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 bp->b_flags = B_BNOCLIP; 2173 bp->b_cmd = BUF_CMD_DONE; 2174 bp->b_vp = NULL; 2175 bp->b_error = 0; 2176 bp->b_resid = 0; 2177 bp->b_bcount = 0; 2178 bp->b_xio.xio_npages = 0; 2179 bp->b_dirtyoff = bp->b_dirtyend = 0; 2180 bp->b_act_count = ACT_INIT; 2181 reinitbufbio(bp); 2182 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 2183 buf_dep_init(bp); 2184 if (blkflags & GETBLK_BHEAVY) 2185 bp->b_flags |= B_HEAVY; 2186 2187 /* 2188 * If we are defragging then free the buffer. 2189 */ 2190 if (defrag) { 2191 bp->b_flags |= B_INVAL; 2192 bfreekva(bp); 2193 brelse(bp); 2194 defrag = 0; 2195 restart_reason = 4; 2196 restart_bp = bp; 2197 goto restart; 2198 } 2199 2200 /* 2201 * If we are overcomitted then recover the buffer and its 2202 * KVM space. This occurs in rare situations when multiple 2203 * processes are blocked in getnewbuf() or allocbuf(). 2204 * 2205 * On 64-bit systems BKVASIZE == MAXBSIZE and overcommit 2206 * should not be possible. 2207 */ 2208 if (bufspace >= hibufspace) 2209 flushingbufs = 1; 2210 if (BKVASIZE != MAXBSIZE) { 2211 if (flushingbufs && bp->b_kvasize != 0) { 2212 bp->b_flags |= B_INVAL; 2213 bfreekva(bp); 2214 brelse(bp); 2215 restart_reason = 5; 2216 restart_bp = bp; 2217 goto restart; 2218 } 2219 } 2220 if (bufspace < lobufspace) 2221 flushingbufs = 0; 2222 2223 /* 2224 * b_refs can transition to a non-zero value while we hold 2225 * the buffer locked due to a findblk(). Our brelvp() above 2226 * interlocked any future possible transitions due to 2227 * findblk()s. 2228 * 2229 * If we find b_refs to be non-zero we can destroy the 2230 * buffer's contents but we cannot yet reuse the buffer. 2231 */ 2232 if (bp->b_refs) { 2233 bp->b_flags |= B_INVAL; 2234 if (BKVASIZE != MAXBSIZE) 2235 bfreekva(bp); 2236 brelse(bp); 2237 restart_reason = 6; 2238 restart_bp = bp; 2239 goto restart; 2240 } 2241 break; 2242 /* NOT REACHED, spin not held */ 2243 } 2244 2245 /* 2246 * If we exhausted our list, iterate other cpus. If that fails, 2247 * sleep as appropriate. We may have to wakeup various daemons 2248 * and write out some dirty buffers. 2249 * 2250 * Generally we are sleeping due to insufficient buffer space. 2251 * 2252 * NOTE: spin is held if bp is NULL, else it is not held. 2253 */ 2254 if (bp == NULL) { 2255 int flags; 2256 char *waitmsg; 2257 2258 spin_unlock(&pcpu->spin); 2259 2260 nqcpu = (nqcpu + 1) % ncpus; 2261 if (nqcpu != mycpu->gd_cpuid) { 2262 restart_reason = 7; 2263 restart_bp = bp; 2264 goto restart; 2265 } 2266 2267 if (defrag) { 2268 flags = VFS_BIO_NEED_BUFSPACE; 2269 waitmsg = "nbufkv"; 2270 } else if (bufspace >= hibufspace) { 2271 waitmsg = "nbufbs"; 2272 flags = VFS_BIO_NEED_BUFSPACE; 2273 } else { 2274 waitmsg = "newbuf"; 2275 flags = VFS_BIO_NEED_ANY; 2276 } 2277 2278 bd_speedup(); /* heeeelp */ 2279 atomic_set_int(&needsbuffer, flags); 2280 while (needsbuffer & flags) { 2281 int value; 2282 2283 tsleep_interlock(&needsbuffer, 0); 2284 value = atomic_fetchadd_int(&needsbuffer, 0); 2285 if (value & flags) { 2286 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags, 2287 waitmsg, slptimeo)) { 2288 return (NULL); 2289 } 2290 } 2291 } 2292 } else { 2293 /* 2294 * We finally have a valid bp. We aren't quite out of the 2295 * woods, we still have to reserve kva space. In order 2296 * to keep fragmentation sane we only allocate kva in 2297 * BKVASIZE chunks. 2298 * 2299 * (spin is not held) 2300 */ 2301 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2302 2303 if (maxsize != bp->b_kvasize) { 2304 vm_offset_t addr = 0; 2305 int count; 2306 2307 bfreekva(bp); 2308 2309 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2310 vm_map_lock(&buffer_map); 2311 2312 if (vm_map_findspace(&buffer_map, 2313 vm_map_min(&buffer_map), maxsize, 2314 maxsize, 0, &addr)) { 2315 /* 2316 * Uh oh. Buffer map is too fragmented. We 2317 * must defragment the map. 2318 */ 2319 vm_map_unlock(&buffer_map); 2320 vm_map_entry_release(count); 2321 ++bufdefragcnt; 2322 defrag = 1; 2323 bp->b_flags |= B_INVAL; 2324 brelse(bp); 2325 restart_reason = 8; 2326 restart_bp = bp; 2327 goto restart; 2328 } 2329 if (addr) { 2330 vm_map_insert(&buffer_map, &count, 2331 NULL, 0, 2332 addr, addr + maxsize, 2333 VM_MAPTYPE_NORMAL, 2334 VM_PROT_ALL, VM_PROT_ALL, 2335 MAP_NOFAULT); 2336 2337 bp->b_kvabase = (caddr_t) addr; 2338 bp->b_kvasize = maxsize; 2339 bufspace += bp->b_kvasize; 2340 ++bufreusecnt; 2341 } 2342 vm_map_unlock(&buffer_map); 2343 vm_map_entry_release(count); 2344 } 2345 bp->b_data = bp->b_kvabase; 2346 } 2347 return(bp); 2348 } 2349 2350 /* 2351 * buf_daemon: 2352 * 2353 * Buffer flushing daemon. Buffers are normally flushed by the 2354 * update daemon but if it cannot keep up this process starts to 2355 * take the load in an attempt to prevent getnewbuf() from blocking. 2356 * 2357 * Once a flush is initiated it does not stop until the number 2358 * of buffers falls below lodirtybuffers, but we will wake up anyone 2359 * waiting at the mid-point. 2360 */ 2361 static struct kproc_desc buf_kp = { 2362 "bufdaemon", 2363 buf_daemon, 2364 &bufdaemon_td 2365 }; 2366 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2367 kproc_start, &buf_kp) 2368 2369 static struct kproc_desc bufhw_kp = { 2370 "bufdaemon_hw", 2371 buf_daemon_hw, 2372 &bufdaemonhw_td 2373 }; 2374 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, 2375 kproc_start, &bufhw_kp) 2376 2377 static void 2378 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long), 2379 int *bd_req) 2380 { 2381 long limit; 2382 struct buf *marker; 2383 2384 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO); 2385 marker->b_flags |= B_MARKER; 2386 marker->b_qindex = BQUEUE_NONE; 2387 marker->b_qcpu = 0; 2388 2389 /* 2390 * This process needs to be suspended prior to shutdown sync. 2391 */ 2392 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 2393 td, SHUTDOWN_PRI_LAST); 2394 curthread->td_flags |= TDF_SYSTHREAD; 2395 2396 /* 2397 * This process is allowed to take the buffer cache to the limit 2398 */ 2399 for (;;) { 2400 kproc_suspend_loop(); 2401 2402 /* 2403 * Do the flush as long as the number of dirty buffers 2404 * (including those running) exceeds lodirtybufspace. 2405 * 2406 * When flushing limit running I/O to hirunningspace 2407 * Do the flush. Limit the amount of in-transit I/O we 2408 * allow to build up, otherwise we would completely saturate 2409 * the I/O system. Wakeup any waiting processes before we 2410 * normally would so they can run in parallel with our drain. 2411 * 2412 * Our aggregate normal+HW lo water mark is lodirtybufspace, 2413 * but because we split the operation into two threads we 2414 * have to cut it in half for each thread. 2415 */ 2416 waitrunningbufspace(); 2417 limit = lodirtybufspace / 2; 2418 while (buf_limit_fn(limit)) { 2419 if (flushbufqueues(marker, queue) == 0) 2420 break; 2421 if (runningbufspace < hirunningspace) 2422 continue; 2423 waitrunningbufspace(); 2424 } 2425 2426 /* 2427 * We reached our low water mark, reset the 2428 * request and sleep until we are needed again. 2429 * The sleep is just so the suspend code works. 2430 */ 2431 tsleep_interlock(bd_req, 0); 2432 if (atomic_swap_int(bd_req, 0) == 0) 2433 tsleep(bd_req, PINTERLOCKED, "psleep", hz); 2434 } 2435 /* NOT REACHED */ 2436 /*kfree(marker, M_BIOBUF);*/ 2437 } 2438 2439 static int 2440 buf_daemon_limit(long limit) 2441 { 2442 return (runningbufspace + dirtykvaspace > limit || 2443 dirtybufcount - dirtybufcounthw >= nbuf / 2); 2444 } 2445 2446 static int 2447 buf_daemon_hw_limit(long limit) 2448 { 2449 return (runningbufspace + dirtykvaspace > limit || 2450 dirtybufcounthw >= nbuf / 2); 2451 } 2452 2453 static void 2454 buf_daemon(void) 2455 { 2456 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit, 2457 &bd_request); 2458 } 2459 2460 static void 2461 buf_daemon_hw(void) 2462 { 2463 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit, 2464 &bd_request_hw); 2465 } 2466 2467 /* 2468 * flushbufqueues: 2469 * 2470 * Try to flush a buffer in the dirty queue. We must be careful to 2471 * free up B_INVAL buffers instead of write them, which NFS is 2472 * particularly sensitive to. 2473 * 2474 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate 2475 * that we really want to try to get the buffer out and reuse it 2476 * due to the write load on the machine. 2477 * 2478 * We must lock the buffer in order to check its validity before we 2479 * can mess with its contents. spin isn't enough. 2480 */ 2481 static int 2482 flushbufqueues(struct buf *marker, bufq_type_t q) 2483 { 2484 struct bufpcpu *pcpu; 2485 struct buf *bp; 2486 int r = 0; 2487 int lcpu = marker->b_qcpu; 2488 2489 KKASSERT(marker->b_qindex == BQUEUE_NONE); 2490 KKASSERT(marker->b_flags & B_MARKER); 2491 2492 again: 2493 /* 2494 * Spinlock needed to perform operations on the queue and may be 2495 * held through a non-blocking BUF_LOCK(), but cannot be held when 2496 * BUF_UNLOCK()ing or through any other major operation. 2497 */ 2498 pcpu = &bufpcpu[marker->b_qcpu]; 2499 spin_lock(&pcpu->spin); 2500 marker->b_qindex = q; 2501 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist); 2502 bp = marker; 2503 2504 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) { 2505 /* 2506 * NOTE: spinlock is always held at the top of the loop 2507 */ 2508 if (bp->b_flags & B_MARKER) 2509 continue; 2510 if ((bp->b_flags & B_DELWRI) == 0) { 2511 kprintf("Unexpected clean buffer %p\n", bp); 2512 continue; 2513 } 2514 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 2515 continue; 2516 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q); 2517 2518 /* 2519 * Once the buffer is locked we will have no choice but to 2520 * unlock the spinlock around a later BUF_UNLOCK and re-set 2521 * bp = marker when looping. Move the marker now to make 2522 * things easier. 2523 */ 2524 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2525 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist); 2526 2527 /* 2528 * Must recheck B_DELWRI after successfully locking 2529 * the buffer. 