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