))) 1992 break; 1993 /* fall through */ 1994 case BQUEUE_CLEAN: 1995 /* 1996 * nbp is NULL. 1997 */ 1998 break; 1999 } 2000 } 2001 2002 /* 2003 * Sanity Checks 2004 */ 2005 KASSERT(bp->b_qindex == qindex, 2006 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2007 2008 /* 2009 * Note: we no longer distinguish between VMIO and non-VMIO 2010 * buffers. 2011 */ 2012 KASSERT((bp->b_flags & B_DELWRI) == 0, 2013 ("delwri buffer %p found in queue %d", bp, qindex)); 2014 2015 /* 2016 * Do not try to reuse a buffer with a non-zero b_refs. 2017 * This is an unsynchronized test. A synchronized test 2018 * is also performed after we lock the buffer. 2019 */ 2020 if (bp->b_refs) 2021 continue; 2022 2023 /* 2024 * If we are defragging then we need a buffer with 2025 * b_kvasize != 0. XXX this situation should no longer 2026 * occur, if defrag is non-zero the buffer's b_kvasize 2027 * should also be non-zero at this point. XXX 2028 */ 2029 if (defrag && bp->b_kvasize == 0) { 2030 kprintf("Warning: defrag empty buffer %p\n", bp); 2031 continue; 2032 } 2033 2034 /* 2035 * Start freeing the bp. This is somewhat involved. nbp 2036 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 2037 * on the clean list must be disassociated from their 2038 * current vnode. Buffers on the empty[kva] lists have 2039 * already been disassociated. 2040 * 2041 * b_refs is checked after locking along with queue changes. 2042 * We must check here to deal with zero->nonzero transitions 2043 * made by the owner of the buffer lock, which is used by 2044 * VFS's to hold the buffer while issuing an unlocked 2045 * uiomove()s. We cannot invalidate the buffer's pages 2046 * for this case. Once we successfully lock a buffer the 2047 * only 0->1 transitions of b_refs will occur via findblk(). 2048 * 2049 * We must also check for queue changes after successful 2050 * locking as the current lock holder may dispose of the 2051 * buffer and change its queue. 2052 */ 2053 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2054 spin_unlock(&bufqspin); 2055 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2056 goto restart; 2057 } 2058 if (bp->b_qindex != qindex || bp->b_refs) { 2059 spin_unlock(&bufqspin); 2060 BUF_UNLOCK(bp); 2061 goto restart; 2062 } 2063 bremfree_locked(bp); 2064 spin_unlock(&bufqspin); 2065 2066 /* 2067 * Dependancies must be handled before we disassociate the 2068 * vnode. 2069 * 2070 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2071 * be immediately disassociated. HAMMER then becomes 2072 * responsible for releasing the buffer. 2073 * 2074 * NOTE: bufqspin is UNLOCKED now. 2075 */ 2076 if (LIST_FIRST(&bp->b_dep) != NULL) { 2077 buf_deallocate(bp); 2078 if (bp->b_flags & B_LOCKED) { 2079 bqrelse(bp); 2080 goto restart; 2081 } 2082 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2083 } 2084 2085 if (qindex == BQUEUE_CLEAN) { 2086 if (bp->b_flags & B_VMIO) 2087 vfs_vmio_release(bp); 2088 if (bp->b_vp) 2089 brelvp(bp); 2090 } 2091 2092 /* 2093 * NOTE: nbp is now entirely invalid. We can only restart 2094 * the scan from this point on. 2095 * 2096 * Get the rest of the buffer freed up. b_kva* is still 2097 * valid after this operation. 2098 */ 2099 KASSERT(bp->b_vp == NULL, 2100 ("bp3 %p flags %08x vnode %p qindex %d " 2101 "unexpectededly still associated!", 2102 bp, bp->b_flags, bp->b_vp, qindex)); 2103 KKASSERT((bp->b_flags & B_HASHED) == 0); 2104 2105 /* 2106 * critical section protection is not required when 2107 * scrapping a buffer's contents because it is already 2108 * wired. 2109 */ 2110 if (bp->b_bufsize) 2111 allocbuf(bp, 0); 2112 2113 bp->b_flags = B_BNOCLIP; 2114 bp->b_cmd = BUF_CMD_DONE; 2115 bp->b_vp = NULL; 2116 bp->b_error = 0; 2117 bp->b_resid = 0; 2118 bp->b_bcount = 0; 2119 bp->b_xio.xio_npages = 0; 2120 bp->b_dirtyoff = bp->b_dirtyend = 0; 2121 bp->b_act_count = ACT_INIT; 2122 reinitbufbio(bp); 2123 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2124 buf_dep_init(bp); 2125 if (blkflags & GETBLK_BHEAVY) 2126 bp->b_flags |= B_HEAVY; 2127 2128 /* 2129 * If we are defragging then free the buffer. 2130 */ 2131 if (defrag) { 2132 bp->b_flags |= B_INVAL; 2133 bfreekva(bp); 2134 brelse(bp); 2135 defrag = 0; 2136 goto restart; 2137 } 2138 2139 /* 2140 * If we are overcomitted then recover the buffer and its 2141 * KVM space. This occurs in rare situations when multiple 2142 * processes are blocked in getnewbuf() or allocbuf(). 2143 */ 2144 if (bufspace >= hibufspace) 2145 flushingbufs = 1; 2146 if (flushingbufs && bp->b_kvasize != 0) { 2147 bp->b_flags |= B_INVAL; 2148 bfreekva(bp); 2149 brelse(bp); 2150 goto restart; 2151 } 2152 if (bufspace < lobufspace) 2153 flushingbufs = 0; 2154 2155 /* 2156 * b_refs can transition to a non-zero value while we hold 2157 * the buffer locked due to a findblk(). Our brelvp() above 2158 * interlocked any future possible transitions due to 2159 * findblk()s. 2160 * 2161 * If we find b_refs to be non-zero we can destroy the 2162 * buffer's contents but we cannot yet reuse the buffer. 2163 */ 2164 if (bp->b_refs) { 2165 bp->b_flags |= B_INVAL; 2166 bfreekva(bp); 2167 brelse(bp); 2168 goto restart; 2169 } 2170 break; 2171 /* NOT REACHED, bufqspin not held */ 2172 } 2173 2174 /* 2175 * If we exhausted our list, sleep as appropriate. We may have to 2176 * wakeup various daemons and write out some dirty buffers. 2177 * 2178 * Generally we are sleeping due to insufficient buffer space. 2179 * 2180 * NOTE: bufqspin is held if bp is NULL, else it is not held. 2181 */ 2182 if (bp == NULL) { 2183 int flags; 2184 char *waitmsg; 2185 2186 spin_unlock(&bufqspin); 2187 if (defrag) { 2188 flags = VFS_BIO_NEED_BUFSPACE; 2189 waitmsg = "nbufkv"; 2190 } else if (bufspace >= hibufspace) { 2191 waitmsg = "nbufbs"; 2192 flags = VFS_BIO_NEED_BUFSPACE; 2193 } else { 2194 waitmsg = "newbuf"; 2195 flags = VFS_BIO_NEED_ANY; 2196 } 2197 2198 bd_speedup(); /* heeeelp */ 2199 spin_lock(&bufcspin); 2200 needsbuffer |= flags; 2201 while (needsbuffer & flags) { 2202 if (ssleep(&needsbuffer, &bufcspin, 2203 slpflags, waitmsg, slptimeo)) { 2204 spin_unlock(&bufcspin); 2205 return (NULL); 2206 } 2207 } 2208 spin_unlock(&bufcspin); 2209 } else { 2210 /* 2211 * We finally have a valid bp. We aren't quite out of the 2212 * woods, we still have to reserve kva space. In order 2213 * to keep fragmentation sane we only allocate kva in 2214 * BKVASIZE chunks. 2215 * 2216 * (bufqspin is not held) 2217 */ 2218 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2219 2220 if (maxsize != bp->b_kvasize) { 2221 vm_offset_t addr = 0; 2222 int count; 2223 2224 bfreekva(bp); 2225 2226 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2227 vm_map_lock(&buffer_map); 2228 2229 if (vm_map_findspace(&buffer_map, 2230 vm_map_min(&buffer_map), maxsize, 2231 maxsize, 0, &addr)) { 2232 /* 2233 * Uh oh. Buffer map is too fragmented. We 2234 * must defragment the map. 2235 */ 2236 vm_map_unlock(&buffer_map); 2237 vm_map_entry_release(count); 2238 ++bufdefragcnt; 2239 defrag = 1; 2240 bp->b_flags |= B_INVAL; 2241 brelse(bp); 2242 goto restart; 2243 } 2244 if (addr) { 2245 vm_map_insert(&buffer_map, &count, 2246 NULL, 0, 2247 addr, addr + maxsize, 2248 VM_MAPTYPE_NORMAL, 2249 VM_PROT_ALL, VM_PROT_ALL, 2250 MAP_NOFAULT); 2251 2252 bp->b_kvabase = (caddr_t) addr; 2253 bp->b_kvasize = maxsize; 2254 bufspace += bp->b_kvasize; 2255 ++bufreusecnt; 2256 } 2257 vm_map_unlock(&buffer_map); 2258 vm_map_entry_release(count); 2259 } 2260 bp->b_data = bp->b_kvabase; 2261 } 2262 return(bp); 2263 } 2264 2265 #if 0 2266 /* 2267 * This routine is called in an emergency to recover VM pages from the 2268 * buffer cache by cashing in clean buffers. The idea is to recover 2269 * enough pages to be able to satisfy a stuck bio_page_alloc(). 2270 * 2271 * XXX Currently not implemented. This function can wind up deadlocking 2272 * against another thread holding one or more of the backing pages busy. 2273 */ 2274 static int 2275 recoverbufpages(void) 2276 { 2277 struct buf *bp; 2278 int bytes = 0; 2279 2280 ++recoverbufcalls; 2281 2282 spin_lock(&bufqspin); 2283 while (bytes < MAXBSIZE) { 2284 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 2285 if (bp == NULL) 2286 break; 2287 2288 /* 2289 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2290 * cycles through the queue twice before being selected. 2291 */ 2292 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) { 2293 bp->b_flags |= B_AGE; 2294 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 2295 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], 2296 bp, b_freelist); 2297 continue; 2298 } 2299 2300 /* 2301 * Sanity Checks 2302 */ 2303 KKASSERT(bp->b_qindex == BQUEUE_CLEAN); 2304 KKASSERT((bp->b_flags & B_DELWRI) == 0); 2305 2306 /* 2307 * Start freeing the bp. This is somewhat involved. 2308 * 2309 * Buffers on the clean list must be disassociated from 2310 * their current vnode 2311 */ 2312 2313 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2314 kprintf("recoverbufpages: warning, locked buf %p, " 2315 "race corrected\n", 2316 bp); 2317 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100); 2318 continue; 2319 } 2320 if (bp->b_qindex != BQUEUE_CLEAN) { 2321 kprintf("recoverbufpages: warning, BUF_LOCK blocked " 2322 "unexpectedly on buf %p index %d, race " 2323 "corrected\n", 2324 bp, bp->b_qindex); 2325 BUF_UNLOCK(bp); 2326 continue; 2327 } 2328 bremfree_locked(bp); 2329 spin_unlock(&bufqspin); 2330 2331 /* 2332 * Sanity check. Only BQUEUE_DIRTY[_HW] employs markers. 2333 */ 2334 KKASSERT((bp->b_flags & B_MARKER) == 0); 2335 2336 /* 2337 * Dependancies must be handled before we disassociate the 2338 * vnode. 2339 * 2340 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2341 * be immediately disassociated. HAMMER then becomes 2342 * responsible for releasing the buffer. 2343 */ 2344 if (LIST_FIRST(&bp->b_dep) != NULL) { 2345 buf_deallocate(bp); 2346 if (bp->b_flags & B_LOCKED) { 2347 bqrelse(bp); 2348 spin_lock(&bufqspin); 2349 continue; 2350 } 2351 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2352 } 2353 2354 bytes += bp->b_bufsize; 2355 2356 if (bp->b_flags & B_VMIO) { 2357 bp->b_flags |= B_DIRECT; /* try to free pages */ 2358 vfs_vmio_release(bp); 2359 } 2360 if (bp->b_vp) 2361 brelvp(bp); 2362 2363 KKASSERT(bp->b_vp == NULL); 2364 KKASSERT((bp->b_flags & B_HASHED) == 0); 2365 2366 /* 2367 * critical section protection is not required when 2368 * scrapping a buffer's contents because it is already 2369 * wired. 2370 */ 2371 if (bp->b_bufsize) 2372 allocbuf(bp, 0); 2373 2374 bp->b_flags = B_BNOCLIP; 2375 bp->b_cmd = BUF_CMD_DONE; 2376 bp->b_vp = NULL; 2377 bp->b_error = 0; 2378 bp->b_resid = 0; 2379 bp->b_bcount = 0; 2380 bp->b_xio.xio_npages = 0; 2381 bp->b_dirtyoff = bp->b_dirtyend = 0; 2382 reinitbufbio(bp); 2383 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2384 buf_dep_init(bp); 2385 bp->b_flags |= B_INVAL; 2386 /* bfreekva(bp); */ 2387 brelse(bp); 2388 spin_lock(&bufqspin); 2389 } 2390 spin_unlock(&bufqspin); 2391 return(bytes); 2392 } 2393 #endif 2394 2395 /* 2396 * buf_daemon: 2397 * 2398 * Buffer flushing daemon. Buffers are normally flushed by the 2399 * update daemon but if it cannot keep up this process starts to 2400 * take the load in an attempt to prevent getnewbuf() from blocking. 