2530 */ 2531 if ((bp->b_flags & B_DELWRI) == 0) { 2532 spin_unlock(&pcpu->spin); 2533 BUF_UNLOCK(bp); 2534 spin_lock(&pcpu->spin); 2535 bp = marker; 2536 continue; 2537 } 2538 2539 /* 2540 * Remove the buffer from its queue. We still own the 2541 * spinlock here. 2542 */ 2543 _bremfree(bp); 2544 2545 /* 2546 * Disposing of an invalid buffer counts as a flush op 2547 */ 2548 if (bp->b_flags & B_INVAL) { 2549 spin_unlock(&pcpu->spin); 2550 brelse(bp); 2551 spin_lock(&pcpu->spin); 2552 ++r; 2553 break; 2554 } 2555 2556 /* 2557 * Release the spinlock for the more complex ops we 2558 * are now going to do. 2559 */ 2560 spin_unlock(&pcpu->spin); 2561 lwkt_yield(); 2562 2563 /* 2564 * This is a bit messy 2565 */ 2566 if (LIST_FIRST(&bp->b_dep) != NULL && 2567 (bp->b_flags & B_DEFERRED) == 0 && 2568 buf_countdeps(bp, 0)) { 2569 spin_lock(&pcpu->spin); 2570 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist); 2571 bp->b_qindex = q; 2572 bp->b_flags |= B_DEFERRED; 2573 spin_unlock(&pcpu->spin); 2574 BUF_UNLOCK(bp); 2575 spin_lock(&pcpu->spin); 2576 bp = marker; 2577 continue; 2578 } 2579 2580 /* 2581 * spinlock not held here. 2582 * 2583 * If the buffer has a dependancy, buf_checkwrite() must 2584 * also return 0 for us to be able to initate the write. 2585 * 2586 * If the buffer is flagged B_ERROR it may be requeued 2587 * over and over again, we try to avoid a live lock. 2588 */ 2589 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 2590 brelse(bp); 2591 } else if (bp->b_flags & B_ERROR) { 2592 tsleep(bp, 0, "bioer", 1); 2593 bp->b_flags &= ~B_AGE; 2594 cluster_awrite(bp); 2595 } else { 2596 bp->b_flags |= B_AGE; 2597 cluster_awrite(bp); 2598 } 2599 spin_lock(&pcpu->spin); 2600 ++r; 2601 break; 2602 } 2603 2604 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist); 2605 marker->b_qindex = BQUEUE_NONE; 2606 spin_unlock(&pcpu->spin); 2607 2608 /* 2609 * Advance the marker to be fair. 2610 */ 2611 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus; 2612 if (bp == NULL) { 2613 if (marker->b_qcpu != lcpu) 2614 goto again; 2615 } 2616 2617 return (r); 2618 } 2619 2620 /* 2621 * inmem: 2622 * 2623 * Returns true if no I/O is needed to access the associated VM object. 2624 * This is like findblk except it also hunts around in the VM system for 2625 * the data. 2626 * 2627 * Note that we ignore vm_page_free() races from interrupts against our 2628 * lookup, since if the caller is not protected our return value will not 2629 * be any more valid then otherwise once we exit the critical section. 2630 */ 2631 int 2632 inmem(struct vnode *vp, off_t loffset) 2633 { 2634 vm_object_t obj; 2635 vm_offset_t toff, tinc, size; 2636 vm_page_t m; 2637 int res = 1; 2638 2639 if (findblk(vp, loffset, FINDBLK_TEST)) 2640 return 1; 2641 if (vp->v_mount == NULL) 2642 return 0; 2643 if ((obj = vp->v_object) == NULL) 2644 return 0; 2645 2646 size = PAGE_SIZE; 2647 if (size > vp->v_mount->mnt_stat.f_iosize) 2648 size = vp->v_mount->mnt_stat.f_iosize; 2649 2650 vm_object_hold(obj); 2651 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2652 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 2653 if (m == NULL) { 2654 res = 0; 2655 break; 2656 } 2657 tinc = size; 2658 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 2659 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 2660 if (vm_page_is_valid(m, 2661 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) { 2662 res = 0; 2663 break; 2664 } 2665 } 2666 vm_object_drop(obj); 2667 return (res); 2668 } 2669 2670 /* 2671 * findblk: 2672 * 2673 * Locate and return the specified buffer. Unless flagged otherwise, 2674 * a locked buffer will be returned if it exists or NULL if it does not. 2675 * 2676 * findblk()'d buffers are still on the bufqueues and if you intend 2677 * to use your (locked NON-TEST) buffer you need to bremfree(bp) 2678 * and possibly do other stuff to it. 2679 * 2680 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible 2681 * for locking the buffer and ensuring that it remains 2682 * the desired buffer after locking. 2683 * 2684 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable 2685 * to acquire the lock we return NULL, even if the 2686 * buffer exists. 2687 * 2688 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal 2689 * reuse by getnewbuf() but does not prevent 2690 * disassociation (B_INVAL). Used to avoid deadlocks 2691 * against random (vp,loffset)s due to reassignment. 2692 * 2693 * (0) - Lock the buffer blocking. 2694 */ 2695 struct buf * 2696 findblk(struct vnode *vp, off_t loffset, int flags) 2697 { 2698 struct buf *bp; 2699 int lkflags; 2700 2701 lkflags = LK_EXCLUSIVE; 2702 if (flags & FINDBLK_NBLOCK) 2703 lkflags |= LK_NOWAIT; 2704 2705 for (;;) { 2706 /* 2707 * Lookup. Ref the buf while holding v_token to prevent 2708 * reuse (but does not prevent diassociation). 2709 */ 2710 lwkt_gettoken_shared(&vp->v_token); 2711 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2712 if (bp == NULL) { 2713 lwkt_reltoken(&vp->v_token); 2714 return(NULL); 2715 } 2716 bqhold(bp); 2717 lwkt_reltoken(&vp->v_token); 2718 2719 /* 2720 * If testing only break and return bp, do not lock. 2721 */ 2722 if (flags & FINDBLK_TEST) 2723 break; 2724 2725 /* 2726 * Lock the buffer, return an error if the lock fails. 2727 * (only FINDBLK_NBLOCK can cause the lock to fail). 2728 */ 2729 if (BUF_LOCK(bp, lkflags)) { 2730 atomic_subtract_int(&bp->b_refs, 1); 2731 /* bp = NULL; not needed */ 2732 return(NULL); 2733 } 2734 2735 /* 2736 * Revalidate the locked buf before allowing it to be 2737 * returned. 2738 */ 2739 if (bp->b_vp == vp && bp->b_loffset == loffset) 2740 break; 2741 atomic_subtract_int(&bp->b_refs, 1); 2742 BUF_UNLOCK(bp); 2743 } 2744 2745 /* 2746 * Success 2747 */ 2748 if ((flags & FINDBLK_REF) == 0) 2749 atomic_subtract_int(&bp->b_refs, 1); 2750 return(bp); 2751 } 2752 2753 /* 2754 * getcacheblk: 2755 * 2756 * Similar to getblk() except only returns the buffer if it is 2757 * B_CACHE and requires no other manipulation. Otherwise NULL 2758 * is returned. NULL is also returned if GETBLK_NOWAIT is set 2759 * and the getblk() would block. 2760 * 2761 * If B_RAM is set the buffer might be just fine, but we return 2762 * NULL anyway because we want the code to fall through to the 2763 * cluster read. Otherwise read-ahead breaks. 2764 * 2765 * If blksize is 0 the buffer cache buffer must already be fully 2766 * cached. 2767 * 2768 * If blksize is non-zero getblk() will be used, allowing a buffer 2769 * to be reinstantiated from its VM backing store. The buffer must 2770 * still be fully cached after reinstantiation to be returned. 2771 */ 2772 struct buf * 2773 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags) 2774 { 2775 struct buf *bp; 2776 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0; 2777 2778 if (blksize) { 2779 bp = getblk(vp, loffset, blksize, blkflags, 0); 2780 if (bp) { 2781 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2782 B_CACHE) { 2783 bp->b_flags &= ~B_AGE; 2784 } else { 2785 brelse(bp); 2786 bp = NULL; 2787 } 2788 } 2789 } else { 2790 bp = findblk(vp, loffset, fndflags); 2791 if (bp) { 2792 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == 2793 B_CACHE) { 2794 bp->b_flags &= ~B_AGE; 2795 bremfree(bp); 2796 } else { 2797 BUF_UNLOCK(bp); 2798 bp = NULL; 2799 } 2800 } 2801 } 2802 return (bp); 2803 } 2804 2805 /* 2806 * getblk: 2807 * 2808 * Get a block given a specified block and offset into a file/device. 2809 * B_INVAL may or may not be set on return. The caller should clear 2810 * B_INVAL prior to initiating a READ. 2811 * 2812 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2813 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2814 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2815 * without doing any of those things the system will likely believe 2816 * the buffer to be valid (especially if it is not B_VMIO), and the 2817 * next getblk() will return the buffer with B_CACHE set. 2818 * 2819 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2820 * an existing buffer. 2821 * 2822 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2823 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2824 * and then cleared based on the backing VM. If the previous buffer is 2825 * non-0-sized but invalid, B_CACHE will be cleared. 2826 * 2827 * If getblk() must create a new buffer, the new buffer is returned with 2828 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2829 * case it is returned with B_INVAL clear and B_CACHE set based on the 2830 * backing VM. 2831 * 2832 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2833 * B_CACHE bit is clear. 2834 * 2835 * What this means, basically, is that the caller should use B_CACHE to 2836 * determine whether the buffer is fully valid or not and should clear 2837 * B_INVAL prior to issuing a read. If the caller intends to validate 2838 * the buffer by loading its data area with something, the caller needs 2839 * to clear B_INVAL. If the caller does this without issuing an I/O, 2840 * the caller should set B_CACHE ( as an optimization ), else the caller 2841 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2842 * a write attempt or if it was a successfull read. If the caller 2843 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2844 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2845 * 2846 * getblk flags: 2847 * 2848 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return 2849 * GETBLK_BHEAVY - heavy-weight buffer cache buffer 2850 */ 2851 struct buf * 2852 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo) 2853 { 2854 struct buf *bp; 2855 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0; 2856 int error; 2857 int lkflags; 2858 2859 if (size > MAXBSIZE) 2860 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2861 if (vp->v_object == NULL) 2862 panic("getblk: vnode %p has no object!", vp); 2863 2864 loop: 2865 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) { 2866 /* 2867 * The buffer was found in the cache, but we need to lock it. 2868 * We must acquire a ref on the bp to prevent reuse, but 2869 * this will not prevent disassociation (brelvp()) so we 2870 * must recheck (vp,loffset) after acquiring the lock. 2871 * 2872 * Without the ref the buffer could potentially be reused 2873 * before we acquire the lock and create a deadlock 2874 * situation between the thread trying to reuse the buffer 2875 * and us due to the fact that we would wind up blocking 2876 * on a random (vp,loffset). 