1 /*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * Copyright (c) 2013 The FreeBSD Foundation 5 * All rights reserved. 6 * 7 * Portions of this software were developed by Konstantin Belousov 8 * under sponsorship from the FreeBSD Foundation. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 */ 31 32 /* 33 * this file contains a new buffer I/O scheme implementing a coherent 34 * VM object and buffer cache scheme. Pains have been taken to make 35 * sure that the performance degradation associated with schemes such 36 * as this is not realized. 37 * 38 * Author: John S. Dyson 39 * Significant help during the development and debugging phases 40 * had been provided by David Greenman, also of the FreeBSD core team. 41 * 42 * see man buf(9) for more info. 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/bio.h> 51 #include <sys/conf.h> 52 #include <sys/buf.h> 53 #include <sys/devicestat.h> 54 #include <sys/eventhandler.h> 55 #include <sys/fail.h> 56 #include <sys/limits.h> 57 #include <sys/lock.h> 58 #include <sys/malloc.h> 59 #include <sys/mount.h> 60 #include <sys/mutex.h> 61 #include <sys/kernel.h> 62 #include <sys/kthread.h> 63 #include <sys/proc.h> 64 #include <sys/resourcevar.h> 65 #include <sys/rwlock.h> 66 #include <sys/sysctl.h> 67 #include <sys/vmem.h> 68 #include <sys/vmmeter.h> 69 #include <sys/vnode.h> 70 #include <geom/geom.h> 71 #include <vm/vm.h> 72 #include <vm/vm_param.h> 73 #include <vm/vm_kern.h> 74 #include <vm/vm_pageout.h> 75 #include <vm/vm_page.h> 76 #include <vm/vm_object.h> 77 #include <vm/vm_extern.h> 78 #include <vm/vm_map.h> 79 #include "opt_compat.h" 80 #include "opt_swap.h" 81 82 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 83 84 struct bio_ops bioops; /* I/O operation notification */ 85 86 struct buf_ops buf_ops_bio = { 87 .bop_name = "buf_ops_bio", 88 .bop_write = bufwrite, 89 .bop_strategy = bufstrategy, 90 .bop_sync = bufsync, 91 .bop_bdflush = bufbdflush, 92 }; 93 94 /* 95 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 96 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 97 */ 98 struct buf *buf; /* buffer header pool */ 99 caddr_t unmapped_buf; 100 101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 102 struct proc *bufdaemonproc; 103 104 static int inmem(struct vnode *vp, daddr_t blkno); 105 static void vm_hold_free_pages(struct buf *bp, int newbsize); 106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 107 vm_offset_t to); 108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 110 vm_page_t m); 111 static void vfs_clean_pages_dirty_buf(struct buf *bp); 112 static void vfs_setdirty_locked_object(struct buf *bp); 113 static void vfs_vmio_release(struct buf *bp); 114 static int vfs_bio_clcheck(struct vnode *vp, int size, 115 daddr_t lblkno, daddr_t blkno); 116 static int buf_flush(struct vnode *vp, int); 117 static int flushbufqueues(struct vnode *, int, int); 118 static void buf_daemon(void); 119 static void bremfreel(struct buf *bp); 120 static __inline void bd_wakeup(void); 121 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 122 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 123 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 124 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 125 #endif 126 127 int vmiodirenable = TRUE; 128 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 129 "Use the VM system for directory writes"); 130 long runningbufspace; 131 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 132 "Amount of presently outstanding async buffer io"); 133 static long bufspace; 134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 135 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 136 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 137 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 138 #else 139 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 140 "Virtual memory used for buffers"); 141 #endif 142 static long unmapped_bufspace; 143 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD, 144 &unmapped_bufspace, 0, 145 "Amount of unmapped buffers, inclusive in the bufspace"); 146 static long maxbufspace; 147 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 148 "Maximum allowed value of bufspace (including buf_daemon)"); 149 static long bufmallocspace; 150 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 151 "Amount of malloced memory for buffers"); 152 static long maxbufmallocspace; 153 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 154 "Maximum amount of malloced memory for buffers"); 155 static long lobufspace; 156 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 157 "Minimum amount of buffers we want to have"); 158 long hibufspace; 159 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 160 "Maximum allowed value of bufspace (excluding buf_daemon)"); 161 static int bufreusecnt; 162 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 163 "Number of times we have reused a buffer"); 164 static int buffreekvacnt; 165 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 166 "Number of times we have freed the KVA space from some buffer"); 167 static int bufdefragcnt; 168 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 169 "Number of times we have had to repeat buffer allocation to defragment"); 170 static long lorunningspace; 171 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 172 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 173 "Minimum preferred space used for in-progress I/O"); 174 static long hirunningspace; 175 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 176 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 177 "Maximum amount of space to use for in-progress I/O"); 178 int dirtybufferflushes; 179 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 180 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 181 int bdwriteskip; 182 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 183 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 184 int altbufferflushes; 185 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 186 0, "Number of fsync flushes to limit dirty buffers"); 187 static int recursiveflushes; 188 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 189 0, "Number of flushes skipped due to being recursive"); 190 static int numdirtybuffers; 191 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 192 "Number of buffers that are dirty (has unwritten changes) at the moment"); 193 static int lodirtybuffers; 194 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 195 "How many buffers we want to have free before bufdaemon can sleep"); 196 static int hidirtybuffers; 197 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 198 "When the number of dirty buffers is considered severe"); 199 int dirtybufthresh; 200 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 201 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 202 static int numfreebuffers; 203 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 204 "Number of free buffers"); 205 static int lofreebuffers; 206 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 207 "XXX Unused"); 208 static int hifreebuffers; 209 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 210 "XXX Complicatedly unused"); 211 static int getnewbufcalls; 212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 213 "Number of calls to getnewbuf"); 214 static int getnewbufrestarts; 215 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 216 "Number of times getnewbuf has had to restart a buffer aquisition"); 217 static int mappingrestarts; 218 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 219 "Number of times getblk has had to restart a buffer mapping for " 220 "unmapped buffer"); 221 static int flushbufqtarget = 100; 222 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 223 "Amount of work to do in flushbufqueues when helping bufdaemon"); 224 static long notbufdflushes; 225 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0, 226 "Number of dirty buffer flushes done by the bufdaemon helpers"); 227 static long barrierwrites; 228 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 229 "Number of barrier writes"); 230 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 231 &unmapped_buf_allowed, 0, 232 "Permit the use of the unmapped i/o"); 233 234 /* 235 * Lock for the non-dirty bufqueues 236 */ 237 static struct mtx_padalign bqclean; 238 239 /* 240 * Lock for the dirty queue. 241 */ 242 static struct mtx_padalign bqdirty; 243 244 /* 245 * This lock synchronizes access to bd_request. 246 */ 247 static struct mtx_padalign bdlock; 248 249 /* 250 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 251 * waitrunningbufspace(). 252 */ 253 static struct mtx_padalign rbreqlock; 254 255 /* 256 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 257 */ 258 static struct rwlock_padalign nblock; 259 260 /* 261 * Lock that protects bdirtywait. 262 */ 263 static struct mtx_padalign bdirtylock; 264 265 /* 266 * Wakeup point for bufdaemon, as well as indicator of whether it is already 267 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 268 * is idling. 269 */ 270 static int bd_request; 271 272 /* 273 * Request for the buf daemon to write more buffers than is indicated by 274 * lodirtybuf. This may be necessary to push out excess dependencies or 275 * defragment the address space where a simple count of the number of dirty 276 * buffers is insufficient to characterize the demand for flushing them. 277 */ 278 static int bd_speedupreq; 279 280 /* 281 * bogus page -- for I/O to/from partially complete buffers 282 * this is a temporary solution to the problem, but it is not 283 * really that bad. it would be better to split the buffer 284 * for input in the case of buffers partially already in memory, 285 * but the code is intricate enough already. 286 */ 287 vm_page_t bogus_page; 288 289 /* 290 * Synchronization (sleep/wakeup) variable for active buffer space requests. 291 * Set when wait starts, cleared prior to wakeup(). 292 * Used in runningbufwakeup() and waitrunningbufspace(). 293 */ 294 static int runningbufreq; 295 296 /* 297 * Synchronization (sleep/wakeup) variable for buffer requests. 298 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 299 * by and/or. 300 * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(), 301 * getnewbuf(), and getblk(). 302 */ 303 static volatile int needsbuffer; 304 305 /* 306 * Synchronization for bwillwrite() waiters. 307 */ 308 static int bdirtywait; 309 310 /* 311 * Definitions for the buffer free lists. 312 */ 313 #define BUFFER_QUEUES 5 /* number of free buffer queues */ 314 315 #define QUEUE_NONE 0 /* on no queue */ 316 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 317 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 318 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 319 #define QUEUE_EMPTY 4 /* empty buffer headers */ 320 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 321 322 /* Queues for free buffers with various properties */ 323 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 324 #ifdef INVARIANTS 325 static int bq_len[BUFFER_QUEUES]; 326 #endif 327 328 /* 329 * Single global constant for BUF_WMESG, to avoid getting multiple references. 330 * buf_wmesg is referred from macros. 331 */ 332 const char *buf_wmesg = BUF_WMESG; 333 334 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 335 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 336 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 337 338 static int 339 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 340 { 341 long value; 342 int error; 343 344 value = *(long *)arg1; 345 error = sysctl_handle_long(oidp, &value, 0, req); 346 if (error != 0 || req->newptr == NULL) 347 return (error); 348 mtx_lock(&rbreqlock); 349 if (arg1 == &hirunningspace) { 350 if (value < lorunningspace) 351 error = EINVAL; 352 else 353 hirunningspace = value; 354 } else { 355 KASSERT(arg1 == &lorunningspace, 356 ("%s: unknown arg1", __func__)); 357 if (value > hirunningspace) 358 error = EINVAL; 359 else 360 lorunningspace = value; 361 } 362 mtx_unlock(&rbreqlock); 363 return (error); 364 } 365 366 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 367 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 368 static int 369 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 370 { 371 long lvalue; 372 int ivalue; 373 374 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 375 return (sysctl_handle_long(oidp, arg1, arg2, req)); 376 lvalue = *(long *)arg1; 377 if (lvalue > INT_MAX) 378 /* On overflow, still write out a long to trigger ENOMEM. */ 379 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 380 ivalue = lvalue; 381 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 382 } 383 #endif 384 385 /* 386 * bqlock: 387 * 388 * Return the appropriate queue lock based on the index. 