2401 * 2402 * Once a flush is initiated it does not stop until the number 2403 * of buffers falls below lodirtybuffers, but we will wake up anyone 2404 * waiting at the mid-point. 2405 */ 2406 static struct kproc_desc buf_kp = { 2407 "bufdaemon", 2408 buf_daemon, 2409 &bufdaemon_td 2410 }; 2411 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2412 kproc_start, &buf_kp) 2413 2414 static struct kproc_desc bufhw_kp = { 2415 "bufdaemon_hw", 2416 buf_daemon_hw, 2417 &bufdaemonhw_td 2418 }; 2419 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2420 kproc_start, &bufhw_kp) 2421 2422 /* 2423 * MPSAFE thread 2424 */ 2425 static void 2426 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long), 2427 int *bd_req) 2428 { 2429 long limit; 2430 struct buf *marker; 2431 2432 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO); 2433 marker->b_flags |= B_MARKER; 2434 marker->b_qindex = BQUEUE_NONE; 2435 2436 /* 2437 * This process needs to be suspended prior to shutdown sync. 2438 */ 2439 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2440 td, SHUTDOWN_PRI_LAST); 2441 curthread->td_flags |= TDF_SYSTHREAD; 2442 2443 /* 2444 * This process is allowed to take the buffer cache to the limit 2445 */ 2446 for (;;) { 2447 kproc_suspend_loop(); 2448 2449 /* 2450 * Do the flush as long as the number of dirty buffers 2451 * (including those running) exceeds lodirtybufspace. 2452 * 2453 * When flushing limit running I/O to hirunningspace 2454 * Do the flush. Limit the amount of in-transit I/O we 2455 * allow to build up, otherwise we would completely saturate 2456 * the I/O system. Wakeup any waiting processes before we 2457 * normally would so they can run in parallel with our drain. 2458 * 2459 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2460 * but because we split the operation into two threads we 2461 * have to cut it in half for each thread. 2462 */ 2463 waitrunningbufspace(); 2464 limit = lodirtybufspace / 2; 2465 while (buf_limit_fn(limit)) { 2466 if (flushbufqueues(marker, queue) == 0) 2467 break; 2468 if (runningbufspace < hirunningspace) 2469 continue; 2470 waitrunningbufspace(); 2471 } 2472 2473 /* 2474 * We reached our low water mark, reset the 2475 * request and sleep until we are needed again. 2476 * The sleep is just so the suspend code works. 2477 */ 2478 spin_lock(&bufcspin); 2479 if (*bd_req == 0) 2480 ssleep(bd_req, &bufcspin, 0, "psleep", hz); 2481 *bd_req = 0; 2482 spin_unlock(&bufcspin); 2483 } 2484 /* NOT REACHED */ 2485 /*kfree(marker, M_BIOBUF);*/ 2486 } 2487 2488 static int 2489 buf_daemon_limit(long limit) 2490 { 2491 return (runningbufspace + dirtykvaspace > limit || 2492 dirtybufcount - dirtybufcounthw >= nbuf / 2); 2493 } 2494 2495 static int 2496 buf_daemon_hw_limit(long limit) 2497 { 2498 return (runningbufspace + dirtykvaspace > limit || 2499 dirtybufcounthw >= nbuf / 2); 2500 } 2501 2502 static void 2503 buf_daemon(void) 2504 { 2505 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit, 2506 &bd_request); 2507 } 2508 2509 static void 2510 buf_daemon_hw(void) 2511 { 2512 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit, 2513 &bd_request_hw); 2514 } 2515 2516 /* 2517 * flushbufqueues: 2518 * 2519 * Try to flush a buffer in the dirty queue. We must be careful to 2520 * free up B_INVAL buffers instead of write them, which NFS is 2521 * particularly sensitive to. 2522 * 2523 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate 2524 * that we really want to try to get the buffer out and reuse it 2525 * due to the write load on the machine. 2526 * 2527 * We must lock the buffer in order to check its validity before we 2528 * can mess with its contents. bufqspin isn't enough. 2529 */ 2530 static int 2531 flushbufqueues(struct buf *marker, bufq_type_t q) 2532 { 2533 struct buf *bp; 2534 int r = 0; 2535 2536 KKASSERT(marker->b_qindex == BQUEUE_NONE); 2537 KKASSERT(marker->b_flags & B_MARKER); 2538 2539 /* 2540 * Spinlock needed to perform operations on the queue and may be 2541 * held through a non-blocking BUF_LOCK(), but cannot be held when 2542 * BUF_UNLOCK()ing or through any other major operation. 2543 */ 2544 spin_lock(&bufqspin); 2545 marker->b_qindex = q; 2546 TAILQ_INSERT_HEAD(&bufqueues[q], marker, b_freelist); 2547 bp = marker; 2548 2549 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) { 2550 /* 2551 * NOTE: spinlock is always held at the top of the loop 2552 */ 2553 if (bp->b_flags & B_MARKER) 2554 continue; 2555 if ((bp->b_flags & B_DELWRI) == 0) { 2556 kprintf("Unexpected clean buffer %p\n", bp); 2557 continue; 2558 } 2559 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 2560 continue; 2561 KKASSERT(bp->b_qindex == q); 2562 2563 /* 2564 * Once the buffer is locked we will have no choice but to 2565 * unlock the spinlock around a later BUF_UNLOCK and re-set 2566 * bp = marker when looping. Move the marker now to make 2567 * things easier. 2568 */ 2569 TAILQ_REMOVE(&bufqueues[q], marker, b_freelist); 2570 TAILQ_INSERT_AFTER(&bufqueues[q], bp, marker, b_freelist); 2571 2572 /* 2573 * Must recheck B_DELWRI after successfully locking 2574 * the buffer. 2575 */ 2576 if ((bp->b_flags & B_DELWRI) == 0) { 2577 spin_unlock(&bufqspin); 2578 BUF_UNLOCK(bp); 2579 spin_lock(&bufqspin); 2580 bp = marker; 2581 continue; 2582 } 2583 2584 /* 2585 * Remove the buffer from its queue. We still own the 2586 * spinlock here. 2587 */ 2588 _bremfree(bp); 2589 2590 /* 2591 * Disposing of an invalid buffer counts as a flush op 2592 */ 2593 if (bp->b_flags & B_INVAL) { 2594 spin_unlock(&bufqspin); 2595 brelse(bp); 2596 spin_lock(&bufqspin); 2597 ++r; 2598 break; 2599 } 2600 2601 /* 2602 * Release the spinlock for the more complex ops we 2603 * are now going to do. 2604 */ 2605 spin_unlock(&bufqspin); 2606 lwkt_yield(); 2607 2608 /* 2609 * This is a bit messy 2610 */ 2611 if (LIST_FIRST(&bp->b_dep) != NULL && 2612 (bp->b_flags & B_DEFERRED) == 0 && 2613 buf_countdeps(bp, 0)) { 2614 spin_lock(&bufqspin); 2615 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist); 2616 bp->b_qindex = q; 2617 bp->b_flags |= B_DEFERRED; 2618 spin_unlock(&bufqspin); 2619 BUF_UNLOCK(bp); 2620 spin_lock(&bufqspin); 2621 bp = marker; 2622 continue; 2623 } 2624 2625 /* 2626 * spinlock not held here. 2627 * 2628 * If the buffer has a dependancy, buf_checkwrite() must 2629 * also return 0 for us to be able to initate the write. 2630 * 2631 * If the buffer is flagged B_ERROR it may be requeued 2632 * over and over again, we try to avoid a live lock. 2633 * 2634 * NOTE: buf_checkwrite is MPSAFE. 2635 */ 2636 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2637 brelse(bp); 2638 } else if (bp->b_flags & B_ERROR) { 2639 tsleep(bp, 0, "bioer", 1); 2640 bp->b_flags &= ~B_AGE; 2641 cluster_awrite(bp); 2642 } else { 2643 bp->b_flags |= B_AGE; 2644 cluster_awrite(bp); 2645 } 2646 spin_lock(&bufqspin); 2647 ++r; 2648 break; 2649 } 2650 TAILQ_REMOVE(&bufqueues[q], marker, b_freelist); 2651 marker->b_qindex = BQUEUE_NONE; 2652 spin_unlock(&bufqspin); 2653 2654 return (r); 2655 } 2656 2657 /* 2658 * inmem: 2659 * 2660 * Returns true if no I/O is needed to access the associated VM object. 2661 * This is like findblk except it also hunts around in the VM system for 2662 * the data. 2663 * 2664 * Note that we ignore vm_page_free() races from interrupts against our 2665 * lookup, since if the caller is not protected our return value will not 2666 * be any more valid then otherwise once we exit the critical section. 2667 */ 2668 int 2669 inmem(struct vnode *vp, off_t loffset) 2670 { 2671 vm_object_t obj; 2672 vm_offset_t toff, tinc, size; 2673 vm_page_t m; 2674 int res = 1; 2675 2676 if (findblk(vp, loffset, FINDBLK_TEST)) 2677 return 1; 2678 if (vp->v_mount == NULL) 2679 return 0; 2680 if ((obj = vp->v_object) == NULL) 2681 return 0; 2682 2683 size = PAGE_SIZE; 2684 if (size > vp->v_mount->mnt_stat.f_iosize) 2685 size = vp->v_mount->mnt_stat.f_iosize; 2686 2687 vm_object_hold(obj); 2688 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2689 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2690 if (m == NULL) { 2691 res = 0; 2692 break; 2693 } 2694 tinc = size; 2695 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2696 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2697 if (vm_page_is_valid(m, 2698 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2699 res = 0; 2700 break; 2701 } 2702 } 2703 vm_object_drop(obj); 2704 return (res); 2705 } 2706 2707 /* 2708 * findblk: 2709 * 2710 * Locate and return the specified buffer. Unless flagged otherwise, 2711 * a locked buffer will be returned if it exists or NULL if it does not. 2712 * 2713 * findblk()'d buffers are still on the bufqueues and if you intend 2714 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2715 * and possibly do other stuff to it. 2716 * 2717 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2718 * for locking the buffer and ensuring that it remains 2719 * the desired buffer after locking. 2720 * 2721 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2722 * to acquire the lock we return NULL, even if the 2723 * buffer exists. 2724 * 2725 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2726 * reuse by getnewbuf() but does not prevent 2727 * disassociation (B_INVAL). Used to avoid deadlocks 2728 * against random (vp,loffset)s due to reassignment. 2729 * 2730 * (0) - Lock the buffer blocking. 2731 * 2732 * MPSAFE 2733 */ 2734 struct buf * 2735 findblk(struct vnode *vp, off_t loffset, int flags) 2736 { 2737 struct buf *bp; 2738 int lkflags; 2739 2740 lkflags = LK_EXCLUSIVE; 2741 if (flags & FINDBLK_NBLOCK) 2742 lkflags |= LK_NOWAIT; 2743 2744 for (;;) { 2745 /* 2746 * Lookup. Ref the buf while holding v_token to prevent 2747 * reuse (but does not prevent diassociation). 2748 */ 2749 lwkt_gettoken_shared(&vp->v_token); 2750 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2751 if (bp == NULL) { 2752 lwkt_reltoken(&vp->v_token); 2753 return(NULL); 2754 } 2755 bqhold(bp); 2756 lwkt_reltoken(&vp->v_token); 2757 2758 /* 2759 * If testing only break and return bp, do not lock. 2760 */ 2761 if (flags & FINDBLK_TEST) 2762 break; 2763 2764 /* 2765 * Lock the buffer, return an error if the lock fails. 2766 * (only FINDBLK_NBLOCK can cause the lock to fail). 2767 */ 2768 if (BUF_LOCK(bp, lkflags)) { 2769 atomic_subtract_int(&bp->b_refs, 1); 2770 /* bp = NULL; not needed */ 2771 return(NULL); 2772 } 2773 2774 /* 2775 * Revalidate the locked buf before allowing it to be 2776 * returned. 