2877 */ 2878 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2879 if (blkflags & GETBLK_NOWAIT) { 2880 bqdrop(bp); 2881 return(NULL); 2882 } 2883 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2884 if (blkflags & GETBLK_PCATCH) 2885 lkflags |= LK_PCATCH; 2886 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo); 2887 if (error) { 2888 bqdrop(bp); 2889 if (error == ENOLCK) 2890 goto loop; 2891 return (NULL); 2892 } 2893 /* buffer may have changed on us */ 2894 } 2895 bqdrop(bp); 2896 2897 /* 2898 * Once the buffer has been locked, make sure we didn't race 2899 * a buffer recyclement. Buffers that are no longer hashed 2900 * will have b_vp == NULL, so this takes care of that check 2901 * as well. 2902 */ 2903 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2904 kprintf("Warning buffer %p (vp %p loffset %lld) " 2905 "was recycled\n", 2906 bp, vp, (long long)loffset); 2907 BUF_UNLOCK(bp); 2908 goto loop; 2909 } 2910 2911 /* 2912 * If SZMATCH any pre-existing buffer must be of the requested 2913 * size or NULL is returned. The caller absolutely does not 2914 * want getblk() to bwrite() the buffer on a size mismatch. 2915 */ 2916 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) { 2917 BUF_UNLOCK(bp); 2918 return(NULL); 2919 } 2920 2921 /* 2922 * All vnode-based buffers must be backed by a VM object. 2923 */ 2924 KKASSERT(bp->b_flags & B_VMIO); 2925 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2926 bp->b_flags &= ~B_AGE; 2927 2928 /* 2929 * Make sure that B_INVAL buffers do not have a cached 2930 * block number translation. 2931 */ 2932 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2933 kprintf("Warning invalid buffer %p (vp %p loffset %lld)" 2934 " did not have cleared bio_offset cache\n", 2935 bp, vp, (long long)loffset); 2936 clearbiocache(&bp->b_bio2); 2937 } 2938 2939 /* 2940 * The buffer is locked. B_CACHE is cleared if the buffer is 2941 * invalid. 2942 */ 2943 if (bp->b_flags & B_INVAL) 2944 bp->b_flags &= ~B_CACHE; 2945 bremfree(bp); 2946 2947 /* 2948 * Any size inconsistancy with a dirty buffer or a buffer 2949 * with a softupdates dependancy must be resolved. Resizing 2950 * the buffer in such circumstances can lead to problems. 2951 * 2952 * Dirty or dependant buffers are written synchronously. 2953 * Other types of buffers are simply released and 2954 * reconstituted as they may be backed by valid, dirty VM 2955 * pages (but not marked B_DELWRI). 2956 * 2957 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized 2958 * and may be left over from a prior truncation (and thus 2959 * no longer represent the actual EOF point), so we 2960 * definitely do not want to B_NOCACHE the backing store. 2961 */ 2962 if (size != bp->b_bcount) { 2963 if (bp->b_flags & B_DELWRI) { 2964 bp->b_flags |= B_RELBUF; 2965 bwrite(bp); 2966 } else if (LIST_FIRST(&bp->b_dep)) { 2967 bp->b_flags |= B_RELBUF; 2968 bwrite(bp); 2969 } else { 2970 bp->b_flags |= B_RELBUF; 2971 brelse(bp); 2972 } 2973 goto loop; 2974 } 2975 KKASSERT(size <= bp->b_kvasize); 2976 KASSERT(bp->b_loffset != NOOFFSET, 2977 ("getblk: no buffer offset")); 2978 2979 /* 2980 * A buffer with B_DELWRI set and B_CACHE clear must 2981 * be committed before we can return the buffer in 2982 * order to prevent the caller from issuing a read 2983 * ( due to B_CACHE not being set ) and overwriting 2984 * it. 2985 * 2986 * Most callers, including NFS and FFS, need this to 2987 * operate properly either because they assume they 2988 * can issue a read if B_CACHE is not set, or because 2989 * ( for example ) an uncached B_DELWRI might loop due 2990 * to softupdates re-dirtying the buffer. In the latter 2991 * case, B_CACHE is set after the first write completes, 2992 * preventing further loops. 2993 * 2994 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2995 * above while extending the buffer, we cannot allow the 2996 * buffer to remain with B_CACHE set after the write 2997 * completes or it will represent a corrupt state. To 2998 * deal with this we set B_NOCACHE to scrap the buffer 2999 * after the write. 3000 * 3001 * XXX Should this be B_RELBUF instead of B_NOCACHE? 3002 * I'm not even sure this state is still possible 3003 * now that getblk() writes out any dirty buffers 3004 * on size changes. 3005 * 3006 * We might be able to do something fancy, like setting 3007 * B_CACHE in bwrite() except if B_DELWRI is already set, 3008 * so the below call doesn't set B_CACHE, but that gets real 3009 * confusing. This is much easier. 3010 */ 3011 3012 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3013 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set " 3014 "and CACHE clear, b_flags %08x\n", 3015 bp, (uintmax_t)bp->b_loffset, bp->b_flags); 3016 bp->b_flags |= B_NOCACHE; 3017 bwrite(bp); 3018 goto loop; 3019 } 3020 } else { 3021 /* 3022 * Buffer is not in-core, create new buffer. The buffer 3023 * returned by getnewbuf() is locked. Note that the returned 3024 * buffer is also considered valid (not marked B_INVAL). 3025 * 3026 * Calculating the offset for the I/O requires figuring out 3027 * the block size. We use DEV_BSIZE for VBLK or VCHR and 3028 * the mount's f_iosize otherwise. If the vnode does not 3029 * have an associated mount we assume that the passed size is 3030 * the block size. 3031 * 3032 * Note that vn_isdisk() cannot be used here since it may 3033 * return a failure for numerous reasons. Note that the 3034 * buffer size may be larger then the block size (the caller 3035 * will use block numbers with the proper multiple). Beware 3036 * of using any v_* fields which are part of unions. In 3037 * particular, in DragonFly the mount point overloading 3038 * mechanism uses the namecache only and the underlying 3039 * directory vnode is not a special case. 3040 */ 3041 int bsize, maxsize; 3042 3043 if (vp->v_type == VBLK || vp->v_type == VCHR) 3044 bsize = DEV_BSIZE; 3045 else if (vp->v_mount) 3046 bsize = vp->v_mount->mnt_stat.f_iosize; 3047 else 3048 bsize = size; 3049 3050 maxsize = size + (loffset & PAGE_MASK); 3051 maxsize = imax(maxsize, bsize); 3052 3053 bp = getnewbuf(blkflags, slptimeo, size, maxsize); 3054 if (bp == NULL) { 3055 if (slpflags || slptimeo) 3056 return NULL; 3057 goto loop; 3058 } 3059 3060 /* 3061 * Atomically insert the buffer into the hash, so that it can 3062 * be found by findblk(). 3063 * 3064 * If bgetvp() returns non-zero a collision occured, and the 3065 * bp will not be associated with the vnode. 3066 * 3067 * Make sure the translation layer has been cleared. 3068 */ 3069 bp->b_loffset = loffset; 3070 bp->b_bio2.bio_offset = NOOFFSET; 3071 /* bp->b_bio2.bio_next = NULL; */ 3072 3073 if (bgetvp(vp, bp, size)) { 3074 bp->b_flags |= B_INVAL; 3075 brelse(bp); 3076 goto loop; 3077 } 3078 3079 /* 3080 * All vnode-based buffers must be backed by a VM object. 3081 */ 3082 KKASSERT(vp->v_object != NULL); 3083 bp->b_flags |= B_VMIO; 3084 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3085 3086 allocbuf(bp, size); 3087 } 3088 KKASSERT(dsched_is_clear_buf_priv(bp)); 3089 return (bp); 3090 } 3091 3092 /* 3093 * regetblk(bp) 3094 * 3095 * Reacquire a buffer that was previously released to the locked queue, 3096 * or reacquire a buffer which is interlocked by having bioops->io_deallocate 3097 * set B_LOCKED (which handles the acquisition race). 3098 * 3099 * To this end, either B_LOCKED must be set or the dependancy list must be 3100 * non-empty. 3101 */ 3102 void 3103 regetblk(struct buf *bp) 3104 { 3105 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL); 3106 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY); 3107 bremfree(bp); 3108 } 3109 3110 /* 3111 * geteblk: 3112 * 3113 * Get an empty, disassociated buffer of given size. The buffer is 3114 * initially set to B_INVAL. 3115 * 3116 * critical section protection is not required for the allocbuf() 3117 * call because races are impossible here. 3118 */ 3119 struct buf * 3120 geteblk(int size) 3121 { 3122 struct buf *bp; 3123 int maxsize; 3124 3125 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3126 3127 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL) 3128 ; 3129 allocbuf(bp, size); 3130 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3131 KKASSERT(dsched_is_clear_buf_priv(bp)); 3132 return (bp); 3133 } 3134 3135 3136 /* 3137 * allocbuf: 3138 * 3139 * This code constitutes the buffer memory from either anonymous system 3140 * memory (in the case of non-VMIO operations) or from an associated 3141 * VM object (in the case of VMIO operations). This code is able to 3142 * resize a buffer up or down. 3143 * 3144 * Note that this code is tricky, and has many complications to resolve 3145 * deadlock or inconsistant data situations. Tread lightly!!! 3146 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3147 * the caller. Calling this code willy nilly can result in the loss of 3148 * data. 3149 * 3150 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3151 * B_CACHE for the non-VMIO case. 3152 * 3153 * This routine does not need to be called from a critical section but you 3154 * must own the buffer. 3155 */ 3156 int 3157 allocbuf(struct buf *bp, int size) 3158 { 3159 int newbsize, mbsize; 3160 int i; 3161 3162 if (BUF_REFCNT(bp) == 0) 3163 panic("allocbuf: buffer not busy"); 3164 3165 if (bp->b_kvasize < size) 3166 panic("allocbuf: buffer too small"); 3167 3168 if ((bp->b_flags & B_VMIO) == 0) { 3169 caddr_t origbuf; 3170 int origbufsize; 3171 /* 3172 * Just get anonymous memory from the kernel. Don't 3173 * mess with B_CACHE. 