389 */ 390 static inline struct mtx * 391 bqlock(int qindex) 392 { 393 394 if (qindex == QUEUE_DIRTY) 395 return (struct mtx *)(&bqdirty); 396 return (struct mtx *)(&bqclean); 397 } 398 399 /* 400 * bdirtywakeup: 401 * 402 * Wakeup any bwillwrite() waiters. 403 */ 404 static void 405 bdirtywakeup(void) 406 { 407 mtx_lock(&bdirtylock); 408 if (bdirtywait) { 409 bdirtywait = 0; 410 wakeup(&bdirtywait); 411 } 412 mtx_unlock(&bdirtylock); 413 } 414 415 /* 416 * bdirtysub: 417 * 418 * Decrement the numdirtybuffers count by one and wakeup any 419 * threads blocked in bwillwrite(). 420 */ 421 static void 422 bdirtysub(void) 423 { 424 425 if (atomic_fetchadd_int(&numdirtybuffers, -1) == 426 (lodirtybuffers + hidirtybuffers) / 2) 427 bdirtywakeup(); 428 } 429 430 /* 431 * bdirtyadd: 432 * 433 * Increment the numdirtybuffers count by one and wakeup the buf 434 * daemon if needed. 435 */ 436 static void 437 bdirtyadd(void) 438 { 439 440 /* 441 * Only do the wakeup once as we cross the boundary. The 442 * buf daemon will keep running until the condition clears. 443 */ 444 if (atomic_fetchadd_int(&numdirtybuffers, 1) == 445 (lodirtybuffers + hidirtybuffers) / 2) 446 bd_wakeup(); 447 } 448 449 /* 450 * bufspacewakeup: 451 * 452 * Called when buffer space is potentially available for recovery. 453 * getnewbuf() will block on this flag when it is unable to free 454 * sufficient buffer space. Buffer space becomes recoverable when 455 * bp's get placed back in the queues. 456 */ 457 458 static __inline void 459 bufspacewakeup(void) 460 { 461 int need_wakeup, on; 462 463 /* 464 * If someone is waiting for BUF space, wake them up. Even 465 * though we haven't freed the kva space yet, the waiting 466 * process will be able to now. 467 */ 468 rw_rlock(&nblock); 469 for (;;) { 470 need_wakeup = 0; 471 on = needsbuffer; 472 if ((on & VFS_BIO_NEED_BUFSPACE) == 0) 473 break; 474 need_wakeup = 1; 475 if (atomic_cmpset_rel_int(&needsbuffer, on, 476 on & ~VFS_BIO_NEED_BUFSPACE)) 477 break; 478 } 479 if (need_wakeup) 480 wakeup(__DEVOLATILE(void *, &needsbuffer)); 481 rw_runlock(&nblock); 482 } 483 484 /* 485 * runningwakeup: 486 * 487 * Wake up processes that are waiting on asynchronous writes to fall 488 * below lorunningspace. 489 */ 490 static void 491 runningwakeup(void) 492 { 493 494 mtx_lock(&rbreqlock); 495 if (runningbufreq) { 496 runningbufreq = 0; 497 wakeup(&runningbufreq); 498 } 499 mtx_unlock(&rbreqlock); 500 } 501 502 /* 503 * runningbufwakeup: 504 * 505 * Decrement the outstanding write count according. 506 */ 507 void 508 runningbufwakeup(struct buf *bp) 509 { 510 long space, bspace; 511 512 bspace = bp->b_runningbufspace; 513 if (bspace == 0) 514 return; 515 space = atomic_fetchadd_long(&runningbufspace, -bspace); 516 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 517 space, bspace)); 518 bp->b_runningbufspace = 0; 519 /* 520 * Only acquire the lock and wakeup on the transition from exceeding 521 * the threshold to falling below it. 522 */ 523 if (space < lorunningspace) 524 return; 525 if (space - bspace > lorunningspace) 526 return; 527 runningwakeup(); 528 } 529 530 /* 531 * bufcountadd: 532 * 533 * Called when a buffer has been added to one of the free queues to 534 * account for the buffer and to wakeup anyone waiting for free buffers. 535 * This typically occurs when large amounts of metadata are being handled 536 * by the buffer cache ( else buffer space runs out first, usually ). 537 */ 538 static __inline void 539 bufcountadd(struct buf *bp) 540 { 541 int mask, need_wakeup, old, on; 542 543 KASSERT((bp->b_flags & B_INFREECNT) == 0, 544 ("buf %p already counted as free", bp)); 545 bp->b_flags |= B_INFREECNT; 546 old = atomic_fetchadd_int(&numfreebuffers, 1); 547 KASSERT(old >= 0 && old < nbuf, 548 ("numfreebuffers climbed to %d", old + 1)); 549 mask = VFS_BIO_NEED_ANY; 550 if (numfreebuffers >= hifreebuffers) 551 mask |= VFS_BIO_NEED_FREE; 552 rw_rlock(&nblock); 553 for (;;) { 554 need_wakeup = 0; 555 on = needsbuffer; 556 if (on == 0) 557 break; 558 need_wakeup = 1; 559 if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask)) 560 break; 561 } 562 if (need_wakeup) 563 wakeup(__DEVOLATILE(void *, &needsbuffer)); 564 rw_runlock(&nblock); 565 } 566 567 /* 568 * bufcountsub: 569 * 570 * Decrement the numfreebuffers count as needed. 571 */ 572 static void 573 bufcountsub(struct buf *bp) 574 { 575 int old; 576 577 /* 578 * Fixup numfreebuffers count. If the buffer is invalid or not 579 * delayed-write, the buffer was free and we must decrement 580 * numfreebuffers. 581 */ 582 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 583 KASSERT((bp->b_flags & B_INFREECNT) != 0, 584 ("buf %p not counted in numfreebuffers", bp)); 585 bp->b_flags &= ~B_INFREECNT; 586 old = atomic_fetchadd_int(&numfreebuffers, -1); 587 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 588 } 589 } 590 591 /* 592 * waitrunningbufspace() 593 * 594 * runningbufspace is a measure of the amount of I/O currently 595 * running. This routine is used in async-write situations to 596 * prevent creating huge backups of pending writes to a device. 597 * Only asynchronous writes are governed by this function. 598 * 599 * This does NOT turn an async write into a sync write. It waits 600 * for earlier writes to complete and generally returns before the 601 * caller's write has reached the device. 602 */ 603 void 604 waitrunningbufspace(void) 605 { 606 607 mtx_lock(&rbreqlock); 608 while (runningbufspace > hirunningspace) { 609 runningbufreq = 1; 610 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 611 } 612 mtx_unlock(&rbreqlock); 613 } 614 615 616 /* 617 * vfs_buf_test_cache: 618 * 619 * Called when a buffer is extended. This function clears the B_CACHE 620 * bit if the newly extended portion of the buffer does not contain 621 * valid data. 622 */ 623 static __inline 624 void 625 vfs_buf_test_cache(struct buf *bp, 626 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 627 vm_page_t m) 628 { 629 630 VM_OBJECT_ASSERT_LOCKED(m->object); 631 if (bp->b_flags & B_CACHE) { 632 int base = (foff + off) & PAGE_MASK; 633 if (vm_page_is_valid(m, base, size) == 0) 634 bp->b_flags &= ~B_CACHE; 635 } 636 } 637 638 /* Wake up the buffer daemon if necessary */ 639 static __inline void 640 bd_wakeup(void) 641 { 642 643 mtx_lock(&bdlock); 644 if (bd_request == 0) { 645 bd_request = 1; 646 wakeup(&bd_request); 647 } 648 mtx_unlock(&bdlock); 649 } 650 651 /* 652 * bd_speedup - speedup the buffer cache flushing code 653 */ 654 void 655 bd_speedup(void) 656 { 657 int needwake; 658 659 mtx_lock(&bdlock); 660 needwake = 0; 661 if (bd_speedupreq == 0 || bd_request == 0) 662 needwake = 1; 663 bd_speedupreq = 1; 664 bd_request = 1; 665 if (needwake) 666 wakeup(&bd_request); 667 mtx_unlock(&bdlock); 668 } 669 670 #ifndef NSWBUF_MIN 671 #define NSWBUF_MIN 16 672 #endif 673 674 #ifdef __i386__ 675 #define TRANSIENT_DENOM 5 676 #else 677 #define TRANSIENT_DENOM 10 678 #endif 679 680 /* 681 * Calculating buffer cache scaling values and reserve space for buffer 682 * headers. This is called during low level kernel initialization and 683 * may be called more then once. We CANNOT write to the memory area 684 * being reserved at this time. 685 */ 686 caddr_t 687 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 688 { 689 int tuned_nbuf; 690 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 691 692 /* 693 * physmem_est is in pages. Convert it to kilobytes (assumes 694 * PAGE_SIZE is >= 1K) 695 */ 696 physmem_est = physmem_est * (PAGE_SIZE / 1024); 697 698 /* 699 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 700 * For the first 64MB of ram nominally allocate sufficient buffers to 701 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 702 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 703 * the buffer cache we limit the eventual kva reservation to 704 * maxbcache bytes. 705 * 706 * factor represents the 1/4 x ram conversion. 707 */ 708 if (nbuf == 0) { 709 int factor = 4 * BKVASIZE / 1024; 710 711 nbuf = 50; 712 if (physmem_est > 4096) 713 nbuf += min((physmem_est - 4096) / factor, 714 65536 / factor); 715 if (physmem_est > 65536) 716 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 717 32 * 1024 * 1024 / (factor * 5)); 718 719 if (maxbcache && nbuf > maxbcache / BKVASIZE) 720 nbuf = maxbcache / BKVASIZE; 721 tuned_nbuf = 1; 722 } else 723 tuned_nbuf = 0; 724 725 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 726 maxbuf = (LONG_MAX / 3) / BKVASIZE; 727 if (nbuf > maxbuf) { 728 if (!tuned_nbuf) 729 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 730 maxbuf); 731 nbuf = maxbuf; 732 } 733 734 /* 735 * Ideal allocation size for the transient bio submap is 10% 736 * of the maximal space buffer map. This roughly corresponds 737 * to the amount of the buffer mapped for typical UFS load. 738 * 739 * Clip the buffer map to reserve space for the transient 740 * BIOs, if its extent is bigger than 90% (80% on i386) of the 741 * maximum buffer map extent on the platform. 742 * 743 * The fall-back to the maxbuf in case of maxbcache unset, 744 * allows to not trim the buffer KVA for the architectures 745 * with ample KVA space. 746 */ 747 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 748 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 749 buf_sz = (long)nbuf * BKVASIZE; 750 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 751 (TRANSIENT_DENOM - 1)) { 752 /* 753 * There is more KVA than memory. Do not 754 * adjust buffer map size, and assign the rest 755 * of maxbuf to transient map. 756 */ 757 biotmap_sz = maxbuf_sz - buf_sz; 758 } else { 759 /* 760 * Buffer map spans all KVA we could afford on 761 * this platform. Give 10% (20% on i386) of 762 * the buffer map to the transient bio map. 763 */ 764 biotmap_sz = buf_sz / TRANSIENT_DENOM; 765 buf_sz -= biotmap_sz; 766 } 767 if (biotmap_sz / INT_MAX > MAXPHYS) 768 bio_transient_maxcnt = INT_MAX; 769 else 770 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 771 /* 772 * Artifically limit to 1024 simultaneous in-flight I/Os 773 * using the transient mapping. 774 */ 775 if (bio_transient_maxcnt > 1024) 776 bio_transient_maxcnt = 1024; 777 if (tuned_nbuf) 778 nbuf = buf_sz / BKVASIZE; 779 } 780 781 /* 782 * swbufs are used as temporary holders for I/O, such as paging I/O. 783 * We have no less then 16 and no more then 256. 784 */ 785 nswbuf = min(nbuf / 4, 256); 786 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf); 787 if (nswbuf < NSWBUF_MIN) 788 nswbuf = NSWBUF_MIN; 789 790 /* 791 * Reserve space for the buffer cache buffers 792 */ 793 swbuf = (void *)v; 794 v = (caddr_t)(swbuf + nswbuf); 795 buf = (void *)v; 796 v = (caddr_t)(buf + nbuf); 797 798 return(v); 799 } 800 801 /* Initialize the buffer subsystem. Called before use of any buffers. */ 802 void 803 bufinit(void) 804 { 805 struct buf *bp; 806 int i; 807 808 CTASSERT(MAXBCACHEBUF >= MAXBSIZE); 809 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF); 810 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF); 811 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 812 rw_init(&nblock, "needsbuffer lock"); 813 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 814 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 815 816 /* next, make a null set of free lists */ 817 for (i = 0; i < BUFFER_QUEUES; i++) 818 TAILQ_INIT(&bufqueues[i]); 819 820 /* finally, initialize each buffer header and stick on empty q */ 821 for (i = 0; i < nbuf; i++) { 822 bp = &buf[i]; 823 bzero(bp, sizeof *bp); 824 bp->b_flags = B_INVAL | B_INFREECNT; 825 bp->b_rcred = NOCRED; 826 bp->b_wcred = NOCRED; 827 bp->b_qindex = QUEUE_EMPTY; 828 bp->b_xflags = 0; 829 LIST_INIT(&bp->b_dep); 830 BUF_LOCKINIT(bp); 831 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 832 #ifdef INVARIANTS 833 bq_len[QUEUE_EMPTY]++; 834 #endif 835 } 836 837 /* 838 * maxbufspace is the absolute maximum amount of buffer space we are 839 * allowed to reserve in KVM and in real terms. The absolute maximum 840 * is nominally used by buf_daemon. hibufspace is the nominal maximum 841 * used by most other processes. The differential is required to 842 * ensure that buf_daemon is able to run when other processes might 843 * be blocked waiting for buffer space. 844 * 845 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 846 * this may result in KVM fragmentation which is not handled optimally 847 * by the system. 848 */ 849 maxbufspace = (long)nbuf * BKVASIZE; 850 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10); 851 lobufspace = hibufspace - MAXBCACHEBUF; 852 853 /* 854 * Note: The 16 MiB upper limit for hirunningspace was chosen 855 * arbitrarily and may need further tuning. It corresponds to 856 * 128 outstanding write IO requests (if IO size is 128 KiB), 857 * which fits with many RAID controllers' tagged queuing limits. 