2777 */ 2778 if (bp->b_vp == vp && bp->b_loffset == loffset) 2779 break; 2780 atomic_subtract_int(&bp->b_refs, 1); 2781 BUF_UNLOCK(bp); 2782 } 2783 2784 /* 2785 * Success 2786 */ 2787 if ((flags & FINDBLK_REF) == 0) 2788 atomic_subtract_int(&bp->b_refs, 1); 2789 return(bp); 2790 } 2791 2792 /* 2793 * getcacheblk: 2794 * 2795 * Similar to getblk() except only returns the buffer if it is 2796 * B_CACHE and requires no other manipulation. Otherwise NULL 2797 * is returned. 2798 * 2799 * If B_RAM is set the buffer might be just fine, but we return 2800 * NULL anyway because we want the code to fall through to the 2801 * cluster read. Otherwise read-ahead breaks. 2802 * 2803 * If blksize is 0 the buffer cache buffer must already be fully 2804 * cached. 2805 * 2806 * If blksize is non-zero getblk() will be used, allowing a buffer 2807 * to be reinstantiated from its VM backing store. The buffer must 2808 * still be fully cached after reinstantiation to be returned. 2809 */ 2810 struct buf * 2811 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags) 2812 { 2813 struct buf *bp; 2814 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0; 2815 2816 if (blksize) { 2817 bp = getblk(vp, loffset, blksize, blkflags, 0); 2818 if (bp) { 2819 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2820 B_CACHE) { 2821 bp->b_flags &= ~B_AGE; 2822 } else { 2823 brelse(bp); 2824 bp = NULL; 2825 } 2826 } 2827 } else { 2828 bp = findblk(vp, loffset, fndflags); 2829 if (bp) { 2830 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2831 B_CACHE) { 2832 bp->b_flags &= ~B_AGE; 2833 bremfree(bp); 2834 } else { 2835 BUF_UNLOCK(bp); 2836 bp = NULL; 2837 } 2838 } 2839 } 2840 return (bp); 2841 } 2842 2843 /* 2844 * getblk: 2845 * 2846 * Get a block given a specified block and offset into a file/device. 2847 * B_INVAL may or may not be set on return. The caller should clear 2848 * B_INVAL prior to initiating a READ. 2849 * 2850 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2851 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2852 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2853 * without doing any of those things the system will likely believe 2854 * the buffer to be valid (especially if it is not B_VMIO), and the 2855 * next getblk() will return the buffer with B_CACHE set. 2856 * 2857 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2858 * an existing buffer. 2859 * 2860 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2861 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2862 * and then cleared based on the backing VM. If the previous buffer is 2863 * non-0-sized but invalid, B_CACHE will be cleared. 2864 * 2865 * If getblk() must create a new buffer, the new buffer is returned with 2866 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2867 * case it is returned with B_INVAL clear and B_CACHE set based on the 2868 * backing VM. 2869 * 2870 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2871 * B_CACHE bit is clear. 2872 * 2873 * What this means, basically, is that the caller should use B_CACHE to 2874 * determine whether the buffer is fully valid or not and should clear 2875 * B_INVAL prior to issuing a read. If the caller intends to validate 2876 * the buffer by loading its data area with something, the caller needs 2877 * to clear B_INVAL. If the caller does this without issuing an I/O, 2878 * the caller should set B_CACHE ( as an optimization ), else the caller 2879 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2880 * a write attempt or if it was a successfull read. If the caller 2881 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2882 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2883 * 2884 * getblk flags: 2885 * 2886 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2887 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2888 * 2889 * MPALMOSTSAFE 2890 */ 2891 struct buf * 2892 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2893 { 2894 struct buf *bp; 2895 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2896 int error; 2897 int lkflags; 2898 2899 if (size > MAXBSIZE) 2900 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2901 if (vp->v_object == NULL) 2902 panic("getblk: vnode %p has no object!", vp); 2903 2904 loop: 2905 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2906 /* 2907 * The buffer was found in the cache, but we need to lock it. 2908 * We must acquire a ref on the bp to prevent reuse, but 2909 * this will not prevent disassociation (brelvp()) so we 2910 * must recheck (vp,loffset) after acquiring the lock. 2911 * 2912 * Without the ref the buffer could potentially be reused 2913 * before we acquire the lock and create a deadlock 2914 * situation between the thread trying to reuse the buffer 2915 * and us due to the fact that we would wind up blocking 2916 * on a random (vp,loffset). 2917 */ 2918 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2919 if (blkflags & GETBLK_NOWAIT) { 2920 bqdrop(bp); 2921 return(NULL); 2922 } 2923 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2924 if (blkflags & GETBLK_PCATCH) 2925 lkflags |= LK_PCATCH; 2926 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2927 if (error) { 2928 bqdrop(bp); 2929 if (error == ENOLCK) 2930 goto loop; 2931 return (NULL); 2932 } 2933 /* buffer may have changed on us */ 2934 } 2935 bqdrop(bp); 2936 2937 /* 2938 * Once the buffer has been locked, make sure we didn't race 2939 * a buffer recyclement. Buffers that are no longer hashed 2940 * will have b_vp == NULL, so this takes care of that check 2941 * as well. 2942 */ 2943 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2944 kprintf("Warning buffer %p (vp %p loffset %lld) " 2945 "was recycled\n", 2946 bp, vp, (long long)loffset); 2947 BUF_UNLOCK(bp); 2948 goto loop; 2949 } 2950 2951 /* 2952 * If SZMATCH any pre-existing buffer must be of the requested 2953 * size or NULL is returned. The caller absolutely does not 2954 * want getblk() to bwrite() the buffer on a size mismatch. 2955 */ 2956 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2957 BUF_UNLOCK(bp); 2958 return(NULL); 2959 } 2960 2961 /* 2962 * All vnode-based buffers must be backed by a VM object. 2963 */ 2964 KKASSERT(bp->b_flags & B_VMIO); 2965 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2966 bp->b_flags &= ~B_AGE; 2967 2968 /* 2969 * Make sure that B_INVAL buffers do not have a cached 2970 * block number translation. 2971 */ 2972 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2973 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2974 " did not have cleared bio_offset cache\n", 2975 bp, vp, (long long)loffset); 2976 clearbiocache(&bp->b_bio2); 2977 } 2978 2979 /* 2980 * The buffer is locked. B_CACHE is cleared if the buffer is 2981 * invalid. 2982 */ 2983 if (bp->b_flags & B_INVAL) 2984 bp->b_flags &= ~B_CACHE; 2985 bremfree(bp); 2986 2987 /* 2988 * Any size inconsistancy with a dirty buffer or a buffer 2989 * with a softupdates dependancy must be resolved. Resizing 2990 * the buffer in such circumstances can lead to problems. 2991 * 2992 * Dirty or dependant buffers are written synchronously. 2993 * Other types of buffers are simply released and 2994 * reconstituted as they may be backed by valid, dirty VM 2995 * pages (but not marked B_DELWRI). 2996 * 2997 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2998 * and may be left over from a prior truncation (and thus 2999 * no longer represent the actual EOF point), so we 3000 * definitely do not want to B_NOCACHE the backing store. 3001 */ 3002 if (size != bp->b_bcount) { 3003 if (bp->b_flags & B_DELWRI) { 3004 bp->b_flags |= B_RELBUF; 3005 bwrite(bp); 3006 } else if (LIST_FIRST(&bp->b_dep)) { 3007 bp->b_flags |= B_RELBUF; 3008 bwrite(bp); 3009 } else { 3010 bp->b_flags |= B_RELBUF; 3011 brelse(bp); 3012 } 3013 goto loop; 3014 } 3015 KKASSERT(size <= bp->b_kvasize); 3016 KASSERT(bp->b_loffset != NOOFFSET, 3017 ("getblk: no buffer offset")); 3018 3019 /* 3020 * A buffer with B_DELWRI set and B_CACHE clear must 3021 * be committed before we can return the buffer in 3022 * order to prevent the caller from issuing a read 3023 * ( due to B_CACHE not being set ) and overwriting 3024 * it. 3025 * 3026 * Most callers, including NFS and FFS, need this to 3027 * operate properly either because they assume they 3028 * can issue a read if B_CACHE is not set, or because 3029 * ( for example ) an uncached B_DELWRI might loop due 3030 * to softupdates re-dirtying the buffer. In the latter 3031 * case, B_CACHE is set after the first write completes, 3032 * preventing further loops. 3033 * 3034 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3035 * above while extending the buffer, we cannot allow the 3036 * buffer to remain with B_CACHE set after the write 3037 * completes or it will represent a corrupt state. To 3038 * deal with this we set B_NOCACHE to scrap the buffer 3039 * after the write. 3040 * 3041 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3042 * I'm not even sure this state is still possible 3043 * now that getblk() writes out any dirty buffers 3044 * on size changes. 3045 * 3046 * We might be able to do something fancy, like setting 3047 * B_CACHE in bwrite() except if B_DELWRI is already set, 3048 * so the below call doesn't set B_CACHE, but that gets real 3049 * confusing. This is much easier. 3050 */ 3051 3052 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3053 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3054 "and CACHE clear, b_flags %08x\n", 3055 bp, (uintmax_t)bp->b_loffset, bp->b_flags); 3056 bp->b_flags |= B_NOCACHE; 3057 bwrite(bp); 3058 goto loop; 3059 } 3060 } else { 3061 /* 3062 * Buffer is not in-core, create new buffer. The buffer 3063 * returned by getnewbuf() is locked. Note that the returned 3064 * buffer is also considered valid (not marked B_INVAL). 3065 * 3066 * Calculating the offset for the I/O requires figuring out 3067 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3068 * the mount's f_iosize otherwise. If the vnode does not 3069 * have an associated mount we assume that the passed size is 3070 * the block size. 3071 * 3072 * Note that vn_isdisk() cannot be used here since it may 3073 * return a failure for numerous reasons. Note that the 3074 * buffer size may be larger then the block size (the caller 3075 * will use block numbers with the proper multiple). Beware 3076 * of using any v_* fields which are part of unions. In 3077 * particular, in DragonFly the mount point overloading 3078 * mechanism uses the namecache only and the underlying 3079 * directory vnode is not a special case. 3080 */ 3081 int bsize, maxsize; 3082 3083 if (vp->v_type == VBLK || vp->v_type == VCHR) 3084 bsize = DEV_BSIZE; 3085 else if (vp->v_mount) 3086 bsize = vp->v_mount->mnt_stat.f_iosize; 3087 else 3088 bsize = size; 3089 3090 maxsize = size + (loffset & PAGE_MASK); 3091 maxsize = imax(maxsize, bsize); 3092 3093 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3094 if (bp == NULL) { 3095 if (slpflags || slptimeo) 3096 return NULL; 3097 goto loop; 3098 } 3099 3100 /* 3101 * Atomically insert the buffer into the hash, so that it can 3102 * be found by findblk(). 