3174 */ 3175 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3176 if (bp->b_flags & B_MALLOC) 3177 newbsize = mbsize; 3178 else 3179 newbsize = round_page(size); 3180 3181 if (newbsize < bp->b_bufsize) { 3182 /* 3183 * Malloced buffers are not shrunk 3184 */ 3185 if (bp->b_flags & B_MALLOC) { 3186 if (newbsize) { 3187 bp->b_bcount = size; 3188 } else { 3189 kfree(bp->b_data, M_BIOBUF); 3190 if (bp->b_bufsize) { 3191 atomic_subtract_long(&bufmallocspace, bp->b_bufsize); 3192 bufspacewakeup(); 3193 bp->b_bufsize = 0; 3194 } 3195 bp->b_data = bp->b_kvabase; 3196 bp->b_bcount = 0; 3197 bp->b_flags &= ~B_MALLOC; 3198 } 3199 return 1; 3200 } 3201 vm_hold_free_pages( 3202 bp, 3203 (vm_offset_t) bp->b_data + newbsize, 3204 (vm_offset_t) bp->b_data + bp->b_bufsize); 3205 } else if (newbsize > bp->b_bufsize) { 3206 /* 3207 * We only use malloced memory on the first allocation. 3208 * and revert to page-allocated memory when the buffer 3209 * grows. 3210 */ 3211 if ((bufmallocspace < maxbufmallocspace) && 3212 (bp->b_bufsize == 0) && 3213 (mbsize <= PAGE_SIZE/2)) { 3214 3215 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK); 3216 bp->b_bufsize = mbsize; 3217 bp->b_bcount = size; 3218 bp->b_flags |= B_MALLOC; 3219 atomic_add_long(&bufmallocspace, mbsize); 3220 return 1; 3221 } 3222 origbuf = NULL; 3223 origbufsize = 0; 3224 /* 3225 * If the buffer is growing on its other-than-first 3226 * allocation, then we revert to the page-allocation 3227 * scheme. 3228 */ 3229 if (bp->b_flags & B_MALLOC) { 3230 origbuf = bp->b_data; 3231 origbufsize = bp->b_bufsize; 3232 bp->b_data = bp->b_kvabase; 3233 if (bp->b_bufsize) { 3234 atomic_subtract_long(&bufmallocspace, 3235 bp->b_bufsize); 3236 bufspacewakeup(); 3237 bp->b_bufsize = 0; 3238 } 3239 bp->b_flags &= ~B_MALLOC; 3240 newbsize = round_page(newbsize); 3241 } 3242 vm_hold_load_pages( 3243 bp, 3244 (vm_offset_t) bp->b_data + bp->b_bufsize, 3245 (vm_offset_t) bp->b_data + newbsize); 3246 if (origbuf) { 3247 bcopy(origbuf, bp->b_data, origbufsize); 3248 kfree(origbuf, M_BIOBUF); 3249 } 3250 } 3251 } else { 3252 vm_page_t m; 3253 int desiredpages; 3254 3255 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3256 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 3257 newbsize + PAGE_MASK) >> PAGE_SHIFT; 3258 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 3259 3260 if (bp->b_flags & B_MALLOC) 3261 panic("allocbuf: VMIO buffer can't be malloced"); 3262 /* 3263 * Set B_CACHE initially if buffer is 0 length or will become 3264 * 0-length. 3265 */ 3266 if (size == 0 || bp->b_bufsize == 0) 3267 bp->b_flags |= B_CACHE; 3268 3269 if (newbsize < bp->b_bufsize) { 3270 /* 3271 * DEV_BSIZE aligned new buffer size is less then the 3272 * DEV_BSIZE aligned existing buffer size. Figure out 3273 * if we have to remove any pages. 3274 */ 3275 if (desiredpages < bp->b_xio.xio_npages) { 3276 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 3277 /* 3278 * the page is not freed here -- it 3279 * is the responsibility of 3280 * vnode_pager_setsize 3281 */ 3282 m = bp->b_xio.xio_pages[i]; 3283 KASSERT(m != bogus_page, 3284 ("allocbuf: bogus page found")); 3285 vm_page_busy_wait(m, TRUE, "biodep"); 3286 bp->b_xio.xio_pages[i] = NULL; 3287 vm_page_unwire(m, 0); 3288 vm_page_wakeup(m); 3289 } 3290 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3291 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 3292 bp->b_xio.xio_npages = desiredpages; 3293 } 3294 } else if (size > bp->b_bcount) { 3295 /* 3296 * We are growing the buffer, possibly in a 3297 * byte-granular fashion. 3298 */ 3299 struct vnode *vp; 3300 vm_object_t obj; 3301 vm_offset_t toff; 3302 vm_offset_t tinc; 3303 3304 /* 3305 * Step 1, bring in the VM pages from the object, 3306 * allocating them if necessary. We must clear 3307 * B_CACHE if these pages are not valid for the 3308 * range covered by the buffer. 3309 * 3310 * critical section protection is required to protect 3311 * against interrupts unbusying and freeing pages 3312 * between our vm_page_lookup() and our 3313 * busycheck/wiring call. 3314 */ 3315 vp = bp->b_vp; 3316 obj = vp->v_object; 3317 3318 vm_object_hold(obj); 3319 while (bp->b_xio.xio_npages < desiredpages) { 3320 vm_page_t m; 3321 vm_pindex_t pi; 3322 int error; 3323 3324 pi = OFF_TO_IDX(bp->b_loffset) + 3325 bp->b_xio.xio_npages; 3326 3327 /* 3328 * Blocking on m->busy might lead to a 3329 * deadlock: 3330 * 3331 * vm_fault->getpages->cluster_read->allocbuf 3332 */ 3333 m = vm_page_lookup_busy_try(obj, pi, FALSE, 3334 &error); 3335 if (error) { 3336 vm_page_sleep_busy(m, FALSE, "pgtblk"); 3337 continue; 3338 } 3339 if (m == NULL) { 3340 /* 3341 * note: must allocate system pages 3342 * since blocking here could intefere 3343 * with paging I/O, no matter which 3344 * process we are. 3345 */ 3346 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages); 3347 if (m) { 3348 vm_page_wire(m); 3349 vm_page_flag_clear(m, PG_ZERO); 3350 vm_page_wakeup(m); 3351 bp->b_flags &= ~B_CACHE; 3352 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3353 ++bp->b_xio.xio_npages; 3354 } 3355 continue; 3356 } 3357 3358 /* 3359 * We found a page and were able to busy it. 3360 */ 3361 vm_page_flag_clear(m, PG_ZERO); 3362 vm_page_wire(m); 3363 vm_page_wakeup(m); 3364 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 3365 ++bp->b_xio.xio_npages; 3366 if (bp->b_act_count < m->act_count) 3367 bp->b_act_count = m->act_count; 3368 } 3369 vm_object_drop(obj); 3370 3371 /* 3372 * Step 2. We've loaded the pages into the buffer, 3373 * we have to figure out if we can still have B_CACHE 3374 * set. Note that B_CACHE is set according to the 3375 * byte-granular range ( bcount and size ), not the 3376 * aligned range ( newbsize ). 3377 * 3378 * The VM test is against m->valid, which is DEV_BSIZE 3379 * aligned. Needless to say, the validity of the data 3380 * needs to also be DEV_BSIZE aligned. Note that this 3381 * fails with NFS if the server or some other client 3382 * extends the file's EOF. If our buffer is resized, 3383 * B_CACHE may remain set! XXX 3384 */ 3385 3386 toff = bp->b_bcount; 3387 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 3388 3389 while ((bp->b_flags & B_CACHE) && toff < size) { 3390 vm_pindex_t pi; 3391 3392 if (tinc > (size - toff)) 3393 tinc = size - toff; 3394 3395 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 3396 PAGE_SHIFT; 3397 3398 vfs_buf_test_cache( 3399 bp, 3400 bp->b_loffset, 3401 toff, 3402 tinc, 3403 bp->b_xio.xio_pages[pi] 3404 ); 3405 toff += tinc; 3406 tinc = PAGE_SIZE; 3407 } 3408 3409 /* 3410 * Step 3, fixup the KVM pmap. Remember that 3411 * bp->b_data is relative to bp->b_loffset, but 3412 * bp->b_loffset may be offset into the first page. 3413 */ 3414 3415 bp->b_data = (caddr_t) 3416 trunc_page((vm_offset_t)bp->b_data); 3417 pmap_qenter( 3418 (vm_offset_t)bp->b_data, 3419 bp->b_xio.xio_pages, 3420 bp->b_xio.xio_npages 3421 ); 3422 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3423 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 3424 } 3425 } 3426 3427 /* adjust space use on already-dirty buffer */ 3428 if (bp->b_flags & B_DELWRI) { 3429 /* dirtykvaspace unchanged */ 3430 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize); 3431 if (bp->b_flags & B_HEAVY) { 3432 atomic_add_long(&dirtybufspacehw, 3433 newbsize - bp->b_bufsize); 3434 } 3435 } 3436 if (newbsize < bp->b_bufsize) 3437 bufspacewakeup(); 3438 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3439 bp->b_bcount = size; /* requested buffer size */ 3440 return 1; 3441 } 3442 3443 /* 3444 * biowait: 3445 * 3446 * Wait for buffer I/O completion, returning error status. B_EINTR 3447 * is converted into an EINTR error but not cleared (since a chain 3448 * of biowait() calls may occur). 3449 * 3450 * On return bpdone() will have been called but the buffer will remain 3451 * locked and will not have been brelse()'d. 3452 * 3453 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is 3454 * likely still in progress on return. 3455 * 3456 * NOTE! This operation is on a BIO, not a BUF. 3457 * 3458 * NOTE! BIO_DONE is cleared by vn_strategy() 3459 */ 3460 static __inline int 3461 _biowait(struct bio *bio, const char *wmesg, int to) 3462 { 3463 struct buf *bp = bio->bio_buf; 3464 u_int32_t flags; 3465 u_int32_t nflags; 3466 int error; 3467 3468 KKASSERT(bio == &bp->b_bio1); 3469 for (;;) { 3470 flags = bio->bio_flags; 3471 if (flags & BIO_DONE) 3472 break; 3473 nflags = flags | BIO_WANT; 3474 tsleep_interlock(bio, 0); 3475 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3476 if (wmesg) 3477 error = tsleep(bio, PINTERLOCKED, wmesg, to); 3478 else if (bp->b_cmd == BUF_CMD_READ) 3479 error = tsleep(bio, PINTERLOCKED, "biord", to); 3480 else 3481 error = tsleep(bio, PINTERLOCKED, "biowr", to); 3482 if (error) { 3483 kprintf("tsleep error biowait %d\n", error); 3484 return (error); 3485 } 3486 } 3487 } 3488 3489 /* 3490 * Finish up. 3491 */ 3492 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 3493 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 3494 if (bp->b_flags & B_EINTR) 3495 return (EINTR); 3496 if (bp->b_flags & B_ERROR) 3497 return (bp->b_error ? bp->b_error : EIO); 3498 return (0); 3499 } 3500 3501 int 3502 biowait(struct bio *bio, const char *wmesg) 3503 { 3504 return(_biowait(bio, wmesg, 0)); 3505 } 3506 3507 int 3508 biowait_timeout(struct bio *bio, const char *wmesg, int to) 3509 { 3510 return(_biowait(bio, wmesg, to)); 3511 } 3512 3513 /* 3514 * This associates a tracking count with an I/O. vn_strategy() and 3515 * dev_dstrategy() do this automatically but there are a few cases 3516 * where a vnode or device layer is bypassed when a block translation 3517 * is cached. In such cases bio_start_transaction() may be called on 3518 * the bypassed layers so the system gets an I/O in progress indication 3519 * for those higher layers. 3520 */ 3521 void 3522 bio_start_transaction(struct bio *bio, struct bio_track *track) 3523 { 3524 bio->bio_track = track; 3525 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3526 dsched_new_buf(bio->bio_buf); 3527 bio_track_ref(track); 3528 } 3529 3530 /* 3531 * Initiate I/O on a vnode. 