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