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); 91 92 struct bufpcpu { 93 struct spinlock spin; 94 struct bqueues bufqueues[BUFFER_QUEUES]; 95 } __cachealign; 96 97 struct bufpcpu bufpcpu[MAXCPU]; 98 99 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 100 101 struct buf *buf; /* buffer header pool */ 102 103 static void vfs_clean_pages(struct buf *bp); 104 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m); 105 #if 0 106 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m); 107 #endif 108 static void vfs_vmio_release(struct buf *bp); 109 static int flushbufqueues(struct buf *marker, bufq_type_t q); 110 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj, 111 vm_pindex_t pg, int deficit); 112 113 static void bd_signal(long totalspace); 114 static void buf_daemon(void); 115 static void buf_daemon_hw(void); 116 117 /* 118 * bogus page -- for I/O to/from partially complete buffers 119 * this is a temporary solution to the problem, but it is not 120 * really that bad. it would be better to split the buffer 121 * for input in the case of buffers partially already in memory, 122 * but the code is intricate enough already. 123 */ 124 vm_page_t bogus_page; 125 126 /* 127 * These are all static, but make the ones we export globals so we do 128 * not need to use compiler magic. 129 */ 130 long bufspace; /* locked by buffer_map */ 131 long maxbufspace; 132 static long bufmallocspace; /* atomic ops */ 133 long maxbufmallocspace, lobufspace, hibufspace; 134 static long bufreusecnt, bufdefragcnt, buffreekvacnt; 135 static long lorunningspace; 136 static long hirunningspace; 137 static long dirtykvaspace; /* atomic */ 138 static long dirtybufspace; /* atomic */ 139 static long dirtybufcount; /* atomic */ 140 static long dirtybufspacehw; /* atomic */ 141 static long dirtybufcounthw; /* atomic */ 142 static long runningbufspace; /* atomic */ 143 static long runningbufcount; /* atomic */ 144 long lodirtybufspace; 145 long hidirtybufspace; 146 static int getnewbufcalls; 147 static int getnewbufrestarts; 148 static int recoverbufcalls; 149 static int needsbuffer; /* atomic */ 150 static int runningbufreq; /* atomic */ 151 static int bd_request; /* atomic */ 152 static int bd_request_hw; /* atomic */ 153 static u_int bd_wake_ary[BD_WAKE_SIZE]; 154 static u_int bd_wake_index; 155 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */ 156 static int debug_commit; 157 static int debug_bufbio; 158 159 static struct thread *bufdaemon_td; 160 static struct thread *bufdaemonhw_td; 161 static u_int lowmempgallocs; 162 static u_int lowmempgfails; 163 164 /* 165 * Sysctls for operational control of the buffer cache. 166 */ 167 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0, 168 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 169 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0, 170 "High watermark used to trigger explicit flushing of dirty buffers"); 171 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 172 "Minimum amount of buffer space required for active I/O"); 173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 174 "Maximum amount of buffer space to usable for active I/O"); 175 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0, 176 "Page allocations done during periods of very low free memory"); 177 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0, 178 "Page allocations which failed during periods of very low free memory"); 179 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0, 180 "Recycle pages to active or inactive queue transition pt 0-64"); 181 /* 182 * Sysctls determining current state of the buffer cache. 183 */ 184 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0, 185 "Total number of buffers in buffer cache"); 186 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0, 187 "KVA reserved by dirty buffers (all)"); 188 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0, 189 "Pending bytes of dirty buffers (all)"); 190 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0, 191 "Pending bytes of dirty buffers (heavy weight)"); 192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0, 193 "Pending number of dirty buffers"); 194 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0, 195 "Pending number of dirty buffers (heavy weight)"); 196 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 197 "I/O bytes currently in progress due to asynchronous writes"); 198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0, 199 "I/O buffers currently in progress due to asynchronous writes"); 200 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 201 "Hard limit on maximum amount of memory usable for buffer space"); 202 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 203 "Soft limit on maximum amount of memory usable for buffer space"); 204 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 205 "Minimum amount of memory to reserve for system buffer space"); 206 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 207 "Amount of memory available for buffers"); 208 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 209 0, "Maximum amount of memory reserved for buffers using malloc"); 210 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 211 "Amount of memory left for buffers using malloc-scheme"); 212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 213 "New buffer header acquisition requests"); 214 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 215 0, "New buffer header acquisition restarts"); 216 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0, 217 "Recover VM space in an emergency"); 218 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 219 "Buffer acquisition restarts due to fragmented buffer map"); 220 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 221 "Amount of time KVA space was deallocated in an arbitrary buffer"); 222 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 223 "Amount of time buffer re-use operations were successful"); 224 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, ""); 225 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, ""); 226 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 227 "sizeof(struct buf)"); 228 229 char *buf_wmesg = BUF_WMESG; 230 231 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 232 #define VFS_BIO_NEED_UNUSED02 0x02 233 #define VFS_BIO_NEED_UNUSED04 0x04 234 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 235 236 /* 237 * bufspacewakeup: 238 * 239 * Called when buffer space is potentially available for recovery. 240 * getnewbuf() will block on this flag when it is unable to free 241 * sufficient buffer space. Buffer space becomes recoverable when 242 * bp's get placed back in the queues. 243 */ 244 static __inline void 245 bufspacewakeup(void) 246 { 247 /* 248 * If someone is waiting for BUF space, wake them up. Even 249 * though we haven't freed the kva space yet, the waiting 250 * process will be able to now. 251 */ 252 for (;;) { 253 int flags = needsbuffer; 254 cpu_ccfence(); 255 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0) 256 break; 257 if (atomic_cmpset_int(&needsbuffer, flags, 258 flags & ~VFS_BIO_NEED_BUFSPACE)) { 259 wakeup(&needsbuffer); 260 break; 261 } 262 /* retry */ 263 } 264 } 265 266 /* 267 * runningbufwakeup: 268 * 269 * Accounting for I/O in progress. 270 * 271 */ 272 static __inline void 273 runningbufwakeup(struct buf *bp) 274 { 275 long totalspace; 276 long limit; 277 long flags; 278 279 if ((totalspace = bp->b_runningbufspace) != 0) { 280 atomic_add_long(&runningbufspace, -totalspace); 281 atomic_add_long(&runningbufcount, -1); 282 bp->b_runningbufspace = 0; 283 284 /* 285 * see waitrunningbufspace() for limit test. 286 */ 287 limit = hirunningspace * 3 / 6; 288 for (;;) { 289 flags = runningbufreq; 290 cpu_ccfence(); 291 if (flags == 0) 292 break; 293 if (atomic_cmpset_int(&runningbufreq, flags, 0)) { 294 wakeup(&runningbufreq); 295 break; 296 } 297 /* retry */ 298 } 299 bd_signal(totalspace); 300 } 301 } 302 303 /* 304 * bufcountwakeup: 305 * 306 * Called when a buffer has been added to one of the free queues to 307 * account for the buffer and to wakeup anyone waiting for free buffers. 308 * This typically occurs when large amounts of metadata are being handled 309 * by the buffer cache ( else buffer space runs out first, usually ). 310 */ 311 static __inline void 312 bufcountwakeup(void) 313 { 314 long flags; 315 316 for (;;) { 317 flags = needsbuffer; 318 if (flags == 0) 319 break; 320 if (atomic_cmpset_int(&needsbuffer, flags, 321 (flags & ~VFS_BIO_NEED_ANY))) { 322 wakeup(&needsbuffer); 323 break; 324 } 325 /* retry */ 326 } 327 } 328 329 /* 330 * waitrunningbufspace() 331 * 332 * If runningbufspace exceeds 4/6 hirunningspace we block until 333 * runningbufspace drops to 3/6 hirunningspace. We also block if another 334 * thread blocked here in order to be fair, even if runningbufspace 335 * is now lower than the limit. 336 * 337 * The caller may be using this function to block in a tight loop, we 338 * must block while runningbufspace is greater than at least 339 * hirunningspace * 3 / 6. 340 */ 341 void 342 waitrunningbufspace(void) 343 { 344 long limit = hirunningspace * 4 / 6; 345 long flags; 346 347 while (runningbufspace > limit || runningbufreq) { 348 tsleep_interlock(&runningbufreq, 0); 349 flags = atomic_fetchadd_int(&runningbufreq, 1); 350 if (runningbufspace > limit || flags) 351 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz); 352 } 353 } 354 355 /* 356 * buf_dirty_count_severe: 357 * 358 * Return true if we have too many dirty buffers. 359 */ 360 int 361 buf_dirty_count_severe(void) 362 { 363 return (runningbufspace + dirtykvaspace >= hidirtybufspace || 364 dirtybufcount >= nbuf / 2); 365 } 366 367 /* 368 * Return true if the amount of running I/O is severe and BIOQ should 369 * start bursting. 370 */ 371 int 372 buf_runningbufspace_severe(void) 373 { 374 return (runningbufspace >= hirunningspace * 4 / 6); 375 } 376 377 /* 378 * vfs_buf_test_cache: 379 * 380 * Called when a buffer is extended. This function clears the B_CACHE 381 * bit if the newly extended portion of the buffer does not contain 382 * valid data. 383 * 384 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer 385 * cache buffers. The VM pages remain dirty, as someone had mmap()'d 386 * them while a clean buffer was present. 387 */ 388 static __inline__ 389 void 390 vfs_buf_test_cache(struct buf *bp, 391 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 392 vm_page_t m) 393 { 394 if (bp->b_flags & B_CACHE) { 395 int base = (foff + off) & PAGE_MASK; 396 if (vm_page_is_valid(m, base, size) == 0) 397 bp->b_flags &= ~B_CACHE; 398 } 399 } 400 401 /* 402 * bd_speedup() 403 * 404 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the 405 * low water mark. 406 */ 407 static __inline__ 408 void 409 bd_speedup(void) 410 { 411 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2) 412 return; 413 414 if (bd_request == 0 && 415 (dirtykvaspace > lodirtybufspace / 2 || 416 dirtybufcount - dirtybufcounthw >= nbuf / 2)) { 417 if (atomic_fetchadd_int(&bd_request, 1) == 0) 418 wakeup(&bd_request); 419 } 420 if (bd_request_hw == 0 && 421 (dirtykvaspace > lodirtybufspace / 2 || 422 dirtybufcounthw >= nbuf / 2)) { 423 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0) 424 wakeup(&bd_request_hw); 425 } 426 } 427 428 /* 429 * bd_heatup() 430 * 431 * Get the buf_daemon heated up when the number of running and dirty 432 * buffers exceeds the mid-point. 433 * 434 * Return the total number of dirty bytes past the second mid point 435 * as a measure of how much excess dirty data there is in the system. 436 */ 437 long 438 bd_heatup(void) 439 { 440 long mid1; 441 long mid2; 442 long totalspace; 443 444 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2; 445 446 totalspace = runningbufspace + dirtykvaspace; 447 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) { 448 bd_speedup(); 449 mid2 = mid1 + (hidirtybufspace - mid1) / 2; 450 if (totalspace >= mid2) 451 return(totalspace - mid2); 452 } 453 return(0); 454 } 455 456 /* 457 * bd_wait() 458 * 459 * Wait for the buffer cache to flush (totalspace) bytes worth of 460 * buffers, then return. 461 * 462 * Regardless this function blocks while the number of dirty buffers 463 * exceeds hidirtybufspace. 464 */ 465 void 466 bd_wait(long totalspace) 467 { 468 u_int i; 469 u_int j; 470 u_int mi; 471 int count; 472 473 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td) 474 return; 475 476 while (totalspace > 0) { 477 bd_heatup(); 478 479 /* 480 * Order is important. Suppliers adjust bd_wake_index after 481 * updating runningbufspace/dirtykvaspace. We want to fetch 482 * bd_wake_index before accessing. Any error should thus 483 * be in our favor. 484 */ 485 i = atomic_fetchadd_int(&bd_wake_index, 0); 486 if (totalspace > runningbufspace + dirtykvaspace) 487 totalspace = runningbufspace + dirtykvaspace; 488 count = totalspace / BKVASIZE; 489 if (count >= BD_WAKE_SIZE / 2) 490 count = BD_WAKE_SIZE / 2; 491 i = i + count; 492 mi = i & BD_WAKE_MASK; 493 494 /* 495 * This is not a strict interlock, so we play a bit loose 496 * with locking access to dirtybufspace*. We have to re-check 497 * bd_wake_index to ensure that it hasn't passed us. 498 */ 499 tsleep_interlock(&bd_wake_ary[mi], 0); 500 atomic_add_int(&bd_wake_ary[mi], 1); 501 j = atomic_fetchadd_int(&bd_wake_index, 0); 502 if ((int)(i - j) >= 0) 503 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz); 504 505 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace; 506 } 507 } 508 509 /* 510 * bd_signal() 511 * 512 * This function is called whenever runningbufspace or dirtykvaspace 513 * is reduced. Track threads waiting for run+dirty buffer I/O 514 * complete. 515 */ 516 static void 517 bd_signal(long totalspace) 518 { 519 u_int i; 520 521 if (totalspace > 0) { 522 if (totalspace > BKVASIZE * BD_WAKE_SIZE) 523 totalspace = BKVASIZE * BD_WAKE_SIZE; 524 while (totalspace > 0) { 525 i = atomic_fetchadd_int(&bd_wake_index, 1); 526 i &= BD_WAKE_MASK; 527 if (atomic_readandclear_int(&bd_wake_ary[i])) 528 wakeup(&bd_wake_ary[i]); 529 totalspace -= BKVASIZE; 530 } 531 } 532 } 533 534 /* 535 * BIO tracking support routines. 536 * 537 * Release a ref on a bio_track. Wakeup requests are atomically released 538 * along with the last reference so bk_active will never wind up set to 539 * only 0x80000000. 540 */ 541 static 542 void 543 bio_track_rel(struct bio_track *track) 544 { 545 int active; 546 int desired; 547 548 /* 549 * Shortcut 550 */ 551 active = track->bk_active; 552 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0)) 553 return; 554 555 /* 556 * Full-on. Note that the wait flag is only atomically released on 557 * the 1->0 count transition. 558 * 559 * We check for a negative count transition using bit 30 since bit 31 560 * has a different meaning. 561 */ 562 for (;;) { 563 desired = (active & 0x7FFFFFFF) - 1; 564 if (desired) 565 desired |= active & 0x80000000; 566 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 567 if (desired & 0x40000000) 568 panic("bio_track_rel: bad count: %p", track); 569 if (active & 0x80000000) 570 wakeup(track); 571 break; 572 } 573 active = track->bk_active; 574 } 575 } 576 577 /* 578 * Wait for the tracking count to reach 0. 579 * 580 * Use atomic ops such that the wait flag is only set atomically when 581 * bk_active is non-zero. 582 */ 583 int 584 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo) 585 { 586 int active; 587 int desired; 588 int error; 589 590 /* 591 * Shortcut 592 */ 593 if (track->bk_active == 0) 594 return(0); 595 596 /* 597 * Full-on. Note that the wait flag may only be atomically set if 598 * the active count is non-zero. 599 * 600 * NOTE: We cannot optimize active == desired since a wakeup could 601 * clear active prior to our tsleep_interlock(). 602 */ 603 error = 0; 604 while ((active = track->bk_active) != 0) { 605 cpu_ccfence(); 606 desired = active | 0x80000000; 607 tsleep_interlock(track, slp_flags); 608 if (atomic_cmpset_int(&track->bk_active, active, desired)) { 609 error = tsleep(track, slp_flags | PINTERLOCKED, 610 "trwait", slp_timo); 611 if (error) 612 break; 613 } 614 } 615 return (error); 616 } 617 618 /* 619 * bufinit: 620 * 621 * Load time initialisation of the buffer cache, called from machine 622 * dependant initialization code. 623 */ 624 static 625 void 626 bufinit(void *dummy __unused) 627 { 628 struct bufpcpu *pcpu; 629 struct buf *bp; 630 vm_offset_t bogus_offset; 631 int i; 632 int j; 633 long n; 634 635 /* next, make a null set of free lists */ 636 for (i = 0; i < ncpus; ++i) { 637 pcpu = &bufpcpu[i]; 638 spin_init(&pcpu->spin); 639 for (j = 0; j < BUFFER_QUEUES; j++) 640 TAILQ_INIT(&pcpu->bufqueues[j]); 641 } 642 643 /* finally, initialize each buffer header and stick on empty q */ 644 i = 0; 645 pcpu = &bufpcpu[i]; 646 647 for (n = 0; n < nbuf; n++) { 648 bp = &buf[n]; 649 bzero(bp, sizeof *bp); 650 bp->b_flags = B_INVAL; /* we're just an empty header */ 651 bp->b_cmd = BUF_CMD_DONE; 652 bp->b_qindex = BQUEUE_EMPTY; 653 bp->b_qcpu = i; 654 initbufbio(bp); 655 xio_init(&bp->b_xio); 656 buf_dep_init(bp); 657 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 658 bp, b_freelist); 659 660 i = (i + 1) % ncpus; 661 pcpu = &bufpcpu[i]; 662 } 663 664 /* 665 * maxbufspace is the absolute maximum amount of buffer space we are 666 * allowed to reserve in KVM and in real terms. The absolute maximum 667 * is nominally used by buf_daemon. hibufspace is the nominal maximum 668 * used by most other processes. The differential is required to 669 * ensure that buf_daemon is able to run when other processes might 670 * be blocked waiting for buffer space. 671 * 672 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 673 * this may result in KVM fragmentation which is not handled optimally 674 * by the system. 