3103 * 3104 * If bgetvp() returns non-zero a collision occured, and the 3105 * bp will not be associated with the vnode. 3106 * 3107 * Make sure the translation layer has been cleared. 3108 */ 3109 bp->b_loffset = loffset; 3110 bp->b_bio2.bio_offset = NOOFFSET; 3111 /* bp->b_bio2.bio_next = NULL; */ 3112 3113 if (bgetvp(vp, bp, size)) { 3114 bp->b_flags |= B_INVAL; 3115 brelse(bp); 3116 goto loop; 3117 } 3118 3119 /* 3120 * All vnode-based buffers must be backed by a VM object. 3121 */ 3122 KKASSERT(vp->v_object != NULL); 3123 bp->b_flags |= B_VMIO; 3124 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3125 3126 allocbuf(bp, size); 3127 } 3128 KKASSERT(dsched_is_clear_buf_priv(bp)); 3129 return (bp); 3130 } 3131 3132 /* 3133 * regetblk(bp) 3134 * 3135 * Reacquire a buffer that was previously released to the locked queue, 3136 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3137 * set B_LOCKED (which handles the acquisition race). 3138 * 3139 * To this end, either B_LOCKED must be set or the dependancy list must be 3140 * non-empty. 3141 * 3142 * MPSAFE 3143 */ 3144 void 3145 regetblk(struct buf *bp) 3146 { 3147 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3148 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3149 bremfree(bp); 3150 } 3151 3152 /* 3153 * geteblk: 3154 * 3155 * Get an empty, disassociated buffer of given size. The buffer is 3156 * initially set to B_INVAL. 3157 * 3158 * critical section protection is not required for the allocbuf() 3159 * call because races are impossible here. 3160 * 3161 * MPALMOSTSAFE 3162 */ 3163 struct buf * 3164 geteblk(int size) 3165 { 3166 struct buf *bp; 3167 int maxsize; 3168 3169 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3170 3171 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL) 3172 ; 3173 allocbuf(bp, size); 3174 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3175 KKASSERT(dsched_is_clear_buf_priv(bp)); 3176 return (bp); 3177 } 3178 3179 3180 /* 3181 * allocbuf: 3182 * 3183 * This code constitutes the buffer memory from either anonymous system 3184 * memory (in the case of non-VMIO operations) or from an associated 3185 * VM object (in the case of VMIO operations). This code is able to 3186 * resize a buffer up or down. 3187 * 3188 * Note that this code is tricky, and has many complications to resolve 3189 * deadlock or inconsistant data situations. Tread lightly!!! 3190 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3191 * the caller. Calling this code willy nilly can result in the loss of 3192 * data. 3193 * 3194 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3195 * B_CACHE for the non-VMIO case. 3196 * 3197 * This routine does not need to be called from a critical section but you 3198 * must own the buffer. 3199 * 3200 * MPSAFE 3201 */ 3202 int 3203 allocbuf(struct buf *bp, int size) 3204 { 3205 int newbsize, mbsize; 3206 int i; 3207 3208 if (BUF_REFCNT(bp) == 0) 3209 panic("allocbuf: buffer not busy"); 3210 3211 if (bp->b_kvasize < size) 3212 panic("allocbuf: buffer too small"); 3213 3214 if ((bp->b_flags & B_VMIO) == 0) { 3215 caddr_t origbuf; 3216 int origbufsize; 3217 /* 3218 * Just get anonymous memory from the kernel. Don't 3219 * mess with B_CACHE. 3220 */ 3221 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3222 if (bp->b_flags & B_MALLOC) 3223 newbsize = mbsize; 3224 else 3225 newbsize = round_page(size); 3226 3227 if (newbsize < bp->b_bufsize) { 3228 /* 3229 * Malloced buffers are not shrunk 3230 */ 3231 if (bp->b_flags & B_MALLOC) { 3232 if (newbsize) { 3233 bp->b_bcount = size; 3234 } else { 3235 kfree(bp->b_data, M_BIOBUF); 3236 if (bp->b_bufsize) { 3237 atomic_subtract_long(&bufmallocspace, bp->b_bufsize); 3238 bufspacewakeup(); 3239 bp->b_bufsize = 0; 3240 } 3241 bp->b_data = bp->b_kvabase; 3242 bp->b_bcount = 0; 3243 bp->b_flags &= ~B_MALLOC; 3244 } 3245 return 1; 3246 } 3247 vm_hold_free_pages( 3248 bp, 3249 (vm_offset_t) bp->b_data + newbsize, 3250 (vm_offset_t) bp->b_data + bp->b_bufsize); 3251 } else if (newbsize > bp->b_bufsize) { 3252 /* 3253 * We only use malloced memory on the first allocation. 3254 * and revert to page-allocated memory when the buffer 3255 * grows. 3256 */ 3257 if ((bufmallocspace < maxbufmallocspace) && 3258 (bp->b_bufsize == 0) && 3259 (mbsize <= PAGE_SIZE/2)) { 3260 3261 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3262 bp->b_bufsize = mbsize; 3263 bp->b_bcount = size; 3264 bp->b_flags |= B_MALLOC; 3265 atomic_add_long(&bufmallocspace, mbsize); 3266 return 1; 3267 } 3268 origbuf = NULL; 3269 origbufsize = 0; 3270 /* 3271 * If the buffer is growing on its other-than-first 3272 * allocation, then we revert to the page-allocation 3273 * scheme. 3274 */ 3275 if (bp->b_flags & B_MALLOC) { 3276 origbuf = bp->b_data; 3277 origbufsize = bp->b_bufsize; 3278 bp->b_data = bp->b_kvabase; 3279 if (bp->b_bufsize) { 3280 atomic_subtract_long(&bufmallocspace, 3281 bp->b_bufsize); 3282 bufspacewakeup(); 3283 bp->b_bufsize = 0; 3284 } 3285 bp->b_flags &= ~B_MALLOC; 3286 newbsize = round_page(newbsize); 3287 } 3288 vm_hold_load_pages( 3289 bp, 3290 (vm_offset_t) bp->b_data + bp->b_bufsize, 3291 (vm_offset_t) bp->b_data + newbsize); 3292 if (origbuf) { 3293 bcopy(origbuf, bp->b_data, origbufsize); 3294 kfree(origbuf, M_BIOBUF); 3295 } 3296 } 3297 } else { 3298 vm_page_t m; 3299 int desiredpages; 3300 3301 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3302 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3303 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3304 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3305 3306 if (bp->b_flags & B_MALLOC) 3307 panic("allocbuf: VMIO buffer can't be malloced"); 3308 /* 3309 * Set B_CACHE initially if buffer is 0 length or will become 3310 * 0-length. 3311 */ 3312 if (size == 0 || bp->b_bufsize == 0) 3313 bp->b_flags |= B_CACHE; 3314 3315 if (newbsize < bp->b_bufsize) { 3316 /* 3317 * DEV_BSIZE aligned new buffer size is less then the 3318 * DEV_BSIZE aligned existing buffer size. Figure out 3319 * if we have to remove any pages. 3320 */ 3321 if (desiredpages < bp->b_xio.xio_npages) { 3322 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3323 /* 3324 * the page is not freed here -- it 3325 * is the responsibility of 3326 * vnode_pager_setsize 3327 */ 3328 m = bp->b_xio.xio_pages[i]; 3329 KASSERT(m != bogus_page, 3330 ("allocbuf: bogus page found")); 3331 vm_page_busy_wait(m, TRUE, "biodep"); 3332 bp->b_xio.xio_pages[i] = NULL; 3333 vm_page_unwire(m, 0); 3334 vm_page_wakeup(m); 3335 } 3336 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3337 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3338 bp->b_xio.xio_npages = desiredpages; 3339 } 3340 } else if (size > bp->b_bcount) { 3341 /* 3342 * We are growing the buffer, possibly in a 3343 * byte-granular fashion. 3344 */ 3345 struct vnode *vp; 3346 vm_object_t obj; 3347 vm_offset_t toff; 3348 vm_offset_t tinc; 3349 3350 /* 3351 * Step 1, bring in the VM pages from the object, 3352 * allocating them if necessary. We must clear 3353 * B_CACHE if these pages are not valid for the 3354 * range covered by the buffer. 3355 * 3356 * critical section protection is required to protect 3357 * against interrupts unbusying and freeing pages 3358 * between our vm_page_lookup() and our 3359 * busycheck/wiring call. 3360 */ 3361 vp = bp->b_vp; 3362 obj = vp->v_object; 3363 3364 vm_object_hold(obj); 3365 while (bp->b_xio.xio_npages < desiredpages) { 3366 vm_page_t m; 3367 vm_pindex_t pi; 3368 int error; 3369 3370 pi = OFF_TO_IDX(bp->b_loffset) + 3371 bp->b_xio.xio_npages; 3372 3373 /* 3374 * Blocking on m->busy might lead to a 3375 * deadlock: 3376 * 3377 * vm_fault->getpages->cluster_read->allocbuf 3378 */ 3379 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3380 &error); 3381 if (error) { 3382 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3383 continue; 3384 } 3385 if (m == NULL) { 3386 /* 3387 * note: must allocate system pages 3388 * since blocking here could intefere 3389 * with paging I/O, no matter which 3390 * process we are. 3391 */ 3392 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages); 3393 if (m) { 3394 vm_page_wire(m); 3395 vm_page_flag_clear(m, PG_ZERO); 3396 vm_page_wakeup(m); 3397 bp->b_flags &= ~B_CACHE; 3398 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3399 ++bp->b_xio.xio_npages; 3400 } 3401 continue; 3402 } 3403 3404 /* 3405 * We found a page and were able to busy it. 3406 */ 3407 vm_page_flag_clear(m, PG_ZERO); 3408 vm_page_wire(m); 3409 vm_page_wakeup(m); 3410 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3411 ++bp->b_xio.xio_npages; 3412 if (bp->b_act_count < m->act_count) 3413 bp->b_act_count = m->act_count; 3414 } 3415 vm_object_drop(obj); 3416 3417 /* 3418 * Step 2. We've loaded the pages into the buffer, 3419 * we have to figure out if we can still have B_CACHE 3420 * set. Note that B_CACHE is set according to the 3421 * byte-granular range ( bcount and size ), not the 3422 * aligned range ( newbsize ). 3423 * 3424 * The VM test is against m->valid, which is DEV_BSIZE 3425 * aligned. Needless to say, the validity of the data 3426 * needs to also be DEV_BSIZE aligned. Note that this 3427 * fails with NFS if the server or some other client 3428 * extends the file's EOF. If our buffer is resized, 3429 * B_CACHE may remain set! XXX 3430 */ 3431 3432 toff = bp->b_bcount; 3433 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3434 3435 while ((bp->b_flags & B_CACHE) && toff < size) { 3436 vm_pindex_t pi; 3437 3438 if (tinc > (size - toff)) 3439 tinc = size - toff; 3440 3441 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3442 PAGE_SHIFT; 3443 3444 vfs_buf_test_cache( 3445 bp, 3446 bp->b_loffset, 3447 toff, 3448 tinc, 3449 bp->b_xio.xio_pages[pi] 3450 ); 3451 toff += tinc; 3452 tinc = PAGE_SIZE; 3453 } 3454 3455 /* 3456 * Step 3, fixup the KVM pmap. Remember that 3457 * bp->b_data is relative to bp->b_loffset, but 3458 * bp->b_loffset may be offset into the first page. 3459 */ 3460 3461 bp->b_data = (caddr_t) 3462 trunc_page((vm_offset_t)bp->b_data); 3463 pmap_qenter( 3464 (vm_offset_t)bp->b_data, 3465 bp->b_xio.xio_pages, 3466 bp->b_xio.xio_npages 3467 ); 3468 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3469 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3470 } 3471 } 3472 3473 /* adjust space use on already-dirty buffer */ 3474 if (bp->b_flags & B_DELWRI) { 3475 spin_lock(&bufcspin); 3476 /* dirtykvaspace unchanged */ 3477 dirtybufspace += newbsize - bp->b_bufsize; 3478 if (bp->b_flags & B_HEAVY) 3479 dirtybufspacehw += newbsize - bp->b_bufsize; 3480 spin_unlock(&bufcspin); 3481 } 3482 if (newbsize < bp->b_bufsize) 3483 bufspacewakeup(); 3484 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3485 bp->b_bcount = size; /* requested buffer size */ 3486 return 1; 3487 } 3488 3489 /* 3490 * biowait: 3491 * 3492 * Wait for buffer I/O completion, returning error status. B_EINTR 3493 * is converted into an EINTR error but not cleared (since a chain 3494 * of biowait() calls may occur). 