3532 * 3533 * SWAPCACHE OPERATION: 3534 * 3535 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately 3536 * devfs also uses b_vp for fake buffers so we also have to check 3537 * that B_PAGING is 0. In this case the passed 'vp' is probably the 3538 * underlying block device. The swap assignments are related to the 3539 * buffer cache buffer's b_vp, not the passed vp. 3540 * 3541 * The passed vp == bp->b_vp only in the case where the strategy call 3542 * is made on the vp itself for its own buffers (a regular file or 3543 * block device vp). The filesystem usually then re-calls vn_strategy() 3544 * after translating the request to an underlying device. 3545 * 3546 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the 3547 * underlying buffer cache buffers. 3548 * 3549 * We can only deal with page-aligned buffers at the moment, because 3550 * we can't tell what the real dirty state for pages straddling a buffer 3551 * are. 3552 * 3553 * In order to call swap_pager_strategy() we must provide the VM object 3554 * and base offset for the underlying buffer cache pages so it can find 3555 * the swap blocks. 3556 */ 3557 void 3558 vn_strategy(struct vnode *vp, struct bio *bio) 3559 { 3560 struct bio_track *track; 3561 struct buf *bp = bio->bio_buf; 3562 3563 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3564 3565 /* 3566 * Set when an I/O is issued on the bp. Cleared by consumers 3567 * (aka HAMMER), allowing the consumer to determine if I/O had 3568 * actually occurred. 3569 */ 3570 bp->b_flags |= B_IODEBUG; 3571 3572 /* 3573 * Handle the swap cache intercept. 3574 */ 3575 if (vn_cache_strategy(vp, bio)) 3576 return; 3577 3578 /* 3579 * Otherwise do the operation through the filesystem 3580 */ 3581 if (bp->b_cmd == BUF_CMD_READ) 3582 track = &vp->v_track_read; 3583 else 3584 track = &vp->v_track_write; 3585 KKASSERT((bio->bio_flags & BIO_DONE) == 0); 3586 bio->bio_track = track; 3587 if (dsched_is_clear_buf_priv(bio->bio_buf)) 3588 dsched_new_buf(bio->bio_buf); 3589 bio_track_ref(track); 3590 vop_strategy(*vp->v_ops, vp, bio); 3591 } 3592 3593 static void vn_cache_strategy_callback(struct bio *bio); 3594 3595 int 3596 vn_cache_strategy(struct vnode *vp, struct bio *bio) 3597 { 3598 struct buf *bp = bio->bio_buf; 3599 struct bio *nbio; 3600 vm_object_t object; 3601 vm_page_t m; 3602 int i; 3603 3604 /* 3605 * Is this buffer cache buffer suitable for reading from 3606 * the swap cache? 3607 */ 3608 if (vm_swapcache_read_enable == 0 || 3609 bp->b_cmd != BUF_CMD_READ || 3610 ((bp->b_flags & B_CLUSTER) == 0 && 3611 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) || 3612 ((int)bp->b_loffset & PAGE_MASK) != 0 || 3613 (bp->b_bcount & PAGE_MASK) != 0) { 3614 return(0); 3615 } 3616 3617 /* 3618 * Figure out the original VM object (it will match the underlying 3619 * VM pages). Note that swap cached data uses page indices relative 3620 * to that object, not relative to bio->bio_offset. 3621 */ 3622 if (bp->b_flags & B_CLUSTER) 3623 object = vp->v_object; 3624 else 3625 object = bp->b_vp->v_object; 3626 3627 /* 3628 * In order to be able to use the swap cache all underlying VM 3629 * pages must be marked as such, and we can't have any bogus pages. 3630 */ 3631 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 3632 m = bp->b_xio.xio_pages[i]; 3633 if ((m->flags & PG_SWAPPED) == 0) 3634 break; 3635 if (m == bogus_page) 3636 break; 3637 } 3638 3639 /* 3640 * If we are good then issue the I/O using swap_pager_strategy(). 3641 * 3642 * We can only do this if the buffer actually supports object-backed 3643 * I/O. If it doesn't npages will be 0. 3644 */ 3645 if (i && i == bp->b_xio.xio_npages) { 3646 m = bp->b_xio.xio_pages[0]; 3647 nbio = push_bio(bio); 3648 nbio->bio_done = vn_cache_strategy_callback; 3649 nbio->bio_offset = ptoa(m->pindex); 3650 KKASSERT(m->object == object); 3651 swap_pager_strategy(object, nbio); 3652 return(1); 3653 } 3654 return(0); 3655 } 3656 3657 /* 3658 * This is a bit of a hack but since the vn_cache_strategy() function can 3659 * override a VFS's strategy function we must make sure that the bio, which 3660 * is probably bio2, doesn't leak an unexpected offset value back to the 3661 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the 3662 * bio went through its own file strategy function and the the bio2 offset 3663 * is a cached disk offset when, in fact, it isn't. 3664 */ 3665 static void 3666 vn_cache_strategy_callback(struct bio *bio) 3667 { 3668 bio->bio_offset = NOOFFSET; 3669 biodone(pop_bio(bio)); 3670 } 3671 3672 /* 3673 * bpdone: 3674 * 3675 * Finish I/O on a buffer after all BIOs have been processed. 3676 * Called when the bio chain is exhausted or by biowait. If called 3677 * by biowait, elseit is typically 0. 3678 * 3679 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp. 3680 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3681 * assuming B_INVAL is clear. 3682 * 3683 * For the VMIO case, we set B_CACHE if the op was a read and no 3684 * read error occured, or if the op was a write. B_CACHE is never 3685 * set if the buffer is invalid or otherwise uncacheable. 3686 * 3687 * bpdone does not mess with B_INVAL, allowing the I/O routine or the 3688 * initiator to leave B_INVAL set to brelse the buffer out of existance 3689 * in the biodone routine. 3690 */ 3691 void 3692 bpdone(struct buf *bp, int elseit) 3693 { 3694 buf_cmd_t cmd; 3695 3696 KASSERT(BUF_REFCNTNB(bp) > 0, 3697 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 3698 KASSERT(bp->b_cmd != BUF_CMD_DONE, 3699 ("biodone: bp %p already done!", bp)); 3700 3701 /* 3702 * No more BIOs are left. All completion functions have been dealt 3703 * with, now we clean up the buffer. 3704 */ 3705 cmd = bp->b_cmd; 3706 bp->b_cmd = BUF_CMD_DONE; 3707 3708 /* 3709 * Only reads and writes are processed past this point. 3710 */ 3711 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 3712 if (cmd == BUF_CMD_FREEBLKS) 3713 bp->b_flags |= B_NOCACHE; 3714 if (elseit) 3715 brelse(bp); 3716 return; 3717 } 3718 3719 /* 3720 * Warning: softupdates may re-dirty the buffer, and HAMMER can do 3721 * a lot worse. XXX - move this above the clearing of b_cmd 3722 */ 3723 if (LIST_FIRST(&bp->b_dep) != NULL) 3724 buf_complete(bp); 3725 3726 /* 3727 * A failed write must re-dirty the buffer unless B_INVAL 3728 * was set. Only applicable to normal buffers (with VPs). 3729 * vinum buffers may not have a vp. 3730 */ 3731 if (cmd == BUF_CMD_WRITE && 3732 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 3733 bp->b_flags &= ~B_NOCACHE; 3734 if (bp->b_vp) 3735 bdirty(bp); 3736 } 3737 3738 if (bp->b_flags & B_VMIO) { 3739 int i; 3740 vm_ooffset_t foff; 3741 vm_page_t m; 3742 vm_object_t obj; 3743 int iosize; 3744 struct vnode *vp = bp->b_vp; 3745 3746 obj = vp->v_object; 3747 3748 #if defined(VFS_BIO_DEBUG) 3749 if (vp->v_auxrefs == 0) 3750 panic("biodone: zero vnode hold count"); 3751 if ((vp->v_flag & VOBJBUF) == 0) 3752 panic("biodone: vnode is not setup for merged cache"); 3753 #endif 3754 3755 foff = bp->b_loffset; 3756 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 3757 KASSERT(obj != NULL, ("biodone: missing VM object")); 3758 3759 #if defined(VFS_BIO_DEBUG) 3760 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3761 kprintf("biodone: paging in progress(%d) < " 3762 "bp->b_xio.xio_npages(%d)\n", 3763 obj->paging_in_progress, 3764 bp->b_xio.xio_npages); 3765 } 3766 #endif 3767 3768 /* 3769 * Set B_CACHE if the op was a normal read and no error 3770 * occured. B_CACHE is set for writes in the b*write() 3771 * routines. 3772 */ 3773 iosize = bp->b_bcount - bp->b_resid; 3774 if (cmd == BUF_CMD_READ && 3775 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 3776 bp->b_flags |= B_CACHE; 3777 } 3778 3779 vm_object_hold(obj); 3780 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3781 int bogusflag = 0; 3782 int resid; 3783 3784 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3785 if (resid > iosize) 3786 resid = iosize; 3787 3788 /* 3789 * cleanup bogus pages, restoring the originals. Since 3790 * the originals should still be wired, we don't have 3791 * to worry about interrupt/freeing races destroying 3792 * the VM object association. 3793 */ 3794 m = bp->b_xio.xio_pages[i]; 3795 if (m == bogus_page) { 3796 bogusflag = 1; 3797 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3798 if (m == NULL) 3799 panic("biodone: page disappeared"); 3800 bp->b_xio.xio_pages[i] = m; 3801 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3802 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3803 } 3804 #if defined(VFS_BIO_DEBUG) 3805 if (OFF_TO_IDX(foff) != m->pindex) { 3806 kprintf("biodone: foff(%lu)/m->pindex(%ld) " 3807 "mismatch\n", 3808 (unsigned long)foff, (long)m->pindex); 3809 } 3810 #endif 3811 3812 /* 3813 * In the write case, the valid and clean bits are 3814 * already changed correctly (see bdwrite()), so we 3815 * only need to do this here in the read case. 3816 */ 3817 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3818 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 3819 vfs_clean_one_page(bp, i, m); 3820 } 3821 vm_page_flag_clear(m, PG_ZERO); 3822 3823 /* 3824 * when debugging new filesystems or buffer I/O 3825 * methods, this is the most common error that pops 3826 * up. if you see this, you have not set the page 3827 * busy flag correctly!!! 3828 */ 3829 if (m->busy == 0) { 3830 kprintf("biodone: page busy < 0, " 3831 "pindex: %d, foff: 0x(%x,%x), " 3832 "resid: %d, index: %d\n", 3833 (int) m->pindex, (int)(foff >> 32), 3834 (int) foff & 0xffffffff, resid, i); 3835 if (!vn_isdisk(vp, NULL)) 3836 kprintf(" iosize: %ld, loffset: %lld, " 3837 "flags: 0x%08x, npages: %d\n", 3838 bp->b_vp->v_mount->mnt_stat.f_iosize, 3839 (long long)bp->b_loffset, 3840 bp->b_flags, bp->b_xio.xio_npages); 3841 else 3842 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 3843 (long long)bp->b_loffset, 3844 bp->b_flags, bp->b_xio.xio_npages); 3845 kprintf(" valid: 0x%x, dirty: 0x%x, " 3846 "wired: %d\n", 3847 m->valid, m->dirty, 3848 m->wire_count); 3849 panic("biodone: page busy < 0"); 3850 } 3851 vm_page_io_finish(m); 3852 vm_page_wakeup(m); 3853 vm_object_pip_wakeup(obj); 3854 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3855 iosize -= resid; 3856 } 3857 bp->b_flags &= ~B_HASBOGUS; 3858 vm_object_drop(obj); 3859 } 3860 3861 /* 3862 * Finish up by releasing the buffer. There are no more synchronous 3863 * or asynchronous completions, those were handled by bio_done 3864 * callbacks. 