675 */ 676 maxbufspace = nbuf * BKVASIZE; 677 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 678 lobufspace = hibufspace - MAXBSIZE; 679 680 lorunningspace = 512 * 1024; 681 /* hirunningspace -- see below */ 682 683 /* 684 * Limit the amount of malloc memory since it is wired permanently 685 * into the kernel space. Even though this is accounted for in 686 * the buffer allocation, we don't want the malloced region to grow 687 * uncontrolled. The malloc scheme improves memory utilization 688 * significantly on average (small) directories. 689 */ 690 maxbufmallocspace = hibufspace / 20; 691 692 /* 693 * Reduce the chance of a deadlock occuring by limiting the number 694 * of delayed-write dirty buffers we allow to stack up. 695 * 696 * We don't want too much actually queued to the device at once 697 * (XXX this needs to be per-mount!), because the buffers will 698 * wind up locked for a very long period of time while the I/O 699 * drains. 700 */ 701 hidirtybufspace = hibufspace / 2; /* dirty + running */ 702 hirunningspace = hibufspace / 16; /* locked & queued to device */ 703 if (hirunningspace < 1024 * 1024) 704 hirunningspace = 1024 * 1024; 705 706 dirtykvaspace = 0; 707 dirtybufspace = 0; 708 dirtybufspacehw = 0; 709 710 lodirtybufspace = hidirtybufspace / 2; 711 712 /* 713 * Maximum number of async ops initiated per buf_daemon loop. This is 714 * somewhat of a hack at the moment, we really need to limit ourselves 715 * based on the number of bytes of I/O in-transit that were initiated 716 * from buf_daemon. 717 */ 718 719 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE); 720 vm_object_hold(&kernel_object); 721 bogus_page = vm_page_alloc(&kernel_object, 722 (bogus_offset >> PAGE_SHIFT), 723 VM_ALLOC_NORMAL); 724 vm_object_drop(&kernel_object); 725 vmstats.v_wire_count++; 726 727 } 728 729 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL); 730 731 /* 732 * Initialize the embedded bio structures, typically used by 733 * deprecated code which tries to allocate its own struct bufs. 734 */ 735 void 736 initbufbio(struct buf *bp) 737 { 738 bp->b_bio1.bio_buf = bp; 739 bp->b_bio1.bio_prev = NULL; 740 bp->b_bio1.bio_offset = NOOFFSET; 741 bp->b_bio1.bio_next = &bp->b_bio2; 742 bp->b_bio1.bio_done = NULL; 743 bp->b_bio1.bio_flags = 0; 744 745 bp->b_bio2.bio_buf = bp; 746 bp->b_bio2.bio_prev = &bp->b_bio1; 747 bp->b_bio2.bio_offset = NOOFFSET; 748 bp->b_bio2.bio_next = NULL; 749 bp->b_bio2.bio_done = NULL; 750 bp->b_bio2.bio_flags = 0; 751 752 BUF_LOCKINIT(bp); 753 } 754 755 /* 756 * Reinitialize the embedded bio structures as well as any additional 757 * translation cache layers. 758 */ 759 void 760 reinitbufbio(struct buf *bp) 761 { 762 struct bio *bio; 763 764 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 765 bio->bio_done = NULL; 766 bio->bio_offset = NOOFFSET; 767 } 768 } 769 770 /* 771 * Undo the effects of an initbufbio(). 772 */ 773 void 774 uninitbufbio(struct buf *bp) 775 { 776 dsched_exit_buf(bp); 777 BUF_LOCKFREE(bp); 778 } 779 780 /* 781 * Push another BIO layer onto an existing BIO and return it. The new 782 * BIO layer may already exist, holding cached translation data. 783 */ 784 struct bio * 785 push_bio(struct bio *bio) 786 { 787 struct bio *nbio; 788 789 if ((nbio = bio->bio_next) == NULL) { 790 int index = bio - &bio->bio_buf->b_bio_array[0]; 791 if (index >= NBUF_BIO - 1) { 792 panic("push_bio: too many layers bp %p", 793 bio->bio_buf); 794 } 795 nbio = &bio->bio_buf->b_bio_array[index + 1]; 796 bio->bio_next = nbio; 797 nbio->bio_prev = bio; 798 nbio->bio_buf = bio->bio_buf; 799 nbio->bio_offset = NOOFFSET; 800 nbio->bio_done = NULL; 801 nbio->bio_next = NULL; 802 } 803 KKASSERT(nbio->bio_done == NULL); 804 return(nbio); 805 } 806 807 /* 808 * Pop a BIO translation layer, returning the previous layer. The 809 * must have been previously pushed. 810 */ 811 struct bio * 812 pop_bio(struct bio *bio) 813 { 814 return(bio->bio_prev); 815 } 816 817 void 818 clearbiocache(struct bio *bio) 819 { 820 while (bio) { 821 bio->bio_offset = NOOFFSET; 822 bio = bio->bio_next; 823 } 824 } 825 826 /* 827 * bfreekva: 828 * 829 * Free the KVA allocation for buffer 'bp'. 830 * 831 * Must be called from a critical section as this is the only locking for 832 * buffer_map. 833 * 834 * Since this call frees up buffer space, we call bufspacewakeup(). 835 */ 836 static void 837 bfreekva(struct buf *bp) 838 { 839 int count; 840 841 if (bp->b_kvasize) { 842 ++buffreekvacnt; 843 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 844 vm_map_lock(&buffer_map); 845 bufspace -= bp->b_kvasize; 846 vm_map_delete(&buffer_map, 847 (vm_offset_t) bp->b_kvabase, 848 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 849 &count 850 ); 851 vm_map_unlock(&buffer_map); 852 vm_map_entry_release(count); 853 bp->b_kvasize = 0; 854 bp->b_kvabase = NULL; 855 bufspacewakeup(); 856 } 857 } 858 859 /* 860 * Remove the buffer from the appropriate free list. 861 * (caller must be locked) 862 */ 863 static __inline void 864 _bremfree(struct buf *bp) 865 { 866 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 867 868 if (bp->b_qindex != BQUEUE_NONE) { 869 KASSERT(BUF_REFCNTNB(bp) == 1, 870 ("bremfree: bp %p not locked",bp)); 871 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist); 872 bp->b_qindex = BQUEUE_NONE; 873 } else { 874 if (BUF_REFCNTNB(bp) <= 1) 875 panic("bremfree: removing a buffer not on a queue"); 876 } 877 } 878 879 /* 880 * bremfree() - must be called with a locked buffer 881 */ 882 void 883 bremfree(struct buf *bp) 884 { 885 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu]; 886 887 spin_lock(&pcpu->spin); 888 _bremfree(bp); 889 spin_unlock(&pcpu->spin); 890 } 891 892 /* 893 * bremfree_locked - must be called with pcpu->spin locked 894 */ 895 static void 896 bremfree_locked(struct buf *bp) 897 { 898 _bremfree(bp); 899 } 900 901 /* 902 * This version of bread issues any required I/O asyncnronously and 903 * makes a callback on completion. 904 * 905 * The callback must check whether BIO_DONE is set in the bio and issue 906 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 907 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 908 */ 909 void 910 breadcb(struct vnode *vp, off_t loffset, int size, 911 void (*func)(struct bio *), void *arg) 912 { 913 struct buf *bp; 914 915 bp = getblk(vp, loffset, size, 0, 0); 916 917 /* if not found in cache, do some I/O */ 918 if ((bp->b_flags & B_CACHE) == 0) { 919 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 920 bp->b_cmd = BUF_CMD_READ; 921 bp->b_bio1.bio_done = func; 922 bp->b_bio1.bio_caller_info1.ptr = arg; 923 vfs_busy_pages(vp, bp); 924 BUF_KERNPROC(bp); 925 vn_strategy(vp, &bp->b_bio1); 926 } else if (func) { 927 /* 928 * Since we are issuing the callback synchronously it cannot 929 * race the BIO_DONE, so no need for atomic ops here. 930 */ 931 /*bp->b_bio1.bio_done = func;*/ 932 bp->b_bio1.bio_caller_info1.ptr = arg; 933 bp->b_bio1.bio_flags |= BIO_DONE; 934 func(&bp->b_bio1); 935 } else { 936 bqrelse(bp); 937 } 938 } 939 940 /* 941 * breadnx() - Terminal function for bread() and breadn(). 942 * 943 * This function will start asynchronous I/O on read-ahead blocks as well 944 * as satisfy the primary request. 945 * 946 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is 947 * set, the buffer is valid and we do not have to do anything. 948 */ 949 int 950 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset, 951 int *rabsize, int cnt, struct buf **bpp) 952 { 953 struct buf *bp, *rabp; 954 int i; 955 int rv = 0, readwait = 0; 956 957 if (*bpp) 958 bp = *bpp; 959 else 960 *bpp = bp = getblk(vp, loffset, size, 0, 0); 961 962 /* if not found in cache, do some I/O */ 963 if ((bp->b_flags & B_CACHE) == 0) { 964 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 965 bp->b_cmd = BUF_CMD_READ; 966 bp->b_bio1.bio_done = biodone_sync; 967 bp->b_bio1.bio_flags |= BIO_SYNC; 968 vfs_busy_pages(vp, bp); 969 vn_strategy(vp, &bp->b_bio1); 970 ++readwait; 971 } 972 973 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 974 if (inmem(vp, *raoffset)) 975 continue; 976 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 977 978 if ((rabp->b_flags & B_CACHE) == 0) { 979 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 980 rabp->b_cmd = BUF_CMD_READ; 981 vfs_busy_pages(vp, rabp); 982 BUF_KERNPROC(rabp); 983 vn_strategy(vp, &rabp->b_bio1); 984 } else { 985 brelse(rabp); 986 } 987 } 988 if (readwait) 989 rv = biowait(&bp->b_bio1, "biord"); 990 return (rv); 991 } 992 993 /* 994 * bwrite: 995 * 996 * Synchronous write, waits for completion. 997 * 998 * Write, release buffer on completion. (Done by iodone 999 * if async). Do not bother writing anything if the buffer 1000 * is invalid. 1001 * 1002 * Note that we set B_CACHE here, indicating that buffer is 1003 * fully valid and thus cacheable. This is true even of NFS 1004 * now so we set it generally. This could be set either here 1005 * or in biodone() since the I/O is synchronous. We put it 1006 * here. 1007 */ 1008 int 1009 bwrite(struct buf *bp) 1010 { 1011 int error; 1012 1013 if (bp->b_flags & B_INVAL) { 1014 brelse(bp); 1015 return (0); 1016 } 1017 if (BUF_REFCNTNB(bp) == 0) 1018 panic("bwrite: buffer is not busy???"); 1019 1020 /* Mark the buffer clean */ 1021 bundirty(bp); 1022 1023 bp->b_flags &= ~(B_ERROR | B_EINTR); 1024 bp->b_flags |= B_CACHE; 1025 bp->b_cmd = BUF_CMD_WRITE; 1026 bp->b_bio1.bio_done = biodone_sync; 1027 bp->b_bio1.bio_flags |= BIO_SYNC; 1028 vfs_busy_pages(bp->b_vp, bp); 1029 1030 /* 1031 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1032 * valid for vnode-backed buffers. 1033 */ 1034 bsetrunningbufspace(bp, bp->b_bufsize); 1035 vn_strategy(bp->b_vp, &bp->b_bio1); 1036 error = biowait(&bp->b_bio1, "biows"); 1037 brelse(bp); 1038 1039 return (error); 1040 } 1041 1042 /* 1043 * bawrite: 1044 * 1045 * Asynchronous write. Start output on a buffer, but do not wait for 1046 * it to complete. The buffer is released when the output completes. 1047 * 1048 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1049 * B_INVAL buffers. Not us. 1050 */ 1051 void 1052 bawrite(struct buf *bp) 1053 { 1054 if (bp->b_flags & B_INVAL) { 1055 brelse(bp); 1056 return; 1057 } 1058 if (BUF_REFCNTNB(bp) == 0) 1059 panic("bwrite: buffer is not busy???"); 1060 1061 /* Mark the buffer clean */ 1062 bundirty(bp); 1063 1064 bp->b_flags &= ~(B_ERROR | B_EINTR); 1065 bp->b_flags |= B_CACHE; 1066 bp->b_cmd = BUF_CMD_WRITE; 1067 KKASSERT(bp->b_bio1.bio_done == NULL); 1068 vfs_busy_pages(bp->b_vp, bp); 1069 1070 /* 1071 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 1072 * valid for vnode-backed buffers. 1073 */ 1074 bsetrunningbufspace(bp, bp->b_bufsize); 1075 BUF_KERNPROC(bp); 1076 vn_strategy(bp->b_vp, &bp->b_bio1); 1077 } 1078 1079 /* 1080 * bowrite: 1081 * 1082 * Ordered write. Start output on a buffer, and flag it so that the 1083 * device will write it in the order it was queued. The buffer is 1084 * released when the output completes. bwrite() ( or the VOP routine 1085 * anyway ) is responsible for handling B_INVAL buffers. 1086 */ 1087 int 1088 bowrite(struct buf *bp) 1089 { 1090 bp->b_flags |= B_ORDERED; 1091 bawrite(bp); 1092 return (0); 1093 } 1094 1095 /* 1096 * bdwrite: 1097 * 1098 * Delayed write. (Buffer is marked dirty). Do not bother writing 1099 * anything if the buffer is marked invalid. 1100 * 1101 * Note that since the buffer must be completely valid, we can safely 1102 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1103 * biodone() in order to prevent getblk from writing the buffer 1104 * out synchronously. 1105 */ 1106 void 1107 bdwrite(struct buf *bp) 1108 { 1109 if (BUF_REFCNTNB(bp) == 0) 1110 panic("bdwrite: buffer is not busy"); 1111 1112 if (bp->b_flags & B_INVAL) { 1113 brelse(bp); 1114 return; 1115 } 1116 bdirty(bp); 1117 1118 if (dsched_is_clear_buf_priv(bp)) 1119 dsched_new_buf(bp); 1120 1121 /* 1122 * Set B_CACHE, indicating that the buffer is fully valid. This is 1123 * true even of NFS now. 1124 */ 1125 bp->b_flags |= B_CACHE; 1126 1127 /* 1128 * This bmap keeps the system from needing to do the bmap later, 1129 * perhaps when the system is attempting to do a sync. Since it 1130 * is likely that the indirect block -- or whatever other datastructure 1131 * that the filesystem needs is still in memory now, it is a good 1132 * thing to do this. Note also, that if the pageout daemon is 1133 * requesting a sync -- there might not be enough memory to do 1134 * the bmap then... So, this is important to do. 1135 */ 1136 if (bp->b_bio2.bio_offset == NOOFFSET) { 1137 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1138 NULL, NULL, BUF_CMD_WRITE); 1139 } 1140 1141 /* 1142 * Because the underlying pages may still be mapped and 1143 * writable trying to set the dirty buffer (b_dirtyoff/end) 1144 * range here will be inaccurate. 1145 * 1146 * However, we must still clean the pages to satisfy the 1147 * vnode_pager and pageout daemon, so theythink the pages 1148 * have been "cleaned". What has really occured is that 1149 * they've been earmarked for later writing by the buffer 1150 * cache. 1151 * 1152 * So we get the b_dirtyoff/end update but will not actually 1153 * depend on it (NFS that is) until the pages are busied for 1154 * writing later on. 1155 */ 1156 vfs_clean_pages(bp); 1157 bqrelse(bp); 1158 1159 /* 1160 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1161 * due to the softdep code. 1162 */ 1163 } 1164 1165 /* 1166 * Fake write - return pages to VM system as dirty, leave the buffer clean. 1167 * This is used by tmpfs. 1168 * 1169 * It is important for any VFS using this routine to NOT use it for 1170 * IO_SYNC or IO_ASYNC operations which occur when the system really 1171 * wants to flush VM pages to backing store. 1172 */ 1173 void 1174 buwrite(struct buf *bp) 1175 { 1176 vm_page_t m; 1177 int i; 1178 1179 /* 1180 * Only works for VMIO buffers. If the buffer is already 1181 * marked for delayed-write we can't avoid the bdwrite(). 1182 */ 1183 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) { 1184 bdwrite(bp); 1185 return; 1186 } 1187 1188 /* 1189 * Mark as needing a commit. 1190 */ 1191 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1192 m = bp->b_xio.xio_pages[i]; 1193 vm_page_need_commit(m); 1194 } 1195 bqrelse(bp); 1196 } 1197 1198 /* 1199 * bdirty: 1200 * 1201 * Turn buffer into delayed write request by marking it B_DELWRI. 1202 * B_RELBUF and B_NOCACHE must be cleared. 1203 * 1204 * We reassign the buffer to itself to properly update it in the 1205 * dirty/clean lists. 1206 * 1207 * Must be called from a critical section. 1208 * The buffer must be on BQUEUE_NONE. 1209 */ 1210 void 1211 bdirty(struct buf *bp) 1212 { 1213 KASSERT(bp->b_qindex == BQUEUE_NONE, 1214 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1215 if (bp->b_flags & B_NOCACHE) { 1216 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 1217 bp->b_flags &= ~B_NOCACHE; 1218 } 1219 if (bp->b_flags & B_INVAL) { 1220 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp); 1221 } 1222 bp->b_flags &= ~B_RELBUF; 1223 1224 if ((bp->b_flags & B_DELWRI) == 0) { 1225 lwkt_gettoken(&bp->b_vp->v_token); 1226 bp->b_flags |= B_DELWRI; 1227 reassignbuf(bp); 1228 lwkt_reltoken(&bp->b_vp->v_token); 1229 1230 atomic_add_long(&dirtybufcount, 1); 1231 atomic_add_long(&dirtykvaspace, bp->b_kvasize); 1232 atomic_add_long(&dirtybufspace, bp->b_bufsize); 1233 if (bp->b_flags & B_HEAVY) { 1234 atomic_add_long(&dirtybufcounthw, 1); 1235 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1236 } 1237 bd_heatup(); 1238 } 1239 } 1240 1241 /* 1242 * Set B_HEAVY, indicating that this is a heavy-weight buffer that 1243 * needs to be flushed with a different buf_daemon thread to avoid 1244 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf(). 1245 */ 1246 void 1247 bheavy(struct buf *bp) 1248 { 1249 if ((bp->b_flags & B_HEAVY) == 0) { 1250 bp->b_flags |= B_HEAVY; 1251 if (bp->b_flags & B_DELWRI) { 1252 atomic_add_long(&dirtybufcounthw, 1); 1253 atomic_add_long(&dirtybufspacehw, bp->b_bufsize); 1254 } 1255 } 1256 } 1257 1258 /* 1259 * bundirty: 1260 * 1261 * Clear B_DELWRI for buffer. 1262 * 1263 * Must be called from a critical section. 1264 * 1265 * The buffer is typically on BQUEUE_NONE but there is one case in 1266 * brelse() that calls this function after placing the buffer on 1267 * a different queue. 1268 */ 1269 void 1270 bundirty(struct buf *bp) 1271 { 1272 if (bp->b_flags & B_DELWRI) { 1273 lwkt_gettoken(&bp->b_vp->v_token); 1274 bp->b_flags &= ~B_DELWRI; 1275 reassignbuf(bp); 1276 lwkt_reltoken(&bp->b_vp->v_token); 1277 1278 atomic_add_long(&dirtybufcount, -1); 1279 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1280 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1281 if (bp->b_flags & B_HEAVY) { 1282 atomic_add_long(&dirtybufcounthw, -1); 1283 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize); 1284 } 1285 bd_signal(bp->b_bufsize); 1286 } 1287 /* 1288 * Since it is now being written, we can clear its deferred write flag. 1289 */ 1290 bp->b_flags &= ~B_DEFERRED; 1291 } 1292 1293 /* 1294 * Set the b_runningbufspace field, used to track how much I/O is 1295 * in progress at any given moment. 1296 */ 1297 void 1298 bsetrunningbufspace(struct buf *bp, int bytes) 1299 { 1300 bp->b_runningbufspace = bytes; 1301 if (bytes) { 1302 atomic_add_long(&runningbufspace, bytes); 1303 atomic_add_long(&runningbufcount, 1); 1304 } 1305 } 1306 1307 /* 1308 * brelse: 1309 * 1310 * Release a busy buffer and, if requested, free its resources. The 1311 * buffer will be stashed in the appropriate bufqueue[] allowing it 1312 * to be accessed later as a cache entity or reused for other purposes. 1313 */ 1314 void 1315 brelse(struct buf *bp) 1316 { 1317 struct bufpcpu *pcpu; 1318 #ifdef INVARIANTS 1319 int saved_flags = bp->b_flags; 1320 #endif 1321 1322 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1323 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1324 1325 /* 1326 * If B_NOCACHE is set we are being asked to destroy the buffer and 1327 * its backing store. Clear B_DELWRI. 