858 * The lower 1 MiB limit is the historical upper limit for 859 * hirunningspace. 860 */ 861 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF), 862 16 * 1024 * 1024), 1024 * 1024); 863 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF); 864 865 /* 866 * Limit the amount of malloc memory since it is wired permanently into 867 * the kernel space. Even though this is accounted for in the buffer 868 * allocation, we don't want the malloced region to grow uncontrolled. 869 * The malloc scheme improves memory utilization significantly on average 870 * (small) directories. 871 */ 872 maxbufmallocspace = hibufspace / 20; 873 874 /* 875 * Reduce the chance of a deadlock occuring by limiting the number 876 * of delayed-write dirty buffers we allow to stack up. 877 */ 878 hidirtybuffers = nbuf / 4 + 20; 879 dirtybufthresh = hidirtybuffers * 9 / 10; 880 numdirtybuffers = 0; 881 /* 882 * To support extreme low-memory systems, make sure hidirtybuffers cannot 883 * eat up all available buffer space. This occurs when our minimum cannot 884 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 885 * BKVASIZE'd buffers. 886 */ 887 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 888 hidirtybuffers >>= 1; 889 } 890 lodirtybuffers = hidirtybuffers / 2; 891 892 /* 893 * Try to keep the number of free buffers in the specified range, 894 * and give special processes (e.g. like buf_daemon) access to an 895 * emergency reserve. 896 */ 897 lofreebuffers = nbuf / 18 + 5; 898 hifreebuffers = 2 * lofreebuffers; 899 numfreebuffers = nbuf; 900 901 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 902 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 903 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 904 } 905 906 #ifdef INVARIANTS 907 static inline void 908 vfs_buf_check_mapped(struct buf *bp) 909 { 910 911 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 912 ("mapped buf %p %x", bp, bp->b_flags)); 913 KASSERT(bp->b_kvabase != unmapped_buf, 914 ("mapped buf: b_kvabase was not updated %p", bp)); 915 KASSERT(bp->b_data != unmapped_buf, 916 ("mapped buf: b_data was not updated %p", bp)); 917 } 918 919 static inline void 920 vfs_buf_check_unmapped(struct buf *bp) 921 { 922 923 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 924 ("unmapped buf %p %x", bp, bp->b_flags)); 925 KASSERT(bp->b_kvabase == unmapped_buf, 926 ("unmapped buf: corrupted b_kvabase %p", bp)); 927 KASSERT(bp->b_data == unmapped_buf, 928 ("unmapped buf: corrupted b_data %p", bp)); 929 } 930 931 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 932 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 933 #else 934 #define BUF_CHECK_MAPPED(bp) do {} while (0) 935 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 936 #endif 937 938 static void 939 bpmap_qenter(struct buf *bp) 940 { 941 942 BUF_CHECK_MAPPED(bp); 943 944 /* 945 * bp->b_data is relative to bp->b_offset, but 946 * bp->b_offset may be offset into the first page. 947 */ 948 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 949 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 950 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 951 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 952 } 953 954 /* 955 * bfreekva() - free the kva allocation for a buffer. 956 * 957 * Since this call frees up buffer space, we call bufspacewakeup(). 958 */ 959 static void 960 bfreekva(struct buf *bp) 961 { 962 963 if (bp->b_kvasize == 0) 964 return; 965 966 atomic_add_int(&buffreekvacnt, 1); 967 atomic_subtract_long(&bufspace, bp->b_kvasize); 968 if ((bp->b_flags & B_UNMAPPED) == 0) { 969 BUF_CHECK_MAPPED(bp); 970 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, 971 bp->b_kvasize); 972 } else { 973 BUF_CHECK_UNMAPPED(bp); 974 if ((bp->b_flags & B_KVAALLOC) != 0) { 975 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc, 976 bp->b_kvasize); 977 } 978 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 979 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 980 } 981 bp->b_kvasize = 0; 982 bufspacewakeup(); 983 } 984 985 /* 986 * binsfree: 987 * 988 * Insert the buffer into the appropriate free list. 989 */ 990 static void 991 binsfree(struct buf *bp, int qindex) 992 { 993 struct mtx *olock, *nlock; 994 995 BUF_ASSERT_XLOCKED(bp); 996 997 olock = bqlock(bp->b_qindex); 998 nlock = bqlock(qindex); 999 mtx_lock(olock); 1000 /* Handle delayed bremfree() processing. */ 1001 if (bp->b_flags & B_REMFREE) 1002 bremfreel(bp); 1003 1004 if (bp->b_qindex != QUEUE_NONE) 1005 panic("binsfree: free buffer onto another queue???"); 1006 1007 bp->b_qindex = qindex; 1008 if (olock != nlock) { 1009 mtx_unlock(olock); 1010 mtx_lock(nlock); 1011 } 1012 if (bp->b_flags & B_AGE) 1013 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1014 else 1015 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1016 #ifdef INVARIANTS 1017 bq_len[bp->b_qindex]++; 1018 #endif 1019 mtx_unlock(nlock); 1020 1021 /* 1022 * Something we can maybe free or reuse. 1023 */ 1024 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1025 bufspacewakeup(); 1026 1027 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1028 bufcountadd(bp); 1029 } 1030 1031 /* 1032 * bremfree: 1033 * 1034 * Mark the buffer for removal from the appropriate free list. 1035 * 1036 */ 1037 void 1038 bremfree(struct buf *bp) 1039 { 1040 1041 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1042 KASSERT((bp->b_flags & B_REMFREE) == 0, 1043 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1044 KASSERT(bp->b_qindex != QUEUE_NONE, 1045 ("bremfree: buffer %p not on a queue.", bp)); 1046 BUF_ASSERT_XLOCKED(bp); 1047 1048 bp->b_flags |= B_REMFREE; 1049 bufcountsub(bp); 1050 } 1051 1052 /* 1053 * bremfreef: 1054 * 1055 * Force an immediate removal from a free list. Used only in nfs when 1056 * it abuses the b_freelist pointer. 1057 */ 1058 void 1059 bremfreef(struct buf *bp) 1060 { 1061 struct mtx *qlock; 1062 1063 qlock = bqlock(bp->b_qindex); 1064 mtx_lock(qlock); 1065 bremfreel(bp); 1066 mtx_unlock(qlock); 1067 } 1068 1069 /* 1070 * bremfreel: 1071 * 1072 * Removes a buffer from the free list, must be called with the 1073 * correct qlock held. 1074 */ 1075 static void 1076 bremfreel(struct buf *bp) 1077 { 1078 1079 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1080 bp, bp->b_vp, bp->b_flags); 1081 KASSERT(bp->b_qindex != QUEUE_NONE, 1082 ("bremfreel: buffer %p not on a queue.", bp)); 1083 BUF_ASSERT_XLOCKED(bp); 1084 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1085 1086 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1087 #ifdef INVARIANTS 1088 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1089 bp->b_qindex)); 1090 bq_len[bp->b_qindex]--; 1091 #endif 1092 bp->b_qindex = QUEUE_NONE; 1093 /* 1094 * If this was a delayed bremfree() we only need to remove the buffer 1095 * from the queue and return the stats are already done. 1096 */ 1097 if (bp->b_flags & B_REMFREE) { 1098 bp->b_flags &= ~B_REMFREE; 1099 return; 1100 } 1101 bufcountsub(bp); 1102 } 1103 1104 /* 1105 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1106 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1107 * the buffer is valid and we do not have to do anything. 1108 */ 1109 void 1110 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 1111 int cnt, struct ucred * cred) 1112 { 1113 struct buf *rabp; 1114 int i; 1115 1116 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1117 if (inmem(vp, *rablkno)) 1118 continue; 1119 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1120 1121 if ((rabp->b_flags & B_CACHE) == 0) { 1122 if (!TD_IS_IDLETHREAD(curthread)) 1123 curthread->td_ru.ru_inblock++; 1124 rabp->b_flags |= B_ASYNC; 1125 rabp->b_flags &= ~B_INVAL; 1126 rabp->b_ioflags &= ~BIO_ERROR; 1127 rabp->b_iocmd = BIO_READ; 1128 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1129 rabp->b_rcred = crhold(cred); 1130 vfs_busy_pages(rabp, 0); 1131 BUF_KERNPROC(rabp); 1132 rabp->b_iooffset = dbtob(rabp->b_blkno); 1133 bstrategy(rabp); 1134 } else { 1135 brelse(rabp); 1136 } 1137 } 1138 } 1139 1140 /* 1141 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1142 * 1143 * Get a buffer with the specified data. Look in the cache first. We 1144 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1145 * is set, the buffer is valid and we do not have to do anything, see 1146 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1147 */ 1148 int 1149 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1150 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 1151 { 1152 struct buf *bp; 1153 int rv = 0, readwait = 0; 1154 1155 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1156 /* 1157 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1158 */ 1159 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1160 if (bp == NULL) 1161 return (EBUSY); 1162 1163 /* if not found in cache, do some I/O */ 1164 if ((bp->b_flags & B_CACHE) == 0) { 1165 if (!TD_IS_IDLETHREAD(curthread)) 1166 curthread->td_ru.ru_inblock++; 1167 bp->b_iocmd = BIO_READ; 1168 bp->b_flags &= ~B_INVAL; 1169 bp->b_ioflags &= ~BIO_ERROR; 1170 if (bp->b_rcred == NOCRED && cred != NOCRED) 1171 bp->b_rcred = crhold(cred); 1172 vfs_busy_pages(bp, 0); 1173 bp->b_iooffset = dbtob(bp->b_blkno); 1174 bstrategy(bp); 1175 ++readwait; 1176 } 1177 1178 breada(vp, rablkno, rabsize, cnt, cred); 1179 1180 if (readwait) { 1181 rv = bufwait(bp); 1182 } 1183 return (rv); 1184 } 1185 1186 /* 1187 * Write, release buffer on completion. (Done by iodone 1188 * if async). Do not bother writing anything if the buffer 1189 * is invalid. 1190 * 1191 * Note that we set B_CACHE here, indicating that buffer is 1192 * fully valid and thus cacheable. This is true even of NFS 1193 * now so we set it generally. This could be set either here 1194 * or in biodone() since the I/O is synchronous. We put it 1195 * here. 1196 */ 1197 int 1198 bufwrite(struct buf *bp) 1199 { 1200 int oldflags; 1201 struct vnode *vp; 1202 long space; 1203 int vp_md; 1204 1205 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1206 if (bp->b_flags & B_INVAL) { 1207 brelse(bp); 1208 return (0); 1209 } 1210 1211 if (bp->b_flags & B_BARRIER) 1212 barrierwrites++; 1213 1214 oldflags = bp->b_flags; 1215 1216 BUF_ASSERT_HELD(bp); 1217 1218 if (bp->b_pin_count > 0) 1219 bunpin_wait(bp); 1220 1221 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1222 ("FFS background buffer should not get here %p", bp)); 1223 1224 vp = bp->b_vp; 1225 if (vp) 1226 vp_md = vp->v_vflag & VV_MD; 1227 else 1228 vp_md = 0; 1229 1230 /* 1231 * Mark the buffer clean. Increment the bufobj write count 1232 * before bundirty() call, to prevent other thread from seeing 1233 * empty dirty list and zero counter for writes in progress, 1234 * falsely indicating that the bufobj is clean. 1235 */ 1236 bufobj_wref(bp->b_bufobj); 1237 bundirty(bp); 1238 1239 bp->b_flags &= ~B_DONE; 1240 bp->b_ioflags &= ~BIO_ERROR; 1241 bp->b_flags |= B_CACHE; 1242 bp->b_iocmd = BIO_WRITE; 1243 1244 vfs_busy_pages(bp, 1); 1245 1246 /* 1247 * Normal bwrites pipeline writes 1248 */ 1249 bp->b_runningbufspace = bp->b_bufsize; 1250 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1251 1252 if (!TD_IS_IDLETHREAD(curthread)) 1253 curthread->td_ru.ru_oublock++; 1254 if (oldflags & B_ASYNC) 1255 BUF_KERNPROC(bp); 1256 bp->b_iooffset = dbtob(bp->b_blkno); 1257 bstrategy(bp); 1258 1259 if ((oldflags & B_ASYNC) == 0) { 1260 int rtval = bufwait(bp); 1261 brelse(bp); 1262 return (rtval); 1263 } else if (space > hirunningspace) { 1264 /* 1265 * don't allow the async write to saturate the I/O 1266 * system. We will not deadlock here because 1267 * we are blocking waiting for I/O that is already in-progress 1268 * to complete. We do not block here if it is the update 1269 * or syncer daemon trying to clean up as that can lead 1270 * to deadlock. 1271 */ 1272 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1273 waitrunningbufspace(); 1274 } 1275 1276 return (0); 1277 } 1278 1279 void 1280 bufbdflush(struct bufobj *bo, struct buf *bp) 1281 { 1282 struct buf *nbp; 1283 1284 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1285 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1286 altbufferflushes++; 1287 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1288 BO_LOCK(bo); 1289 /* 1290 * Try to find a buffer to flush. 