3865 */ 3866 if (elseit) { 3867 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF)) 3868 brelse(bp); 3869 else 3870 bqrelse(bp); 3871 } 3872 } 3873 3874 /* 3875 * Normal biodone. 3876 */ 3877 void 3878 biodone(struct bio *bio) 3879 { 3880 struct buf *bp = bio->bio_buf; 3881 3882 runningbufwakeup(bp); 3883 3884 /* 3885 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 3886 */ 3887 while (bio) { 3888 biodone_t *done_func; 3889 struct bio_track *track; 3890 3891 /* 3892 * BIO tracking. Most but not all BIOs are tracked. 3893 */ 3894 if ((track = bio->bio_track) != NULL) { 3895 bio_track_rel(track); 3896 bio->bio_track = NULL; 3897 } 3898 3899 /* 3900 * A bio_done function terminates the loop. The function 3901 * will be responsible for any further chaining and/or 3902 * buffer management. 3903 * 3904 * WARNING! The done function can deallocate the buffer! 3905 */ 3906 if ((done_func = bio->bio_done) != NULL) { 3907 bio->bio_done = NULL; 3908 done_func(bio); 3909 return; 3910 } 3911 bio = bio->bio_prev; 3912 } 3913 3914 /* 3915 * If we've run out of bio's do normal [a]synchronous completion. 3916 */ 3917 bpdone(bp, 1); 3918 } 3919 3920 /* 3921 * Synchronous biodone - this terminates a synchronous BIO. 3922 * 3923 * bpdone() is called with elseit=FALSE, leaving the buffer completed 3924 * but still locked. The caller must brelse() the buffer after waiting 3925 * for completion. 3926 */ 3927 void 3928 biodone_sync(struct bio *bio) 3929 { 3930 struct buf *bp = bio->bio_buf; 3931 int flags; 3932 int nflags; 3933 3934 KKASSERT(bio == &bp->b_bio1); 3935 bpdone(bp, 0); 3936 3937 for (;;) { 3938 flags = bio->bio_flags; 3939 nflags = (flags | BIO_DONE) & ~BIO_WANT; 3940 3941 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) { 3942 if (flags & BIO_WANT) 3943 wakeup(bio); 3944 break; 3945 } 3946 } 3947 } 3948 3949 /* 3950 * vfs_unbusy_pages: 3951 * 3952 * This routine is called in lieu of iodone in the case of 3953 * incomplete I/O. This keeps the busy status for pages 3954 * consistant. 3955 */ 3956 void 3957 vfs_unbusy_pages(struct buf *bp) 3958 { 3959 int i; 3960 3961 runningbufwakeup(bp); 3962 3963 if (bp->b_flags & B_VMIO) { 3964 struct vnode *vp = bp->b_vp; 3965 vm_object_t obj; 3966 3967 obj = vp->v_object; 3968 vm_object_hold(obj); 3969 3970 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3971 vm_page_t m = bp->b_xio.xio_pages[i]; 3972 3973 /* 3974 * When restoring bogus changes the original pages 3975 * should still be wired, so we are in no danger of 3976 * losing the object association and do not need 3977 * critical section protection particularly. 3978 */ 3979 if (m == bogus_page) { 3980 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 3981 if (!m) { 3982 panic("vfs_unbusy_pages: page missing"); 3983 } 3984 bp->b_xio.xio_pages[i] = m; 3985 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3986 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3987 } 3988 vm_page_busy_wait(m, FALSE, "bpdpgw"); 3989 vm_page_flag_clear(m, PG_ZERO); 3990 vm_page_io_finish(m); 3991 vm_page_wakeup(m); 3992 vm_object_pip_wakeup(obj); 3993 } 3994 bp->b_flags &= ~B_HASBOGUS; 3995 vm_object_drop(obj); 3996 } 3997 } 3998 3999 /* 4000 * vfs_busy_pages: 4001 * 4002 * This routine is called before a device strategy routine. 4003 * It is used to tell the VM system that paging I/O is in 4004 * progress, and treat the pages associated with the buffer 4005 * almost as being PG_BUSY. Also the object 'paging_in_progress' 4006 * flag is handled to make sure that the object doesn't become 4007 * inconsistant. 4008 * 4009 * Since I/O has not been initiated yet, certain buffer flags 4010 * such as B_ERROR or B_INVAL may be in an inconsistant state 4011 * and should be ignored. 4012 */ 4013 void 4014 vfs_busy_pages(struct vnode *vp, struct buf *bp) 4015 { 4016 int i, bogus; 4017 struct lwp *lp = curthread->td_lwp; 4018 4019 /* 4020 * The buffer's I/O command must already be set. If reading, 4021 * B_CACHE must be 0 (double check against callers only doing 4022 * I/O when B_CACHE is 0). 4023 */ 4024 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4025 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 4026 4027 if (bp->b_flags & B_VMIO) { 4028 vm_object_t obj; 4029 4030 obj = vp->v_object; 4031 KASSERT(bp->b_loffset != NOOFFSET, 4032 ("vfs_busy_pages: no buffer offset")); 4033 4034 /* 4035 * Busy all the pages. We have to busy them all at once 4036 * to avoid deadlocks. 4037 */ 4038 retry: 4039 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4040 vm_page_t m = bp->b_xio.xio_pages[i]; 4041 4042 if (vm_page_busy_try(m, FALSE)) { 4043 vm_page_sleep_busy(m, FALSE, "vbpage"); 4044 while (--i >= 0) 4045 vm_page_wakeup(bp->b_xio.xio_pages[i]); 4046 goto retry; 4047 } 4048 } 4049 4050 /* 4051 * Setup for I/O, soft-busy the page right now because 4052 * the next loop may block. 4053 */ 4054 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4055 vm_page_t m = bp->b_xio.xio_pages[i]; 4056 4057 vm_page_flag_clear(m, PG_ZERO); 4058 if ((bp->b_flags & B_CLUSTER) == 0) { 4059 vm_object_pip_add(obj, 1); 4060 vm_page_io_start(m); 4061 } 4062 } 4063 4064 /* 4065 * Adjust protections for I/O and do bogus-page mapping. 4066 * Assume that vm_page_protect() can block (it can block 4067 * if VM_PROT_NONE, don't take any chances regardless). 4068 * 4069 * In particular note that for writes we must incorporate 4070 * page dirtyness from the VM system into the buffer's 4071 * dirty range. 4072 * 4073 * For reads we theoretically must incorporate page dirtyness 4074 * from the VM system to determine if the page needs bogus 4075 * replacement, but we shortcut the test by simply checking 4076 * that all m->valid bits are set, indicating that the page 4077 * is fully valid and does not need to be re-read. For any 4078 * VM system dirtyness the page will also be fully valid 4079 * since it was mapped at one point. 4080 */ 4081 bogus = 0; 4082 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4083 vm_page_t m = bp->b_xio.xio_pages[i]; 4084 4085 vm_page_flag_clear(m, PG_ZERO); /* XXX */ 4086 if (bp->b_cmd == BUF_CMD_WRITE) { 4087 /* 4088 * When readying a vnode-backed buffer for 4089 * a write we must zero-fill any invalid 4090 * portions of the backing VM pages, mark 4091 * it valid and clear related dirty bits. 4092 * 4093 * vfs_clean_one_page() incorporates any 4094 * VM dirtyness and updates the b_dirtyoff 4095 * range (after we've made the page RO). 4096 * 4097 * It is also expected that the pmap modified 4098 * bit has already been cleared by the 4099 * vm_page_protect(). We may not be able 4100 * to clear all dirty bits for a page if it 4101 * was also memory mapped (NFS). 4102 * 4103 * Finally be sure to unassign any swap-cache 4104 * backing store as it is now stale. 4105 */ 4106 vm_page_protect(m, VM_PROT_READ); 4107 vfs_clean_one_page(bp, i, m); 4108 swap_pager_unswapped(m); 4109 } else if (m->valid == VM_PAGE_BITS_ALL) { 4110 /* 4111 * When readying a vnode-backed buffer for 4112 * read we must replace any dirty pages with 4113 * a bogus page so dirty data is not destroyed 4114 * when filling gaps. 4115 * 4116 * To avoid testing whether the page is 4117 * dirty we instead test that the page was 4118 * at some point mapped (m->valid fully 4119 * valid) with the understanding that 4120 * this also covers the dirty case. 4121 */ 4122 bp->b_xio.xio_pages[i] = bogus_page; 4123 bp->b_flags |= B_HASBOGUS; 4124 bogus++; 4125 } else if (m->valid & m->dirty) { 4126 /* 4127 * This case should not occur as partial 4128 * dirtyment can only happen if the buffer 4129 * is B_CACHE, and this code is not entered 4130 * if the buffer is B_CACHE. 4131 */ 4132 kprintf("Warning: vfs_busy_pages - page not " 4133 "fully valid! loff=%jx bpf=%08x " 4134 "idx=%d val=%02x dir=%02x\n", 4135 (uintmax_t)bp->b_loffset, bp->b_flags, 4136 i, m->valid, m->dirty); 4137 vm_page_protect(m, VM_PROT_NONE); 4138 } else { 4139 /* 4140 * The page is not valid and can be made 4141 * part of the read. 4142 */ 4143 vm_page_protect(m, VM_PROT_NONE); 4144 } 4145 vm_page_wakeup(m); 4146 } 4147 if (bogus) { 4148 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4149 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 4150 } 4151 } 4152 4153 /* 4154 * This is the easiest place to put the process accounting for the I/O 4155 * for now. 4156 */ 4157 if (lp != NULL) { 4158 if (bp->b_cmd == BUF_CMD_READ) 4159 lp->lwp_ru.ru_inblock++; 4160 else 4161 lp->lwp_ru.ru_oublock++; 4162 } 4163 } 4164 4165 /* 4166 * Tell the VM system that the pages associated with this buffer 4167 * are clean. This is used for delayed writes where the data is 4168 * going to go to disk eventually without additional VM intevention. 4169 * 4170 * NOTE: While we only really need to clean through to b_bcount, we 4171 * just go ahead and clean through to b_bufsize. 4172 */ 4173 static void 4174 vfs_clean_pages(struct buf *bp) 4175 { 4176 vm_page_t m; 4177 int i; 4178 4179 if ((bp->b_flags & B_VMIO) == 0) 4180 return; 4181 4182 KASSERT(bp->b_loffset != NOOFFSET, 4183 ("vfs_clean_pages: no buffer offset")); 4184 4185 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4186 m = bp->b_xio.xio_pages[i]; 4187 vfs_clean_one_page(bp, i, m); 4188 } 4189 } 4190 4191 /* 4192 * vfs_clean_one_page: 4193 * 4194 * Set the valid bits and clear the dirty bits in a page within a 4195 * buffer. The range is restricted to the buffer's size and the 4196 * buffer's logical offset might index into the first page. 4197 * 4198 * The caller has busied or soft-busied the page and it is not mapped, 4199 * test and incorporate the dirty bits into b_dirtyoff/end before 4200 * clearing them. Note that we need to clear the pmap modified bits 4201 * after determining the the page was dirty, vm_page_set_validclean() 4202 * does not do it for us. 4203 * 4204 * This routine is typically called after a read completes (dirty should 4205 * be zero in that case as we are not called on bogus-replace pages), 4206 * or before a write is initiated. 