1328 * 1329 * B_NOCACHE is set in two cases: (1) when the caller really wants 1330 * to destroy the buffer and backing store and (2) when the caller 1331 * wants to destroy the buffer and backing store after a write 1332 * completes. 1333 */ 1334 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 1335 bundirty(bp); 1336 } 1337 1338 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) { 1339 /* 1340 * A re-dirtied buffer is only subject to destruction 1341 * by B_INVAL. B_ERROR and B_NOCACHE are ignored. 1342 */ 1343 /* leave buffer intact */ 1344 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1345 (bp->b_bufsize <= 0)) { 1346 /* 1347 * Either a failed read or we were asked to free or not 1348 * cache the buffer. This path is reached with B_DELWRI 1349 * set only if B_INVAL is already set. B_NOCACHE governs 1350 * backing store destruction. 1351 * 1352 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the 1353 * buffer cannot be immediately freed. 1354 */ 1355 bp->b_flags |= B_INVAL; 1356 if (LIST_FIRST(&bp->b_dep) != NULL) 1357 buf_deallocate(bp); 1358 if (bp->b_flags & B_DELWRI) { 1359 atomic_add_long(&dirtybufcount, -1); 1360 atomic_add_long(&dirtykvaspace, -bp->b_kvasize); 1361 atomic_add_long(&dirtybufspace, -bp->b_bufsize); 1362 if (bp->b_flags & B_HEAVY) { 1363 atomic_add_long(&dirtybufcounthw, -1); 1364 atomic_add_long(&dirtybufspacehw, 1365 -bp->b_bufsize); 1366 } 1367 bd_signal(bp->b_bufsize); 1368 } 1369 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1370 } 1371 1372 /* 1373 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set, 1374 * or if b_refs is non-zero. 1375 * 1376 * If vfs_vmio_release() is called with either bit set, the 1377 * underlying pages may wind up getting freed causing a previous 1378 * write (bdwrite()) to get 'lost' because pages associated with 1379 * a B_DELWRI bp are marked clean. Pages associated with a 1380 * B_LOCKED buffer may be mapped by the filesystem. 1381 * 1382 * If we want to release the buffer ourselves (rather then the 1383 * originator asking us to release it), give the originator a 1384 * chance to countermand the release by setting B_LOCKED. 1385 * 1386 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1387 * if B_DELWRI is set. 1388 * 1389 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1390 * on pages to return pages to the VM page queues. 1391 */ 1392 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) { 1393 bp->b_flags &= ~B_RELBUF; 1394 } else if (vm_page_count_min(0)) { 1395 if (LIST_FIRST(&bp->b_dep) != NULL) 1396 buf_deallocate(bp); /* can set B_LOCKED */ 1397 if (bp->b_flags & (B_DELWRI | B_LOCKED)) 1398 bp->b_flags &= ~B_RELBUF; 1399 else 1400 bp->b_flags |= B_RELBUF; 1401 } 1402 1403 /* 1404 * Make sure b_cmd is clear. It may have already been cleared by 1405 * biodone(). 1406 * 1407 * At this point destroying the buffer is governed by the B_INVAL 1408 * or B_RELBUF flags. 1409 */ 1410 bp->b_cmd = BUF_CMD_DONE; 1411 dsched_exit_buf(bp); 1412 1413 /* 1414 * VMIO buffer rundown. Make sure the VM page array is restored 1415 * after an I/O may have replaces some of the pages with bogus pages 1416 * in order to not destroy dirty pages in a fill-in read. 1417 * 1418 * Note that due to the code above, if a buffer is marked B_DELWRI 1419 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1420 * B_INVAL may still be set, however. 1421 * 1422 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1423 * but not the backing store. B_NOCACHE will destroy the backing 1424 * store. 1425 * 1426 * Note that dirty NFS buffers contain byte-granular write ranges 1427 * and should not be destroyed w/ B_INVAL even if the backing store 1428 * is left intact. 1429 */ 1430 if (bp->b_flags & B_VMIO) { 1431 /* 1432 * Rundown for VMIO buffers which are not dirty NFS buffers. 1433 */ 1434 int i, j, resid; 1435 vm_page_t m; 1436 off_t foff; 1437 vm_pindex_t poff; 1438 vm_object_t obj; 1439 struct vnode *vp; 1440 1441 vp = bp->b_vp; 1442 1443 /* 1444 * Get the base offset and length of the buffer. Note that 1445 * in the VMIO case if the buffer block size is not 1446 * page-aligned then b_data pointer may not be page-aligned. 1447 * But our b_xio.xio_pages array *IS* page aligned. 1448 * 1449 * block sizes less then DEV_BSIZE (usually 512) are not 1450 * supported due to the page granularity bits (m->valid, 1451 * m->dirty, etc...). 1452 * 1453 * See man buf(9) for more information 1454 */ 1455 1456 resid = bp->b_bufsize; 1457 foff = bp->b_loffset; 1458 1459 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1460 m = bp->b_xio.xio_pages[i]; 1461 vm_page_flag_clear(m, PG_ZERO); 1462 /* 1463 * If we hit a bogus page, fixup *all* of them 1464 * now. Note that we left these pages wired 1465 * when we removed them so they had better exist, 1466 * and they cannot be ripped out from under us so 1467 * no critical section protection is necessary. 1468 */ 1469 if (m == bogus_page) { 1470 obj = vp->v_object; 1471 poff = OFF_TO_IDX(bp->b_loffset); 1472 1473 vm_object_hold(obj); 1474 for (j = i; j < bp->b_xio.xio_npages; j++) { 1475 vm_page_t mtmp; 1476 1477 mtmp = bp->b_xio.xio_pages[j]; 1478 if (mtmp == bogus_page) { 1479 mtmp = vm_page_lookup(obj, poff + j); 1480 if (!mtmp) { 1481 panic("brelse: page missing"); 1482 } 1483 bp->b_xio.xio_pages[j] = mtmp; 1484 } 1485 } 1486 bp->b_flags &= ~B_HASBOGUS; 1487 vm_object_drop(obj); 1488 1489 if ((bp->b_flags & B_INVAL) == 0) { 1490 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1491 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1492 } 1493 m = bp->b_xio.xio_pages[i]; 1494 } 1495 1496 /* 1497 * Invalidate the backing store if B_NOCACHE is set 1498 * (e.g. used with vinvalbuf()). If this is NFS 1499 * we impose a requirement that the block size be 1500 * a multiple of PAGE_SIZE and create a temporary 1501 * hack to basically invalidate the whole page. The 1502 * problem is that NFS uses really odd buffer sizes 1503 * especially when tracking piecemeal writes and 1504 * it also vinvalbuf()'s a lot, which would result 1505 * in only partial page validation and invalidation 1506 * here. If the file page is mmap()'d, however, 1507 * all the valid bits get set so after we invalidate 1508 * here we would end up with weird m->valid values 1509 * like 0xfc. nfs_getpages() can't handle this so 1510 * we clear all the valid bits for the NFS case 1511 * instead of just some of them. 1512 * 1513 * The real bug is the VM system having to set m->valid 1514 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1515 * itself is an artifact of the whole 512-byte 1516 * granular mess that exists to support odd block 1517 * sizes and UFS meta-data block sizes (e.g. 6144). 1518 * A complete rewrite is required. 1519 * 1520 * XXX 1521 */ 1522 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1523 int poffset = foff & PAGE_MASK; 1524 int presid; 1525 1526 presid = PAGE_SIZE - poffset; 1527 if (bp->b_vp->v_tag == VT_NFS && 1528 bp->b_vp->v_type == VREG) { 1529 ; /* entire page */ 1530 } else if (presid > resid) { 1531 presid = resid; 1532 } 1533 KASSERT(presid >= 0, ("brelse: extra page")); 1534 vm_page_set_invalid(m, poffset, presid); 1535 1536 /* 1537 * Also make sure any swap cache is removed 1538 * as it is now stale (HAMMER in particular 1539 * uses B_NOCACHE to deal with buffer 1540 * aliasing). 1541 */ 1542 swap_pager_unswapped(m); 1543 } 1544 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1545 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1546 } 1547 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1548 vfs_vmio_release(bp); 1549 } else { 1550 /* 1551 * Rundown for non-VMIO buffers. 1552 */ 1553 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1554 if (bp->b_bufsize) 1555 allocbuf(bp, 0); 1556 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1557 if (bp->b_vp) 1558 brelvp(bp); 1559 } 1560 } 1561 1562 if (bp->b_qindex != BQUEUE_NONE) 1563 panic("brelse: free buffer onto another queue???"); 1564 if (BUF_REFCNTNB(bp) > 1) { 1565 /* Temporary panic to verify exclusive locking */ 1566 /* This panic goes away when we allow shared refs */ 1567 panic("brelse: multiple refs"); 1568 /* NOT REACHED */ 1569 return; 1570 } 1571 1572 /* 1573 * Figure out the correct queue to place the cleaned up buffer on. 1574 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1575 * disassociated from their vnode. 1576 * 1577 * Return the buffer to its original pcpu area 1578 */ 1579 pcpu = &bufpcpu[bp->b_qcpu]; 1580 spin_lock(&pcpu->spin); 1581 1582 if (bp->b_flags & B_LOCKED) { 1583 /* 1584 * Buffers that are locked are placed in the locked queue 1585 * immediately, regardless of their state. 1586 */ 1587 bp->b_qindex = BQUEUE_LOCKED; 1588 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1589 bp, b_freelist); 1590 } else if (bp->b_bufsize == 0) { 1591 /* 1592 * Buffers with no memory. Due to conditionals near the top 1593 * of brelse() such buffers should probably already be 1594 * marked B_INVAL and disassociated from their vnode. 1595 */ 1596 bp->b_flags |= B_INVAL; 1597 KASSERT(bp->b_vp == NULL, 1598 ("bp1 %p flags %08x/%08x vnode %p " 1599 "unexpectededly still associated!", 1600 bp, saved_flags, bp->b_flags, bp->b_vp)); 1601 KKASSERT((bp->b_flags & B_HASHED) == 0); 1602 if (bp->b_kvasize) { 1603 bp->b_qindex = BQUEUE_EMPTYKVA; 1604 } else { 1605 bp->b_qindex = BQUEUE_EMPTY; 1606 } 1607 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1608 bp, b_freelist); 1609 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) { 1610 /* 1611 * Buffers with junk contents. Again these buffers had better 1612 * already be disassociated from their vnode. 1613 */ 1614 KASSERT(bp->b_vp == NULL, 1615 ("bp2 %p flags %08x/%08x vnode %p unexpectededly " 1616 "still associated!", 1617 bp, saved_flags, bp->b_flags, bp->b_vp)); 1618 KKASSERT((bp->b_flags & B_HASHED) == 0); 1619 bp->b_flags |= B_INVAL; 1620 bp->b_qindex = BQUEUE_CLEAN; 1621 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex], 1622 bp, b_freelist); 1623 } else { 1624 /* 1625 * Remaining buffers. These buffers are still associated with 1626 * their vnode. 1627 */ 1628 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) { 1629 case B_DELWRI: 1630 bp->b_qindex = BQUEUE_DIRTY; 1631 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1632 bp, b_freelist); 1633 break; 1634 case B_DELWRI | B_HEAVY: 1635 bp->b_qindex = BQUEUE_DIRTY_HW; 1636 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1637 bp, b_freelist); 1638 break; 1639 default: 1640 /* 1641 * NOTE: Buffers are always placed at the end of the 1642 * queue. If B_AGE is not set the buffer will cycle 1643 * through the queue twice. 1644 */ 1645 bp->b_qindex = BQUEUE_CLEAN; 1646 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1647 bp, b_freelist); 1648 break; 1649 } 1650 } 1651 spin_unlock(&pcpu->spin); 1652 1653 /* 1654 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1655 * on the correct queue but we have not yet unlocked it. 1656 */ 1657 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1658 bundirty(bp); 1659 1660 /* 1661 * The bp is on an appropriate queue unless locked. If it is not 1662 * locked or dirty we can wakeup threads waiting for buffer space. 1663 * 1664 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1665 * if B_INVAL is set ). 1666 */ 1667 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) 1668 bufcountwakeup(); 1669 1670 /* 1671 * Something we can maybe free or reuse 1672 */ 1673 if (bp->b_bufsize || bp->b_kvasize) 1674 bufspacewakeup(); 1675 1676 /* 1677 * Clean up temporary flags and unlock the buffer. 1678 */ 1679 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT); 1680 BUF_UNLOCK(bp); 1681 } 1682 1683 /* 1684 * bqrelse: 1685 * 1686 * Release a buffer back to the appropriate queue but do not try to free 1687 * it. The buffer is expected to be used again soon. 1688 * 1689 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1690 * biodone() to requeue an async I/O on completion. It is also used when 1691 * known good buffers need to be requeued but we think we may need the data 1692 * again soon. 1693 * 1694 * XXX we should be able to leave the B_RELBUF hint set on completion. 1695 */ 1696 void 1697 bqrelse(struct buf *bp) 1698 { 1699 struct bufpcpu *pcpu; 1700 1701 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1702 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1703 1704 if (bp->b_qindex != BQUEUE_NONE) 1705 panic("bqrelse: free buffer onto another queue???"); 1706 if (BUF_REFCNTNB(bp) > 1) { 1707 /* do not release to free list */ 1708 panic("bqrelse: multiple refs"); 1709 return; 1710 } 1711 1712 buf_act_advance(bp); 1713 1714 pcpu = &bufpcpu[bp->b_qcpu]; 1715 spin_lock(&pcpu->spin); 1716 1717 if (bp->b_flags & B_LOCKED) { 1718 /* 1719 * Locked buffers are released to the locked queue. However, 1720 * if the buffer is dirty it will first go into the dirty 1721 * queue and later on after the I/O completes successfully it 1722 * will be released to the locked queue. 1723 */ 1724 bp->b_qindex = BQUEUE_LOCKED; 1725 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1726 bp, b_freelist); 1727 } else if (bp->b_flags & B_DELWRI) { 1728 bp->b_qindex = (bp->b_flags & B_HEAVY) ? 1729 BQUEUE_DIRTY_HW : BQUEUE_DIRTY; 1730 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1731 bp, b_freelist); 1732 } else if (vm_page_count_min(0)) { 1733 /* 1734 * We are too low on memory, we have to try to free the 1735 * buffer (most importantly: the wired pages making up its 1736 * backing store) *now*. 1737 */ 1738 spin_unlock(&pcpu->spin); 1739 brelse(bp); 1740 return; 1741 } else { 1742 bp->b_qindex = BQUEUE_CLEAN; 1743 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex], 1744 bp, b_freelist); 1745 } 1746 spin_unlock(&pcpu->spin); 1747 1748 /* 1749 * We have now placed the buffer on the proper queue, but have yet 1750 * to unlock it. 1751 */ 1752 if ((bp->b_flags & B_LOCKED) == 0 && 1753 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) { 1754 bufcountwakeup(); 1755 } 1756 1757 /* 1758 * Something we can maybe free or reuse. 1759 */ 1760 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1761 bufspacewakeup(); 1762 1763 /* 1764 * Final cleanup and unlock. Clear bits that are only used while a 1765 * buffer is actively locked. 1766 */ 1767 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF); 1768 dsched_exit_buf(bp); 1769 BUF_UNLOCK(bp); 1770 } 1771 1772 /* 1773 * Hold a buffer, preventing it from being reused. This will prevent 1774 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL 1775 * operations. If a B_INVAL operation occurs the buffer will remain held 1776 * but the underlying pages may get ripped out. 1777 * 1778 * These functions are typically used in VOP_READ/VOP_WRITE functions 1779 * to hold a buffer during a copyin or copyout, preventing deadlocks 1780 * or recursive lock panics when read()/write() is used over mmap()'d 1781 * space. 1782 * 1783 * NOTE: bqhold() requires that the buffer be locked at the time of the 1784 * hold. bqdrop() has no requirements other than the buffer having 1785 * previously been held. 1786 */ 1787 void 1788 bqhold(struct buf *bp) 1789 { 1790 atomic_add_int(&bp->b_refs, 1); 1791 } 1792 1793 void 1794 bqdrop(struct buf *bp) 1795 { 1796 KKASSERT(bp->b_refs > 0); 1797 atomic_add_int(&bp->b_refs, -1); 1798 } 1799 1800 /* 1801 * Return backing pages held by the buffer 'bp' back to the VM system. 1802 * This routine is called when the bp is invalidated, released, or 1803 * reused. 1804 * 1805 * The KVA mapping (b_data) for the underlying pages is removed by 1806 * this function. 1807 * 1808 * WARNING! This routine is integral to the low memory critical path 1809 * when a buffer is B_RELBUF'd. If the system has a severe page 1810 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE 1811 * queues so they can be reused in the current pageout daemon 1812 * pass. 1813 */ 1814 static void 1815 vfs_vmio_release(struct buf *bp) 1816 { 1817 int i; 1818 vm_page_t m; 1819 1820 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1821 m = bp->b_xio.xio_pages[i]; 1822 bp->b_xio.xio_pages[i] = NULL; 1823 1824 /* 1825 * We need to own the page in order to safely unwire it. 1826 */ 1827 vm_page_busy_wait(m, FALSE, "vmiopg"); 1828 1829 /* 1830 * The VFS is telling us this is not a meta-data buffer 1831 * even if it is backed by a block device. 1832 */ 1833 if (bp->b_flags & B_NOTMETA) 1834 vm_page_flag_set(m, PG_NOTMETA); 1835 1836 /* 1837 * This is a very important bit of code. We try to track 1838 * VM page use whether the pages are wired into the buffer 1839 * cache or not. While wired into the buffer cache the 1840 * bp tracks the act_count. 1841 * 1842 * We can choose to place unwired pages on the inactive 1843 * queue (0) or active queue (1). If we place too many 1844 * on the active queue the queue will cycle the act_count 1845 * on pages we'd like to keep, just from single-use pages 1846 * (such as when doing a tar-up or file scan). 1847 */ 1848 if (bp->b_act_count < vm_cycle_point) 1849 vm_page_unwire(m, 0); 1850 else 1851 vm_page_unwire(m, 1); 1852 1853 /* 1854 * If the wire_count has dropped to 0 we may need to take 1855 * further action before unbusying the page. 1856 * 1857 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us. 1858 */ 1859 if (m->wire_count == 0) { 1860 vm_page_flag_clear(m, PG_ZERO); 1861 1862 if (bp->b_flags & B_DIRECT) { 1863 /* 1864 * Attempt to free the page if B_DIRECT is 1865 * set, the caller does not desire the page 1866 * to be cached. 1867 */ 1868 vm_page_wakeup(m); 1869 vm_page_try_to_free(m); 1870 } else if ((bp->b_flags & B_NOTMETA) || 1871 vm_page_count_min(0)) { 1872 /* 1873 * Attempt to move the page to PQ_CACHE 1874 * if B_NOTMETA is set. This flag is set 1875 * by HAMMER to remove one of the two pages 1876 * present when double buffering is enabled. 