1291 */ 1292 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1293 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1294 BUF_LOCK(nbp, 1295 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1296 continue; 1297 if (bp == nbp) 1298 panic("bdwrite: found ourselves"); 1299 BO_UNLOCK(bo); 1300 /* Don't countdeps with the bo lock held. */ 1301 if (buf_countdeps(nbp, 0)) { 1302 BO_LOCK(bo); 1303 BUF_UNLOCK(nbp); 1304 continue; 1305 } 1306 if (nbp->b_flags & B_CLUSTEROK) { 1307 vfs_bio_awrite(nbp); 1308 } else { 1309 bremfree(nbp); 1310 bawrite(nbp); 1311 } 1312 dirtybufferflushes++; 1313 break; 1314 } 1315 if (nbp == NULL) 1316 BO_UNLOCK(bo); 1317 } 1318 } 1319 1320 /* 1321 * Delayed write. (Buffer is marked dirty). Do not bother writing 1322 * anything if the buffer is marked invalid. 1323 * 1324 * Note that since the buffer must be completely valid, we can safely 1325 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1326 * biodone() in order to prevent getblk from writing the buffer 1327 * out synchronously. 1328 */ 1329 void 1330 bdwrite(struct buf *bp) 1331 { 1332 struct thread *td = curthread; 1333 struct vnode *vp; 1334 struct bufobj *bo; 1335 1336 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1337 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1338 KASSERT((bp->b_flags & B_BARRIER) == 0, 1339 ("Barrier request in delayed write %p", bp)); 1340 BUF_ASSERT_HELD(bp); 1341 1342 if (bp->b_flags & B_INVAL) { 1343 brelse(bp); 1344 return; 1345 } 1346 1347 /* 1348 * If we have too many dirty buffers, don't create any more. 1349 * If we are wildly over our limit, then force a complete 1350 * cleanup. Otherwise, just keep the situation from getting 1351 * out of control. Note that we have to avoid a recursive 1352 * disaster and not try to clean up after our own cleanup! 1353 */ 1354 vp = bp->b_vp; 1355 bo = bp->b_bufobj; 1356 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1357 td->td_pflags |= TDP_INBDFLUSH; 1358 BO_BDFLUSH(bo, bp); 1359 td->td_pflags &= ~TDP_INBDFLUSH; 1360 } else 1361 recursiveflushes++; 1362 1363 bdirty(bp); 1364 /* 1365 * Set B_CACHE, indicating that the buffer is fully valid. This is 1366 * true even of NFS now. 1367 */ 1368 bp->b_flags |= B_CACHE; 1369 1370 /* 1371 * This bmap keeps the system from needing to do the bmap later, 1372 * perhaps when the system is attempting to do a sync. Since it 1373 * is likely that the indirect block -- or whatever other datastructure 1374 * that the filesystem needs is still in memory now, it is a good 1375 * thing to do this. Note also, that if the pageout daemon is 1376 * requesting a sync -- there might not be enough memory to do 1377 * the bmap then... So, this is important to do. 1378 */ 1379 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1380 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1381 } 1382 1383 /* 1384 * Set the *dirty* buffer range based upon the VM system dirty 1385 * pages. 1386 * 1387 * Mark the buffer pages as clean. We need to do this here to 1388 * satisfy the vnode_pager and the pageout daemon, so that it 1389 * thinks that the pages have been "cleaned". Note that since 1390 * the pages are in a delayed write buffer -- the VFS layer 1391 * "will" see that the pages get written out on the next sync, 1392 * or perhaps the cluster will be completed. 1393 */ 1394 vfs_clean_pages_dirty_buf(bp); 1395 bqrelse(bp); 1396 1397 /* 1398 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1399 * due to the softdep code. 1400 */ 1401 } 1402 1403 /* 1404 * bdirty: 1405 * 1406 * Turn buffer into delayed write request. We must clear BIO_READ and 1407 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1408 * itself to properly update it in the dirty/clean lists. We mark it 1409 * B_DONE to ensure that any asynchronization of the buffer properly 1410 * clears B_DONE ( else a panic will occur later ). 1411 * 1412 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1413 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1414 * should only be called if the buffer is known-good. 1415 * 1416 * Since the buffer is not on a queue, we do not update the numfreebuffers 1417 * count. 1418 * 1419 * The buffer must be on QUEUE_NONE. 1420 */ 1421 void 1422 bdirty(struct buf *bp) 1423 { 1424 1425 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1426 bp, bp->b_vp, bp->b_flags); 1427 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1428 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1429 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1430 BUF_ASSERT_HELD(bp); 1431 bp->b_flags &= ~(B_RELBUF); 1432 bp->b_iocmd = BIO_WRITE; 1433 1434 if ((bp->b_flags & B_DELWRI) == 0) { 1435 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1436 reassignbuf(bp); 1437 bdirtyadd(); 1438 } 1439 } 1440 1441 /* 1442 * bundirty: 1443 * 1444 * Clear B_DELWRI for buffer. 1445 * 1446 * Since the buffer is not on a queue, we do not update the numfreebuffers 1447 * count. 1448 * 1449 * The buffer must be on QUEUE_NONE. 1450 */ 1451 1452 void 1453 bundirty(struct buf *bp) 1454 { 1455 1456 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1457 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1458 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1459 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1460 BUF_ASSERT_HELD(bp); 1461 1462 if (bp->b_flags & B_DELWRI) { 1463 bp->b_flags &= ~B_DELWRI; 1464 reassignbuf(bp); 1465 bdirtysub(); 1466 } 1467 /* 1468 * Since it is now being written, we can clear its deferred write flag. 1469 */ 1470 bp->b_flags &= ~B_DEFERRED; 1471 } 1472 1473 /* 1474 * bawrite: 1475 * 1476 * Asynchronous write. Start output on a buffer, but do not wait for 1477 * it to complete. The buffer is released when the output completes. 1478 * 1479 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1480 * B_INVAL buffers. Not us. 1481 */ 1482 void 1483 bawrite(struct buf *bp) 1484 { 1485 1486 bp->b_flags |= B_ASYNC; 1487 (void) bwrite(bp); 1488 } 1489 1490 /* 1491 * babarrierwrite: 1492 * 1493 * Asynchronous barrier write. Start output on a buffer, but do not 1494 * wait for it to complete. Place a write barrier after this write so 1495 * that this buffer and all buffers written before it are committed to 1496 * the disk before any buffers written after this write are committed 1497 * to the disk. The buffer is released when the output completes. 1498 */ 1499 void 1500 babarrierwrite(struct buf *bp) 1501 { 1502 1503 bp->b_flags |= B_ASYNC | B_BARRIER; 1504 (void) bwrite(bp); 1505 } 1506 1507 /* 1508 * bbarrierwrite: 1509 * 1510 * Synchronous barrier write. Start output on a buffer and wait for 1511 * it to complete. Place a write barrier after this write so that 1512 * this buffer and all buffers written before it are committed to 1513 * the disk before any buffers written after this write are committed 1514 * to the disk. The buffer is released when the output completes. 1515 */ 1516 int 1517 bbarrierwrite(struct buf *bp) 1518 { 1519 1520 bp->b_flags |= B_BARRIER; 1521 return (bwrite(bp)); 1522 } 1523 1524 /* 1525 * bwillwrite: 1526 * 1527 * Called prior to the locking of any vnodes when we are expecting to 1528 * write. We do not want to starve the buffer cache with too many 1529 * dirty buffers so we block here. By blocking prior to the locking 1530 * of any vnodes we attempt to avoid the situation where a locked vnode 1531 * prevents the various system daemons from flushing related buffers. 1532 */ 1533 void 1534 bwillwrite(void) 1535 { 1536 1537 if (numdirtybuffers >= hidirtybuffers) { 1538 mtx_lock(&bdirtylock); 1539 while (numdirtybuffers >= hidirtybuffers) { 1540 bdirtywait = 1; 1541 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 1542 "flswai", 0); 1543 } 1544 mtx_unlock(&bdirtylock); 1545 } 1546 } 1547 1548 /* 1549 * Return true if we have too many dirty buffers. 1550 */ 1551 int 1552 buf_dirty_count_severe(void) 1553 { 1554 1555 return(numdirtybuffers >= hidirtybuffers); 1556 } 1557 1558 static __noinline int 1559 buf_vm_page_count_severe(void) 1560 { 1561 1562 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1563 1564 return vm_page_count_severe(); 1565 } 1566 1567 /* 1568 * brelse: 1569 * 1570 * Release a busy buffer and, if requested, free its resources. The 1571 * buffer will be stashed in the appropriate bufqueue[] allowing it 1572 * to be accessed later as a cache entity or reused for other purposes. 1573 */ 1574 void 1575 brelse(struct buf *bp) 1576 { 1577 int qindex; 1578 1579 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1580 bp, bp->b_vp, bp->b_flags); 1581 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1582 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1583 1584 if (BUF_LOCKRECURSED(bp)) { 1585 /* 1586 * Do not process, in particular, do not handle the 1587 * B_INVAL/B_RELBUF and do not release to free list. 1588 */ 1589 BUF_UNLOCK(bp); 1590 return; 1591 } 1592 1593 if (bp->b_flags & B_MANAGED) { 1594 bqrelse(bp); 1595 return; 1596 } 1597 1598 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1599 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1600 /* 1601 * Failed write, redirty. Must clear BIO_ERROR to prevent 1602 * pages from being scrapped. If the error is anything 1603 * other than an I/O error (EIO), assume that retrying 1604 * is futile. 1605 */ 1606 bp->b_ioflags &= ~BIO_ERROR; 1607 bdirty(bp); 1608 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1609 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1610 /* 1611 * Either a failed I/O or we were asked to free or not 1612 * cache the buffer. 1613 */ 1614 bp->b_flags |= B_INVAL; 1615 if (!LIST_EMPTY(&bp->b_dep)) 1616 buf_deallocate(bp); 1617 if (bp->b_flags & B_DELWRI) 1618 bdirtysub(); 1619 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1620 if ((bp->b_flags & B_VMIO) == 0) { 1621 if (bp->b_bufsize) 1622 allocbuf(bp, 0); 1623 if (bp->b_vp) 1624 brelvp(bp); 1625 } 1626 } 1627 1628 /* 1629 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1630 * is called with B_DELWRI set, the underlying pages may wind up 1631 * getting freed causing a previous write (bdwrite()) to get 'lost' 1632 * because pages associated with a B_DELWRI bp are marked clean. 1633 * 1634 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1635 * if B_DELWRI is set. 1636 * 1637 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1638 * on pages to return pages to the VM page queues. 1639 */ 1640 if (bp->b_flags & B_DELWRI) 1641 bp->b_flags &= ~B_RELBUF; 1642 else if (buf_vm_page_count_severe()) { 1643 /* 1644 * BKGRDINPROG can only be set with the buf and bufobj 1645 * locks both held. We tolerate a race to clear it here. 1646 */ 1647 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1648 bp->b_flags |= B_RELBUF; 1649 } 1650 1651 /* 1652 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1653 * constituted, not even NFS buffers now. Two flags effect this. If 1654 * B_INVAL, the struct buf is invalidated but the VM object is kept 1655 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1656 * 1657 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1658 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1659 * buffer is also B_INVAL because it hits the re-dirtying code above. 1660 * 1661 * Normally we can do this whether a buffer is B_DELWRI or not. If 1662 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1663 * the commit state and we cannot afford to lose the buffer. If the 1664 * buffer has a background write in progress, we need to keep it 1665 * around to prevent it from being reconstituted and starting a second 1666 * background write. 1667 */ 1668 if ((bp->b_flags & B_VMIO) 1669 && !(bp->b_vp->v_mount != NULL && 1670 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1671 !vn_isdisk(bp->b_vp, NULL) && 1672 (bp->b_flags & B_DELWRI)) 1673 ) { 1674 1675 int i, j, resid; 1676 vm_page_t m; 1677 off_t foff; 1678 vm_pindex_t poff; 1679 vm_object_t obj; 1680 1681 obj = bp->b_bufobj->bo_object; 1682 1683 /* 1684 * Get the base offset and length of the buffer. Note that 1685 * in the VMIO case if the buffer block size is not 1686 * page-aligned then b_data pointer may not be page-aligned. 1687 * But our b_pages[] array *IS* page aligned. 1688 * 1689 * block sizes less then DEV_BSIZE (usually 512) are not 1690 * supported due to the page granularity bits (m->valid, 1691 * m->dirty, etc...). 1692 * 1693 * See man buf(9) for more information 1694 */ 1695 resid = bp->b_bufsize; 1696 foff = bp->b_offset; 1697 for (i = 0; i < bp->b_npages; i++) { 1698 int had_bogus = 0; 1699 1700 m = bp->b_pages[i]; 1701 1702 /* 1703 * If we hit a bogus page, fixup *all* the bogus pages 1704 * now. 1705 */ 1706 if (m == bogus_page) { 1707 poff = OFF_TO_IDX(bp->b_offset); 1708 had_bogus = 1; 1709 1710 VM_OBJECT_RLOCK(obj); 1711 for (j = i; j < bp->b_npages; j++) { 1712 vm_page_t mtmp; 1713 mtmp = bp->b_pages[j]; 1714 if (mtmp == bogus_page) { 1715 mtmp = vm_page_lookup(obj, poff + j); 1716 if (!mtmp) { 1717 panic("brelse: page missing\n"); 1718 } 1719 bp->b_pages[j] = mtmp; 1720 } 1721 } 1722 VM_OBJECT_RUNLOCK(obj); 1723 1724 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1725 BUF_CHECK_MAPPED(bp); 1726 pmap_qenter( 1727 trunc_page((vm_offset_t)bp->b_data), 1728 bp->b_pages, bp->b_npages); 1729 } 1730 m = bp->b_pages[i]; 1731 } 1732 if ((bp->b_flags & B_NOCACHE) || 1733 (bp->b_ioflags & BIO_ERROR && 1734 bp->b_iocmd == BIO_READ)) { 1735 int poffset = foff & PAGE_MASK; 1736 int presid = resid > (PAGE_SIZE - poffset) ? 