4207 */ 4208 static void 4209 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m) 4210 { 4211 int bcount; 4212 int xoff; 4213 int soff; 4214 int eoff; 4215 4216 /* 4217 * Calculate offset range within the page but relative to buffer's 4218 * loffset. loffset might be offset into the first page. 4219 */ 4220 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4221 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4222 4223 if (pageno == 0) { 4224 soff = xoff; 4225 eoff = PAGE_SIZE; 4226 } else { 4227 soff = (pageno << PAGE_SHIFT); 4228 eoff = soff + PAGE_SIZE; 4229 } 4230 if (eoff > bcount) 4231 eoff = bcount; 4232 if (soff >= eoff) 4233 return; 4234 4235 /* 4236 * Test dirty bits and adjust b_dirtyoff/end. 4237 * 4238 * If dirty pages are incorporated into the bp any prior 4239 * B_NEEDCOMMIT state (NFS) must be cleared because the 4240 * caller has not taken into account the new dirty data. 4241 * 4242 * If the page was memory mapped the dirty bits might go beyond the 4243 * end of the buffer, but we can't really make the assumption that 4244 * a file EOF straddles the buffer (even though this is the case for 4245 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing 4246 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer. 4247 * This also saves some console spam. 4248 * 4249 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK, 4250 * NFS can handle huge commits but not huge writes. 4251 */ 4252 vm_page_test_dirty(m); 4253 if (m->dirty) { 4254 if ((bp->b_flags & B_NEEDCOMMIT) && 4255 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) { 4256 if (debug_commit) 4257 kprintf("Warning: vfs_clean_one_page: bp %p " 4258 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT" 4259 " cmd %d vd %02x/%02x x/s/e %d %d %d " 4260 "doff/end %d %d\n", 4261 bp, (uintmax_t)bp->b_loffset, bp->b_bcount, 4262 bp->b_flags, bp->b_cmd, 4263 m->valid, m->dirty, xoff, soff, eoff, 4264 bp->b_dirtyoff, bp->b_dirtyend); 4265 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 4266 if (debug_commit) 4267 print_backtrace(-1); 4268 } 4269 /* 4270 * Only clear the pmap modified bits if ALL the dirty bits 4271 * are set, otherwise the system might mis-clear portions 4272 * of a page. 4273 */ 4274 if (m->dirty == VM_PAGE_BITS_ALL && 4275 (bp->b_flags & B_NEEDCOMMIT) == 0) { 4276 pmap_clear_modify(m); 4277 } 4278 if (bp->b_dirtyoff > soff - xoff) 4279 bp->b_dirtyoff = soff - xoff; 4280 if (bp->b_dirtyend < eoff - xoff) 4281 bp->b_dirtyend = eoff - xoff; 4282 } 4283 4284 /* 4285 * Set related valid bits, clear related dirty bits. 4286 * Does not mess with the pmap modified bit. 4287 * 4288 * WARNING! We cannot just clear all of m->dirty here as the 4289 * buffer cache buffers may use a DEV_BSIZE'd aligned 4290 * block size, or have an odd size (e.g. NFS at file EOF). 4291 * The putpages code can clear m->dirty to 0. 4292 * 4293 * If a VOP_WRITE generates a buffer cache buffer which 4294 * covers the same space as mapped writable pages the 4295 * buffer flush might not be able to clear all the dirty 4296 * bits and still require a putpages from the VM system 4297 * to finish it off. 4298 * 4299 * WARNING! vm_page_set_validclean() currently assumes vm_token 4300 * is held. The page might not be busied (bdwrite() case). 4301 * XXX remove this comment once we've validated that this 4302 * is no longer an issue. 4303 */ 4304 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff); 4305 } 4306 4307 #if 0 4308 /* 4309 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty. 4310 * The page data is assumed to be valid (there is no zeroing here). 4311 */ 4312 static void 4313 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m) 4314 { 4315 int bcount; 4316 int xoff; 4317 int soff; 4318 int eoff; 4319 4320 /* 4321 * Calculate offset range within the page but relative to buffer's 4322 * loffset. loffset might be offset into the first page. 4323 */ 4324 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */ 4325 bcount = bp->b_bcount + xoff; /* offset adjusted */ 4326 4327 if (pageno == 0) { 4328 soff = xoff; 4329 eoff = PAGE_SIZE; 4330 } else { 4331 soff = (pageno << PAGE_SHIFT); 4332 eoff = soff + PAGE_SIZE; 4333 } 4334 if (eoff > bcount) 4335 eoff = bcount; 4336 if (soff >= eoff) 4337 return; 4338 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff); 4339 } 4340 #endif 4341 4342 /* 4343 * vfs_bio_clrbuf: 4344 * 4345 * Clear a buffer. This routine essentially fakes an I/O, so we need 4346 * to clear B_ERROR and B_INVAL. 4347 * 4348 * Note that while we only theoretically need to clear through b_bcount, 4349 * we go ahead and clear through b_bufsize. 4350 */ 4351 4352 void 4353 vfs_bio_clrbuf(struct buf *bp) 4354 { 4355 int i, mask = 0; 4356 caddr_t sa, ea; 4357 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 4358 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR); 4359 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4360 (bp->b_loffset & PAGE_MASK) == 0) { 4361 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4362 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 4363 bp->b_resid = 0; 4364 return; 4365 } 4366 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 4367 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 4368 bzero(bp->b_data, bp->b_bufsize); 4369 bp->b_xio.xio_pages[0]->valid |= mask; 4370 bp->b_resid = 0; 4371 return; 4372 } 4373 } 4374 sa = bp->b_data; 4375 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 4376 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 4377 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 4378 ea = (caddr_t)(vm_offset_t)ulmin( 4379 (u_long)(vm_offset_t)ea, 4380 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 4381 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4382 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 4383 continue; 4384 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 4385 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 4386 bzero(sa, ea - sa); 4387 } 4388 } else { 4389 for (; sa < ea; sa += DEV_BSIZE, j++) { 4390 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 4391 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 4392 bzero(sa, DEV_BSIZE); 4393 } 4394 } 4395 bp->b_xio.xio_pages[i]->valid |= mask; 4396 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 4397 } 4398 bp->b_resid = 0; 4399 } else { 4400 clrbuf(bp); 4401 } 4402 } 4403 4404 /* 4405 * vm_hold_load_pages: 4406 * 4407 * Load pages into the buffer's address space. The pages are 4408 * allocated from the kernel object in order to reduce interference 4409 * with the any VM paging I/O activity. The range of loaded 4410 * pages will be wired. 4411 * 4412 * If a page cannot be allocated, the 'pagedaemon' is woken up to 4413 * retrieve the full range (to - from) of pages. 4414 */ 4415 void 4416 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4417 { 4418 vm_offset_t pg; 4419 vm_page_t p; 4420 int index; 4421 4422 to = round_page(to); 4423 from = round_page(from); 4424 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4425 4426 pg = from; 4427 while (pg < to) { 4428 /* 4429 * Note: must allocate system pages since blocking here 4430 * could intefere with paging I/O, no matter which 4431 * process we are. 4432 */ 4433 vm_object_hold(&kernel_object); 4434 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT, 4435 (vm_pindex_t)((to - pg) >> PAGE_SHIFT)); 4436 vm_object_drop(&kernel_object); 4437 if (p) { 4438 vm_page_wire(p); 4439 p->valid = VM_PAGE_BITS_ALL; 4440 vm_page_flag_clear(p, PG_ZERO); 4441 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 4442 bp->b_xio.xio_pages[index] = p; 4443 vm_page_wakeup(p); 4444 4445 pg += PAGE_SIZE; 4446 ++index; 4447 } 4448 } 4449 bp->b_xio.xio_npages = index; 4450 } 4451 4452 /* 4453 * Allocate a page for a buffer cache buffer. 4454 * 4455 * If NULL is returned the caller is expected to retry (typically check if 4456 * the page already exists on retry before trying to allocate one). 4457 * 4458 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This 4459 * function will use the system reserve with the hope that the page 4460 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller 4461 * is done with the buffer. 4462 * 4463 * NOTE! However, TMPFS is a special case because flushing a dirty buffer 4464 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon 4465 * is capable of retiring pages (to swap). For TMPFS we don't dig 4466 * into the system reserve because doing so could stall out pretty 4467 * much every process running on the system. 4468 */ 4469 static 4470 vm_page_t 4471 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit) 4472 { 4473 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK; 4474 vm_page_t p; 4475 4476 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj)); 4477 4478 /* 4479 * Try a normal allocation first. 4480 */ 4481 p = vm_page_alloc(obj, pg, vmflags); 4482 if (p) 4483 return(p); 4484 if (vm_page_lookup(obj, pg)) 4485 return(NULL); 4486 vm_pageout_deficit += deficit; 4487 4488 /* 4489 * Try again, digging into the system reserve. 