3495 * 3496 * On return bpdone() will have been called but the buffer will remain 3497 * locked and will not have been brelse()'d. 3498 * 3499 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3500 * likely still in progress on return. 3501 * 3502 * NOTE! This operation is on a BIO, not a BUF. 3503 * 3504 * NOTE! BIO_DONE is cleared by vn_strategy() 3505 * 3506 * MPSAFE 3507 */ 3508 static __inline int 3509 _biowait(struct bio *bio, const char *wmesg, int to) 3510 { 3511 struct buf *bp = bio->bio_buf; 3512 u_int32_t flags; 3513 u_int32_t nflags; 3514 int error; 3515 3516 KKASSERT(bio == &bp->b_bio1); 3517 for (;;) { 3518 flags = bio->bio_flags; 3519 if (flags & BIO_DONE) 3520 break; 3521 nflags = flags | BIO_WANT; 3522 tsleep_interlock(bio, 0); 3523 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3524 if (wmesg) 3525 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3526 else if (bp->b_cmd == BUF_CMD_READ) 3527 error = tsleep(bio, PINTERLOCKED, "biord", to); 3528 else 3529 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3530 if (error) { 3531 kprintf("tsleep error biowait %d\n", error); 3532 return (error); 3533 } 3534 } 3535 } 3536 3537 /* 3538 * Finish up. 3539 */ 3540 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3541 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3542 if (bp->b_flags & B_EINTR) 3543 return (EINTR); 3544 if (bp->b_flags & B_ERROR) 3545 return (bp->b_error ? bp->b_error : EIO); 3546 return (0); 3547 } 3548 3549 int 3550 biowait(struct bio *bio, const char *wmesg) 3551 { 3552 return(_biowait(bio, wmesg, 0)); 3553 } 3554 3555 int 3556 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3557 { 3558 return(_biowait(bio, wmesg, to)); 3559 } 3560 3561 /* 3562 * This associates a tracking count with an I/O. vn_strategy() and 3563 * dev_dstrategy() do this automatically but there are a few cases 3564 * where a vnode or device layer is bypassed when a block translation 3565 * is cached. In such cases bio_start_transaction() may be called on 3566 * the bypassed layers so the system gets an I/O in progress indication 3567 * for those higher layers. 3568 */ 3569 void 3570 bio_start_transaction(struct bio *bio, struct bio_track *track) 3571 { 3572 bio->bio_track = track; 3573 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3574 dsched_new_buf(bio->bio_buf); 3575 bio_track_ref(track); 3576 } 3577 3578 /* 3579 * Initiate I/O on a vnode. 3580 * 3581 * SWAPCACHE OPERATION: 3582 * 3583 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3584 * devfs also uses b_vp for fake buffers so we also have to check 3585 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3586 * underlying block device. The swap assignments are related to the 3587 * buffer cache buffer's b_vp, not the passed vp. 3588 * 3589 * The passed vp == bp->b_vp only in the case where the strategy call 3590 * is made on the vp itself for its own buffers (a regular file or 3591 * block device vp). The filesystem usually then re-calls vn_strategy() 3592 * after translating the request to an underlying device. 3593 * 3594 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3595 * underlying buffer cache buffers. 3596 * 3597 * We can only deal with page-aligned buffers at the moment, because 3598 * we can't tell what the real dirty state for pages straddling a buffer 3599 * are. 3600 * 3601 * In order to call swap_pager_strategy() we must provide the VM object 3602 * and base offset for the underlying buffer cache pages so it can find 3603 * the swap blocks. 3604 */ 3605 void 3606 vn_strategy(struct vnode *vp, struct bio *bio) 3607 { 3608 struct bio_track *track; 3609 struct buf *bp = bio->bio_buf; 3610 3611 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3612 3613 /* 3614 * Set when an I/O is issued on the bp. Cleared by consumers 3615 * (aka HAMMER), allowing the consumer to determine if I/O had 3616 * actually occurred. 3617 */ 3618 bp->b_flags |= B_IODEBUG; 3619 3620 /* 3621 * Handle the swap cache intercept. 3622 */ 3623 if (vn_cache_strategy(vp, bio)) 3624 return; 3625 3626 /* 3627 * Otherwise do the operation through the filesystem 3628 */ 3629 if (bp->b_cmd == BUF_CMD_READ) 3630 track = &vp->v_track_read; 3631 else 3632 track = &vp->v_track_write; 3633 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3634 bio->bio_track = track; 3635 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3636 dsched_new_buf(bio->bio_buf); 3637 bio_track_ref(track); 3638 vop_strategy(*vp->v_ops, vp, bio); 3639 } 3640 3641 static void vn_cache_strategy_callback(struct bio *bio); 3642 3643 int 3644 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3645 { 3646 struct buf *bp = bio->bio_buf; 3647 struct bio *nbio; 3648 vm_object_t object; 3649 vm_page_t m; 3650 int i; 3651 3652 /* 3653 * Is this buffer cache buffer suitable for reading from 3654 * the swap cache? 3655 */ 3656 if (vm_swapcache_read_enable == 0 || 3657 bp->b_cmd != BUF_CMD_READ || 3658 ((bp->b_flags & B_CLUSTER) == 0 && 3659 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3660 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3661 (bp->b_bcount & PAGE_MASK) != 0) { 3662 return(0); 3663 } 3664 3665 /* 3666 * Figure out the original VM object (it will match the underlying 3667 * VM pages). Note that swap cached data uses page indices relative 3668 * to that object, not relative to bio->bio_offset. 3669 */ 3670 if (bp->b_flags & B_CLUSTER) 3671 object = vp->v_object; 3672 else 3673 object = bp->b_vp->v_object; 3674 3675 /* 3676 * In order to be able to use the swap cache all underlying VM 3677 * pages must be marked as such, and we can't have any bogus pages. 3678 */ 3679 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3680 m = bp->b_xio.xio_pages[i]; 3681 if ((m->flags & PG_SWAPPED) == 0) 3682 break; 3683 if (m == bogus_page) 3684 break; 3685 } 3686 3687 /* 3688 * If we are good then issue the I/O using swap_pager_strategy(). 3689 * 3690 * We can only do this if the buffer actually supports object-backed 3691 * I/O. If it doesn't npages will be 0. 3692 */ 3693 if (i && i == bp->b_xio.xio_npages) { 3694 m = bp->b_xio.xio_pages[0]; 3695 nbio = push_bio(bio); 3696 nbio->bio_done = vn_cache_strategy_callback; 3697 nbio->bio_offset = ptoa(m->pindex); 3698 KKASSERT(m->object == object); 3699 swap_pager_strategy(object, nbio); 3700 return(1); 3701 } 3702 return(0); 3703 } 3704 3705 /* 3706 * This is a bit of a hack but since the vn_cache_strategy() function can 3707 * override a VFS's strategy function we must make sure that the bio, which 3708 * is probably bio2, doesn't leak an unexpected offset value back to the 3709 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3710 * bio went through its own file strategy function and the the bio2 offset 3711 * is a cached disk offset when, in fact, it isn't. 3712 */ 3713 static void 3714 vn_cache_strategy_callback(struct bio *bio) 3715 { 3716 bio->bio_offset = NOOFFSET; 3717 biodone(pop_bio(bio)); 3718 } 3719 3720 /* 3721 * bpdone: 3722 * 3723 * Finish I/O on a buffer after all BIOs have been processed. 3724 * Called when the bio chain is exhausted or by biowait. If called 3725 * by biowait, elseit is typically 0. 3726 * 3727 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3728 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3729 * assuming B_INVAL is clear. 3730 * 3731 * For the VMIO case, we set B_CACHE if the op was a read and no 3732 * read error occured, or if the op was a write. B_CACHE is never 3733 * set if the buffer is invalid or otherwise uncacheable. 3734 * 3735 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3736 * initiator to leave B_INVAL set to brelse the buffer out of existance 3737 * in the biodone routine. 3738 */ 3739 void 3740 bpdone(struct buf *bp, int elseit) 3741 { 3742 buf_cmd_t cmd; 3743 3744 KASSERT(BUF_REFCNTNB(bp) > 0, 3745 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3746 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3747 ("biodone: bp %p already done!", bp)); 3748 3749 /* 3750 * No more BIOs are left. All completion functions have been dealt 3751 * with, now we clean up the buffer. 3752 */ 3753 cmd = bp->b_cmd; 3754 bp->b_cmd = BUF_CMD_DONE; 3755 3756 /* 3757 * Only reads and writes are processed past this point. 3758 */ 3759 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3760 if (cmd == BUF_CMD_FREEBLKS) 3761 bp->b_flags |= B_NOCACHE; 3762 if (elseit) 3763 brelse(bp); 3764 return; 3765 } 3766 3767 /* 3768 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3769 * a lot worse. XXX - move this above the clearing of b_cmd 3770 */ 3771 if (LIST_FIRST(&bp->b_dep) != NULL) 3772 buf_complete(bp); /* MPSAFE */ 3773 3774 /* 3775 * A failed write must re-dirty the buffer unless B_INVAL 3776 * was set. Only applicable to normal buffers (with VPs). 3777 * vinum buffers may not have a vp. 3778 */ 3779 if (cmd == BUF_CMD_WRITE && 3780 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3781 bp->b_flags &= ~B_NOCACHE; 3782 if (bp->b_vp) 3783 bdirty(bp); 3784 } 3785 3786 if (bp->b_flags & B_VMIO) { 3787 int i; 3788 vm_ooffset_t foff; 3789 vm_page_t m; 3790 vm_object_t obj; 3791 int iosize; 3792 struct vnode *vp = bp->b_vp; 3793 3794 obj = vp->v_object; 3795 3796 #if defined(VFS_BIO_DEBUG) 3797 if (vp->v_auxrefs == 0) 3798 panic("biodone: zero vnode hold count"); 3799 if ((vp->v_flag & VOBJBUF) == 0) 3800 panic("biodone: vnode is not setup for merged cache"); 3801 #endif 3802 3803 foff = bp->b_loffset; 3804 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 3805 KASSERT(obj != NULL, ("biodone: missing VM object")); 3806 3807 #if defined(VFS_BIO_DEBUG) 3808 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3809 kprintf("biodone: paging in progress(%d) < " 3810 "bp->b_xio.xio_npages(%d)\n", 3811 obj->paging_in_progress, 3812 bp->b_xio.xio_npages); 3813 } 3814 #endif 3815 3816 /* 3817 * Set B_CACHE if the op was a normal read and no error 3818 * occured. B_CACHE is set for writes in the b*write() 3819 * routines. 3820 */ 3821 iosize = bp->b_bcount - bp->b_resid; 3822 if (cmd == BUF_CMD_READ && 3823 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3824 bp->b_flags |= B_CACHE; 3825 } 3826 3827 vm_object_hold(obj); 3828 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3829 int bogusflag = 0; 3830 int resid; 3831 3832 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3833 if (resid > iosize) 3834 resid = iosize; 3835 3836 /* 3837 * cleanup bogus pages, restoring the originals. Since 3838 * the originals should still be wired, we don't have 3839 * to worry about interrupt/freeing races destroying 3840 * the VM object association. 