1737 (PAGE_SIZE - poffset) : resid; 1738 1739 KASSERT(presid >= 0, ("brelse: extra page")); 1740 VM_OBJECT_WLOCK(obj); 1741 while (vm_page_xbusied(m)) { 1742 vm_page_lock(m); 1743 VM_OBJECT_WUNLOCK(obj); 1744 vm_page_busy_sleep(m, "mbncsh"); 1745 VM_OBJECT_WLOCK(obj); 1746 } 1747 if (pmap_page_wired_mappings(m) == 0) 1748 vm_page_set_invalid(m, poffset, presid); 1749 VM_OBJECT_WUNLOCK(obj); 1750 if (had_bogus) 1751 printf("avoided corruption bug in bogus_page/brelse code\n"); 1752 } 1753 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1754 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1755 } 1756 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1757 vfs_vmio_release(bp); 1758 1759 } else if (bp->b_flags & B_VMIO) { 1760 1761 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1762 vfs_vmio_release(bp); 1763 } 1764 1765 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1766 if (bp->b_bufsize != 0) 1767 allocbuf(bp, 0); 1768 if (bp->b_vp != NULL) 1769 brelvp(bp); 1770 } 1771 1772 /* 1773 * If the buffer has junk contents signal it and eventually 1774 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1775 * doesn't find it. 1776 */ 1777 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1778 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1779 bp->b_flags |= B_INVAL; 1780 if (bp->b_flags & B_INVAL) { 1781 if (bp->b_flags & B_DELWRI) 1782 bundirty(bp); 1783 if (bp->b_vp) 1784 brelvp(bp); 1785 } 1786 1787 /* buffers with no memory */ 1788 if (bp->b_bufsize == 0) { 1789 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1790 if (bp->b_vflags & BV_BKGRDINPROG) 1791 panic("losing buffer 1"); 1792 if (bp->b_kvasize) 1793 qindex = QUEUE_EMPTYKVA; 1794 else 1795 qindex = QUEUE_EMPTY; 1796 bp->b_flags |= B_AGE; 1797 /* buffers with junk contents */ 1798 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1799 (bp->b_ioflags & BIO_ERROR)) { 1800 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1801 if (bp->b_vflags & BV_BKGRDINPROG) 1802 panic("losing buffer 2"); 1803 qindex = QUEUE_CLEAN; 1804 bp->b_flags |= B_AGE; 1805 /* remaining buffers */ 1806 } else if (bp->b_flags & B_DELWRI) 1807 qindex = QUEUE_DIRTY; 1808 else 1809 qindex = QUEUE_CLEAN; 1810 1811 binsfree(bp, qindex); 1812 1813 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1814 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1815 panic("brelse: not dirty"); 1816 /* unlock */ 1817 BUF_UNLOCK(bp); 1818 } 1819 1820 /* 1821 * Release a buffer back to the appropriate queue but do not try to free 1822 * it. The buffer is expected to be used again soon. 1823 * 1824 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1825 * biodone() to requeue an async I/O on completion. It is also used when 1826 * known good buffers need to be requeued but we think we may need the data 1827 * again soon. 1828 * 1829 * XXX we should be able to leave the B_RELBUF hint set on completion. 1830 */ 1831 void 1832 bqrelse(struct buf *bp) 1833 { 1834 int qindex; 1835 1836 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1837 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1838 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1839 1840 if (BUF_LOCKRECURSED(bp)) { 1841 /* do not release to free list */ 1842 BUF_UNLOCK(bp); 1843 return; 1844 } 1845 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1846 1847 if (bp->b_flags & B_MANAGED) { 1848 if (bp->b_flags & B_REMFREE) 1849 bremfreef(bp); 1850 goto out; 1851 } 1852 1853 /* buffers with stale but valid contents */ 1854 if (bp->b_flags & B_DELWRI) { 1855 qindex = QUEUE_DIRTY; 1856 } else { 1857 if ((bp->b_flags & B_DELWRI) == 0 && 1858 (bp->b_xflags & BX_VNDIRTY)) 1859 panic("bqrelse: not dirty"); 1860 /* 1861 * BKGRDINPROG can only be set with the buf and bufobj 1862 * locks both held. We tolerate a race to clear it here. 1863 */ 1864 if (buf_vm_page_count_severe() && 1865 (bp->b_vflags & BV_BKGRDINPROG) == 0) { 1866 /* 1867 * We are too low on memory, we have to try to free 1868 * the buffer (most importantly: the wired pages 1869 * making up its backing store) *now*. 1870 */ 1871 brelse(bp); 1872 return; 1873 } 1874 qindex = QUEUE_CLEAN; 1875 } 1876 binsfree(bp, qindex); 1877 1878 out: 1879 /* unlock */ 1880 BUF_UNLOCK(bp); 1881 } 1882 1883 /* Give pages used by the bp back to the VM system (where possible) */ 1884 static void 1885 vfs_vmio_release(struct buf *bp) 1886 { 1887 vm_object_t obj; 1888 vm_page_t m; 1889 int i; 1890 1891 if ((bp->b_flags & B_UNMAPPED) == 0) { 1892 BUF_CHECK_MAPPED(bp); 1893 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1894 } else 1895 BUF_CHECK_UNMAPPED(bp); 1896 obj = bp->b_bufobj->bo_object; 1897 if (obj != NULL) 1898 VM_OBJECT_WLOCK(obj); 1899 for (i = 0; i < bp->b_npages; i++) { 1900 m = bp->b_pages[i]; 1901 bp->b_pages[i] = NULL; 1902 /* 1903 * In order to keep page LRU ordering consistent, put 1904 * everything on the inactive queue. 1905 */ 1906 vm_page_lock(m); 1907 vm_page_unwire(m, PQ_INACTIVE); 1908 1909 /* 1910 * Might as well free the page if we can and it has 1911 * no valid data. We also free the page if the 1912 * buffer was used for direct I/O 1913 */ 1914 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1915 if (m->wire_count == 0 && !vm_page_busied(m)) 1916 vm_page_free(m); 1917 } else if (bp->b_flags & B_DIRECT) 1918 vm_page_try_to_free(m); 1919 else if (buf_vm_page_count_severe()) 1920 vm_page_try_to_cache(m); 1921 vm_page_unlock(m); 1922 } 1923 if (obj != NULL) 1924 VM_OBJECT_WUNLOCK(obj); 1925 1926 if (bp->b_bufsize) { 1927 bufspacewakeup(); 1928 bp->b_bufsize = 0; 1929 } 1930 bp->b_npages = 0; 1931 bp->b_flags &= ~B_VMIO; 1932 if (bp->b_vp) 1933 brelvp(bp); 1934 } 1935 1936 /* 1937 * Check to see if a block at a particular lbn is available for a clustered 1938 * write. 1939 */ 1940 static int 1941 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1942 { 1943 struct buf *bpa; 1944 int match; 1945 1946 match = 0; 1947 1948 /* If the buf isn't in core skip it */ 1949 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1950 return (0); 1951 1952 /* If the buf is busy we don't want to wait for it */ 1953 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1954 return (0); 1955 1956 /* Only cluster with valid clusterable delayed write buffers */ 1957 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1958 (B_DELWRI | B_CLUSTEROK)) 1959 goto done; 1960 1961 if (bpa->b_bufsize != size) 1962 goto done; 1963 1964 /* 1965 * Check to see if it is in the expected place on disk and that the 1966 * block has been mapped. 1967 */ 1968 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1969 match = 1; 1970 done: 1971 BUF_UNLOCK(bpa); 1972 return (match); 1973 } 1974 1975 /* 1976 * vfs_bio_awrite: 1977 * 1978 * Implement clustered async writes for clearing out B_DELWRI buffers. 1979 * This is much better then the old way of writing only one buffer at 1980 * a time. Note that we may not be presented with the buffers in the 1981 * correct order, so we search for the cluster in both directions. 1982 */ 1983 int 1984 vfs_bio_awrite(struct buf *bp) 1985 { 1986 struct bufobj *bo; 1987 int i; 1988 int j; 1989 daddr_t lblkno = bp->b_lblkno; 1990 struct vnode *vp = bp->b_vp; 1991 int ncl; 1992 int nwritten; 1993 int size; 1994 int maxcl; 1995 int gbflags; 1996 1997 bo = &vp->v_bufobj; 1998 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1999 /* 2000 * right now we support clustered writing only to regular files. If 2001 * we find a clusterable block we could be in the middle of a cluster 2002 * rather then at the beginning. 2003 */ 2004 if ((vp->v_type == VREG) && 2005 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 2006 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 2007 2008 size = vp->v_mount->mnt_stat.f_iosize; 2009 maxcl = MAXPHYS / size; 2010 2011 BO_RLOCK(bo); 2012 for (i = 1; i < maxcl; i++) 2013 if (vfs_bio_clcheck(vp, size, lblkno + i, 2014 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 2015 break; 2016 2017 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 2018 if (vfs_bio_clcheck(vp, size, lblkno - j, 2019 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 2020 break; 2021 BO_RUNLOCK(bo); 2022 --j; 2023 ncl = i + j; 2024 /* 2025 * this is a possible cluster write 2026 */ 2027 if (ncl != 1) { 2028 BUF_UNLOCK(bp); 2029 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 2030 gbflags); 2031 return (nwritten); 2032 } 2033 } 2034 bremfree(bp); 2035 bp->b_flags |= B_ASYNC; 2036 /* 2037 * default (old) behavior, writing out only one block 2038 * 2039 * XXX returns b_bufsize instead of b_bcount for nwritten? 2040 */ 2041 nwritten = bp->b_bufsize; 2042 (void) bwrite(bp); 2043 2044 return (nwritten); 2045 } 2046 2047 static void 2048 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 2049 { 2050 2051 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2052 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 2053 if ((gbflags & GB_UNMAPPED) == 0) { 2054 bp->b_kvabase = (caddr_t)addr; 2055 } else if ((gbflags & GB_KVAALLOC) != 0) { 2056 KASSERT((gbflags & GB_UNMAPPED) != 0, 2057 ("GB_KVAALLOC without GB_UNMAPPED")); 2058 bp->b_kvaalloc = (caddr_t)addr; 2059 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2060 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2061 } 2062 bp->b_kvasize = maxsize; 2063 } 2064 2065 /* 2066 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 2067 * needed. 2068 */ 2069 static int 2070 allocbufkva(struct buf *bp, int maxsize, int gbflags) 2071 { 2072 vm_offset_t addr; 2073 2074 bfreekva(bp); 2075 addr = 0; 2076 2077 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) { 2078 /* 2079 * Buffer map is too fragmented. Request the caller 2080 * to defragment the map. 2081 */ 2082 atomic_add_int(&bufdefragcnt, 1); 2083 return (1); 2084 } 2085 setbufkva(bp, addr, maxsize, gbflags); 2086 atomic_add_long(&bufspace, bp->b_kvasize); 2087 return (0); 2088 } 2089 2090 /* 2091 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2092 * locked vnode is supplied. 2093 */ 2094 static void 2095 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2096 int defrag) 2097 { 2098 struct thread *td; 2099 char *waitmsg; 2100 int error, fl, flags, norunbuf; 2101 2102 mtx_assert(&bqclean, MA_OWNED); 2103 2104 if (defrag) { 2105 flags = VFS_BIO_NEED_BUFSPACE; 2106 waitmsg = "nbufkv"; 2107 } else if (bufspace >= hibufspace) { 2108 waitmsg = "nbufbs"; 2109 flags = VFS_BIO_NEED_BUFSPACE; 2110 } else { 2111 waitmsg = "newbuf"; 2112 flags = VFS_BIO_NEED_ANY; 2113 } 2114 atomic_set_int(&needsbuffer, flags); 2115 mtx_unlock(&bqclean); 2116 2117 bd_speedup(); /* heeeelp */ 2118 if ((gbflags & GB_NOWAIT_BD) != 0) 2119 return; 2120 2121 td = curthread; 2122 rw_wlock(&nblock); 2123 while ((needsbuffer & flags) != 0) { 2124 if (vp != NULL && vp->v_type != VCHR && 2125 (td->td_pflags & TDP_BUFNEED) == 0) { 2126 rw_wunlock(&nblock); 2127 /* 2128 * getblk() is called with a vnode locked, and 2129 * some majority of the dirty buffers may as 2130 * well belong to the vnode. Flushing the 2131 * buffers there would make a progress that 2132 * cannot be achieved by the buf_daemon, that 2133 * cannot lock the vnode. 2134 */ 2135 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2136 (td->td_pflags & TDP_NORUNNINGBUF); 2137 2138 /* 2139 * Play bufdaemon. The getnewbuf() function 2140 * may be called while the thread owns lock 2141 * for another dirty buffer for the same 2142 * vnode, which makes it impossible to use 2143 * VOP_FSYNC() there, due to the buffer lock 2144 * recursion. 2145 */ 2146 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2147 fl = buf_flush(vp, flushbufqtarget); 2148 td->td_pflags &= norunbuf; 2149 rw_wlock(&nblock); 2150 if (fl != 0) 2151 continue; 2152 if ((needsbuffer & flags) == 0) 2153 break; 2154 } 2155 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 2156 (PRIBIO + 4) | slpflag, waitmsg, slptimeo); 2157 if (error != 0) 2158 break; 2159 } 2160 rw_wunlock(&nblock); 2161 } 2162 2163 static void 2164 getnewbuf_reuse_bp(struct buf *bp, int qindex) 2165 { 2166 2167 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2168 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2169 bp->b_kvasize, bp->b_bufsize, qindex); 2170 mtx_assert(&bqclean, MA_NOTOWNED); 2171 2172 /* 2173 * Note: we no longer distinguish between VMIO and non-VMIO 2174 * buffers. 