4490 * 4491 * Trying to recover pages from the buffer cache here can deadlock 4492 * against other threads trying to busy underlying pages so we 4493 * depend on the code in brelse() and bqrelse() to free/cache the 4494 * underlying buffer cache pages when memory is low. 4495 */ 4496 if (curthread->td_flags & TDF_SYSTHREAD) 4497 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT; 4498 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS) 4499 vmflags |= 0; 4500 else 4501 vmflags |= VM_ALLOC_SYSTEM; 4502 4503 /*recoverbufpages();*/ 4504 p = vm_page_alloc(obj, pg, vmflags); 4505 if (p) 4506 return(p); 4507 if (vm_page_lookup(obj, pg)) 4508 return(NULL); 4509 4510 /* 4511 * Wait for memory to free up and try again 4512 */ 4513 if (vm_page_count_severe()) 4514 ++lowmempgallocs; 4515 vm_wait(hz / 20 + 1); 4516 4517 p = vm_page_alloc(obj, pg, vmflags); 4518 if (p) 4519 return(p); 4520 if (vm_page_lookup(obj, pg)) 4521 return(NULL); 4522 4523 /* 4524 * Ok, now we are really in trouble. 4525 */ 4526 { 4527 static struct krate biokrate = { .freq = 1 }; 4528 krateprintf(&biokrate, 4529 "Warning: bio_page_alloc: memory exhausted " 4530 "during bufcache page allocation from %s\n", 4531 curthread->td_comm); 4532 } 4533 if (curthread->td_flags & TDF_SYSTHREAD) 4534 vm_wait(hz / 20 + 1); 4535 else 4536 vm_wait(hz / 2 + 1); 4537 return (NULL); 4538 } 4539 4540 /* 4541 * vm_hold_free_pages: 4542 * 4543 * Return pages associated with the buffer back to the VM system. 4544 * 4545 * The range of pages underlying the buffer's address space will 4546 * be unmapped and un-wired. 4547 */ 4548 void 4549 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4550 { 4551 vm_offset_t pg; 4552 vm_page_t p; 4553 int index, newnpages; 4554 4555 from = round_page(from); 4556 to = round_page(to); 4557 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4558 newnpages = index; 4559 4560 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4561 p = bp->b_xio.xio_pages[index]; 4562 if (p && (index < bp->b_xio.xio_npages)) { 4563 if (p->busy) { 4564 kprintf("vm_hold_free_pages: doffset: %lld, " 4565 "loffset: %lld\n", 4566 (long long)bp->b_bio2.bio_offset, 4567 (long long)bp->b_loffset); 4568 } 4569 bp->b_xio.xio_pages[index] = NULL; 4570 pmap_kremove(pg); 4571 vm_page_busy_wait(p, FALSE, "vmhldpg"); 4572 vm_page_unwire(p, 0); 4573 vm_page_free(p); 4574 } 4575 } 4576 bp->b_xio.xio_npages = newnpages; 4577 } 4578 4579 /* 4580 * vmapbuf: 4581 * 4582 * Map a user buffer into KVM via a pbuf. On return the buffer's 4583 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 4584 * initialized. 4585 */ 4586 int 4587 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 4588 { 4589 caddr_t addr; 4590 vm_offset_t va; 4591 vm_page_t m; 4592 int vmprot; 4593 int error; 4594 int pidx; 4595 int i; 4596 4597 /* 4598 * bp had better have a command and it better be a pbuf. 4599 */ 4600 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 4601 KKASSERT(bp->b_flags & B_PAGING); 4602 KKASSERT(bp->b_kvabase); 4603 4604 if (bytes < 0) 4605 return (-1); 4606 4607 /* 4608 * Map the user data into KVM. Mappings have to be page-aligned. 4609 */ 4610 addr = (caddr_t)trunc_page((vm_offset_t)udata); 4611 pidx = 0; 4612 4613 vmprot = VM_PROT_READ; 4614 if (bp->b_cmd == BUF_CMD_READ) 4615 vmprot |= VM_PROT_WRITE; 4616 4617 while (addr < udata + bytes) { 4618 /* 4619 * Do the vm_fault if needed; do the copy-on-write thing 4620 * when reading stuff off device into memory. 4621 * 4622 * vm_fault_page*() returns a held VM page. 4623 */ 4624 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata; 4625 va = trunc_page(va); 4626 4627 m = vm_fault_page_quick(va, vmprot, &error); 4628 if (m == NULL) { 4629 for (i = 0; i < pidx; ++i) { 4630 vm_page_unhold(bp->b_xio.xio_pages[i]); 4631 bp->b_xio.xio_pages[i] = NULL; 4632 } 4633 return(-1); 4634 } 4635 bp->b_xio.xio_pages[pidx] = m; 4636 addr += PAGE_SIZE; 4637 ++pidx; 4638 } 4639 4640 /* 4641 * Map the page array and set the buffer fields to point to 4642 * the mapped data buffer. 4643 */ 4644 if (pidx > btoc(MAXPHYS)) 4645 panic("vmapbuf: mapped more than MAXPHYS"); 4646 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 4647 4648 bp->b_xio.xio_npages = pidx; 4649 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 4650 bp->b_bcount = bytes; 4651 bp->b_bufsize = bytes; 4652 return(0); 4653 } 4654 4655 /* 4656 * vunmapbuf: 4657 * 4658 * Free the io map PTEs associated with this IO operation. 4659 * We also invalidate the TLB entries and restore the original b_addr. 4660 */ 4661 void 4662 vunmapbuf(struct buf *bp) 4663 { 4664 int pidx; 4665 int npages; 4666 4667 KKASSERT(bp->b_flags & B_PAGING); 4668 4669 npages = bp->b_xio.xio_npages; 4670 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4671 for (pidx = 0; pidx < npages; ++pidx) { 4672 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 4673 bp->b_xio.xio_pages[pidx] = NULL; 4674 } 4675 bp->b_xio.xio_npages = 0; 4676 bp->b_data = bp->b_kvabase; 4677 } 4678 4679 /* 4680 * Scan all buffers in the system and issue the callback. 4681 */ 4682 int 4683 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 4684 { 4685 int count = 0; 4686 int error; 4687 long n; 4688 4689 for (n = 0; n < nbuf; ++n) { 4690 if ((error = callback(&buf[n], info)) < 0) { 4691 count = error; 4692 break; 4693 } 4694 count += error; 4695 } 4696 return (count); 4697 } 4698 4699 /* 4700 * nestiobuf_iodone: biodone callback for nested buffers and propagate 4701 * completion to the master buffer. 4702 */ 4703 static void 4704 nestiobuf_iodone(struct bio *bio) 4705 { 4706 struct bio *mbio; 4707 struct buf *mbp, *bp; 4708 struct devstat *stats; 4709 int error; 4710 int donebytes; 4711 4712 bp = bio->bio_buf; 4713 mbio = bio->bio_caller_info1.ptr; 4714 stats = bio->bio_caller_info2.ptr; 4715 mbp = mbio->bio_buf; 4716 4717 KKASSERT(bp->b_bcount <= bp->b_bufsize); 4718 KKASSERT(mbp != bp); 4719 4720 error = bp->b_error; 4721 if (bp->b_error == 0 && 4722 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) { 4723 /* 4724 * Not all got transfered, raise an error. We have no way to 4725 * propagate these conditions to mbp. 4726 */ 4727 error = EIO; 4728 } 4729 4730 donebytes = bp->b_bufsize; 4731 4732 relpbuf(bp, NULL); 4733 4734 nestiobuf_done(mbio, donebytes, error, stats); 4735 } 4736 4737 void 4738 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats) 4739 { 4740 struct buf *mbp; 4741 4742 mbp = mbio->bio_buf; 4743 4744 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0); 4745 4746 /* 4747 * If an error occured, propagate it to the master buffer. 4748 * 4749 * Several biodone()s may wind up running concurrently so 4750 * use an atomic op to adjust b_flags. 4751 */ 4752 if (error) { 4753 mbp->b_error = error; 4754 atomic_set_int(&mbp->b_flags, B_ERROR); 4755 } 4756 4757 /* 4758 * Decrement the operations in progress counter and terminate the 4759 * I/O if this was the last bit. 4760 */ 4761 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4762 mbp->b_resid = 0; 4763 if (stats) 4764 devstat_end_transaction_buf(stats, mbp); 4765 biodone(mbio); 4766 } 4767 } 4768 4769 /* 4770 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent 4771 * the mbio from being biodone()'d while we are still adding sub-bios to 4772 * it. 4773 */ 4774 void 4775 nestiobuf_init(struct bio *bio) 4776 { 4777 bio->bio_driver_info = (void *)1; 4778 } 4779 4780 /* 4781 * The BIOs added to the nestedio have already been started, remove the 4782 * count that placeheld our mbio and biodone() it if the count would 4783 * transition to 0. 4784 */ 4785 void 4786 nestiobuf_start(struct bio *mbio) 4787 { 4788 struct buf *mbp = mbio->bio_buf; 4789 4790 /* 4791 * Decrement the operations in progress counter and terminate the 4792 * I/O if this was the last bit. 4793 */ 4794 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) { 4795 if (mbp->b_flags & B_ERROR) 4796 mbp->b_resid = mbp->b_bcount; 4797 else 4798 mbp->b_resid = 0; 4799 biodone(mbio); 4800 } 4801 } 4802 4803 /* 4804 * Set an intermediate error prior to calling nestiobuf_start() 4805 */ 4806 void 4807 nestiobuf_error(struct bio *mbio, int error) 4808 { 4809 struct buf *mbp = mbio->bio_buf; 4810 4811 if (error) { 4812 mbp->b_error = error; 4813 atomic_set_int(&mbp->b_flags, B_ERROR); 4814 } 4815 } 4816 4817 /* 4818 * nestiobuf_add: setup a "nested" buffer. 4819 * 4820 * => 'mbp' is a "master" buffer which is being divided into sub pieces. 4821 * => 'bp' should be a buffer allocated by getiobuf. 4822 * => 'offset' is a byte offset in the master buffer. 4823 * => 'size' is a size in bytes of this nested buffer. 4824 */ 4825 void 4826 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats) 4827 { 4828 struct buf *mbp = mbio->bio_buf; 4829 struct vnode *vp = mbp->b_vp; 4830 4831 KKASSERT(mbp->b_bcount >= offset + size); 4832 4833 atomic_add_int((int *)&mbio->bio_driver_info, 1); 4834 4835 /* kernel needs to own the lock for it to be released in biodone */ 4836 BUF_KERNPROC(bp); 4837 bp->b_vp = vp; 4838 bp->b_cmd = mbp->b_cmd; 4839 bp->b_bio1.bio_done = nestiobuf_iodone; 4840 bp->b_data = (char *)mbp->b_data + offset; 4841 bp->b_resid = bp->b_bcount = size; 4842 bp->b_bufsize = bp->b_bcount; 4843 4844 bp->b_bio1.bio_track = NULL; 4845 bp->b_bio1.bio_caller_info1.ptr = mbio; 4846 bp->b_bio1.bio_caller_info2.ptr = stats; 4847 } 4848 4849 #ifdef DDB 4850 4851 DB_SHOW_COMMAND(buffer, db_show_buffer) 4852 { 4853 /* get args */ 4854 struct buf *bp = (struct buf *)addr; 4855 4856 if (!have_addr) { 4857 db_printf("usage: show buffer <addr>\n"); 4858 return; 4859 } 4860 4861 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4862 db_printf("b_cmd = %d\n", bp->b_cmd); 4863 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 4864 "b_resid = %d\n, b_data = %p, " 4865 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 4866 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4867 bp->b_data, 4868 (long long)bp->b_bio2.bio_offset, 4869 (long long)(bp->b_bio2.bio_next ? 4870 bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 4871 if (bp->b_xio.xio_npages) { 4872 int i; 4873 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 4874 bp->b_xio.xio_npages); 4875 for (i = 0; i < bp->b_xio.xio_npages; i++) { 4876 vm_page_t m; 4877 m = bp->b_xio.xio_pages[i]; 4878 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4879 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4880 if ((i + 1) < bp->b_xio.xio_npages) 4881 db_printf(","); 4882 } 4883 db_printf("\n"); 4884 } 4885 } 4886 #endif /* DDB */ 4887