3841 */ 3842 m = bp->b_xio.xio_pages[i]; 3843 if (m == bogus_page) { 3844 bogusflag = 1; 3845 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3846 if (m == NULL) 3847 panic("biodone: page disappeared"); 3848 bp->b_xio.xio_pages[i] = m; 3849 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3850 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3851 } 3852 #if defined(VFS_BIO_DEBUG) 3853 if (OFF_TO_IDX(foff) != m->pindex) { 3854 kprintf("biodone: foff(%lu)/m->pindex(%ld) " 3855 "mismatch\n", 3856 (unsigned long)foff, (long)m->pindex); 3857 } 3858 #endif 3859 3860 /* 3861 * In the write case, the valid and clean bits are 3862 * already changed correctly (see bdwrite()), so we 3863 * only need to do this here in the read case. 3864 */ 3865 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3866 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3867 vfs_clean_one_page(bp, i, m); 3868 } 3869 vm_page_flag_clear(m, PG_ZERO); 3870 3871 /* 3872 * when debugging new filesystems or buffer I/O 3873 * methods, this is the most common error that pops 3874 * up. if you see this, you have not set the page 3875 * busy flag correctly!!! 3876 */ 3877 if (m->busy == 0) { 3878 kprintf("biodone: page busy < 0, " 3879 "pindex: %d, foff: 0x(%x,%x), " 3880 "resid: %d, index: %d\n", 3881 (int) m->pindex, (int)(foff >> 32), 3882 (int) foff & 0xffffffff, resid, i); 3883 if (!vn_isdisk(vp, NULL)) 3884 kprintf(" iosize: %ld, loffset: %lld, " 3885 "flags: 0x%08x, npages: %d\n", 3886 bp->b_vp->v_mount->mnt_stat.f_iosize, 3887 (long long)bp->b_loffset, 3888 bp->b_flags, bp->b_xio.xio_npages); 3889 else 3890 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3891 (long long)bp->b_loffset, 3892 bp->b_flags, bp->b_xio.xio_npages); 3893 kprintf(" valid: 0x%x, dirty: 0x%x, " 3894 "wired: %d\n", 3895 m->valid, m->dirty, 3896 m->wire_count); 3897 panic("biodone: page busy < 0"); 3898 } 3899 vm_page_io_finish(m); 3900 vm_page_wakeup(m); 3901 vm_object_pip_wakeup(obj); 3902 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3903 iosize -= resid; 3904 } 3905 bp->b_flags &= ~B_HASBOGUS; 3906 vm_object_drop(obj); 3907 } 3908 3909 /* 3910 * Finish up by releasing the buffer. There are no more synchronous 3911 * or asynchronous completions, those were handled by bio_done 3912 * callbacks. 3913 */ 3914 if (elseit) { 3915 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3916 brelse(bp); 3917 else 3918 bqrelse(bp); 3919 } 3920 } 3921 3922 /* 3923 * Normal biodone. 3924 */ 3925 void 3926 biodone(struct bio *bio) 3927 { 3928 struct buf *bp = bio->bio_buf; 3929 3930 runningbufwakeup(bp); 3931 3932 /* 3933 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3934 */ 3935 while (bio) { 3936 biodone_t *done_func; 3937 struct bio_track *track; 3938 3939 /* 3940 * BIO tracking. Most but not all BIOs are tracked. 3941 */ 3942 if ((track = bio->bio_track) != NULL) { 3943 bio_track_rel(track); 3944 bio->bio_track = NULL; 3945 } 3946 3947 /* 3948 * A bio_done function terminates the loop. The function 3949 * will be responsible for any further chaining and/or 3950 * buffer management. 3951 * 3952 * WARNING! The done function can deallocate the buffer! 3953 */ 3954 if ((done_func = bio->bio_done) != NULL) { 3955 bio->bio_done = NULL; 3956 done_func(bio); 3957 return; 3958 } 3959 bio = bio->bio_prev; 3960 } 3961 3962 /* 3963 * If we've run out of bio's do normal [a]synchronous completion. 3964 */ 3965 bpdone(bp, 1); 3966 } 3967 3968 /* 3969 * Synchronous biodone - this terminates a synchronous BIO. 3970 * 3971 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3972 * but still locked. The caller must brelse() the buffer after waiting 3973 * for completion. 3974 */ 3975 void 3976 biodone_sync(struct bio *bio) 3977 { 3978 struct buf *bp = bio->bio_buf; 3979 int flags; 3980 int nflags; 3981 3982 KKASSERT(bio == &bp->b_bio1); 3983 bpdone(bp, 0); 3984 3985 for (;;) { 3986 flags = bio->bio_flags; 3987 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3988 3989 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3990 if (flags & BIO_WANT) 3991 wakeup(bio); 3992 break; 3993 } 3994 } 3995 } 3996 3997 /* 3998 * vfs_unbusy_pages: 3999 * 4000 * This routine is called in lieu of iodone in the case of 4001 * incomplete I/O. This keeps the busy status for pages 4002 * consistant. 4003 */ 4004 void 4005 vfs_unbusy_pages(struct buf *bp) 4006 { 4007 int i; 4008 4009 runningbufwakeup(bp); 4010 4011 if (bp->b_flags & B_VMIO) { 4012 struct vnode *vp = bp->b_vp; 4013 vm_object_t obj; 4014 4015 obj = vp->v_object; 4016 vm_object_hold(obj); 4017 4018 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4019 vm_page_t m = bp->b_xio.xio_pages[i]; 4020 4021 /* 4022 * When restoring bogus changes the original pages 4023 * should still be wired, so we are in no danger of 4024 * losing the object association and do not need 4025 * critical section protection particularly. 4026 */ 4027 if (m == bogus_page) { 4028 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 4029 if (!m) { 4030 panic("vfs_unbusy_pages: page missing"); 4031 } 4032 bp->b_xio.xio_pages[i] = m; 4033 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4034 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4035 } 4036 vm_page_busy_wait(m, FALSE, "bpdpgw"); 4037 vm_page_flag_clear(m, PG_ZERO); 4038 vm_page_io_finish(m); 4039 vm_page_wakeup(m); 4040 vm_object_pip_wakeup(obj); 4041 } 4042 bp->b_flags &= ~B_HASBOGUS; 4043 vm_object_drop(obj); 4044 } 4045 } 4046 4047 /* 4048 * vfs_busy_pages: 4049 * 4050 * This routine is called before a device strategy routine. 4051 * It is used to tell the VM system that paging I/O is in 4052 * progress, and treat the pages associated with the buffer 4053 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4054 * flag is handled to make sure that the object doesn't become 4055 * inconsistant. 4056 * 4057 * Since I/O has not been initiated yet, certain buffer flags 4058 * such as B_ERROR or B_INVAL may be in an inconsistant state 4059 * and should be ignored. 4060 * 4061 * MPSAFE 4062 */ 4063 void 4064 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4065 { 4066 int i, bogus; 4067 struct lwp *lp = curthread->td_lwp; 4068 4069 /* 4070 * The buffer's I/O command must already be set. If reading, 4071 * B_CACHE must be 0 (double check against callers only doing 4072 * I/O when B_CACHE is 0). 4073 */ 4074 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4075 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4076 4077 if (bp->b_flags & B_VMIO) { 4078 vm_object_t obj; 4079 4080 obj = vp->v_object; 4081 KASSERT(bp->b_loffset != NOOFFSET, 4082 ("vfs_busy_pages: no buffer offset")); 4083 4084 /* 4085 * Busy all the pages. We have to busy them all at once 4086 * to avoid deadlocks. 4087 */ 4088 retry: 4089 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4090 vm_page_t m = bp->b_xio.xio_pages[i]; 4091 4092 if (vm_page_busy_try(m, FALSE)) { 4093 vm_page_sleep_busy(m, FALSE, "vbpage"); 4094 while (--i >= 0) 4095 vm_page_wakeup(bp->b_xio.xio_pages[i]); 4096 goto retry; 4097 } 4098 } 4099 4100 /* 4101 * Setup for I/O, soft-busy the page right now because 4102 * the next loop may block. 4103 */ 4104 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4105 vm_page_t m = bp->b_xio.xio_pages[i]; 4106 4107 vm_page_flag_clear(m, PG_ZERO); 4108 if ((bp->b_flags & B_CLUSTER) == 0) { 4109 vm_object_pip_add(obj, 1); 4110 vm_page_io_start(m); 4111 } 4112 } 4113 4114 /* 4115 * Adjust protections for I/O and do bogus-page mapping. 4116 * Assume that vm_page_protect() can block (it can block 4117 * if VM_PROT_NONE, don't take any chances regardless). 4118 * 4119 * In particular note that for writes we must incorporate 4120 * page dirtyness from the VM system into the buffer's 4121 * dirty range. 4122 * 4123 * For reads we theoretically must incorporate page dirtyness 4124 * from the VM system to determine if the page needs bogus 4125 * replacement, but we shortcut the test by simply checking 4126 * that all m->valid bits are set, indicating that the page 4127 * is fully valid and does not need to be re-read. For any 4128 * VM system dirtyness the page will also be fully valid 4129 * since it was mapped at one point. 4130 */ 4131 bogus = 0; 4132 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4133 vm_page_t m = bp->b_xio.xio_pages[i]; 4134 4135 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4136 if (bp->b_cmd == BUF_CMD_WRITE) { 4137 /* 4138 * When readying a vnode-backed buffer for 4139 * a write we must zero-fill any invalid 4140 * portions of the backing VM pages, mark 4141 * it valid and clear related dirty bits. 4142 * 4143 * vfs_clean_one_page() incorporates any 4144 * VM dirtyness and updates the b_dirtyoff 4145 * range (after we've made the page RO). 4146 * 4147 * It is also expected that the pmap modified 4148 * bit has already been cleared by the 4149 * vm_page_protect(). We may not be able 4150 * to clear all dirty bits for a page if it 4151 * was also memory mapped (NFS). 4152 * 4153 * Finally be sure to unassign any swap-cache 4154 * backing store as it is now stale. 4155 */ 4156 vm_page_protect(m, VM_PROT_READ); 4157 vfs_clean_one_page(bp, i, m); 4158 swap_pager_unswapped(m); 4159 } else if (m->valid == VM_PAGE_BITS_ALL) { 4160 /* 4161 * When readying a vnode-backed buffer for 4162 * read we must replace any dirty pages with 4163 * a bogus page so dirty data is not destroyed 4164 * when filling gaps. 4165 * 4166 * To avoid testing whether the page is 4167 * dirty we instead test that the page was 4168 * at some point mapped (m->valid fully 4169 * valid) with the understanding that 4170 * this also covers the dirty case. 4171 */ 4172 bp->b_xio.xio_pages[i] = bogus_page; 4173 bp->b_flags |= B_HASBOGUS; 4174 bogus++; 4175 } else if (m->valid & m->dirty) { 4176 /* 4177 * This case should not occur as partial 4178 * dirtyment can only happen if the buffer 4179 * is B_CACHE, and this code is not entered 4180 * if the buffer is B_CACHE. 4181 */ 4182 kprintf("Warning: vfs_busy_pages - page not " 4183 "fully valid! loff=%jx bpf=%08x " 4184 "idx=%d val=%02x dir=%02x\n", 4185 (uintmax_t)bp->b_loffset, bp->b_flags, 4186 i, m->valid, m->dirty); 4187 vm_page_protect(m, VM_PROT_NONE); 4188 } else { 4189 /* 4190 * The page is not valid and can be made 4191 * part of the read. 4192 */ 4193 vm_page_protect(m, VM_PROT_NONE); 4194 } 4195 vm_page_wakeup(m); 4196 } 4197 if (bogus) { 4198 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4199 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4200 } 4201 } 4202 4203 /* 4204 * This is the easiest place to put the process accounting for the I/O 4205 * for now. 4206 */ 4207 if (lp != NULL) { 4208 if (bp->b_cmd == BUF_CMD_READ) 4209 lp->lwp_ru.ru_inblock++; 4210 else 4211 lp->lwp_ru.ru_oublock++; 4212 } 4213 } 4214 4215 /* 4216 * Tell the VM system that the pages associated with this buffer 4217 * are clean. This is used for delayed writes where the data is 4218 * going to go to disk eventually without additional VM intevention. 