1877 * 1878 * Attempt to move the page to PQ_CACHE 1879 * If we have a severe page deficit. This 1880 * will cause buffer cache operations related 1881 * to pageouts to recycle the related pages 1882 * in order to avoid a low memory deadlock. 1883 */ 1884 m->act_count = bp->b_act_count; 1885 vm_page_wakeup(m); 1886 vm_page_try_to_cache(m); 1887 } else { 1888 /* 1889 * Nominal case, leave the page on the 1890 * queue the original unwiring placed it on 1891 * (active or inactive). 1892 */ 1893 m->act_count = bp->b_act_count; 1894 vm_page_wakeup(m); 1895 } 1896 } else { 1897 vm_page_wakeup(m); 1898 } 1899 } 1900 1901 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1902 bp->b_xio.xio_npages); 1903 if (bp->b_bufsize) { 1904 bufspacewakeup(); 1905 bp->b_bufsize = 0; 1906 } 1907 bp->b_xio.xio_npages = 0; 1908 bp->b_flags &= ~B_VMIO; 1909 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL); 1910 if (bp->b_vp) 1911 brelvp(bp); 1912 } 1913 1914 /* 1915 * Find and initialize a new buffer header, freeing up existing buffers 1916 * in the bufqueues as necessary. The new buffer is returned locked. 1917 * 1918 * Important: B_INVAL is not set. If the caller wishes to throw the 1919 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1920 * 1921 * We block if: 1922 * We have insufficient buffer headers 1923 * We have insufficient buffer space 1924 * buffer_map is too fragmented ( space reservation fails ) 1925 * If we have to flush dirty buffers ( but we try to avoid this ) 1926 * 1927 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1928 * Instead we ask the buf daemon to do it for us. We attempt to 1929 * avoid piecemeal wakeups of the pageout daemon. 1930 */ 1931 struct buf * 1932 getnewbuf(int blkflags, int slptimeo, int size, int maxsize) 1933 { 1934 struct bufpcpu *pcpu; 1935 struct buf *bp; 1936 struct buf *nbp; 1937 int defrag = 0; 1938 int nqindex; 1939 int nqcpu; 1940 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 1941 int maxloops = 200000; 1942 int restart_reason = 0; 1943 struct buf *restart_bp = NULL; 1944 static int flushingbufs; 1945 1946 /* 1947 * We can't afford to block since we might be holding a vnode lock, 1948 * which may prevent system daemons from running. We deal with 1949 * low-memory situations by proactively returning memory and running 1950 * async I/O rather then sync I/O. 1951 */ 1952 1953 ++getnewbufcalls; 1954 --getnewbufrestarts; 1955 nqcpu = mycpu->gd_cpuid; 1956 restart: 1957 ++getnewbufrestarts; 1958 1959 if (debug_bufbio && --maxloops == 0) 1960 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)", 1961 mycpu->gd_cpuid, restart_reason, restart_bp); 1962 1963 /* 1964 * Setup for scan. If we do not have enough free buffers, 1965 * we setup a degenerate case that immediately fails. Note 1966 * that if we are specially marked process, we are allowed to 1967 * dip into our reserves. 1968 * 1969 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1970 * 1971 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1972 * However, there are a number of cases (defragging, reusing, ...) 1973 * where we cannot backup. 1974 */ 1975 pcpu = &bufpcpu[nqcpu]; 1976 nqindex = BQUEUE_EMPTYKVA; 1977 spin_lock(&pcpu->spin); 1978 1979 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]); 1980 1981 if (nbp == NULL) { 1982 /* 1983 * If no EMPTYKVA buffers and we are either 1984 * defragging or reusing, locate a CLEAN buffer 1985 * to free or reuse. If bufspace useage is low 1986 * skip this step so we can allocate a new buffer. 1987 */ 1988 if (defrag || bufspace >= lobufspace) { 1989 nqindex = BQUEUE_CLEAN; 1990 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]); 1991 } 1992 1993 /* 1994 * If we could not find or were not allowed to reuse a 1995 * CLEAN buffer, check to see if it is ok to use an EMPTY 1996 * buffer. We can only use an EMPTY buffer if allocating 1997 * its KVA would not otherwise run us out of buffer space. 1998 */ 1999 if (nbp == NULL && defrag == 0 && 2000 bufspace + maxsize < hibufspace) { 2001 nqindex = BQUEUE_EMPTY; 2002 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]); 2003 } 2004 } 2005 2006 /* 2007 * Run scan, possibly freeing data and/or kva mappings on the fly 2008 * depending. 2009 * 2010 * WARNING! spin is held! 2011 */ 2012 while ((bp = nbp) != NULL) { 2013 int qindex = nqindex; 2014 2015 nbp = TAILQ_NEXT(bp, b_freelist); 2016 2017 /* 2018 * BQUEUE_CLEAN - B_AGE special case. If not set the bp 2019 * cycles through the queue twice before being selected. 2020 */ 2021 if (qindex == BQUEUE_CLEAN && 2022 (bp->b_flags & B_AGE) == 0 && nbp) { 2023 bp->b_flags |= B_AGE; 2024 TAILQ_REMOVE(&pcpu->bufqueues[qindex], 2025 bp, b_freelist); 2026 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex], 2027 bp, b_freelist); 2028 continue; 2029 } 2030 2031 /* 2032 * Calculate next bp ( we can only use it if we do not block 2033 * or do other fancy things ). 2034 */ 2035 if (nbp == NULL) { 2036 switch(qindex) { 2037 case BQUEUE_EMPTY: 2038 nqindex = BQUEUE_EMPTYKVA; 2039 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]))) 2040 break; 2041 /* fall through */ 2042 case BQUEUE_EMPTYKVA: 2043 nqindex = BQUEUE_CLEAN; 2044 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]))) 2045 break; 2046 /* fall through */ 2047 case BQUEUE_CLEAN: 2048 /* 2049 * nbp is NULL. 2050 */ 2051 break; 2052 } 2053 } 2054 2055 /* 2056 * Sanity Checks 2057 */ 2058 KASSERT(bp->b_qindex == qindex, 2059 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2060 2061 /* 2062 * Note: we no longer distinguish between VMIO and non-VMIO 2063 * buffers. 2064 */ 2065 KASSERT((bp->b_flags & B_DELWRI) == 0, 2066 ("delwri buffer %p found in queue %d", bp, qindex)); 2067 2068 /* 2069 * Do not try to reuse a buffer with a non-zero b_refs. 2070 * This is an unsynchronized test. A synchronized test 2071 * is also performed after we lock the buffer. 2072 */ 2073 if (bp->b_refs) 2074 continue; 2075 2076 /* 2077 * If we are defragging then we need a buffer with 2078 * b_kvasize != 0. XXX this situation should no longer 2079 * occur, if defrag is non-zero the buffer's b_kvasize 2080 * should also be non-zero at this point. XXX 2081 */ 2082 if (defrag && bp->b_kvasize == 0) { 2083 kprintf("Warning: defrag empty buffer %p\n", bp); 2084 continue; 2085 } 2086 2087 /* 2088 * Start freeing the bp. This is somewhat involved. nbp 2089 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 2090 * on the clean list must be disassociated from their 2091 * current vnode. Buffers on the empty[kva] lists have 2092 * already been disassociated. 2093 * 2094 * b_refs is checked after locking along with queue changes. 2095 * We must check here to deal with zero->nonzero transitions 2096 * made by the owner of the buffer lock, which is used by 2097 * VFS's to hold the buffer while issuing an unlocked 2098 * uiomove()s. We cannot invalidate the buffer's pages 2099 * for this case. Once we successfully lock a buffer the 2100 * only 0->1 transitions of b_refs will occur via findblk(). 2101 * 2102 * We must also check for queue changes after successful 2103 * locking as the current lock holder may dispose of the 2104 * buffer and change its queue. 2105 */ 2106 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 2107 spin_unlock(&pcpu->spin); 2108 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100); 2109 restart_reason = 1; 2110 restart_bp = bp; 2111 goto restart; 2112 } 2113 if (bp->b_qindex != qindex || bp->b_refs) { 2114 spin_unlock(&pcpu->spin); 2115 BUF_UNLOCK(bp); 2116 restart_reason = 2; 2117 restart_bp = bp; 2118 goto restart; 2119 } 2120 bremfree_locked(bp); 2121 spin_unlock(&pcpu->spin); 2122 2123 /* 2124 * Dependancies must be handled before we disassociate the 2125 * vnode. 2126 * 2127 * NOTE: HAMMER will set B_LOCKED if the buffer cannot 2128 * be immediately disassociated. HAMMER then becomes 2129 * responsible for releasing the buffer. 2130 * 2131 * NOTE: spin is UNLOCKED now. 2132 */ 2133 if (LIST_FIRST(&bp->b_dep) != NULL) { 2134 buf_deallocate(bp); 2135 if (bp->b_flags & B_LOCKED) { 2136 bqrelse(bp); 2137 restart_reason = 3; 2138 restart_bp = bp; 2139 goto restart; 2140 } 2141 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2142 } 2143 2144 if (qindex == BQUEUE_CLEAN) { 2145 if (bp->b_flags & B_VMIO) 2146 vfs_vmio_release(bp); 2147 if (bp->b_vp) 2148 brelvp(bp); 2149 } 2150 2151 /* 2152 * NOTE: nbp is now entirely invalid. We can only restart 2153 * the scan from this point on. 2154 * 2155 * Get the rest of the buffer freed up. b_kva* is still 2156 * valid after this operation. 2157 */ 2158 KASSERT(bp->b_vp == NULL, 2159 ("bp3 %p flags %08x vnode %p qindex %d " 2160 "unexpectededly still associated!", 2161 bp, bp->b_flags, bp->b_vp, qindex)); 2162 KKASSERT((bp->b_flags & B_HASHED) == 0); 2163 2164 /* 2165 * critical section protection is not required when 2166 * scrapping a buffer's contents because it is already 2167 * wired. 2168 */ 2169 if (bp->b_bufsize) 2170 allocbuf(bp, 0); 2171 2172 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) { 2173 kprintf("getnewbuf: caught bug vp queue " 2174 "%p/%08x qidx %d\n", 2175 bp, bp->b_flags, qindex); 2176 brelvp(bp); 2177 } 2178 bp->b_flags = B_BNOCLIP; 2179 bp->b_cmd = BUF_CMD_DONE; 2180 bp->b_vp = NULL; 2181 bp->b_error = 0; 2182 bp->b_resid = 0; 2183 bp->b_bcount = 0; 2184 bp->b_xio.xio_npages = 0; 2185 bp->b_dirtyoff = bp->b_dirtyend = 0; 2186 bp->b_act_count = ACT_INIT; 2187 reinitbufbio(bp); 2188 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2189 buf_dep_init(bp); 2190 if (blkflags & GETBLK_BHEAVY) 2191 bp->b_flags |= B_HEAVY; 2192 2193 /* 2194 * If we are defragging then free the buffer. 2195 */ 2196 if (defrag) { 2197 bp->b_flags |= B_INVAL; 2198 bfreekva(bp); 2199 brelse(bp); 2200 defrag = 0; 2201 restart_reason = 4; 2202 restart_bp = bp; 2203 goto restart; 2204 } 2205 2206 /* 2207 * If we are overcomitted then recover the buffer and its 2208 * KVM space. This occurs in rare situations when multiple 2209 * processes are blocked in getnewbuf() or allocbuf(). 2210 * 2211 * On 64-bit systems BKVASIZE == MAXBSIZE and overcommit 2212 * should not be possible. 2213 */ 2214 if (bufspace >= hibufspace) 2215 flushingbufs = 1; 2216 if (BKVASIZE != MAXBSIZE) { 2217 if (flushingbufs && bp->b_kvasize != 0) { 2218 bp->b_flags |= B_INVAL; 2219 bfreekva(bp); 2220 brelse(bp); 2221 restart_reason = 5; 2222 restart_bp = bp; 2223 goto restart; 2224 } 2225 } 2226 if (bufspace < lobufspace) 2227 flushingbufs = 0; 2228 2229 /* 2230 * b_refs can transition to a non-zero value while we hold 2231 * the buffer locked due to a findblk(). Our brelvp() above 2232 * interlocked any future possible transitions due to 2233 * findblk()s. 2234 * 2235 * If we find b_refs to be non-zero we can destroy the 2236 * buffer's contents but we cannot yet reuse the buffer. 2237 */ 2238 if (bp->b_refs) { 2239 bp->b_flags |= B_INVAL; 2240 if (BKVASIZE != MAXBSIZE) 2241 bfreekva(bp); 2242 brelse(bp); 2243 restart_reason = 6; 2244 restart_bp = bp; 2245 goto restart; 2246 } 2247 break; 2248 /* NOT REACHED, spin not held */ 2249 } 2250 2251 /* 2252 * If we exhausted our list, iterate other cpus. If that fails, 2253 * sleep as appropriate. We may have to wakeup various daemons 2254 * and write out some dirty buffers. 2255 * 2256 * Generally we are sleeping due to insufficient buffer space. 2257 * 2258 * NOTE: spin is held if bp is NULL, else it is not held. 2259 */ 2260 if (bp == NULL) { 2261 int flags; 2262 char *waitmsg; 2263 2264 spin_unlock(&pcpu->spin); 2265 2266 nqcpu = (nqcpu + 1) % ncpus; 2267 if (nqcpu != mycpu->gd_cpuid) { 2268 restart_reason = 7; 2269 restart_bp = bp; 2270 goto restart; 2271 } 2272 2273 if (defrag) { 2274 flags = VFS_BIO_NEED_BUFSPACE; 2275 waitmsg = "nbufkv"; 2276 } else if (bufspace >= hibufspace) { 2277 waitmsg = "nbufbs"; 2278 flags = VFS_BIO_NEED_BUFSPACE; 2279 } else { 2280 waitmsg = "newbuf"; 2281 flags = VFS_BIO_NEED_ANY; 2282 } 2283 2284 bd_speedup(); /* heeeelp */ 2285 atomic_set_int(&needsbuffer, flags); 2286 while (needsbuffer & flags) { 2287 int value; 2288 2289 tsleep_interlock(&needsbuffer, 0); 2290 value = atomic_fetchadd_int(&needsbuffer, 0); 2291 if (value & flags) { 2292 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags, 2293 waitmsg, slptimeo)) { 2294 return (NULL); 2295 } 2296 } 2297 } 2298 } else { 2299 /* 2300 * We finally have a valid bp. We aren't quite out of the 2301 * woods, we still have to reserve kva space. In order 2302 * to keep fragmentation sane we only allocate kva in 2303 * BKVASIZE chunks. 2304 * 2305 * (spin is not held) 2306 */ 2307 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2308 2309 if (maxsize != bp->b_kvasize) { 2310 vm_offset_t addr = 0; 2311 int count; 2312 2313 bfreekva(bp); 2314 2315 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2316 vm_map_lock(&buffer_map); 2317 2318 if (vm_map_findspace(&buffer_map, 2319 vm_map_min(&buffer_map), maxsize, 2320 maxsize, 0, &addr)) { 2321 /* 2322 * Uh oh. Buffer map is too fragmented. We 2323 * must defragment the map. 2324 */ 2325 vm_map_unlock(&buffer_map); 2326 vm_map_entry_release(count); 2327 ++bufdefragcnt; 2328 defrag = 1; 2329 bp->b_flags |= B_INVAL; 2330 brelse(bp); 2331 restart_reason = 8; 2332 restart_bp = bp; 2333 goto restart; 2334 } 2335 if (addr) { 2336 vm_map_insert(&buffer_map, &count, 2337 NULL, 0, 2338 addr, addr + maxsize, 2339 VM_MAPTYPE_NORMAL, 2340 VM_PROT_ALL, VM_PROT_ALL, 2341 MAP_NOFAULT); 2342 2343 bp->b_kvabase = (caddr_t) addr; 2344 bp->b_kvasize = maxsize; 2345 bufspace += bp->b_kvasize; 2346 ++bufreusecnt; 2347 } 2348 vm_map_unlock(&buffer_map); 2349 vm_map_entry_release(count); 2350 } 2351 bp->b_data = bp->b_kvabase; 2352 } 2353 return(bp); 2354 } 2355 2356 /* 2357 * buf_daemon: 2358 * 2359 * Buffer flushing daemon. Buffers are normally flushed by the 2360 * update daemon but if it cannot keep up this process starts to 2361 * take the load in an attempt to prevent getnewbuf() from blocking. 2362 * 2363 * Once a flush is initiated it does not stop until the number 2364 * of buffers falls below lodirtybuffers, but we will wake up anyone 2365 * waiting at the mid-point. 2366 */ 2367 static struct kproc_desc buf_kp = { 2368 "bufdaemon", 2369 buf_daemon, 2370 &bufdaemon_td 2371 }; 2372 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2373 kproc_start, &buf_kp) 2374 2375 static struct kproc_desc bufhw_kp = { 2376 "bufdaemon_hw", 2377 buf_daemon_hw, 2378 &bufdaemonhw_td 2379 }; 2380 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2381 kproc_start, &bufhw_kp) 2382 2383 static void 2384 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long), 2385 int *bd_req) 2386 { 2387 long limit; 2388 struct buf *marker; 2389 2390 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO); 2391 marker->b_flags |= B_MARKER; 2392 marker->b_qindex = BQUEUE_NONE; 2393 marker->b_qcpu = 0; 2394 2395 /* 2396 * This process needs to be suspended prior to shutdown sync. 2397 */ 2398 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2399 td, SHUTDOWN_PRI_LAST); 2400 curthread->td_flags |= TDF_SYSTHREAD; 2401 2402 /* 2403 * This process is allowed to take the buffer cache to the limit 2404 */ 2405 for (;;) { 2406 kproc_suspend_loop(); 2407 2408 /* 2409 * Do the flush as long as the number of dirty buffers 2410 * (including those running) exceeds lodirtybufspace. 2411 * 2412 * When flushing limit running I/O to hirunningspace 2413 * Do the flush. Limit the amount of in-transit I/O we 2414 * allow to build up, otherwise we would completely saturate 2415 * the I/O system. Wakeup any waiting processes before we 2416 * normally would so they can run in parallel with our drain. 2417 * 2418 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2419 * but because we split the operation into two threads we 2420 * have to cut it in half for each thread. 2421 */ 2422 waitrunningbufspace(); 2423 limit = lodirtybufspace / 2; 2424 while (buf_limit_fn(limit)) { 2425 if (flushbufqueues(marker, queue) == 0) 2426 break; 2427 if (runningbufspace < hirunningspace) 2428 continue; 2429 waitrunningbufspace(); 2430 } 2431 2432 /* 2433 * We reached our low water mark, reset the 2434 * request and sleep until we are needed again. 2435 * The sleep is just so the suspend code works. 2436 */ 2437 tsleep_interlock(bd_req, 0); 2438 if (atomic_swap_int(bd_req, 0) == 0) 2439 tsleep(bd_req, PINTERLOCKED, "psleep", hz); 2440 } 2441 /* NOT REACHED */ 2442 /*kfree(marker, M_BIOBUF);*/ 2443 } 2444 2445 static int 2446 buf_daemon_limit(long limit) 2447 { 2448 return (runningbufspace + dirtykvaspace > limit || 2449 dirtybufcount - dirtybufcounthw >= nbuf / 2); 2450 } 2451 2452 static int 2453 buf_daemon_hw_limit(long limit) 2454 { 2455 return (runningbufspace + dirtykvaspace > limit || 2456 dirtybufcounthw >= nbuf / 2); 2457 } 2458 2459 static void 2460 buf_daemon(void) 2461 { 2462 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit, 2463 &bd_request); 2464 } 2465 2466 static void 2467 buf_daemon_hw(void) 2468 { 2469 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit, 2470 &bd_request_hw); 2471 } 2472 2473 /* 2474 * flushbufqueues: 2475 * 2476 * Try to flush a buffer in the dirty queue. We must be careful to 2477 * free up B_INVAL buffers instead of write them, which NFS is 2478 * particularly sensitive to. 2479 * 2480 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate 2481 * that we really want to try to get the buffer out and reuse it 2482 * due to the write load on the machine. 2483 * 2484 * We must lock the buffer in order to check its validity before we 2485 * can mess with its contents. spin isn't enough. 