2175 */ 2176 KASSERT((bp->b_flags & B_DELWRI) == 0, 2177 ("delwri buffer %p found in queue %d", bp, qindex)); 2178 2179 if (qindex == QUEUE_CLEAN) { 2180 if (bp->b_flags & B_VMIO) { 2181 bp->b_flags &= ~B_ASYNC; 2182 vfs_vmio_release(bp); 2183 } 2184 if (bp->b_vp != NULL) 2185 brelvp(bp); 2186 } 2187 2188 /* 2189 * Get the rest of the buffer freed up. b_kva* is still valid 2190 * after this operation. 2191 */ 2192 2193 if (bp->b_rcred != NOCRED) { 2194 crfree(bp->b_rcred); 2195 bp->b_rcred = NOCRED; 2196 } 2197 if (bp->b_wcred != NOCRED) { 2198 crfree(bp->b_wcred); 2199 bp->b_wcred = NOCRED; 2200 } 2201 if (!LIST_EMPTY(&bp->b_dep)) 2202 buf_deallocate(bp); 2203 if (bp->b_vflags & BV_BKGRDINPROG) 2204 panic("losing buffer 3"); 2205 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2206 bp, bp->b_vp, qindex)); 2207 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2208 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2209 2210 if (bp->b_bufsize) 2211 allocbuf(bp, 0); 2212 2213 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2214 bp->b_ioflags = 0; 2215 bp->b_xflags = 0; 2216 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2217 ("buf %p still counted as free?", bp)); 2218 bp->b_vflags = 0; 2219 bp->b_vp = NULL; 2220 bp->b_blkno = bp->b_lblkno = 0; 2221 bp->b_offset = NOOFFSET; 2222 bp->b_iodone = 0; 2223 bp->b_error = 0; 2224 bp->b_resid = 0; 2225 bp->b_bcount = 0; 2226 bp->b_npages = 0; 2227 bp->b_dirtyoff = bp->b_dirtyend = 0; 2228 bp->b_bufobj = NULL; 2229 bp->b_pin_count = 0; 2230 bp->b_fsprivate1 = NULL; 2231 bp->b_fsprivate2 = NULL; 2232 bp->b_fsprivate3 = NULL; 2233 2234 LIST_INIT(&bp->b_dep); 2235 } 2236 2237 static int flushingbufs; 2238 2239 static struct buf * 2240 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2241 { 2242 struct buf *bp, *nbp; 2243 int nqindex, qindex, pass; 2244 2245 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2246 2247 pass = 1; 2248 restart: 2249 atomic_add_int(&getnewbufrestarts, 1); 2250 2251 /* 2252 * Setup for scan. If we do not have enough free buffers, 2253 * we setup a degenerate case that immediately fails. Note 2254 * that if we are specially marked process, we are allowed to 2255 * dip into our reserves. 2256 * 2257 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2258 * for the allocation of the mapped buffer. For unmapped, the 2259 * easiest is to start with EMPTY outright. 2260 * 2261 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2262 * However, there are a number of cases (defragging, reusing, ...) 2263 * where we cannot backup. 2264 */ 2265 nbp = NULL; 2266 mtx_lock(&bqclean); 2267 if (!defrag && unmapped) { 2268 nqindex = QUEUE_EMPTY; 2269 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2270 } 2271 if (nbp == NULL) { 2272 nqindex = QUEUE_EMPTYKVA; 2273 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2274 } 2275 2276 /* 2277 * If no EMPTYKVA buffers and we are either defragging or 2278 * reusing, locate a CLEAN buffer to free or reuse. If 2279 * bufspace useage is low skip this step so we can allocate a 2280 * new buffer. 2281 */ 2282 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2283 nqindex = QUEUE_CLEAN; 2284 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2285 } 2286 2287 /* 2288 * If we could not find or were not allowed to reuse a CLEAN 2289 * buffer, check to see if it is ok to use an EMPTY buffer. 2290 * We can only use an EMPTY buffer if allocating its KVA would 2291 * not otherwise run us out of buffer space. No KVA is needed 2292 * for the unmapped allocation. 2293 */ 2294 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2295 metadata)) { 2296 nqindex = QUEUE_EMPTY; 2297 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2298 } 2299 2300 /* 2301 * All available buffers might be clean, retry ignoring the 2302 * lobufspace as the last resort. 2303 */ 2304 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2305 nqindex = QUEUE_CLEAN; 2306 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2307 } 2308 2309 /* 2310 * Run scan, possibly freeing data and/or kva mappings on the fly 2311 * depending. 2312 */ 2313 while ((bp = nbp) != NULL) { 2314 qindex = nqindex; 2315 2316 /* 2317 * Calculate next bp (we can only use it if we do not 2318 * block or do other fancy things). 2319 */ 2320 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2321 switch (qindex) { 2322 case QUEUE_EMPTY: 2323 nqindex = QUEUE_EMPTYKVA; 2324 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2325 if (nbp != NULL) 2326 break; 2327 /* FALLTHROUGH */ 2328 case QUEUE_EMPTYKVA: 2329 nqindex = QUEUE_CLEAN; 2330 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2331 if (nbp != NULL) 2332 break; 2333 /* FALLTHROUGH */ 2334 case QUEUE_CLEAN: 2335 if (metadata && pass == 1) { 2336 pass = 2; 2337 nqindex = QUEUE_EMPTY; 2338 nbp = TAILQ_FIRST( 2339 &bufqueues[QUEUE_EMPTY]); 2340 } 2341 /* 2342 * nbp is NULL. 2343 */ 2344 break; 2345 } 2346 } 2347 /* 2348 * If we are defragging then we need a buffer with 2349 * b_kvasize != 0. XXX this situation should no longer 2350 * occur, if defrag is non-zero the buffer's b_kvasize 2351 * should also be non-zero at this point. XXX 2352 */ 2353 if (defrag && bp->b_kvasize == 0) { 2354 printf("Warning: defrag empty buffer %p\n", bp); 2355 continue; 2356 } 2357 2358 /* 2359 * Start freeing the bp. This is somewhat involved. nbp 2360 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2361 */ 2362 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2363 continue; 2364 /* 2365 * BKGRDINPROG can only be set with the buf and bufobj 2366 * locks both held. We tolerate a race to clear it here. 2367 */ 2368 if (bp->b_vflags & BV_BKGRDINPROG) { 2369 BUF_UNLOCK(bp); 2370 continue; 2371 } 2372 2373 KASSERT(bp->b_qindex == qindex, 2374 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2375 2376 bremfreel(bp); 2377 mtx_unlock(&bqclean); 2378 /* 2379 * NOTE: nbp is now entirely invalid. We can only restart 2380 * the scan from this point on. 2381 */ 2382 2383 getnewbuf_reuse_bp(bp, qindex); 2384 mtx_assert(&bqclean, MA_NOTOWNED); 2385 2386 /* 2387 * If we are defragging then free the buffer. 2388 */ 2389 if (defrag) { 2390 bp->b_flags |= B_INVAL; 2391 bfreekva(bp); 2392 brelse(bp); 2393 defrag = 0; 2394 goto restart; 2395 } 2396 2397 /* 2398 * Notify any waiters for the buffer lock about 2399 * identity change by freeing the buffer. 2400 */ 2401 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2402 bp->b_flags |= B_INVAL; 2403 bfreekva(bp); 2404 brelse(bp); 2405 goto restart; 2406 } 2407 2408 if (metadata) 2409 break; 2410 2411 /* 2412 * If we are overcomitted then recover the buffer and its 2413 * KVM space. This occurs in rare situations when multiple 2414 * processes are blocked in getnewbuf() or allocbuf(). 2415 */ 2416 if (bufspace >= hibufspace) 2417 flushingbufs = 1; 2418 if (flushingbufs && bp->b_kvasize != 0) { 2419 bp->b_flags |= B_INVAL; 2420 bfreekva(bp); 2421 brelse(bp); 2422 goto restart; 2423 } 2424 if (bufspace < lobufspace) 2425 flushingbufs = 0; 2426 break; 2427 } 2428 return (bp); 2429 } 2430 2431 /* 2432 * getnewbuf: 2433 * 2434 * Find and initialize a new buffer header, freeing up existing buffers 2435 * in the bufqueues as necessary. The new buffer is returned locked. 2436 * 2437 * Important: B_INVAL is not set. If the caller wishes to throw the 2438 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2439 * 2440 * We block if: 2441 * We have insufficient buffer headers 2442 * We have insufficient buffer space 2443 * buffer_arena is too fragmented ( space reservation fails ) 2444 * If we have to flush dirty buffers ( but we try to avoid this ) 2445 */ 2446 static struct buf * 2447 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2448 int gbflags) 2449 { 2450 struct buf *bp; 2451 int defrag, metadata; 2452 2453 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2454 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2455 if (!unmapped_buf_allowed) 2456 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2457 2458 defrag = 0; 2459 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2460 vp->v_type == VCHR) 2461 metadata = 1; 2462 else 2463 metadata = 0; 2464 /* 2465 * We can't afford to block since we might be holding a vnode lock, 2466 * which may prevent system daemons from running. We deal with 2467 * low-memory situations by proactively returning memory and running 2468 * async I/O rather then sync I/O. 2469 */ 2470 atomic_add_int(&getnewbufcalls, 1); 2471 atomic_subtract_int(&getnewbufrestarts, 1); 2472 restart: 2473 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2474 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2475 if (bp != NULL) 2476 defrag = 0; 2477 2478 /* 2479 * If we exhausted our list, sleep as appropriate. We may have to 2480 * wakeup various daemons and write out some dirty buffers. 2481 * 2482 * Generally we are sleeping due to insufficient buffer space. 2483 */ 2484 if (bp == NULL) { 2485 mtx_assert(&bqclean, MA_OWNED); 2486 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2487 mtx_assert(&bqclean, MA_NOTOWNED); 2488 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2489 mtx_assert(&bqclean, MA_NOTOWNED); 2490 2491 bfreekva(bp); 2492 bp->b_flags |= B_UNMAPPED; 2493 bp->b_kvabase = bp->b_data = unmapped_buf; 2494 bp->b_kvasize = maxsize; 2495 atomic_add_long(&bufspace, bp->b_kvasize); 2496 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2497 atomic_add_int(&bufreusecnt, 1); 2498 } else { 2499 mtx_assert(&bqclean, MA_NOTOWNED); 2500 2501 /* 2502 * We finally have a valid bp. We aren't quite out of the 2503 * woods, we still have to reserve kva space. In order 2504 * to keep fragmentation sane we only allocate kva in 2505 * BKVASIZE chunks. 2506 */ 2507 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2508 2509 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2510 B_KVAALLOC)) == B_UNMAPPED) { 2511 if (allocbufkva(bp, maxsize, gbflags)) { 2512 defrag = 1; 2513 bp->b_flags |= B_INVAL; 2514 brelse(bp); 2515 goto restart; 2516 } 2517 atomic_add_int(&bufreusecnt, 1); 2518 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2519 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2520 /* 2521 * If the reused buffer has KVA allocated, 2522 * reassign b_kvaalloc to b_kvabase. 2523 */ 2524 bp->b_kvabase = bp->b_kvaalloc; 2525 bp->b_flags &= ~B_KVAALLOC; 2526 atomic_subtract_long(&unmapped_bufspace, 2527 bp->b_kvasize); 2528 atomic_add_int(&bufreusecnt, 1); 2529 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2530 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2531 GB_KVAALLOC)) { 2532 /* 2533 * The case of reused buffer already have KVA 2534 * mapped, but the request is for unmapped 2535 * buffer with KVA allocated. 2536 */ 2537 bp->b_kvaalloc = bp->b_kvabase; 2538 bp->b_data = bp->b_kvabase = unmapped_buf; 2539 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2540 atomic_add_long(&unmapped_bufspace, 2541 bp->b_kvasize); 2542 atomic_add_int(&bufreusecnt, 1); 2543 } 2544 if ((gbflags & GB_UNMAPPED) == 0) { 2545 bp->b_saveaddr = bp->b_kvabase; 2546 bp->b_data = bp->b_saveaddr; 2547 bp->b_flags &= ~B_UNMAPPED; 2548 BUF_CHECK_MAPPED(bp); 2549 } 2550 } 2551 return (bp); 2552 } 2553 2554 /* 2555 * buf_daemon: 2556 * 2557 * buffer flushing daemon. Buffers are normally flushed by the 2558 * update daemon but if it cannot keep up this process starts to 2559 * take the load in an attempt to prevent getnewbuf() from blocking. 2560 */ 2561 2562 static struct kproc_desc buf_kp = { 2563 "bufdaemon", 2564 buf_daemon, 2565 &bufdaemonproc 2566 }; 2567 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2568 2569 static int 2570 buf_flush(struct vnode *vp, int target) 2571 { 2572 int flushed; 2573 2574 flushed = flushbufqueues(vp, target, 0); 2575 if (flushed == 0) { 2576 /* 2577 * Could not find any buffers without rollback 2578 * dependencies, so just write the first one 2579 * in the hopes of eventually making progress. 2580 */ 2581 if (vp != NULL && target > 2) 2582 target /= 2; 2583 flushbufqueues(vp, target, 1); 2584 } 2585 return (flushed); 2586 } 2587 2588 static void 2589 buf_daemon() 2590 { 2591 int lodirty; 2592 2593 /* 2594 * This process needs to be suspended prior to shutdown sync. 2595 */ 2596 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2597 SHUTDOWN_PRI_LAST); 2598 2599 /* 2600 * This process is allowed to take the buffer cache to the limit 2601 */ 2602 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2603 mtx_lock(&bdlock); 2604 for (;;) { 2605 bd_request = 0; 2606 mtx_unlock(&bdlock); 2607 2608 kproc_suspend_check(bufdaemonproc); 2609 lodirty = lodirtybuffers; 2610 if (bd_speedupreq) { 2611 lodirty = numdirtybuffers / 2; 2612 bd_speedupreq = 0; 2613 } 2614 /* 2615 * Do the flush. Limit the amount of in-transit I/O we 2616 * allow to build up, otherwise we would completely saturate 2617 * the I/O system. 