4219 * 4220 * NOTE: While we only really need to clean through to b_bcount, we 4221 * just go ahead and clean through to b_bufsize. 4222 */ 4223 static void 4224 vfs_clean_pages(struct buf *bp) 4225 { 4226 vm_page_t m; 4227 int i; 4228 4229 if ((bp->b_flags & B_VMIO) == 0) 4230 return; 4231 4232 KASSERT(bp->b_loffset != NOOFFSET, 4233 ("vfs_clean_pages: no buffer offset")); 4234 4235 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4236 m = bp->b_xio.xio_pages[i]; 4237 vfs_clean_one_page(bp, i, m); 4238 } 4239 } 4240 4241 /* 4242 * vfs_clean_one_page: 4243 * 4244 * Set the valid bits and clear the dirty bits in a page within a 4245 * buffer. The range is restricted to the buffer's size and the 4246 * buffer's logical offset might index into the first page. 4247 * 4248 * The caller has busied or soft-busied the page and it is not mapped, 4249 * test and incorporate the dirty bits into b_dirtyoff/end before 4250 * clearing them. Note that we need to clear the pmap modified bits 4251 * after determining the the page was dirty, vm_page_set_validclean() 4252 * does not do it for us. 4253 * 4254 * This routine is typically called after a read completes (dirty should 4255 * be zero in that case as we are not called on bogus-replace pages), 4256 * or before a write is initiated. 4257 */ 4258 static void 4259 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4260 { 4261 int bcount; 4262 int xoff; 4263 int soff; 4264 int eoff; 4265 4266 /* 4267 * Calculate offset range within the page but relative to buffer's 4268 * loffset. loffset might be offset into the first page. 4269 */ 4270 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4271 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4272 4273 if (pageno == 0) { 4274 soff = xoff; 4275 eoff = PAGE_SIZE; 4276 } else { 4277 soff = (pageno << PAGE_SHIFT); 4278 eoff = soff + PAGE_SIZE; 4279 } 4280 if (eoff > bcount) 4281 eoff = bcount; 4282 if (soff >= eoff) 4283 return; 4284 4285 /* 4286 * Test dirty bits and adjust b_dirtyoff/end. 4287 * 4288 * If dirty pages are incorporated into the bp any prior 4289 * B_NEEDCOMMIT state (NFS) must be cleared because the 4290 * caller has not taken into account the new dirty data. 4291 * 4292 * If the page was memory mapped the dirty bits might go beyond the 4293 * end of the buffer, but we can't really make the assumption that 4294 * a file EOF straddles the buffer (even though this is the case for 4295 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4296 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4297 * This also saves some console spam. 4298 * 4299 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4300 * NFS can handle huge commits but not huge writes. 4301 */ 4302 vm_page_test_dirty(m); 4303 if (m->dirty) { 4304 if ((bp->b_flags & B_NEEDCOMMIT) && 4305 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4306 if (debug_commit) 4307 kprintf("Warning: vfs_clean_one_page: bp %p " 4308 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4309 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4310 "doff/end %d %d\n", 4311 bp, (uintmax_t)bp->b_loffset, bp->b_bcount, 4312 bp->b_flags, bp->b_cmd, 4313 m->valid, m->dirty, xoff, soff, eoff, 4314 bp->b_dirtyoff, bp->b_dirtyend); 4315 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4316 if (debug_commit) 4317 print_backtrace(-1); 4318 } 4319 /* 4320 * Only clear the pmap modified bits if ALL the dirty bits 4321 * are set, otherwise the system might mis-clear portions 4322 * of a page. 4323 */ 4324 if (m->dirty == VM_PAGE_BITS_ALL && 4325 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4326 pmap_clear_modify(m); 4327 } 4328 if (bp->b_dirtyoff > soff - xoff) 4329 bp->b_dirtyoff = soff - xoff; 4330 if (bp->b_dirtyend < eoff - xoff) 4331 bp->b_dirtyend = eoff - xoff; 4332 } 4333 4334 /* 4335 * Set related valid bits, clear related dirty bits. 4336 * Does not mess with the pmap modified bit. 4337 * 4338 * WARNING! We cannot just clear all of m->dirty here as the 4339 * buffer cache buffers may use a DEV_BSIZE'd aligned 4340 * block size, or have an odd size (e.g. NFS at file EOF). 4341 * The putpages code can clear m->dirty to 0. 4342 * 4343 * If a VOP_WRITE generates a buffer cache buffer which 4344 * covers the same space as mapped writable pages the 4345 * buffer flush might not be able to clear all the dirty 4346 * bits and still require a putpages from the VM system 4347 * to finish it off. 4348 * 4349 * WARNING! vm_page_set_validclean() currently assumes vm_token 4350 * is held. The page might not be busied (bdwrite() case). 4351 * XXX remove this comment once we've validated that this 4352 * is no longer an issue. 4353 */ 4354 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4355 } 4356 4357 #if 0 4358 /* 4359 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4360 * The page data is assumed to be valid (there is no zeroing here). 4361 */ 4362 static void 4363 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4364 { 4365 int bcount; 4366 int xoff; 4367 int soff; 4368 int eoff; 4369 4370 /* 4371 * Calculate offset range within the page but relative to buffer's 4372 * loffset. loffset might be offset into the first page. 4373 */ 4374 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4375 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4376 4377 if (pageno == 0) { 4378 soff = xoff; 4379 eoff = PAGE_SIZE; 4380 } else { 4381 soff = (pageno << PAGE_SHIFT); 4382 eoff = soff + PAGE_SIZE; 4383 } 4384 if (eoff > bcount) 4385 eoff = bcount; 4386 if (soff >= eoff) 4387 return; 4388 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4389 } 4390 #endif 4391 4392 /* 4393 * vfs_bio_clrbuf: 4394 * 4395 * Clear a buffer. This routine essentially fakes an I/O, so we need 4396 * to clear B_ERROR and B_INVAL. 4397 * 4398 * Note that while we only theoretically need to clear through b_bcount, 4399 * we go ahead and clear through b_bufsize. 4400 */ 4401 4402 void 4403 vfs_bio_clrbuf(struct buf *bp) 4404 { 4405 int i, mask = 0; 4406 caddr_t sa, ea; 4407 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4408 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4409 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4410 (bp->b_loffset & PAGE_MASK) == 0) { 4411 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4412 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4413 bp->b_resid = 0; 4414 return; 4415 } 4416 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4417 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4418 bzero(bp->b_data, bp->b_bufsize); 4419 bp->b_xio.xio_pages[0]->valid |= mask; 4420 bp->b_resid = 0; 4421 return; 4422 } 4423 } 4424 sa = bp->b_data; 4425 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4426 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4427 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4428 ea = (caddr_t)(vm_offset_t)ulmin( 4429 (u_long)(vm_offset_t)ea, 4430 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4431 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4432 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4433 continue; 4434 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4435 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4436 bzero(sa, ea - sa); 4437 } 4438 } else { 4439 for (; sa < ea; sa += DEV_BSIZE, j++) { 4440 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4441 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4442 bzero(sa, DEV_BSIZE); 4443 } 4444 } 4445 bp->b_xio.xio_pages[i]->valid |= mask; 4446 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4447 } 4448 bp->b_resid = 0; 4449 } else { 4450 clrbuf(bp); 4451 } 4452 } 4453 4454 /* 4455 * vm_hold_load_pages: 4456 * 4457 * Load pages into the buffer's address space. The pages are 4458 * allocated from the kernel object in order to reduce interference 4459 * with the any VM paging I/O activity. The range of loaded 4460 * pages will be wired. 4461 * 4462 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4463 * retrieve the full range (to - from) of pages. 4464 * 4465 * MPSAFE 4466 */ 4467 void 4468 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4469 { 4470 vm_offset_t pg; 4471 vm_page_t p; 4472 int index; 4473 4474 to = round_page(to); 4475 from = round_page(from); 4476 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4477 4478 pg = from; 4479 while (pg < to) { 4480 /* 4481 * Note: must allocate system pages since blocking here 4482 * could intefere with paging I/O, no matter which 4483 * process we are. 4484 */ 4485 vm_object_hold(&kernel_object); 4486 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT, 4487 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4488 vm_object_drop(&kernel_object); 4489 if (p) { 4490 vm_page_wire(p); 4491 p->valid = VM_PAGE_BITS_ALL; 4492 vm_page_flag_clear(p, PG_ZERO); 4493 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 4494 bp->b_xio.xio_pages[index] = p; 4495 vm_page_wakeup(p); 4496 4497 pg += PAGE_SIZE; 4498 ++index; 4499 } 4500 } 4501 bp->b_xio.xio_npages = index; 4502 } 4503 4504 /* 4505 * Allocate a page for a buffer cache buffer. 4506 * 4507 * If NULL is returned the caller is expected to retry (typically check if 4508 * the page already exists on retry before trying to allocate one). 4509 * 4510 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This 4511 * function will use the system reserve with the hope that the page 4512 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller 4513 * is done with the buffer. 4514 */ 4515 static 4516 vm_page_t 4517 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit) 4518 { 4519 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK; 4520 vm_page_t p; 4521 4522 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4523 4524 /* 4525 * Try a normal allocation first. 4526 */ 4527 p = vm_page_alloc(obj, pg, vmflags); 4528 if (p) 4529 return(p); 4530 if (vm_page_lookup(obj, pg)) 4531 return(NULL); 4532 vm_pageout_deficit += deficit; 4533 4534 /* 4535 * Try again, digging into the system reserve. 4536 * 4537 * Trying to recover pages from the buffer cache here can deadlock 4538 * against other threads trying to busy underlying pages so we 4539 * depend on the code in brelse() and bqrelse() to free/cache the 4540 * underlying buffer cache pages when memory is low. 4541 */ 4542 if (curthread->td_flags & TDF_SYSTHREAD) 4543 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT; 4544 else 4545 vmflags |= VM_ALLOC_SYSTEM; 4546 4547 /*recoverbufpages();*/ 4548 p = vm_page_alloc(obj, pg, vmflags); 4549 if (p) 4550 return(p); 4551 if (vm_page_lookup(obj, pg)) 4552 return(NULL); 4553 4554 /* 4555 * Wait for memory to free up and try again 4556 */ 4557 if (vm_page_count_severe()) 4558 ++lowmempgallocs; 4559 vm_wait(hz / 20 + 1); 4560 4561 p = vm_page_alloc(obj, pg, vmflags); 4562 if (p) 4563 return(p); 4564 if (vm_page_lookup(obj, pg)) 4565 return(NULL); 4566 4567 /* 4568 * Ok, now we are really in trouble. 