2486 */ 2487 static int 2488 flushbufqueues(struct buf *marker, bufq_type_t q) 2489 { 2490 struct bufpcpu *pcpu; 2491 struct buf *bp; 2492 int r = 0; 2493 int lcpu = marker->b_qcpu; 2494 2495 KKASSERT(marker->b_qindex == BQUEUE_NONE); 2496 KKASSERT(marker->b_flags & B_MARKER); 2497 2498 again: 2499 /* 2500 * Spinlock needed to perform operations on the queue and may be 2501 * held through a non-blocking BUF_LOCK(), but cannot be held when 2502 * BUF_UNLOCK()ing or through any other major operation. 2503 */ 2504 pcpu = &bufpcpu[marker->b_qcpu]; 2505 spin_lock(&pcpu->spin); 2506 marker->b_qindex = q; 2507 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist); 2508 bp = marker; 2509 2510 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) { 2511 /* 2512 * NOTE: spinlock is always held at the top of the loop 2513 */ 2514 if (bp->b_flags & B_MARKER) 2515 continue; 2516 if ((bp->b_flags & B_DELWRI) == 0) { 2517 kprintf("Unexpected clean buffer %p\n", bp); 2518 continue; 2519 } 2520 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 2521 continue; 2522 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q); 2523 2524 /* 2525 * Once the buffer is locked we will have no choice but to 2526 * unlock the spinlock around a later BUF_UNLOCK and re-set 2527 * bp = marker when looping. Move the marker now to make 2528 * things easier. 2529 */ 2530 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2531 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist); 2532 2533 /* 2534 * Must recheck B_DELWRI after successfully locking 2535 * the buffer. 2536 */ 2537 if ((bp->b_flags & B_DELWRI) == 0) { 2538 spin_unlock(&pcpu->spin); 2539 BUF_UNLOCK(bp); 2540 spin_lock(&pcpu->spin); 2541 bp = marker; 2542 continue; 2543 } 2544 2545 /* 2546 * Remove the buffer from its queue. We still own the 2547 * spinlock here. 2548 */ 2549 _bremfree(bp); 2550 2551 /* 2552 * Disposing of an invalid buffer counts as a flush op 2553 */ 2554 if (bp->b_flags & B_INVAL) { 2555 spin_unlock(&pcpu->spin); 2556 brelse(bp); 2557 spin_lock(&pcpu->spin); 2558 ++r; 2559 break; 2560 } 2561 2562 /* 2563 * Release the spinlock for the more complex ops we 2564 * are now going to do. 2565 */ 2566 spin_unlock(&pcpu->spin); 2567 lwkt_yield(); 2568 2569 /* 2570 * This is a bit messy 2571 */ 2572 if (LIST_FIRST(&bp->b_dep) != NULL && 2573 (bp->b_flags & B_DEFERRED) == 0 && 2574 buf_countdeps(bp, 0)) { 2575 spin_lock(&pcpu->spin); 2576 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist); 2577 bp->b_qindex = q; 2578 bp->b_flags |= B_DEFERRED; 2579 spin_unlock(&pcpu->spin); 2580 BUF_UNLOCK(bp); 2581 spin_lock(&pcpu->spin); 2582 bp = marker; 2583 continue; 2584 } 2585 2586 /* 2587 * spinlock not held here. 2588 * 2589 * If the buffer has a dependancy, buf_checkwrite() must 2590 * also return 0 for us to be able to initate the write. 2591 * 2592 * If the buffer is flagged B_ERROR it may be requeued 2593 * over and over again, we try to avoid a live lock. 2594 */ 2595 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2596 brelse(bp); 2597 } else if (bp->b_flags & B_ERROR) { 2598 tsleep(bp, 0, "bioer", 1); 2599 bp->b_flags &= ~B_AGE; 2600 cluster_awrite(bp); 2601 } else { 2602 bp->b_flags |= B_AGE; 2603 cluster_awrite(bp); 2604 } 2605 spin_lock(&pcpu->spin); 2606 ++r; 2607 break; 2608 } 2609 2610 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2611 marker->b_qindex = BQUEUE_NONE; 2612 spin_unlock(&pcpu->spin); 2613 2614 /* 2615 * Advance the marker to be fair. 2616 */ 2617 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus; 2618 if (bp == NULL) { 2619 if (marker->b_qcpu != lcpu) 2620 goto again; 2621 } 2622 2623 return (r); 2624 } 2625 2626 /* 2627 * inmem: 2628 * 2629 * Returns true if no I/O is needed to access the associated VM object. 2630 * This is like findblk except it also hunts around in the VM system for 2631 * the data. 2632 * 2633 * Note that we ignore vm_page_free() races from interrupts against our 2634 * lookup, since if the caller is not protected our return value will not 2635 * be any more valid then otherwise once we exit the critical section. 2636 */ 2637 int 2638 inmem(struct vnode *vp, off_t loffset) 2639 { 2640 vm_object_t obj; 2641 vm_offset_t toff, tinc, size; 2642 vm_page_t m; 2643 int res = 1; 2644 2645 if (findblk(vp, loffset, FINDBLK_TEST)) 2646 return 1; 2647 if (vp->v_mount == NULL) 2648 return 0; 2649 if ((obj = vp->v_object) == NULL) 2650 return 0; 2651 2652 size = PAGE_SIZE; 2653 if (size > vp->v_mount->mnt_stat.f_iosize) 2654 size = vp->v_mount->mnt_stat.f_iosize; 2655 2656 vm_object_hold(obj); 2657 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2658 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2659 if (m == NULL) { 2660 res = 0; 2661 break; 2662 } 2663 tinc = size; 2664 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2665 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2666 if (vm_page_is_valid(m, 2667 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2668 res = 0; 2669 break; 2670 } 2671 } 2672 vm_object_drop(obj); 2673 return (res); 2674 } 2675 2676 /* 2677 * findblk: 2678 * 2679 * Locate and return the specified buffer. Unless flagged otherwise, 2680 * a locked buffer will be returned if it exists or NULL if it does not. 2681 * 2682 * findblk()'d buffers are still on the bufqueues and if you intend 2683 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2684 * and possibly do other stuff to it. 2685 * 2686 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2687 * for locking the buffer and ensuring that it remains 2688 * the desired buffer after locking. 2689 * 2690 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2691 * to acquire the lock we return NULL, even if the 2692 * buffer exists. 2693 * 2694 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2695 * reuse by getnewbuf() but does not prevent 2696 * disassociation (B_INVAL). Used to avoid deadlocks 2697 * against random (vp,loffset)s due to reassignment. 2698 * 2699 * (0) - Lock the buffer blocking. 2700 */ 2701 struct buf * 2702 findblk(struct vnode *vp, off_t loffset, int flags) 2703 { 2704 struct buf *bp; 2705 int lkflags; 2706 2707 lkflags = LK_EXCLUSIVE; 2708 if (flags & FINDBLK_NBLOCK) 2709 lkflags |= LK_NOWAIT; 2710 2711 for (;;) { 2712 /* 2713 * Lookup. Ref the buf while holding v_token to prevent 2714 * reuse (but does not prevent diassociation). 2715 */ 2716 lwkt_gettoken_shared(&vp->v_token); 2717 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2718 if (bp == NULL) { 2719 lwkt_reltoken(&vp->v_token); 2720 return(NULL); 2721 } 2722 bqhold(bp); 2723 lwkt_reltoken(&vp->v_token); 2724 2725 /* 2726 * If testing only break and return bp, do not lock. 2727 */ 2728 if (flags & FINDBLK_TEST) 2729 break; 2730 2731 /* 2732 * Lock the buffer, return an error if the lock fails. 2733 * (only FINDBLK_NBLOCK can cause the lock to fail). 2734 */ 2735 if (BUF_LOCK(bp, lkflags)) { 2736 atomic_subtract_int(&bp->b_refs, 1); 2737 /* bp = NULL; not needed */ 2738 return(NULL); 2739 } 2740 2741 /* 2742 * Revalidate the locked buf before allowing it to be 2743 * returned. 2744 */ 2745 if (bp->b_vp == vp && bp->b_loffset == loffset) 2746 break; 2747 atomic_subtract_int(&bp->b_refs, 1); 2748 BUF_UNLOCK(bp); 2749 } 2750 2751 /* 2752 * Success 2753 */ 2754 if ((flags & FINDBLK_REF) == 0) 2755 atomic_subtract_int(&bp->b_refs, 1); 2756 return(bp); 2757 } 2758 2759 /* 2760 * getcacheblk: 2761 * 2762 * Similar to getblk() except only returns the buffer if it is 2763 * B_CACHE and requires no other manipulation. Otherwise NULL 2764 * is returned. NULL is also returned if GETBLK_NOWAIT is set 2765 * and the getblk() would block. 2766 * 2767 * If B_RAM is set the buffer might be just fine, but we return 2768 * NULL anyway because we want the code to fall through to the 2769 * cluster read. Otherwise read-ahead breaks. 2770 * 2771 * If blksize is 0 the buffer cache buffer must already be fully 2772 * cached. 2773 * 2774 * If blksize is non-zero getblk() will be used, allowing a buffer 2775 * to be reinstantiated from its VM backing store. The buffer must 2776 * still be fully cached after reinstantiation to be returned. 2777 */ 2778 struct buf * 2779 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags) 2780 { 2781 struct buf *bp; 2782 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0; 2783 2784 if (blksize) { 2785 bp = getblk(vp, loffset, blksize, blkflags, 0); 2786 if (bp) { 2787 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2788 B_CACHE) { 2789 bp->b_flags &= ~B_AGE; 2790 } else { 2791 brelse(bp); 2792 bp = NULL; 2793 } 2794 } 2795 } else { 2796 bp = findblk(vp, loffset, fndflags); 2797 if (bp) { 2798 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2799 B_CACHE) { 2800 bp->b_flags &= ~B_AGE; 2801 bremfree(bp); 2802 } else { 2803 BUF_UNLOCK(bp); 2804 bp = NULL; 2805 } 2806 } 2807 } 2808 return (bp); 2809 } 2810 2811 /* 2812 * getblk: 2813 * 2814 * Get a block given a specified block and offset into a file/device. 2815 * B_INVAL may or may not be set on return. The caller should clear 2816 * B_INVAL prior to initiating a READ. 2817 * 2818 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2819 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2820 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2821 * without doing any of those things the system will likely believe 2822 * the buffer to be valid (especially if it is not B_VMIO), and the 2823 * next getblk() will return the buffer with B_CACHE set. 2824 * 2825 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2826 * an existing buffer. 2827 * 2828 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2829 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2830 * and then cleared based on the backing VM. If the previous buffer is 2831 * non-0-sized but invalid, B_CACHE will be cleared. 2832 * 2833 * If getblk() must create a new buffer, the new buffer is returned with 2834 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2835 * case it is returned with B_INVAL clear and B_CACHE set based on the 2836 * backing VM. 2837 * 2838 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2839 * B_CACHE bit is clear. 2840 * 2841 * What this means, basically, is that the caller should use B_CACHE to 2842 * determine whether the buffer is fully valid or not and should clear 2843 * B_INVAL prior to issuing a read. If the caller intends to validate 2844 * the buffer by loading its data area with something, the caller needs 2845 * to clear B_INVAL. If the caller does this without issuing an I/O, 2846 * the caller should set B_CACHE ( as an optimization ), else the caller 2847 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2848 * a write attempt or if it was a successfull read. If the caller 2849 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2850 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2851 * 2852 * getblk flags: 2853 * 2854 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2855 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2856 */ 2857 struct buf * 2858 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2859 { 2860 struct buf *bp; 2861 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2862 int error; 2863 int lkflags; 2864 2865 if (size > MAXBSIZE) 2866 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2867 if (vp->v_object == NULL) 2868 panic("getblk: vnode %p has no object!", vp); 2869 2870 loop: 2871 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2872 /* 2873 * The buffer was found in the cache, but we need to lock it. 2874 * We must acquire a ref on the bp to prevent reuse, but 2875 * this will not prevent disassociation (brelvp()) so we 2876 * must recheck (vp,loffset) after acquiring the lock. 2877 * 2878 * Without the ref the buffer could potentially be reused 2879 * before we acquire the lock and create a deadlock 2880 * situation between the thread trying to reuse the buffer 2881 * and us due to the fact that we would wind up blocking 2882 * on a random (vp,loffset). 2883 */ 2884 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2885 if (blkflags & GETBLK_NOWAIT) { 2886 bqdrop(bp); 2887 return(NULL); 2888 } 2889 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2890 if (blkflags & GETBLK_PCATCH) 2891 lkflags |= LK_PCATCH; 2892 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2893 if (error) { 2894 bqdrop(bp); 2895 if (error == ENOLCK) 2896 goto loop; 2897 return (NULL); 2898 } 2899 /* buffer may have changed on us */ 2900 } 2901 bqdrop(bp); 2902 2903 /* 2904 * Once the buffer has been locked, make sure we didn't race 2905 * a buffer recyclement. Buffers that are no longer hashed 2906 * will have b_vp == NULL, so this takes care of that check 2907 * as well. 2908 */ 2909 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2910 kprintf("Warning buffer %p (vp %p loffset %lld) " 2911 "was recycled\n", 2912 bp, vp, (long long)loffset); 2913 BUF_UNLOCK(bp); 2914 goto loop; 2915 } 2916 2917 /* 2918 * If SZMATCH any pre-existing buffer must be of the requested 2919 * size or NULL is returned. The caller absolutely does not 2920 * want getblk() to bwrite() the buffer on a size mismatch. 2921 */ 2922 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2923 BUF_UNLOCK(bp); 2924 return(NULL); 2925 } 2926 2927 /* 2928 * All vnode-based buffers must be backed by a VM object. 2929 */ 2930 KKASSERT(bp->b_flags & B_VMIO); 2931 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2932 bp->b_flags &= ~B_AGE; 2933 2934 /* 2935 * Make sure that B_INVAL buffers do not have a cached 2936 * block number translation. 2937 */ 2938 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2939 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2940 " did not have cleared bio_offset cache\n", 2941 bp, vp, (long long)loffset); 2942 clearbiocache(&bp->b_bio2); 2943 } 2944 2945 /* 2946 * The buffer is locked. B_CACHE is cleared if the buffer is 2947 * invalid. 2948 */ 2949 if (bp->b_flags & B_INVAL) 2950 bp->b_flags &= ~B_CACHE; 2951 bremfree(bp); 2952 2953 /* 2954 * Any size inconsistancy with a dirty buffer or a buffer 2955 * with a softupdates dependancy must be resolved. Resizing 2956 * the buffer in such circumstances can lead to problems. 2957 * 2958 * Dirty or dependant buffers are written synchronously. 2959 * Other types of buffers are simply released and 2960 * reconstituted as they may be backed by valid, dirty VM 2961 * pages (but not marked B_DELWRI). 2962 * 2963 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2964 * and may be left over from a prior truncation (and thus 2965 * no longer represent the actual EOF point), so we 2966 * definitely do not want to B_NOCACHE the backing store. 2967 */ 2968 if (size != bp->b_bcount) { 2969 if (bp->b_flags & B_DELWRI) { 2970 bp->b_flags |= B_RELBUF; 2971 bwrite(bp); 2972 } else if (LIST_FIRST(&bp->b_dep)) { 2973 bp->b_flags |= B_RELBUF; 2974 bwrite(bp); 2975 } else { 2976 bp->b_flags |= B_RELBUF; 2977 brelse(bp); 2978 } 2979 goto loop; 2980 } 2981 KKASSERT(size <= bp->b_kvasize); 2982 KASSERT(bp->b_loffset != NOOFFSET, 2983 ("getblk: no buffer offset")); 2984 2985 /* 2986 * A buffer with B_DELWRI set and B_CACHE clear must 2987 * be committed before we can return the buffer in 2988 * order to prevent the caller from issuing a read 2989 * ( due to B_CACHE not being set ) and overwriting 2990 * it. 2991 * 2992 * Most callers, including NFS and FFS, need this to 2993 * operate properly either because they assume they 2994 * can issue a read if B_CACHE is not set, or because 2995 * ( for example ) an uncached B_DELWRI might loop due 2996 * to softupdates re-dirtying the buffer. In the latter 2997 * case, B_CACHE is set after the first write completes, 2998 * preventing further loops. 2999 * 3000 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3001 * above while extending the buffer, we cannot allow the 3002 * buffer to remain with B_CACHE set after the write 3003 * completes or it will represent a corrupt state. To 3004 * deal with this we set B_NOCACHE to scrap the buffer 3005 * after the write. 3006 * 3007 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3008 * I'm not even sure this state is still possible 3009 * now that getblk() writes out any dirty buffers 3010 * on size changes. 3011 * 3012 * We might be able to do something fancy, like setting 3013 * B_CACHE in bwrite() except if B_DELWRI is already set, 3014 * so the below call doesn't set B_CACHE, but that gets real 3015 * confusing. This is much easier. 3016 */ 3017 3018 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3019 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3020 "and CACHE clear, b_flags %08x\n", 3021 bp, (uintmax_t)bp->b_loffset, bp->b_flags); 3022 bp->b_flags |= B_NOCACHE; 3023 bwrite(bp); 3024 goto loop; 3025 } 3026 } else { 3027 /* 3028 * Buffer is not in-core, create new buffer. The buffer 3029 * returned by getnewbuf() is locked. Note that the returned 3030 * buffer is also considered valid (not marked B_INVAL). 3031 * 3032 * Calculating the offset for the I/O requires figuring out 3033 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3034 * the mount's f_iosize otherwise. If the vnode does not 3035 * have an associated mount we assume that the passed size is 3036 * the block size. 3037 * 3038 * Note that vn_isdisk() cannot be used here since it may 3039 * return a failure for numerous reasons. Note that the 3040 * buffer size may be larger then the block size (the caller 3041 * will use block numbers with the proper multiple). Beware 3042 * of using any v_* fields which are part of unions. In 3043 * particular, in DragonFly the mount point overloading 3044 * mechanism uses the namecache only and the underlying 3045 * directory vnode is not a special case. 3046 */ 3047 int bsize, maxsize; 3048 3049 if (vp->v_type == VBLK || vp->v_type == VCHR) 3050 bsize = DEV_BSIZE; 3051 else if (vp->v_mount) 3052 bsize = vp->v_mount->mnt_stat.f_iosize; 3053 else 3054 bsize = size; 3055 3056 maxsize = size + (loffset & PAGE_MASK); 3057 maxsize = imax(maxsize, bsize); 3058 3059 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3060 if (bp == NULL) { 3061 if (slpflags || slptimeo) 3062 return NULL; 3063 goto loop; 3064 } 3065 3066 /* 3067 * Atomically insert the buffer into the hash, so that it can 3068 * be found by findblk(). 