2618 */ 2619 while (numdirtybuffers > lodirty) { 2620 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0) 2621 break; 2622 kern_yield(PRI_USER); 2623 } 2624 2625 /* 2626 * Only clear bd_request if we have reached our low water 2627 * mark. The buf_daemon normally waits 1 second and 2628 * then incrementally flushes any dirty buffers that have 2629 * built up, within reason. 2630 * 2631 * If we were unable to hit our low water mark and couldn't 2632 * find any flushable buffers, we sleep for a short period 2633 * to avoid endless loops on unlockable buffers. 2634 */ 2635 mtx_lock(&bdlock); 2636 if (numdirtybuffers <= lodirtybuffers) { 2637 /* 2638 * We reached our low water mark, reset the 2639 * request and sleep until we are needed again. 2640 * The sleep is just so the suspend code works. 2641 */ 2642 bd_request = 0; 2643 /* 2644 * Do an extra wakeup in case dirty threshold 2645 * changed via sysctl and the explicit transition 2646 * out of shortfall was missed. 2647 */ 2648 bdirtywakeup(); 2649 if (runningbufspace <= lorunningspace) 2650 runningwakeup(); 2651 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2652 } else { 2653 /* 2654 * We couldn't find any flushable dirty buffers but 2655 * still have too many dirty buffers, we 2656 * have to sleep and try again. (rare) 2657 */ 2658 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2659 } 2660 } 2661 } 2662 2663 /* 2664 * flushbufqueues: 2665 * 2666 * Try to flush a buffer in the dirty queue. We must be careful to 2667 * free up B_INVAL buffers instead of write them, which NFS is 2668 * particularly sensitive to. 2669 */ 2670 static int flushwithdeps = 0; 2671 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2672 0, "Number of buffers flushed with dependecies that require rollbacks"); 2673 2674 static int 2675 flushbufqueues(struct vnode *lvp, int target, int flushdeps) 2676 { 2677 struct buf *sentinel; 2678 struct vnode *vp; 2679 struct mount *mp; 2680 struct buf *bp; 2681 int hasdeps; 2682 int flushed; 2683 int queue; 2684 int error; 2685 bool unlock; 2686 2687 flushed = 0; 2688 queue = QUEUE_DIRTY; 2689 bp = NULL; 2690 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2691 sentinel->b_qindex = QUEUE_SENTINEL; 2692 mtx_lock(&bqdirty); 2693 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2694 mtx_unlock(&bqdirty); 2695 while (flushed != target) { 2696 maybe_yield(); 2697 mtx_lock(&bqdirty); 2698 bp = TAILQ_NEXT(sentinel, b_freelist); 2699 if (bp != NULL) { 2700 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2701 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2702 b_freelist); 2703 } else { 2704 mtx_unlock(&bqdirty); 2705 break; 2706 } 2707 /* 2708 * Skip sentinels inserted by other invocations of the 2709 * flushbufqueues(), taking care to not reorder them. 2710 * 2711 * Only flush the buffers that belong to the 2712 * vnode locked by the curthread. 2713 */ 2714 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 2715 bp->b_vp != lvp)) { 2716 mtx_unlock(&bqdirty); 2717 continue; 2718 } 2719 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 2720 mtx_unlock(&bqdirty); 2721 if (error != 0) 2722 continue; 2723 if (bp->b_pin_count > 0) { 2724 BUF_UNLOCK(bp); 2725 continue; 2726 } 2727 /* 2728 * BKGRDINPROG can only be set with the buf and bufobj 2729 * locks both held. We tolerate a race to clear it here. 2730 */ 2731 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2732 (bp->b_flags & B_DELWRI) == 0) { 2733 BUF_UNLOCK(bp); 2734 continue; 2735 } 2736 if (bp->b_flags & B_INVAL) { 2737 bremfreef(bp); 2738 brelse(bp); 2739 flushed++; 2740 continue; 2741 } 2742 2743 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2744 if (flushdeps == 0) { 2745 BUF_UNLOCK(bp); 2746 continue; 2747 } 2748 hasdeps = 1; 2749 } else 2750 hasdeps = 0; 2751 /* 2752 * We must hold the lock on a vnode before writing 2753 * one of its buffers. Otherwise we may confuse, or 2754 * in the case of a snapshot vnode, deadlock the 2755 * system. 2756 * 2757 * The lock order here is the reverse of the normal 2758 * of vnode followed by buf lock. This is ok because 2759 * the NOWAIT will prevent deadlock. 2760 */ 2761 vp = bp->b_vp; 2762 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2763 BUF_UNLOCK(bp); 2764 continue; 2765 } 2766 if (lvp == NULL) { 2767 unlock = true; 2768 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 2769 } else { 2770 ASSERT_VOP_LOCKED(vp, "getbuf"); 2771 unlock = false; 2772 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 2773 vn_lock(vp, LK_TRYUPGRADE); 2774 } 2775 if (error == 0) { 2776 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2777 bp, bp->b_vp, bp->b_flags); 2778 if (curproc == bufdaemonproc) { 2779 vfs_bio_awrite(bp); 2780 } else { 2781 bremfree(bp); 2782 bwrite(bp); 2783 notbufdflushes++; 2784 } 2785 vn_finished_write(mp); 2786 if (unlock) 2787 VOP_UNLOCK(vp, 0); 2788 flushwithdeps += hasdeps; 2789 flushed++; 2790 2791 /* 2792 * Sleeping on runningbufspace while holding 2793 * vnode lock leads to deadlock. 2794 */ 2795 if (curproc == bufdaemonproc && 2796 runningbufspace > hirunningspace) 2797 waitrunningbufspace(); 2798 continue; 2799 } 2800 vn_finished_write(mp); 2801 BUF_UNLOCK(bp); 2802 } 2803 mtx_lock(&bqdirty); 2804 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2805 mtx_unlock(&bqdirty); 2806 free(sentinel, M_TEMP); 2807 return (flushed); 2808 } 2809 2810 /* 2811 * Check to see if a block is currently memory resident. 2812 */ 2813 struct buf * 2814 incore(struct bufobj *bo, daddr_t blkno) 2815 { 2816 struct buf *bp; 2817 2818 BO_RLOCK(bo); 2819 bp = gbincore(bo, blkno); 2820 BO_RUNLOCK(bo); 2821 return (bp); 2822 } 2823 2824 /* 2825 * Returns true if no I/O is needed to access the 2826 * associated VM object. This is like incore except 2827 * it also hunts around in the VM system for the data. 2828 */ 2829 2830 static int 2831 inmem(struct vnode * vp, daddr_t blkno) 2832 { 2833 vm_object_t obj; 2834 vm_offset_t toff, tinc, size; 2835 vm_page_t m; 2836 vm_ooffset_t off; 2837 2838 ASSERT_VOP_LOCKED(vp, "inmem"); 2839 2840 if (incore(&vp->v_bufobj, blkno)) 2841 return 1; 2842 if (vp->v_mount == NULL) 2843 return 0; 2844 obj = vp->v_object; 2845 if (obj == NULL) 2846 return (0); 2847 2848 size = PAGE_SIZE; 2849 if (size > vp->v_mount->mnt_stat.f_iosize) 2850 size = vp->v_mount->mnt_stat.f_iosize; 2851 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2852 2853 VM_OBJECT_RLOCK(obj); 2854 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2855 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2856 if (!m) 2857 goto notinmem; 2858 tinc = size; 2859 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2860 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2861 if (vm_page_is_valid(m, 2862 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2863 goto notinmem; 2864 } 2865 VM_OBJECT_RUNLOCK(obj); 2866 return 1; 2867 2868 notinmem: 2869 VM_OBJECT_RUNLOCK(obj); 2870 return (0); 2871 } 2872 2873 /* 2874 * Set the dirty range for a buffer based on the status of the dirty 2875 * bits in the pages comprising the buffer. The range is limited 2876 * to the size of the buffer. 2877 * 2878 * Tell the VM system that the pages associated with this buffer 2879 * are clean. This is used for delayed writes where the data is 2880 * going to go to disk eventually without additional VM intevention. 2881 * 2882 * Note that while we only really need to clean through to b_bcount, we 2883 * just go ahead and clean through to b_bufsize. 2884 */ 2885 static void 2886 vfs_clean_pages_dirty_buf(struct buf *bp) 2887 { 2888 vm_ooffset_t foff, noff, eoff; 2889 vm_page_t m; 2890 int i; 2891 2892 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2893 return; 2894 2895 foff = bp->b_offset; 2896 KASSERT(bp->b_offset != NOOFFSET, 2897 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2898 2899 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2900 vfs_drain_busy_pages(bp); 2901 vfs_setdirty_locked_object(bp); 2902 for (i = 0; i < bp->b_npages; i++) { 2903 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2904 eoff = noff; 2905 if (eoff > bp->b_offset + bp->b_bufsize) 2906 eoff = bp->b_offset + bp->b_bufsize; 2907 m = bp->b_pages[i]; 2908 vfs_page_set_validclean(bp, foff, m); 2909 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2910 foff = noff; 2911 } 2912 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2913 } 2914 2915 static void 2916 vfs_setdirty_locked_object(struct buf *bp) 2917 { 2918 vm_object_t object; 2919 int i; 2920 2921 object = bp->b_bufobj->bo_object; 2922 VM_OBJECT_ASSERT_WLOCKED(object); 2923 2924 /* 2925 * We qualify the scan for modified pages on whether the 2926 * object has been flushed yet. 2927 */ 2928 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2929 vm_offset_t boffset; 2930 vm_offset_t eoffset; 2931 2932 /* 2933 * test the pages to see if they have been modified directly 2934 * by users through the VM system. 2935 */ 2936 for (i = 0; i < bp->b_npages; i++) 2937 vm_page_test_dirty(bp->b_pages[i]); 2938 2939 /* 2940 * Calculate the encompassing dirty range, boffset and eoffset, 2941 * (eoffset - boffset) bytes. 2942 */ 2943 2944 for (i = 0; i < bp->b_npages; i++) { 2945 if (bp->b_pages[i]->dirty) 2946 break; 2947 } 2948 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2949 2950 for (i = bp->b_npages - 1; i >= 0; --i) { 2951 if (bp->b_pages[i]->dirty) { 2952 break; 2953 } 2954 } 2955 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2956 2957 /* 2958 * Fit it to the buffer. 2959 */ 2960 2961 if (eoffset > bp->b_bcount) 2962 eoffset = bp->b_bcount; 2963 2964 /* 2965 * If we have a good dirty range, merge with the existing 2966 * dirty range. 2967 */ 2968 2969 if (boffset < eoffset) { 2970 if (bp->b_dirtyoff > boffset) 2971 bp->b_dirtyoff = boffset; 2972 if (bp->b_dirtyend < eoffset) 2973 bp->b_dirtyend = eoffset; 2974 } 2975 } 2976 } 2977 2978 /* 2979 * Allocate the KVA mapping for an existing buffer. It handles the 2980 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2981 * KVA which is not mapped (B_KVAALLOC). 2982 */ 2983 static void 2984 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2985 { 2986 struct buf *scratch_bp; 2987 int bsize, maxsize, need_mapping, need_kva; 2988 off_t offset; 2989 2990 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2991 (gbflags & GB_UNMAPPED) == 0; 2992 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2993 (gbflags & GB_KVAALLOC) != 0; 2994 if (!need_mapping && !need_kva) 2995 return; 2996 2997 BUF_CHECK_UNMAPPED(bp); 2998 2999 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 3000 /* 3001 * Buffer is not mapped, but the KVA was already 3002 * reserved at the time of the instantiation. Use the 3003 * allocated space. 3004 */ 3005 bp->b_flags &= ~B_KVAALLOC; 3006 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 3007 bp->b_kvabase = bp->b_kvaalloc; 3008 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 3009 goto has_addr; 3010 } 3011 3012 /* 3013 * Calculate the amount of the address space we would reserve 3014 * if the buffer was mapped. 3015 */ 3016 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3017 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3018 offset = blkno * bsize; 3019 maxsize = size + (offset & PAGE_MASK); 3020 maxsize = imax(maxsize, bsize); 3021 3022 mapping_loop: 3023 if (allocbufkva(bp, maxsize, gbflags)) { 3024 /* 3025 * Request defragmentation. getnewbuf() returns us the 3026 * allocated space by the scratch buffer KVA. 3027 */ 3028 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 3029 (GB_UNMAPPED | GB_KVAALLOC)); 3030 if (scratch_bp == NULL) { 3031 if ((gbflags & GB_NOWAIT_BD) != 0) { 3032 /* 3033 * XXXKIB: defragmentation cannot 3034 * succeed, not sure what else to do. 3035 */ 3036 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 3037 } 3038 atomic_add_int(&mappingrestarts, 1); 3039 goto mapping_loop; 3040 } 3041 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 3042 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 3043 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 3044 scratch_bp->b_kvasize, gbflags); 3045 3046 /* Get rid of the scratch buffer. */ 3047 scratch_bp->b_kvasize = 0; 3048 scratch_bp->b_flags |= B_INVAL; 3049 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 3050 brelse(scratch_bp); 3051 } 3052 if (!need_mapping) 3053 return; 3054 3055 has_addr: 3056 bp->b_saveaddr = bp->b_kvabase; 3057 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 3058 bp->b_flags &= ~B_UNMAPPED; 3059 BUF_CHECK_MAPPED(bp); 3060 bpmap_qenter(bp); 3061 } 3062 3063 /* 3064 * getblk: 3065 * 3066 * Get a block given a specified block and offset into a file/device. 