4569 */ 4570 { 4571 static struct krate biokrate = { .freq = 1 }; 4572 krateprintf(&biokrate, 4573 "Warning: bio_page_alloc: memory exhausted " 4574 "during bufcache page allocation from %s\n", 4575 curthread->td_comm); 4576 } 4577 if (curthread->td_flags & TDF_SYSTHREAD) 4578 vm_wait(hz / 20 + 1); 4579 else 4580 vm_wait(hz / 2 + 1); 4581 return (NULL); 4582 } 4583 4584 /* 4585 * vm_hold_free_pages: 4586 * 4587 * Return pages associated with the buffer back to the VM system. 4588 * 4589 * The range of pages underlying the buffer's address space will 4590 * be unmapped and un-wired. 4591 * 4592 * MPSAFE 4593 */ 4594 void 4595 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4596 { 4597 vm_offset_t pg; 4598 vm_page_t p; 4599 int index, newnpages; 4600 4601 from = round_page(from); 4602 to = round_page(to); 4603 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4604 newnpages = index; 4605 4606 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4607 p = bp->b_xio.xio_pages[index]; 4608 if (p && (index < bp->b_xio.xio_npages)) { 4609 if (p->busy) { 4610 kprintf("vm_hold_free_pages: doffset: %lld, " 4611 "loffset: %lld\n", 4612 (long long)bp->b_bio2.bio_offset, 4613 (long long)bp->b_loffset); 4614 } 4615 bp->b_xio.xio_pages[index] = NULL; 4616 pmap_kremove(pg); 4617 vm_page_busy_wait(p, FALSE, "vmhldpg"); 4618 vm_page_unwire(p, 0); 4619 vm_page_free(p); 4620 } 4621 } 4622 bp->b_xio.xio_npages = newnpages; 4623 } 4624 4625 /* 4626 * vmapbuf: 4627 * 4628 * Map a user buffer into KVM via a pbuf. On return the buffer's 4629 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4630 * initialized. 4631 */ 4632 int 4633 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4634 { 4635 caddr_t addr; 4636 vm_offset_t va; 4637 vm_page_t m; 4638 int vmprot; 4639 int error; 4640 int pidx; 4641 int i; 4642 4643 /* 4644 * bp had better have a command and it better be a pbuf. 4645 */ 4646 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4647 KKASSERT(bp->b_flags & B_PAGING); 4648 KKASSERT(bp->b_kvabase); 4649 4650 if (bytes < 0) 4651 return (-1); 4652 4653 /* 4654 * Map the user data into KVM. Mappings have to be page-aligned. 4655 */ 4656 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4657 pidx = 0; 4658 4659 vmprot = VM_PROT_READ; 4660 if (bp->b_cmd == BUF_CMD_READ) 4661 vmprot |= VM_PROT_WRITE; 4662 4663 while (addr < udata + bytes) { 4664 /* 4665 * Do the vm_fault if needed; do the copy-on-write thing 4666 * when reading stuff off device into memory. 4667 * 4668 * vm_fault_page*() returns a held VM page. 4669 */ 4670 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4671 va = trunc_page(va); 4672 4673 m = vm_fault_page_quick(va, vmprot, &error); 4674 if (m == NULL) { 4675 for (i = 0; i < pidx; ++i) { 4676 vm_page_unhold(bp->b_xio.xio_pages[i]); 4677 bp->b_xio.xio_pages[i] = NULL; 4678 } 4679 return(-1); 4680 } 4681 bp->b_xio.xio_pages[pidx] = m; 4682 addr += PAGE_SIZE; 4683 ++pidx; 4684 } 4685 4686 /* 4687 * Map the page array and set the buffer fields to point to 4688 * the mapped data buffer. 4689 */ 4690 if (pidx > btoc(MAXPHYS)) 4691 panic("vmapbuf: mapped more than MAXPHYS"); 4692 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4693 4694 bp->b_xio.xio_npages = pidx; 4695 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4696 bp->b_bcount = bytes; 4697 bp->b_bufsize = bytes; 4698 return(0); 4699 } 4700 4701 /* 4702 * vunmapbuf: 4703 * 4704 * Free the io map PTEs associated with this IO operation. 4705 * We also invalidate the TLB entries and restore the original b_addr. 4706 */ 4707 void 4708 vunmapbuf(struct buf *bp) 4709 { 4710 int pidx; 4711 int npages; 4712 4713 KKASSERT(bp->b_flags & B_PAGING); 4714 4715 npages = bp->b_xio.xio_npages; 4716 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4717 for (pidx = 0; pidx < npages; ++pidx) { 4718 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4719 bp->b_xio.xio_pages[pidx] = NULL; 4720 } 4721 bp->b_xio.xio_npages = 0; 4722 bp->b_data = bp->b_kvabase; 4723 } 4724 4725 /* 4726 * Scan all buffers in the system and issue the callback. 4727 */ 4728 int 4729 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4730 { 4731 int count = 0; 4732 int error; 4733 long n; 4734 4735 for (n = 0; n < nbuf; ++n) { 4736 if ((error = callback(&buf[n], info)) < 0) { 4737 count = error; 4738 break; 4739 } 4740 count += error; 4741 } 4742 return (count); 4743 } 4744 4745 /* 4746 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4747 * completion to the master buffer. 4748 */ 4749 static void 4750 nestiobuf_iodone(struct bio *bio) 4751 { 4752 struct bio *mbio; 4753 struct buf *mbp, *bp; 4754 struct devstat *stats; 4755 int error; 4756 int donebytes; 4757 4758 bp = bio->bio_buf; 4759 mbio = bio->bio_caller_info1.ptr; 4760 stats = bio->bio_caller_info2.ptr; 4761 mbp = mbio->bio_buf; 4762 4763 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4764 KKASSERT(mbp != bp); 4765 4766 error = bp->b_error; 4767 if (bp->b_error == 0 && 4768 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4769 /* 4770 * Not all got transfered, raise an error. We have no way to 4771 * propagate these conditions to mbp. 4772 */ 4773 error = EIO; 4774 } 4775 4776 donebytes = bp->b_bufsize; 4777 4778 relpbuf(bp, NULL); 4779 4780 nestiobuf_done(mbio, donebytes, error, stats); 4781 } 4782 4783 void 4784 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4785 { 4786 struct buf *mbp; 4787 4788 mbp = mbio->bio_buf; 4789 4790 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4791 4792 /* 4793 * If an error occured, propagate it to the master buffer. 4794 * 4795 * Several biodone()s may wind up running concurrently so 4796 * use an atomic op to adjust b_flags. 4797 */ 4798 if (error) { 4799 mbp->b_error = error; 4800 atomic_set_int(&mbp->b_flags, B_ERROR); 4801 } 4802 4803 /* 4804 * Decrement the operations in progress counter and terminate the 4805 * I/O if this was the last bit. 4806 */ 4807 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4808 mbp->b_resid = 0; 4809 if (stats) 4810 devstat_end_transaction_buf(stats, mbp); 4811 biodone(mbio); 4812 } 4813 } 4814 4815 /* 4816 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4817 * the mbio from being biodone()'d while we are still adding sub-bios to 4818 * it. 4819 */ 4820 void 4821 nestiobuf_init(struct bio *bio) 4822 { 4823 bio->bio_driver_info = (void *)1; 4824 } 4825 4826 /* 4827 * The BIOs added to the nestedio have already been started, remove the 4828 * count that placeheld our mbio and biodone() it if the count would 4829 * transition to 0. 4830 */ 4831 void 4832 nestiobuf_start(struct bio *mbio) 4833 { 4834 struct buf *mbp = mbio->bio_buf; 4835 4836 /* 4837 * Decrement the operations in progress counter and terminate the 4838 * I/O if this was the last bit. 4839 */ 4840 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4841 if (mbp->b_flags & B_ERROR) 4842 mbp->b_resid = mbp->b_bcount; 4843 else 4844 mbp->b_resid = 0; 4845 biodone(mbio); 4846 } 4847 } 4848 4849 /* 4850 * Set an intermediate error prior to calling nestiobuf_start() 4851 */ 4852 void 4853 nestiobuf_error(struct bio *mbio, int error) 4854 { 4855 struct buf *mbp = mbio->bio_buf; 4856 4857 if (error) { 4858 mbp->b_error = error; 4859 atomic_set_int(&mbp->b_flags, B_ERROR); 4860 } 4861 } 4862 4863 /* 4864 * nestiobuf_add: setup a "nested" buffer. 4865 * 4866 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4867 * => 'bp' should be a buffer allocated by getiobuf. 4868 * => 'offset' is a byte offset in the master buffer. 4869 * => 'size' is a size in bytes of this nested buffer. 4870 */ 4871 void 4872 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4873 { 4874 struct buf *mbp = mbio->bio_buf; 4875 struct vnode *vp = mbp->b_vp; 4876 4877 KKASSERT(mbp->b_bcount >= offset + size); 4878 4879 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4880 4881 /* kernel needs to own the lock for it to be released in biodone */ 4882 BUF_KERNPROC(bp); 4883 bp->b_vp = vp; 4884 bp->b_cmd = mbp->b_cmd; 4885 bp->b_bio1.bio_done = nestiobuf_iodone; 4886 bp->b_data = (char *)mbp->b_data + offset; 4887 bp->b_resid = bp->b_bcount = size; 4888 bp->b_bufsize = bp->b_bcount; 4889 4890 bp->b_bio1.bio_track = NULL; 4891 bp->b_bio1.bio_caller_info1.ptr = mbio; 4892 bp->b_bio1.bio_caller_info2.ptr = stats; 4893 } 4894 4895 /* 4896 * print out statistics from the current status of the buffer pool 4897 * this can be toggeled by the system control option debug.syncprt 4898 */ 4899 #ifdef DEBUG 4900 void 4901 vfs_bufstats(void) 4902 { 4903 int i, j, count; 4904 struct buf *bp; 4905 struct bqueues *dp; 4906 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 4907 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 4908 4909 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 4910 count = 0; 4911 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4912 counts[j] = 0; 4913 4914 spin_lock(&bufqspin); 4915 TAILQ_FOREACH(bp, dp, b_freelist) { 4916 if (bp->b_flags & B_MARKER) 4917 continue; 4918 counts[bp->b_bufsize/PAGE_SIZE]++; 4919 count++; 4920 } 4921 spin_unlock(&bufqspin); 4922 4923 kprintf("%s: total-%d", bname[i], count); 4924 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 4925 if (counts[j] != 0) 4926 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]); 4927 kprintf("\n"); 4928 } 4929 } 4930 #endif 4931 4932 #ifdef DDB 4933 4934 DB_SHOW_COMMAND(buffer, db_show_buffer) 4935 { 4936 /* get args */ 4937 struct buf *bp = (struct buf *)addr; 4938 4939 if (!have_addr) { 4940 db_printf("usage: show buffer <addr>\n"); 4941 return; 4942 } 4943 4944 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4945 db_printf("b_cmd = %d\n", bp->b_cmd); 4946 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4947 "b_resid = %d\n, b_data = %p, " 4948 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4949 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4950 bp->b_data, 4951 (long long)bp->b_bio2.bio_offset, 4952 (long long)(bp->b_bio2.bio_next ? 4953 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4954 if (bp->b_xio.xio_npages) { 4955 int i; 4956 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4957 bp->b_xio.xio_npages); 4958 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4959 vm_page_t m; 4960 m = bp->b_xio.xio_pages[i]; 4961 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4962 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4963 if ((i + 1) < bp->b_xio.xio_npages) 4964 db_printf(","); 4965 } 4966 db_printf("\n"); 4967 } 4968 } 4969 #endif /* DDB */ 4970