3069 * 3070 * If bgetvp() returns non-zero a collision occured, and the 3071 * bp will not be associated with the vnode. 3072 * 3073 * Make sure the translation layer has been cleared. 3074 */ 3075 bp->b_loffset = loffset; 3076 bp->b_bio2.bio_offset = NOOFFSET; 3077 /* bp->b_bio2.bio_next = NULL; */ 3078 3079 if (bgetvp(vp, bp, size)) { 3080 bp->b_flags |= B_INVAL; 3081 brelse(bp); 3082 goto loop; 3083 } 3084 3085 /* 3086 * All vnode-based buffers must be backed by a VM object. 3087 */ 3088 KKASSERT(vp->v_object != NULL); 3089 bp->b_flags |= B_VMIO; 3090 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3091 3092 allocbuf(bp, size); 3093 } 3094 KKASSERT(dsched_is_clear_buf_priv(bp)); 3095 return (bp); 3096 } 3097 3098 /* 3099 * regetblk(bp) 3100 * 3101 * Reacquire a buffer that was previously released to the locked queue, 3102 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3103 * set B_LOCKED (which handles the acquisition race). 3104 * 3105 * To this end, either B_LOCKED must be set or the dependancy list must be 3106 * non-empty. 3107 */ 3108 void 3109 regetblk(struct buf *bp) 3110 { 3111 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3112 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3113 bremfree(bp); 3114 } 3115 3116 /* 3117 * geteblk: 3118 * 3119 * Get an empty, disassociated buffer of given size. The buffer is 3120 * initially set to B_INVAL. 3121 * 3122 * critical section protection is not required for the allocbuf() 3123 * call because races are impossible here. 3124 */ 3125 struct buf * 3126 geteblk(int size) 3127 { 3128 struct buf *bp; 3129 int maxsize; 3130 3131 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3132 3133 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL) 3134 ; 3135 allocbuf(bp, size); 3136 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3137 KKASSERT(dsched_is_clear_buf_priv(bp)); 3138 return (bp); 3139 } 3140 3141 3142 /* 3143 * allocbuf: 3144 * 3145 * This code constitutes the buffer memory from either anonymous system 3146 * memory (in the case of non-VMIO operations) or from an associated 3147 * VM object (in the case of VMIO operations). This code is able to 3148 * resize a buffer up or down. 3149 * 3150 * Note that this code is tricky, and has many complications to resolve 3151 * deadlock or inconsistant data situations. Tread lightly!!! 3152 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3153 * the caller. Calling this code willy nilly can result in the loss of 3154 * data. 3155 * 3156 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3157 * B_CACHE for the non-VMIO case. 3158 * 3159 * This routine does not need to be called from a critical section but you 3160 * must own the buffer. 3161 */ 3162 int 3163 allocbuf(struct buf *bp, int size) 3164 { 3165 int newbsize, mbsize; 3166 int i; 3167 3168 if (BUF_REFCNT(bp) == 0) 3169 panic("allocbuf: buffer not busy"); 3170 3171 if (bp->b_kvasize < size) 3172 panic("allocbuf: buffer too small"); 3173 3174 if ((bp->b_flags & B_VMIO) == 0) { 3175 caddr_t origbuf; 3176 int origbufsize; 3177 /* 3178 * Just get anonymous memory from the kernel. Don't 3179 * mess with B_CACHE. 3180 */ 3181 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3182 if (bp->b_flags & B_MALLOC) 3183 newbsize = mbsize; 3184 else 3185 newbsize = round_page(size); 3186 3187 if (newbsize < bp->b_bufsize) { 3188 /* 3189 * Malloced buffers are not shrunk 3190 */ 3191 if (bp->b_flags & B_MALLOC) { 3192 if (newbsize) { 3193 bp->b_bcount = size; 3194 } else { 3195 kfree(bp->b_data, M_BIOBUF); 3196 if (bp->b_bufsize) { 3197 atomic_subtract_long(&bufmallocspace, bp->b_bufsize); 3198 bufspacewakeup(); 3199 bp->b_bufsize = 0; 3200 } 3201 bp->b_data = bp->b_kvabase; 3202 bp->b_bcount = 0; 3203 bp->b_flags &= ~B_MALLOC; 3204 } 3205 return 1; 3206 } 3207 vm_hold_free_pages( 3208 bp, 3209 (vm_offset_t) bp->b_data + newbsize, 3210 (vm_offset_t) bp->b_data + bp->b_bufsize); 3211 } else if (newbsize > bp->b_bufsize) { 3212 /* 3213 * We only use malloced memory on the first allocation. 3214 * and revert to page-allocated memory when the buffer 3215 * grows. 3216 */ 3217 if ((bufmallocspace < maxbufmallocspace) && 3218 (bp->b_bufsize == 0) && 3219 (mbsize <= PAGE_SIZE/2)) { 3220 3221 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3222 bp->b_bufsize = mbsize; 3223 bp->b_bcount = size; 3224 bp->b_flags |= B_MALLOC; 3225 atomic_add_long(&bufmallocspace, mbsize); 3226 return 1; 3227 } 3228 origbuf = NULL; 3229 origbufsize = 0; 3230 /* 3231 * If the buffer is growing on its other-than-first 3232 * allocation, then we revert to the page-allocation 3233 * scheme. 3234 */ 3235 if (bp->b_flags & B_MALLOC) { 3236 origbuf = bp->b_data; 3237 origbufsize = bp->b_bufsize; 3238 bp->b_data = bp->b_kvabase; 3239 if (bp->b_bufsize) { 3240 atomic_subtract_long(&bufmallocspace, 3241 bp->b_bufsize); 3242 bufspacewakeup(); 3243 bp->b_bufsize = 0; 3244 } 3245 bp->b_flags &= ~B_MALLOC; 3246 newbsize = round_page(newbsize); 3247 } 3248 vm_hold_load_pages( 3249 bp, 3250 (vm_offset_t) bp->b_data + bp->b_bufsize, 3251 (vm_offset_t) bp->b_data + newbsize); 3252 if (origbuf) { 3253 bcopy(origbuf, bp->b_data, origbufsize); 3254 kfree(origbuf, M_BIOBUF); 3255 } 3256 } 3257 } else { 3258 vm_page_t m; 3259 int desiredpages; 3260 3261 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3262 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3263 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3264 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3265 3266 if (bp->b_flags & B_MALLOC) 3267 panic("allocbuf: VMIO buffer can't be malloced"); 3268 /* 3269 * Set B_CACHE initially if buffer is 0 length or will become 3270 * 0-length. 3271 */ 3272 if (size == 0 || bp->b_bufsize == 0) 3273 bp->b_flags |= B_CACHE; 3274 3275 if (newbsize < bp->b_bufsize) { 3276 /* 3277 * DEV_BSIZE aligned new buffer size is less then the 3278 * DEV_BSIZE aligned existing buffer size. Figure out 3279 * if we have to remove any pages. 3280 */ 3281 if (desiredpages < bp->b_xio.xio_npages) { 3282 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3283 /* 3284 * the page is not freed here -- it 3285 * is the responsibility of 3286 * vnode_pager_setsize 3287 */ 3288 m = bp->b_xio.xio_pages[i]; 3289 KASSERT(m != bogus_page, 3290 ("allocbuf: bogus page found")); 3291 vm_page_busy_wait(m, TRUE, "biodep"); 3292 bp->b_xio.xio_pages[i] = NULL; 3293 vm_page_unwire(m, 0); 3294 vm_page_wakeup(m); 3295 } 3296 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3297 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3298 bp->b_xio.xio_npages = desiredpages; 3299 } 3300 } else if (size > bp->b_bcount) { 3301 /* 3302 * We are growing the buffer, possibly in a 3303 * byte-granular fashion. 3304 */ 3305 struct vnode *vp; 3306 vm_object_t obj; 3307 vm_offset_t toff; 3308 vm_offset_t tinc; 3309 3310 /* 3311 * Step 1, bring in the VM pages from the object, 3312 * allocating them if necessary. We must clear 3313 * B_CACHE if these pages are not valid for the 3314 * range covered by the buffer. 3315 * 3316 * critical section protection is required to protect 3317 * against interrupts unbusying and freeing pages 3318 * between our vm_page_lookup() and our 3319 * busycheck/wiring call. 3320 */ 3321 vp = bp->b_vp; 3322 obj = vp->v_object; 3323 3324 vm_object_hold(obj); 3325 while (bp->b_xio.xio_npages < desiredpages) { 3326 vm_page_t m; 3327 vm_pindex_t pi; 3328 int error; 3329 3330 pi = OFF_TO_IDX(bp->b_loffset) + 3331 bp->b_xio.xio_npages; 3332 3333 /* 3334 * Blocking on m->busy might lead to a 3335 * deadlock: 3336 * 3337 * vm_fault->getpages->cluster_read->allocbuf 3338 */ 3339 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3340 &error); 3341 if (error) { 3342 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3343 continue; 3344 } 3345 if (m == NULL) { 3346 /* 3347 * note: must allocate system pages 3348 * since blocking here could intefere 3349 * with paging I/O, no matter which 3350 * process we are. 3351 */ 3352 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages); 3353 if (m) { 3354 vm_page_wire(m); 3355 vm_page_flag_clear(m, PG_ZERO); 3356 vm_page_wakeup(m); 3357 bp->b_flags &= ~B_CACHE; 3358 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3359 ++bp->b_xio.xio_npages; 3360 } 3361 continue; 3362 } 3363 3364 /* 3365 * We found a page and were able to busy it. 3366 */ 3367 vm_page_flag_clear(m, PG_ZERO); 3368 vm_page_wire(m); 3369 vm_page_wakeup(m); 3370 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3371 ++bp->b_xio.xio_npages; 3372 if (bp->b_act_count < m->act_count) 3373 bp->b_act_count = m->act_count; 3374 } 3375 vm_object_drop(obj); 3376 3377 /* 3378 * Step 2. We've loaded the pages into the buffer, 3379 * we have to figure out if we can still have B_CACHE 3380 * set. Note that B_CACHE is set according to the 3381 * byte-granular range ( bcount and size ), not the 3382 * aligned range ( newbsize ). 3383 * 3384 * The VM test is against m->valid, which is DEV_BSIZE 3385 * aligned. Needless to say, the validity of the data 3386 * needs to also be DEV_BSIZE aligned. Note that this 3387 * fails with NFS if the server or some other client 3388 * extends the file's EOF. If our buffer is resized, 3389 * B_CACHE may remain set! XXX 3390 */ 3391 3392 toff = bp->b_bcount; 3393 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3394 3395 while ((bp->b_flags & B_CACHE) && toff < size) { 3396 vm_pindex_t pi; 3397 3398 if (tinc > (size - toff)) 3399 tinc = size - toff; 3400 3401 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3402 PAGE_SHIFT; 3403 3404 vfs_buf_test_cache( 3405 bp, 3406 bp->b_loffset, 3407 toff, 3408 tinc, 3409 bp->b_xio.xio_pages[pi] 3410 ); 3411 toff += tinc; 3412 tinc = PAGE_SIZE; 3413 } 3414 3415 /* 3416 * Step 3, fixup the KVM pmap. Remember that 3417 * bp->b_data is relative to bp->b_loffset, but 3418 * bp->b_loffset may be offset into the first page. 3419 */ 3420 3421 bp->b_data = (caddr_t) 3422 trunc_page((vm_offset_t)bp->b_data); 3423 pmap_qenter( 3424 (vm_offset_t)bp->b_data, 3425 bp->b_xio.xio_pages, 3426 bp->b_xio.xio_npages 3427 ); 3428 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3429 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3430 } 3431 } 3432 3433 /* adjust space use on already-dirty buffer */ 3434 if (bp->b_flags & B_DELWRI) { 3435 /* dirtykvaspace unchanged */ 3436 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize); 3437 if (bp->b_flags & B_HEAVY) { 3438 atomic_add_long(&dirtybufspacehw, 3439 newbsize - bp->b_bufsize); 3440 } 3441 } 3442 if (newbsize < bp->b_bufsize) 3443 bufspacewakeup(); 3444 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3445 bp->b_bcount = size; /* requested buffer size */ 3446 return 1; 3447 } 3448 3449 /* 3450 * biowait: 3451 * 3452 * Wait for buffer I/O completion, returning error status. B_EINTR 3453 * is converted into an EINTR error but not cleared (since a chain 3454 * of biowait() calls may occur). 3455 * 3456 * On return bpdone() will have been called but the buffer will remain 3457 * locked and will not have been brelse()'d. 3458 * 3459 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3460 * likely still in progress on return. 3461 * 3462 * NOTE! This operation is on a BIO, not a BUF. 3463 * 3464 * NOTE! BIO_DONE is cleared by vn_strategy() 3465 */ 3466 static __inline int 3467 _biowait(struct bio *bio, const char *wmesg, int to) 3468 { 3469 struct buf *bp = bio->bio_buf; 3470 u_int32_t flags; 3471 u_int32_t nflags; 3472 int error; 3473 3474 KKASSERT(bio == &bp->b_bio1); 3475 for (;;) { 3476 flags = bio->bio_flags; 3477 if (flags & BIO_DONE) 3478 break; 3479 nflags = flags | BIO_WANT; 3480 tsleep_interlock(bio, 0); 3481 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3482 if (wmesg) 3483 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3484 else if (bp->b_cmd == BUF_CMD_READ) 3485 error = tsleep(bio, PINTERLOCKED, "biord", to); 3486 else 3487 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3488 if (error) { 3489 kprintf("tsleep error biowait %d\n", error); 3490 return (error); 3491 } 3492 } 3493 } 3494 3495 /* 3496 * Finish up. 3497 */ 3498 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3499 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3500 if (bp->b_flags & B_EINTR) 3501 return (EINTR); 3502 if (bp->b_flags & B_ERROR) 3503 return (bp->b_error ? bp->b_error : EIO); 3504 return (0); 3505 } 3506 3507 int 3508 biowait(struct bio *bio, const char *wmesg) 3509 { 3510 return(_biowait(bio, wmesg, 0)); 3511 } 3512 3513 int 3514 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3515 { 3516 return(_biowait(bio, wmesg, to)); 3517 } 3518 3519 /* 3520 * This associates a tracking count with an I/O. vn_strategy() and 3521 * dev_dstrategy() do this automatically but there are a few cases 3522 * where a vnode or device layer is bypassed when a block translation 3523 * is cached. In such cases bio_start_transaction() may be called on 3524 * the bypassed layers so the system gets an I/O in progress indication 3525 * for those higher layers. 3526 */ 3527 void 3528 bio_start_transaction(struct bio *bio, struct bio_track *track) 3529 { 3530 bio->bio_track = track; 3531 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3532 dsched_new_buf(bio->bio_buf); 3533 bio_track_ref(track); 3534 } 3535 3536 /* 3537 * Initiate I/O on a vnode. 3538 * 3539 * SWAPCACHE OPERATION: 3540 * 3541 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3542 * devfs also uses b_vp for fake buffers so we also have to check 3543 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3544 * underlying block device. The swap assignments are related to the 3545 * buffer cache buffer's b_vp, not the passed vp. 3546 * 3547 * The passed vp == bp->b_vp only in the case where the strategy call 3548 * is made on the vp itself for its own buffers (a regular file or 3549 * block device vp). The filesystem usually then re-calls vn_strategy() 3550 * after translating the request to an underlying device. 3551 * 3552 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3553 * underlying buffer cache buffers. 3554 * 3555 * We can only deal with page-aligned buffers at the moment, because 3556 * we can't tell what the real dirty state for pages straddling a buffer 3557 * are. 3558 * 3559 * In order to call swap_pager_strategy() we must provide the VM object 3560 * and base offset for the underlying buffer cache pages so it can find 3561 * the swap blocks. 3562 */ 3563 void 3564 vn_strategy(struct vnode *vp, struct bio *bio) 3565 { 3566 struct bio_track *track; 3567 struct buf *bp = bio->bio_buf; 3568 3569 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3570 3571 /* 3572 * Set when an I/O is issued on the bp. Cleared by consumers 3573 * (aka HAMMER), allowing the consumer to determine if I/O had 3574 * actually occurred. 3575 */ 3576 bp->b_flags |= B_IODEBUG; 3577 3578 /* 3579 * Handle the swap cache intercept. 3580 */ 3581 if (vn_cache_strategy(vp, bio)) 3582 return; 3583 3584 /* 3585 * Otherwise do the operation through the filesystem 3586 */ 3587 if (bp->b_cmd == BUF_CMD_READ) 3588 track = &vp->v_track_read; 3589 else 3590 track = &vp->v_track_write; 3591 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3592 bio->bio_track = track; 3593 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3594 dsched_new_buf(bio->bio_buf); 3595 bio_track_ref(track); 3596 vop_strategy(*vp->v_ops, vp, bio); 3597 } 3598 3599 static void vn_cache_strategy_callback(struct bio *bio); 3600 3601 int 3602 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3603 { 3604 struct buf *bp = bio->bio_buf; 3605 struct bio *nbio; 3606 vm_object_t object; 3607 vm_page_t m; 3608 int i; 3609 3610 /* 3611 * Is this buffer cache buffer suitable for reading from 3612 * the swap cache? 3613 */ 3614 if (vm_swapcache_read_enable == 0 || 3615 bp->b_cmd != BUF_CMD_READ || 3616 ((bp->b_flags & B_CLUSTER) == 0 && 3617 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3618 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3619 (bp->b_bcount & PAGE_MASK) != 0) { 3620 return(0); 3621 } 3622 3623 /* 3624 * Figure out the original VM object (it will match the underlying 3625 * VM pages). Note that swap cached data uses page indices relative 3626 * to that object, not relative to bio->bio_offset. 3627 */ 3628 if (bp->b_flags & B_CLUSTER) 3629 object = vp->v_object; 3630 else 3631 object = bp->b_vp->v_object; 3632 3633 /* 3634 * In order to be able to use the swap cache all underlying VM 3635 * pages must be marked as such, and we can't have any bogus pages. 3636 */ 3637 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3638 m = bp->b_xio.xio_pages[i]; 3639 if ((m->flags & PG_SWAPPED) == 0) 3640 break; 3641 if (m == bogus_page) 3642 break; 3643 } 3644 3645 /* 3646 * If we are good then issue the I/O using swap_pager_strategy(). 3647 * 3648 * We can only do this if the buffer actually supports object-backed 3649 * I/O. If it doesn't npages will be 0. 3650 */ 3651 if (i && i == bp->b_xio.xio_npages) { 3652 m = bp->b_xio.xio_pages[0]; 3653 nbio = push_bio(bio); 3654 nbio->bio_done = vn_cache_strategy_callback; 3655 nbio->bio_offset = ptoa(m->pindex); 3656 KKASSERT(m->object == object); 3657 swap_pager_strategy(object, nbio); 3658 return(1); 3659 } 3660 return(0); 3661 } 3662 3663 /* 3664 * This is a bit of a hack but since the vn_cache_strategy() function can 3665 * override a VFS's strategy function we must make sure that the bio, which 3666 * is probably bio2, doesn't leak an unexpected offset value back to the 3667 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3668 * bio went through its own file strategy function and the the bio2 offset 3669 * is a cached disk offset when, in fact, it isn't. 3670 */ 3671 static void 3672 vn_cache_strategy_callback(struct bio *bio) 3673 { 3674 bio->bio_offset = NOOFFSET; 3675 biodone(pop_bio(bio)); 3676 } 3677 3678 /* 3679 * bpdone: 3680 * 3681 * Finish I/O on a buffer after all BIOs have been processed. 3682 * Called when the bio chain is exhausted or by biowait. If called 3683 * by biowait, elseit is typically 0. 3684 * 3685 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3686 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3687 * assuming B_INVAL is clear. 3688 * 3689 * For the VMIO case, we set B_CACHE if the op was a read and no 3690 * read error occured, or if the op was a write. B_CACHE is never 3691 * set if the buffer is invalid or otherwise uncacheable. 3692 * 3693 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3694 * initiator to leave B_INVAL set to brelse the buffer out of existance 3695 * in the biodone routine. 3696 */ 3697 void 3698 bpdone(struct buf *bp, int elseit) 3699 { 3700 buf_cmd_t cmd; 3701 3702 KASSERT(BUF_REFCNTNB(bp) > 0, 3703 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3704 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3705 ("biodone: bp %p already done!", bp)); 3706 3707 /* 3708 * No more BIOs are left. All completion functions have been dealt 3709 * with, now we clean up the buffer. 3710 */ 3711 cmd = bp->b_cmd; 3712 bp->b_cmd = BUF_CMD_DONE; 3713 3714 /* 3715 * Only reads and writes are processed past this point. 3716 */ 3717 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3718 if (cmd == BUF_CMD_FREEBLKS) 3719 bp->b_flags |= B_NOCACHE; 3720 if (elseit) 3721 brelse(bp); 3722 return; 3723 } 3724 3725 /* 3726 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3727 * a lot worse. XXX - move this above the clearing of b_cmd 3728 */ 3729 if (LIST_FIRST(&bp->b_dep) != NULL) 3730 buf_complete(bp); 3731 3732 /* 3733 * A failed write must re-dirty the buffer unless B_INVAL 3734 * was set. Only applicable to normal buffers (with VPs). 3735 * vinum buffers may not have a vp. 3736 */ 3737 if (cmd == BUF_CMD_WRITE && 3738 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3739 bp->b_flags &= ~B_NOCACHE; 3740 if (bp->b_vp) 3741 bdirty(bp); 3742 } 3743 3744 if (bp->b_flags & B_VMIO) { 3745 int i; 3746 vm_ooffset_t foff; 3747 vm_page_t m; 3748 vm_object_t obj; 3749 int iosize; 3750 struct vnode *vp = bp->b_vp; 3751 3752 obj = vp->v_object; 3753 3754 #if defined(VFS_BIO_DEBUG) 3755 if (vp->v_auxrefs == 0) 3756 panic("biodone: zero vnode hold count"); 3757 if ((vp->v_flag & VOBJBUF) == 0) 3758 panic("biodone: vnode is not setup for merged cache"); 3759 #endif 3760 3761 foff = bp->b_loffset; 3762 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 3763 KASSERT(obj != NULL, ("biodone: missing VM object")); 3764 3765 #if defined(VFS_BIO_DEBUG) 3766 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3767 kprintf("biodone: paging in progress(%d) < " 3768 "bp->b_xio.xio_npages(%d)\n", 3769 obj->paging_in_progress, 3770 bp->b_xio.xio_npages); 3771 } 3772 #endif 3773 3774 /* 3775 * Set B_CACHE if the op was a normal read and no error 3776 * occured. B_CACHE is set for writes in the b*write() 3777 * routines. 3778 */ 3779 iosize = bp->b_bcount - bp->b_resid; 3780 if (cmd == BUF_CMD_READ && 3781 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3782 bp->b_flags |= B_CACHE; 3783 } 3784 3785 vm_object_hold(obj); 3786 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3787 int bogusflag = 0; 3788 int resid; 3789 3790 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3791 if (resid > iosize) 3792 resid = iosize; 3793 3794 /* 3795 * cleanup bogus pages, restoring the originals. Since 3796 * the originals should still be wired, we don't have 3797 * to worry about interrupt/freeing races destroying 3798 * the VM object association. 3799 */ 3800 m = bp->b_xio.xio_pages[i]; 3801 if (m == bogus_page) { 3802 bogusflag = 1; 3803 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3804 if (m == NULL) 3805 panic("biodone: page disappeared"); 3806 bp->b_xio.xio_pages[i] = m; 3807 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3808 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3809 } 3810 #if defined(VFS_BIO_DEBUG) 3811 if (OFF_TO_IDX(foff) != m->pindex) { 3812 kprintf("biodone: foff(%lu)/m->pindex(%ld) " 3813 "mismatch\n", 3814 (unsigned long)foff, (long)m->pindex); 3815 } 3816 #endif 3817 3818 /* 3819 * In the write case, the valid and clean bits are 3820 * already changed correctly (see bdwrite()), so we 3821 * only need to do this here in the read case. 3822 */ 3823 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3824 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3825 vfs_clean_one_page(bp, i, m); 3826 } 3827 vm_page_flag_clear(m, PG_ZERO); 3828 3829 /* 3830 * when debugging new filesystems or buffer I/O 3831 * methods, this is the most common error that pops 3832 * up. if you see this, you have not set the page 3833 * busy flag correctly!!! 3834 */ 3835 if (m->busy == 0) { 3836 kprintf("biodone: page busy < 0, " 3837 "pindex: %d, foff: 0x(%x,%x), " 3838 "resid: %d, index: %d\n", 3839 (int) m->pindex, (int)(foff >> 32), 3840 (int) foff & 0xffffffff, resid, i); 3841 if (!vn_isdisk(vp, NULL)) 3842 kprintf(" iosize: %ld, loffset: %lld, " 3843 "flags: 0x%08x, npages: %d\n", 3844 bp->b_vp->v_mount->mnt_stat.f_iosize, 3845 (long long)bp->b_loffset, 3846 bp->b_flags, bp->b_xio.xio_npages); 3847 else 3848 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3849 (long long)bp->b_loffset, 3850 bp->b_flags, bp->b_xio.xio_npages); 3851 kprintf(" valid: 0x%x, dirty: 0x%x, " 3852 "wired: %d\n", 3853 m->valid, m->dirty, 3854 m->wire_count); 3855 panic("biodone: page busy < 0"); 3856 } 3857 vm_page_io_finish(m); 3858 vm_page_wakeup(m); 3859 vm_object_pip_wakeup(obj); 3860 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3861 iosize -= resid; 3862 } 3863 bp->b_flags &= ~B_HASBOGUS; 3864 vm_object_drop(obj); 3865 } 3866 3867 /* 3868 * Finish up by releasing the buffer. There are no more synchronous 3869 * or asynchronous completions, those were handled by bio_done 3870 * callbacks. 3871 */ 3872 if (elseit) { 3873 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3874 brelse(bp); 3875 else 3876 bqrelse(bp); 3877 } 3878 } 3879 3880 /* 3881 * Normal biodone. 3882 */ 3883 void 3884 biodone(struct bio *bio) 3885 { 3886 struct buf *bp = bio->bio_buf; 3887 3888 runningbufwakeup(bp); 3889 3890 /* 3891 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3892 */ 3893 while (bio) { 3894 biodone_t *done_func; 3895 struct bio_track *track; 3896 3897 /* 3898 * BIO tracking. Most but not all BIOs are tracked. 3899 */ 3900 if ((track = bio->bio_track) != NULL) { 3901 bio_track_rel(track); 3902 bio->bio_track = NULL; 3903 } 3904 3905 /* 3906 * A bio_done function terminates the loop. The function 3907 * will be responsible for any further chaining and/or 3908 * buffer management. 3909 * 3910 * WARNING! The done function can deallocate the buffer! 3911 */ 3912 if ((done_func = bio->bio_done) != NULL) { 3913 bio->bio_done = NULL; 3914 done_func(bio); 3915 return; 3916 } 3917 bio = bio->bio_prev; 3918 } 3919 3920 /* 3921 * If we've run out of bio's do normal [a]synchronous completion. 3922 */ 3923 bpdone(bp, 1); 3924 } 3925 3926 /* 3927 * Synchronous biodone - this terminates a synchronous BIO. 3928 * 3929 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3930 * but still locked. The caller must brelse() the buffer after waiting 3931 * for completion. 3932 */ 3933 void 3934 biodone_sync(struct bio *bio) 3935 { 3936 struct buf *bp = bio->bio_buf; 3937 int flags; 3938 int nflags; 3939 3940 KKASSERT(bio == &bp->b_bio1); 3941 bpdone(bp, 0); 3942 3943 for (;;) { 3944 flags = bio->bio_flags; 3945 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3946 3947 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3948 if (flags & BIO_WANT) 3949 wakeup(bio); 3950 break; 3951 } 3952 } 3953 } 3954 3955 /* 3956 * vfs_unbusy_pages: 3957 * 3958 * This routine is called in lieu of iodone in the case of 3959 * incomplete I/O. This keeps the busy status for pages 3960 * consistant. 3961 */ 3962 void 3963 vfs_unbusy_pages(struct buf *bp) 3964 { 3965 int i; 3966 3967 runningbufwakeup(bp); 3968 3969 if (bp->b_flags & B_VMIO) { 3970 struct vnode *vp = bp->b_vp; 3971 vm_object_t obj; 3972 3973 obj = vp->v_object; 3974 vm_object_hold(obj); 3975 3976 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3977 vm_page_t m = bp->b_xio.xio_pages[i]; 3978 3979 /* 3980 * When restoring bogus changes the original pages 3981 * should still be wired, so we are in no danger of 3982 * losing the object association and do not need 3983 * critical section protection particularly. 3984 */ 3985 if (m == bogus_page) { 3986 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 3987 if (!m) { 3988 panic("vfs_unbusy_pages: page missing"); 3989 } 3990 bp->b_xio.xio_pages[i] = m; 3991 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3992 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3993 } 3994 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3995 vm_page_flag_clear(m, PG_ZERO); 3996 vm_page_io_finish(m); 3997 vm_page_wakeup(m); 3998 vm_object_pip_wakeup(obj); 3999 } 4000 bp->b_flags &= ~B_HASBOGUS; 4001 vm_object_drop(obj); 4002 } 4003 } 4004 4005 /* 4006 * vfs_busy_pages: 4007 * 4008 * This routine is called before a device strategy routine. 4009 * It is used to tell the VM system that paging I/O is in 4010 * progress, and treat the pages associated with the buffer 4011 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4012 * flag is handled to make sure that the object doesn't become 4013 * inconsistant. 4014 * 4015 * Since I/O has not been initiated yet, certain buffer flags 4016 * such as B_ERROR or B_INVAL may be in an inconsistant state 4017 * and should be ignored. 4018 */ 4019 void 4020 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4021 { 4022 int i, bogus; 4023 struct lwp *lp = curthread->td_lwp; 4024 4025 /* 4026 * The buffer's I/O command must already be set. If reading, 4027 * B_CACHE must be 0 (double check against callers only doing 4028 * I/O when B_CACHE is 0). 4029 */ 4030 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4031 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4032 4033 if (bp->b_flags & B_VMIO) { 4034 vm_object_t obj; 4035 4036 obj = vp->v_object; 4037 KASSERT(bp->b_loffset != NOOFFSET, 4038 ("vfs_busy_pages: no buffer offset")); 4039 4040 /* 4041 * Busy all the pages. We have to busy them all at once 4042 * to avoid deadlocks. 4043 */ 4044 retry: 4045 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4046 vm_page_t m = bp->b_xio.xio_pages[i]; 4047 4048 if (vm_page_busy_try(m, FALSE)) { 4049 vm_page_sleep_busy(m, FALSE, "vbpage"); 4050 while (--i >= 0) 4051 vm_page_wakeup(bp->b_xio.xio_pages[i]); 4052 goto retry; 4053 } 4054 } 4055 4056 /* 4057 * Setup for I/O, soft-busy the page right now because 4058 * the next loop may block. 4059 */ 4060 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4061 vm_page_t m = bp->b_xio.xio_pages[i]; 4062 4063 vm_page_flag_clear(m, PG_ZERO); 4064 if ((bp->b_flags & B_CLUSTER) == 0) { 4065 vm_object_pip_add(obj, 1); 4066 vm_page_io_start(m); 4067 } 4068 } 4069 4070 /* 4071 * Adjust protections for I/O and do bogus-page mapping. 4072 * Assume that vm_page_protect() can block (it can block 4073 * if VM_PROT_NONE, don't take any chances regardless). 4074 * 4075 * In particular note that for writes we must incorporate 4076 * page dirtyness from the VM system into the buffer's 4077 * dirty range. 4078 * 4079 * For reads we theoretically must incorporate page dirtyness 4080 * from the VM system to determine if the page needs bogus 4081 * replacement, but we shortcut the test by simply checking 4082 * that all m->valid bits are set, indicating that the page 4083 * is fully valid and does not need to be re-read. For any 4084 * VM system dirtyness the page will also be fully valid 4085 * since it was mapped at one point. 4086 */ 4087 bogus = 0; 4088 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4089 vm_page_t m = bp->b_xio.xio_pages[i]; 4090 4091 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4092 if (bp->b_cmd == BUF_CMD_WRITE) { 4093 /* 4094 * When readying a vnode-backed buffer for 4095 * a write we must zero-fill any invalid 4096 * portions of the backing VM pages, mark 4097 * it valid and clear related dirty bits. 4098 * 4099 * vfs_clean_one_page() incorporates any 4100 * VM dirtyness and updates the b_dirtyoff 4101 * range (after we've made the page RO). 4102 * 4103 * It is also expected that the pmap modified 4104 * bit has already been cleared by the 4105 * vm_page_protect(). We may not be able 4106 * to clear all dirty bits for a page if it 4107 * was also memory mapped (NFS). 4108 * 4109 * Finally be sure to unassign any swap-cache 4110 * backing store as it is now stale. 4111 */ 4112 vm_page_protect(m, VM_PROT_READ); 4113 vfs_clean_one_page(bp, i, m); 4114 swap_pager_unswapped(m); 4115 } else if (m->valid == VM_PAGE_BITS_ALL) { 4116 /* 4117 * When readying a vnode-backed buffer for 4118 * read we must replace any dirty pages with 4119 * a bogus page so dirty data is not destroyed 4120 * when filling gaps. 4121 * 4122 * To avoid testing whether the page is 4123 * dirty we instead test that the page was 4124 * at some point mapped (m->valid fully 4125 * valid) with the understanding that 4126 * this also covers the dirty case. 4127 */ 4128 bp->b_xio.xio_pages[i] = bogus_page; 4129 bp->b_flags |= B_HASBOGUS; 4130 bogus++; 4131 } else if (m->valid & m->dirty) { 4132 /* 4133 * This case should not occur as partial 4134 * dirtyment can only happen if the buffer 4135 * is B_CACHE, and this code is not entered 4136 * if the buffer is B_CACHE. 4137 */ 4138 kprintf("Warning: vfs_busy_pages - page not " 4139 "fully valid! loff=%jx bpf=%08x " 4140 "idx=%d val=%02x dir=%02x\n", 4141 (uintmax_t)bp->b_loffset, bp->b_flags, 4142 i, m->valid, m->dirty); 4143 vm_page_protect(m, VM_PROT_NONE); 4144 } else { 4145 /* 4146 * The page is not valid and can be made 4147 * part of the read. 4148 */ 4149 vm_page_protect(m, VM_PROT_NONE); 4150 } 4151 vm_page_wakeup(m); 4152 } 4153 if (bogus) { 4154 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4155 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4156 } 4157 } 4158 4159 /* 4160 * This is the easiest place to put the process accounting for the I/O 4161 * for now. 4162 */ 4163 if (lp != NULL) { 4164 if (bp->b_cmd == BUF_CMD_READ) 4165 lp->lwp_ru.ru_inblock++; 4166 else 4167 lp->lwp_ru.ru_oublock++; 4168 } 4169 } 4170 4171 /* 4172 * Tell the VM system that the pages associated with this buffer 4173 * are clean. This is used for delayed writes where the data is 4174 * going to go to disk eventually without additional VM intevention. 4175 * 4176 * NOTE: While we only really need to clean through to b_bcount, we 4177 * just go ahead and clean through to b_bufsize. 4178 */ 4179 static void 4180 vfs_clean_pages(struct buf *bp) 4181 { 4182 vm_page_t m; 4183 int i; 4184 4185 if ((bp->b_flags & B_VMIO) == 0) 4186 return; 4187 4188 KASSERT(bp->b_loffset != NOOFFSET, 4189 ("vfs_clean_pages: no buffer offset")); 4190 4191 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4192 m = bp->b_xio.xio_pages[i]; 4193 vfs_clean_one_page(bp, i, m); 4194 } 4195 } 4196 4197 /* 4198 * vfs_clean_one_page: 4199 * 4200 * Set the valid bits and clear the dirty bits in a page within a 4201 * buffer. The range is restricted to the buffer's size and the 4202 * buffer's logical offset might index into the first page. 4203 * 4204 * The caller has busied or soft-busied the page and it is not mapped, 4205 * test and incorporate the dirty bits into b_dirtyoff/end before 4206 * clearing them. Note that we need to clear the pmap modified bits 4207 * after determining the the page was dirty, vm_page_set_validclean() 4208 * does not do it for us. 4209 * 4210 * This routine is typically called after a read completes (dirty should 4211 * be zero in that case as we are not called on bogus-replace pages), 4212 * or before a write is initiated. 4213 */ 4214 static void 4215 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4216 { 4217 int bcount; 4218 int xoff; 4219 int soff; 4220 int eoff; 4221 4222 /* 4223 * Calculate offset range within the page but relative to buffer's 4224 * loffset. loffset might be offset into the first page. 4225 */ 4226 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4227 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4228 4229 if (pageno == 0) { 4230 soff = xoff; 4231 eoff = PAGE_SIZE; 4232 } else { 4233 soff = (pageno << PAGE_SHIFT); 4234 eoff = soff + PAGE_SIZE; 4235 } 4236 if (eoff > bcount) 4237 eoff = bcount; 4238 if (soff >= eoff) 4239 return; 4240 4241 /* 4242 * Test dirty bits and adjust b_dirtyoff/end. 4243 * 4244 * If dirty pages are incorporated into the bp any prior 4245 * B_NEEDCOMMIT state (NFS) must be cleared because the 4246 * caller has not taken into account the new dirty data. 4247 * 4248 * If the page was memory mapped the dirty bits might go beyond the 4249 * end of the buffer, but we can't really make the assumption that 4250 * a file EOF straddles the buffer (even though this is the case for 4251 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4252 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4253 * This also saves some console spam. 4254 * 4255 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4256 * NFS can handle huge commits but not huge writes. 4257 */ 4258 vm_page_test_dirty(m); 4259 if (m->dirty) { 4260 if ((bp->b_flags & B_NEEDCOMMIT) && 4261 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4262 if (debug_commit) 4263 kprintf("Warning: vfs_clean_one_page: bp %p " 4264 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4265 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4266 "doff/end %d %d\n", 4267 bp, (uintmax_t)bp->b_loffset, bp->b_bcount, 4268 bp->b_flags, bp->b_cmd, 4269 m->valid, m->dirty, xoff, soff, eoff, 4270 bp->b_dirtyoff, bp->b_dirtyend); 4271 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4272 if (debug_commit) 4273 print_backtrace(-1); 4274 } 4275 /* 4276 * Only clear the pmap modified bits if ALL the dirty bits 4277 * are set, otherwise the system might mis-clear portions 4278 * of a page. 4279 */ 4280 if (m->dirty == VM_PAGE_BITS_ALL && 4281 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4282 pmap_clear_modify(m); 4283 } 4284 if (bp->b_dirtyoff > soff - xoff) 4285 bp->b_dirtyoff = soff - xoff; 4286 if (bp->b_dirtyend < eoff - xoff) 4287 bp->b_dirtyend = eoff - xoff; 4288 } 4289 4290 /* 4291 * Set related valid bits, clear related dirty bits. 4292 * Does not mess with the pmap modified bit. 4293 * 4294 * WARNING! We cannot just clear all of m->dirty here as the 4295 * buffer cache buffers may use a DEV_BSIZE'd aligned 4296 * block size, or have an odd size (e.g. NFS at file EOF). 4297 * The putpages code can clear m->dirty to 0. 4298 * 4299 * If a VOP_WRITE generates a buffer cache buffer which 4300 * covers the same space as mapped writable pages the 4301 * buffer flush might not be able to clear all the dirty 4302 * bits and still require a putpages from the VM system 4303 * to finish it off. 4304 * 4305 * WARNING! vm_page_set_validclean() currently assumes vm_token 4306 * is held. The page might not be busied (bdwrite() case). 4307 * XXX remove this comment once we've validated that this 4308 * is no longer an issue. 4309 */ 4310 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4311 } 4312 4313 #if 0 4314 /* 4315 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4316 * The page data is assumed to be valid (there is no zeroing here). 4317 */ 4318 static void 4319 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4320 { 4321 int bcount; 4322 int xoff; 4323 int soff; 4324 int eoff; 4325 4326 /* 4327 * Calculate offset range within the page but relative to buffer's 4328 * loffset. loffset might be offset into the first page. 4329 */ 4330 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4331 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4332 4333 if (pageno == 0) { 4334 soff = xoff; 4335 eoff = PAGE_SIZE; 4336 } else { 4337 soff = (pageno << PAGE_SHIFT); 4338 eoff = soff + PAGE_SIZE; 4339 } 4340 if (eoff > bcount) 4341 eoff = bcount; 4342 if (soff >= eoff) 4343 return; 4344 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4345 } 4346 #endif 4347 4348 /* 4349 * vfs_bio_clrbuf: 4350 * 4351 * Clear a buffer. This routine essentially fakes an I/O, so we need 4352 * to clear B_ERROR and B_INVAL. 4353 * 4354 * Note that while we only theoretically need to clear through b_bcount, 4355 * we go ahead and clear through b_bufsize. 4356 */ 4357 4358 void 4359 vfs_bio_clrbuf(struct buf *bp) 4360 { 4361 int i, mask = 0; 4362 caddr_t sa, ea; 4363 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4364 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4365 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4366 (bp->b_loffset & PAGE_MASK) == 0) { 4367 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4368 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4369 bp->b_resid = 0; 4370 return; 4371 } 4372 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4373 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4374 bzero(bp->b_data, bp->b_bufsize); 4375 bp->b_xio.xio_pages[0]->valid |= mask; 4376 bp->b_resid = 0; 4377 return; 4378 } 4379 } 4380 sa = bp->b_data; 4381 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4382 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4383 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4384 ea = (caddr_t)(vm_offset_t)ulmin( 4385 (u_long)(vm_offset_t)ea, 4386 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4387 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4388 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4389 continue; 4390 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4391 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4392 bzero(sa, ea - sa); 4393 } 4394 } else { 4395 for (; sa < ea; sa += DEV_BSIZE, j++) { 4396 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4397 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4398 bzero(sa, DEV_BSIZE); 4399 } 4400 } 4401 bp->b_xio.xio_pages[i]->valid |= mask; 4402 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4403 } 4404 bp->b_resid = 0; 4405 } else { 4406 clrbuf(bp); 4407 } 4408 } 4409 4410 /* 4411 * vm_hold_load_pages: 4412 * 4413 * Load pages into the buffer's address space. The pages are 4414 * allocated from the kernel object in order to reduce interference 4415 * with the any VM paging I/O activity. The range of loaded 4416 * pages will be wired. 4417 * 4418 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4419 * retrieve the full range (to - from) of pages. 4420 */ 4421 void 4422 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4423 { 4424 vm_offset_t pg; 4425 vm_page_t p; 4426 int index; 4427 4428 to = round_page(to); 4429 from = round_page(from); 4430 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4431 4432 pg = from; 4433 while (pg < to) { 4434 /* 4435 * Note: must allocate system pages since blocking here 4436 * could intefere with paging I/O, no matter which 4437 * process we are. 4438 */ 4439 vm_object_hold(&kernel_object); 4440 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT, 4441 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4442 vm_object_drop(&kernel_object); 4443 if (p) { 4444 vm_page_wire(p); 4445 p->valid = VM_PAGE_BITS_ALL; 4446 vm_page_flag_clear(p, PG_ZERO); 4447 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 4448 bp->b_xio.xio_pages[index] = p; 4449 vm_page_wakeup(p); 4450 4451 pg += PAGE_SIZE; 4452 ++index; 4453 } 4454 } 4455 bp->b_xio.xio_npages = index; 4456 } 4457 4458 /* 4459 * Allocate a page for a buffer cache buffer. 4460 * 4461 * If NULL is returned the caller is expected to retry (typically check if 4462 * the page already exists on retry before trying to allocate one). 4463 * 4464 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This 4465 * function will use the system reserve with the hope that the page 4466 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller 4467 * is done with the buffer. 4468 * 4469 * NOTE! However, TMPFS is a special case because flushing a dirty buffer 4470 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon 4471 * is capable of retiring pages (to swap). For TMPFS we don't dig 4472 * into the system reserve because doing so could stall out pretty 4473 * much every process running on the system. 4474 */ 4475 static 4476 vm_page_t 4477 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit) 4478 { 4479 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK; 4480 vm_page_t p; 4481 4482 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4483 4484 /* 4485 * Try a normal allocation first. 4486 */ 4487 p = vm_page_alloc(obj, pg, vmflags); 4488 if (p) 4489 return(p); 4490 if (vm_page_lookup(obj, pg)) 4491 return(NULL); 4492 vm_pageout_deficit += deficit; 4493 4494 /* 4495 * Try again, digging into the system reserve. 4496 * 4497 * Trying to recover pages from the buffer cache here can deadlock 4498 * against other threads trying to busy underlying pages so we 4499 * depend on the code in brelse() and bqrelse() to free/cache the 4500 * underlying buffer cache pages when memory is low. 4501 */ 4502 if (curthread->td_flags & TDF_SYSTHREAD) 4503 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT; 4504 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS) 4505 vmflags |= 0; 4506 else 4507 vmflags |= VM_ALLOC_SYSTEM; 4508 4509 /*recoverbufpages();*/ 4510 p = vm_page_alloc(obj, pg, vmflags); 4511 if (p) 4512 return(p); 4513 if (vm_page_lookup(obj, pg)) 4514 return(NULL); 4515 4516 /* 4517 * Wait for memory to free up and try again 4518 */ 4519 if (vm_page_count_severe()) 4520 ++lowmempgallocs; 4521 vm_wait(hz / 20 + 1); 4522 4523 p = vm_page_alloc(obj, pg, vmflags); 4524 if (p) 4525 return(p); 4526 if (vm_page_lookup(obj, pg)) 4527 return(NULL); 4528 4529 /* 4530 * Ok, now we are really in trouble. 4531 */ 4532 { 4533 static struct krate biokrate = { .freq = 1 }; 4534 krateprintf(&biokrate, 4535 "Warning: bio_page_alloc: memory exhausted " 4536 "during bufcache page allocation from %s\n", 4537 curthread->td_comm); 4538 } 4539 if (curthread->td_flags & TDF_SYSTHREAD) 4540 vm_wait(hz / 20 + 1); 4541 else 4542 vm_wait(hz / 2 + 1); 4543 return (NULL); 4544 } 4545 4546 /* 4547 * vm_hold_free_pages: 4548 * 4549 * Return pages associated with the buffer back to the VM system. 4550 * 4551 * The range of pages underlying the buffer's address space will 4552 * be unmapped and un-wired. 4553 */ 4554 void 4555 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4556 { 4557 vm_offset_t pg; 4558 vm_page_t p; 4559 int index, newnpages; 4560 4561 from = round_page(from); 4562 to = round_page(to); 4563 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4564 newnpages = index; 4565 4566 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4567 p = bp->b_xio.xio_pages[index]; 4568 if (p && (index < bp->b_xio.xio_npages)) { 4569 if (p->busy) { 4570 kprintf("vm_hold_free_pages: doffset: %lld, " 4571 "loffset: %lld\n", 4572 (long long)bp->b_bio2.bio_offset, 4573 (long long)bp->b_loffset); 4574 } 4575 bp->b_xio.xio_pages[index] = NULL; 4576 pmap_kremove(pg); 4577 vm_page_busy_wait(p, FALSE, "vmhldpg"); 4578 vm_page_unwire(p, 0); 4579 vm_page_free(p); 4580 } 4581 } 4582 bp->b_xio.xio_npages = newnpages; 4583 } 4584 4585 /* 4586 * vmapbuf: 4587 * 4588 * Map a user buffer into KVM via a pbuf. On return the buffer's 4589 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4590 * initialized. 4591 */ 4592 int 4593 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4594 { 4595 caddr_t addr; 4596 vm_offset_t va; 4597 vm_page_t m; 4598 int vmprot; 4599 int error; 4600 int pidx; 4601 int i; 4602 4603 /* 4604 * bp had better have a command and it better be a pbuf. 4605 */ 4606 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4607 KKASSERT(bp->b_flags & B_PAGING); 4608 KKASSERT(bp->b_kvabase); 4609 4610 if (bytes < 0) 4611 return (-1); 4612 4613 /* 4614 * Map the user data into KVM. Mappings have to be page-aligned. 4615 */ 4616 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4617 pidx = 0; 4618 4619 vmprot = VM_PROT_READ; 4620 if (bp->b_cmd == BUF_CMD_READ) 4621 vmprot |= VM_PROT_WRITE; 4622 4623 while (addr < udata + bytes) { 4624 /* 4625 * Do the vm_fault if needed; do the copy-on-write thing 4626 * when reading stuff off device into memory. 4627 * 4628 * vm_fault_page*() returns a held VM page. 4629 */ 4630 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4631 va = trunc_page(va); 4632 4633 m = vm_fault_page_quick(va, vmprot, &error); 4634 if (m == NULL) { 4635 for (i = 0; i < pidx; ++i) { 4636 vm_page_unhold(bp->b_xio.xio_pages[i]); 4637 bp->b_xio.xio_pages[i] = NULL; 4638 } 4639 return(-1); 4640 } 4641 bp->b_xio.xio_pages[pidx] = m; 4642 addr += PAGE_SIZE; 4643 ++pidx; 4644 } 4645 4646 /* 4647 * Map the page array and set the buffer fields to point to 4648 * the mapped data buffer. 4649 */ 4650 if (pidx > btoc(MAXPHYS)) 4651 panic("vmapbuf: mapped more than MAXPHYS"); 4652 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4653 4654 bp->b_xio.xio_npages = pidx; 4655 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4656 bp->b_bcount = bytes; 4657 bp->b_bufsize = bytes; 4658 return(0); 4659 } 4660 4661 /* 4662 * vunmapbuf: 4663 * 4664 * Free the io map PTEs associated with this IO operation. 4665 * We also invalidate the TLB entries and restore the original b_addr. 4666 */ 4667 void 4668 vunmapbuf(struct buf *bp) 4669 { 4670 int pidx; 4671 int npages; 4672 4673 KKASSERT(bp->b_flags & B_PAGING); 4674 4675 npages = bp->b_xio.xio_npages; 4676 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4677 for (pidx = 0; pidx < npages; ++pidx) { 4678 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4679 bp->b_xio.xio_pages[pidx] = NULL; 4680 } 4681 bp->b_xio.xio_npages = 0; 4682 bp->b_data = bp->b_kvabase; 4683 } 4684 4685 /* 4686 * Scan all buffers in the system and issue the callback. 4687 */ 4688 int 4689 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4690 { 4691 int count = 0; 4692 int error; 4693 long n; 4694 4695 for (n = 0; n < nbuf; ++n) { 4696 if ((error = callback(&buf[n], info)) < 0) { 4697 count = error; 4698 break; 4699 } 4700 count += error; 4701 } 4702 return (count); 4703 } 4704 4705 /* 4706 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4707 * completion to the master buffer. 4708 */ 4709 static void 4710 nestiobuf_iodone(struct bio *bio) 4711 { 4712 struct bio *mbio; 4713 struct buf *mbp, *bp; 4714 struct devstat *stats; 4715 int error; 4716 int donebytes; 4717 4718 bp = bio->bio_buf; 4719 mbio = bio->bio_caller_info1.ptr; 4720 stats = bio->bio_caller_info2.ptr; 4721 mbp = mbio->bio_buf; 4722 4723 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4724 KKASSERT(mbp != bp); 4725 4726 error = bp->b_error; 4727 if (bp->b_error == 0 && 4728 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4729 /* 4730 * Not all got transfered, raise an error. We have no way to 4731 * propagate these conditions to mbp. 4732 */ 4733 error = EIO; 4734 } 4735 4736 donebytes = bp->b_bufsize; 4737 4738 relpbuf(bp, NULL); 4739 4740 nestiobuf_done(mbio, donebytes, error, stats); 4741 } 4742 4743 void 4744 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4745 { 4746 struct buf *mbp; 4747 4748 mbp = mbio->bio_buf; 4749 4750 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4751 4752 /* 4753 * If an error occured, propagate it to the master buffer. 4754 * 4755 * Several biodone()s may wind up running concurrently so 4756 * use an atomic op to adjust b_flags. 4757 */ 4758 if (error) { 4759 mbp->b_error = error; 4760 atomic_set_int(&mbp->b_flags, B_ERROR); 4761 } 4762 4763 /* 4764 * Decrement the operations in progress counter and terminate the 4765 * I/O if this was the last bit. 4766 */ 4767 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4768 mbp->b_resid = 0; 4769 if (stats) 4770 devstat_end_transaction_buf(stats, mbp); 4771 biodone(mbio); 4772 } 4773 } 4774 4775 /* 4776 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4777 * the mbio from being biodone()'d while we are still adding sub-bios to 4778 * it. 4779 */ 4780 void 4781 nestiobuf_init(struct bio *bio) 4782 { 4783 bio->bio_driver_info = (void *)1; 4784 } 4785 4786 /* 4787 * The BIOs added to the nestedio have already been started, remove the 4788 * count that placeheld our mbio and biodone() it if the count would 4789 * transition to 0. 4790 */ 4791 void 4792 nestiobuf_start(struct bio *mbio) 4793 { 4794 struct buf *mbp = mbio->bio_buf; 4795 4796 /* 4797 * Decrement the operations in progress counter and terminate the 4798 * I/O if this was the last bit. 4799 */ 4800 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4801 if (mbp->b_flags & B_ERROR) 4802 mbp->b_resid = mbp->b_bcount; 4803 else 4804 mbp->b_resid = 0; 4805 biodone(mbio); 4806 } 4807 } 4808 4809 /* 4810 * Set an intermediate error prior to calling nestiobuf_start() 4811 */ 4812 void 4813 nestiobuf_error(struct bio *mbio, int error) 4814 { 4815 struct buf *mbp = mbio->bio_buf; 4816 4817 if (error) { 4818 mbp->b_error = error; 4819 atomic_set_int(&mbp->b_flags, B_ERROR); 4820 } 4821 } 4822 4823 /* 4824 * nestiobuf_add: setup a "nested" buffer. 4825 * 4826 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4827 * => 'bp' should be a buffer allocated by getiobuf. 4828 * => 'offset' is a byte offset in the master buffer. 4829 * => 'size' is a size in bytes of this nested buffer. 4830 */ 4831 void 4832 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4833 { 4834 struct buf *mbp = mbio->bio_buf; 4835 struct vnode *vp = mbp->b_vp; 4836 4837 KKASSERT(mbp->b_bcount >= offset + size); 4838 4839 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4840 4841 /* kernel needs to own the lock for it to be released in biodone */ 4842 BUF_KERNPROC(bp); 4843 bp->b_vp = vp; 4844 bp->b_cmd = mbp->b_cmd; 4845 bp->b_bio1.bio_done = nestiobuf_iodone; 4846 bp->b_data = (char *)mbp->b_data + offset; 4847 bp->b_resid = bp->b_bcount = size; 4848 bp->b_bufsize = bp->b_bcount; 4849 4850 bp->b_bio1.bio_track = NULL; 4851 bp->b_bio1.bio_caller_info1.ptr = mbio; 4852 bp->b_bio1.bio_caller_info2.ptr = stats; 4853 } 4854 4855 #ifdef DDB 4856 4857 DB_SHOW_COMMAND(buffer, db_show_buffer) 4858 { 4859 /* get args */ 4860 struct buf *bp = (struct buf *)addr; 4861 4862 if (!have_addr) { 4863 db_printf("usage: show buffer <addr>\n"); 4864 return; 4865 } 4866 4867 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4868 db_printf("b_cmd = %d\n", bp->b_cmd); 4869 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4870 "b_resid = %d\n, b_data = %p, " 4871 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4872 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4873 bp->b_data, 4874 (long long)bp->b_bio2.bio_offset, 4875 (long long)(bp->b_bio2.bio_next ? 4876 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4877 if (bp->b_xio.xio_npages) { 4878 int i; 4879 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4880 bp->b_xio.xio_npages); 4881 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4882 vm_page_t m; 4883 m = bp->b_xio.xio_pages[i]; 4884 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4885 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4886 if ((i + 1) < bp->b_xio.xio_npages) 4887 db_printf(","); 4888 } 4889 db_printf("\n"); 4890 } 4891 } 4892 #endif /* DDB */ 4893