3067 * The buffers B_DONE bit will be cleared on return, making it almost 3068 * ready for an I/O initiation. B_INVAL may or may not be set on 3069 * return. The caller should clear B_INVAL prior to initiating a 3070 * READ. 3071 * 3072 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3073 * an existing buffer. 3074 * 3075 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3076 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3077 * and then cleared based on the backing VM. If the previous buffer is 3078 * non-0-sized but invalid, B_CACHE will be cleared. 3079 * 3080 * If getblk() must create a new buffer, the new buffer is returned with 3081 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3082 * case it is returned with B_INVAL clear and B_CACHE set based on the 3083 * backing VM. 3084 * 3085 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3086 * B_CACHE bit is clear. 3087 * 3088 * What this means, basically, is that the caller should use B_CACHE to 3089 * determine whether the buffer is fully valid or not and should clear 3090 * B_INVAL prior to issuing a read. If the caller intends to validate 3091 * the buffer by loading its data area with something, the caller needs 3092 * to clear B_INVAL. If the caller does this without issuing an I/O, 3093 * the caller should set B_CACHE ( as an optimization ), else the caller 3094 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3095 * a write attempt or if it was a successfull read. If the caller 3096 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3097 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3098 */ 3099 struct buf * 3100 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3101 int flags) 3102 { 3103 struct buf *bp; 3104 struct bufobj *bo; 3105 int bsize, error, maxsize, vmio; 3106 off_t offset; 3107 3108 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3109 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3110 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3111 ASSERT_VOP_LOCKED(vp, "getblk"); 3112 if (size > MAXBCACHEBUF) 3113 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size, 3114 MAXBCACHEBUF); 3115 if (!unmapped_buf_allowed) 3116 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3117 3118 bo = &vp->v_bufobj; 3119 loop: 3120 BO_RLOCK(bo); 3121 bp = gbincore(bo, blkno); 3122 if (bp != NULL) { 3123 int lockflags; 3124 /* 3125 * Buffer is in-core. If the buffer is not busy nor managed, 3126 * it must be on a queue. 3127 */ 3128 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3129 3130 if (flags & GB_LOCK_NOWAIT) 3131 lockflags |= LK_NOWAIT; 3132 3133 error = BUF_TIMELOCK(bp, lockflags, 3134 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3135 3136 /* 3137 * If we slept and got the lock we have to restart in case 3138 * the buffer changed identities. 3139 */ 3140 if (error == ENOLCK) 3141 goto loop; 3142 /* We timed out or were interrupted. */ 3143 else if (error) 3144 return (NULL); 3145 /* If recursed, assume caller knows the rules. */ 3146 else if (BUF_LOCKRECURSED(bp)) 3147 goto end; 3148 3149 /* 3150 * The buffer is locked. B_CACHE is cleared if the buffer is 3151 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3152 * and for a VMIO buffer B_CACHE is adjusted according to the 3153 * backing VM cache. 3154 */ 3155 if (bp->b_flags & B_INVAL) 3156 bp->b_flags &= ~B_CACHE; 3157 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3158 bp->b_flags |= B_CACHE; 3159 if (bp->b_flags & B_MANAGED) 3160 MPASS(bp->b_qindex == QUEUE_NONE); 3161 else 3162 bremfree(bp); 3163 3164 /* 3165 * check for size inconsistencies for non-VMIO case. 3166 */ 3167 if (bp->b_bcount != size) { 3168 if ((bp->b_flags & B_VMIO) == 0 || 3169 (size > bp->b_kvasize)) { 3170 if (bp->b_flags & B_DELWRI) { 3171 /* 3172 * If buffer is pinned and caller does 3173 * not want sleep waiting for it to be 3174 * unpinned, bail out 3175 * */ 3176 if (bp->b_pin_count > 0) { 3177 if (flags & GB_LOCK_NOWAIT) { 3178 bqrelse(bp); 3179 return (NULL); 3180 } else { 3181 bunpin_wait(bp); 3182 } 3183 } 3184 bp->b_flags |= B_NOCACHE; 3185 bwrite(bp); 3186 } else { 3187 if (LIST_EMPTY(&bp->b_dep)) { 3188 bp->b_flags |= B_RELBUF; 3189 brelse(bp); 3190 } else { 3191 bp->b_flags |= B_NOCACHE; 3192 bwrite(bp); 3193 } 3194 } 3195 goto loop; 3196 } 3197 } 3198 3199 /* 3200 * Handle the case of unmapped buffer which should 3201 * become mapped, or the buffer for which KVA 3202 * reservation is requested. 3203 */ 3204 bp_unmapped_get_kva(bp, blkno, size, flags); 3205 3206 /* 3207 * If the size is inconsistant in the VMIO case, we can resize 3208 * the buffer. This might lead to B_CACHE getting set or 3209 * cleared. If the size has not changed, B_CACHE remains 3210 * unchanged from its previous state. 3211 */ 3212 if (bp->b_bcount != size) 3213 allocbuf(bp, size); 3214 3215 KASSERT(bp->b_offset != NOOFFSET, 3216 ("getblk: no buffer offset")); 3217 3218 /* 3219 * A buffer with B_DELWRI set and B_CACHE clear must 3220 * be committed before we can return the buffer in 3221 * order to prevent the caller from issuing a read 3222 * ( due to B_CACHE not being set ) and overwriting 3223 * it. 3224 * 3225 * Most callers, including NFS and FFS, need this to 3226 * operate properly either because they assume they 3227 * can issue a read if B_CACHE is not set, or because 3228 * ( for example ) an uncached B_DELWRI might loop due 3229 * to softupdates re-dirtying the buffer. In the latter 3230 * case, B_CACHE is set after the first write completes, 3231 * preventing further loops. 3232 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3233 * above while extending the buffer, we cannot allow the 3234 * buffer to remain with B_CACHE set after the write 3235 * completes or it will represent a corrupt state. To 3236 * deal with this we set B_NOCACHE to scrap the buffer 3237 * after the write. 3238 * 3239 * We might be able to do something fancy, like setting 3240 * B_CACHE in bwrite() except if B_DELWRI is already set, 3241 * so the below call doesn't set B_CACHE, but that gets real 3242 * confusing. This is much easier. 3243 */ 3244 3245 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3246 bp->b_flags |= B_NOCACHE; 3247 bwrite(bp); 3248 goto loop; 3249 } 3250 bp->b_flags &= ~B_DONE; 3251 } else { 3252 /* 3253 * Buffer is not in-core, create new buffer. The buffer 3254 * returned by getnewbuf() is locked. Note that the returned 3255 * buffer is also considered valid (not marked B_INVAL). 3256 */ 3257 BO_RUNLOCK(bo); 3258 /* 3259 * If the user does not want us to create the buffer, bail out 3260 * here. 3261 */ 3262 if (flags & GB_NOCREAT) 3263 return NULL; 3264 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3265 return NULL; 3266 3267 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3268 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3269 offset = blkno * bsize; 3270 vmio = vp->v_object != NULL; 3271 if (vmio) { 3272 maxsize = size + (offset & PAGE_MASK); 3273 } else { 3274 maxsize = size; 3275 /* Do not allow non-VMIO notmapped buffers. */ 3276 flags &= ~GB_UNMAPPED; 3277 } 3278 maxsize = imax(maxsize, bsize); 3279 3280 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3281 if (bp == NULL) { 3282 if (slpflag || slptimeo) 3283 return NULL; 3284 goto loop; 3285 } 3286 3287 /* 3288 * This code is used to make sure that a buffer is not 3289 * created while the getnewbuf routine is blocked. 3290 * This can be a problem whether the vnode is locked or not. 3291 * If the buffer is created out from under us, we have to 3292 * throw away the one we just created. 3293 * 3294 * Note: this must occur before we associate the buffer 3295 * with the vp especially considering limitations in 3296 * the splay tree implementation when dealing with duplicate 3297 * lblkno's. 3298 */ 3299 BO_LOCK(bo); 3300 if (gbincore(bo, blkno)) { 3301 BO_UNLOCK(bo); 3302 bp->b_flags |= B_INVAL; 3303 brelse(bp); 3304 goto loop; 3305 } 3306 3307 /* 3308 * Insert the buffer into the hash, so that it can 3309 * be found by incore. 3310 */ 3311 bp->b_blkno = bp->b_lblkno = blkno; 3312 bp->b_offset = offset; 3313 bgetvp(vp, bp); 3314 BO_UNLOCK(bo); 3315 3316 /* 3317 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3318 * buffer size starts out as 0, B_CACHE will be set by 3319 * allocbuf() for the VMIO case prior to it testing the 3320 * backing store for validity. 3321 */ 3322 3323 if (vmio) { 3324 bp->b_flags |= B_VMIO; 3325 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3326 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3327 bp, vp->v_object, bp->b_bufobj->bo_object)); 3328 } else { 3329 bp->b_flags &= ~B_VMIO; 3330 KASSERT(bp->b_bufobj->bo_object == NULL, 3331 ("ARGH! has b_bufobj->bo_object %p %p\n", 3332 bp, bp->b_bufobj->bo_object)); 3333 BUF_CHECK_MAPPED(bp); 3334 } 3335 3336 allocbuf(bp, size); 3337 bp->b_flags &= ~B_DONE; 3338 } 3339 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3340 BUF_ASSERT_HELD(bp); 3341 end: 3342 KASSERT(bp->b_bufobj == bo, 3343 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3344 return (bp); 3345 } 3346 3347 /* 3348 * Get an empty, disassociated buffer of given size. The buffer is initially 3349 * set to B_INVAL. 3350 */ 3351 struct buf * 3352 geteblk(int size, int flags) 3353 { 3354 struct buf *bp; 3355 int maxsize; 3356 3357 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3358 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3359 if ((flags & GB_NOWAIT_BD) && 3360 (curthread->td_pflags & TDP_BUFNEED) != 0) 3361 return (NULL); 3362 } 3363 allocbuf(bp, size); 3364 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3365 BUF_ASSERT_HELD(bp); 3366 return (bp); 3367 } 3368 3369 3370 /* 3371 * This code constitutes the buffer memory from either anonymous system 3372 * memory (in the case of non-VMIO operations) or from an associated 3373 * VM object (in the case of VMIO operations). This code is able to 3374 * resize a buffer up or down. 3375 * 3376 * Note that this code is tricky, and has many complications to resolve 3377 * deadlock or inconsistant data situations. Tread lightly!!! 3378 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3379 * the caller. Calling this code willy nilly can result in the loss of data. 3380 * 3381 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3382 * B_CACHE for the non-VMIO case. 3383 */ 3384 3385 int 3386 allocbuf(struct buf *bp, int size) 3387 { 3388 int newbsize, mbsize; 3389 int i; 3390 3391 BUF_ASSERT_HELD(bp); 3392 3393 if (bp->b_kvasize < size) 3394 panic("allocbuf: buffer too small"); 3395 3396 if ((bp->b_flags & B_VMIO) == 0) { 3397 caddr_t origbuf; 3398 int origbufsize; 3399 /* 3400 * Just get anonymous memory from the kernel. Don't 3401 * mess with B_CACHE. 3402 */ 3403 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3404 if (bp->b_flags & B_MALLOC) 3405 newbsize = mbsize; 3406 else 3407 newbsize = round_page(size); 3408 3409 if (newbsize < bp->b_bufsize) { 3410 /* 3411 * malloced buffers are not shrunk 3412 */ 3413 if (bp->b_flags & B_MALLOC) { 3414 if (newbsize) { 3415 bp->b_bcount = size; 3416 } else { 3417 free(bp->b_data, M_BIOBUF); 3418 if (bp->b_bufsize) { 3419 atomic_subtract_long( 3420 &bufmallocspace, 3421 bp->b_bufsize); 3422 bufspacewakeup(); 3423 bp->b_bufsize = 0; 3424 } 3425 bp->b_saveaddr = bp->b_kvabase; 3426 bp->b_data = bp->b_saveaddr; 3427 bp->b_bcount = 0; 3428 bp->b_flags &= ~B_MALLOC; 3429 } 3430 return 1; 3431 } 3432 vm_hold_free_pages(bp, newbsize); 3433 } else if (newbsize > bp->b_bufsize) { 3434 /* 3435 * We only use malloced memory on the first allocation. 3436 * and revert to page-allocated memory when the buffer 3437 * grows. 3438 */ 3439 /* 3440 * There is a potential smp race here that could lead 3441 * to bufmallocspace slightly passing the max. It 3442 * is probably extremely rare and not worth worrying 3443 * over. 3444 */ 3445 if ( (bufmallocspace < maxbufmallocspace) && 3446 (bp->b_bufsize == 0) && 3447 (mbsize <= PAGE_SIZE/2)) { 3448 3449 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3450 bp->b_bufsize = mbsize; 3451 bp->b_bcount = size; 3452 bp->b_flags |= B_MALLOC; 3453 atomic_add_long(&bufmallocspace, mbsize); 3454 return 1; 3455 } 3456 origbuf = NULL; 3457 origbufsize = 0; 3458 /* 3459 * If the buffer is growing on its other-than-first allocation, 3460 * then we revert to the page-allocation scheme. 3461 */ 3462 if (bp->b_flags & B_MALLOC) { 3463 origbuf = bp->b_data; 3464 origbufsize = bp->b_bufsize; 3465 bp->b_data = bp->b_kvabase; 3466 if (bp->b_bufsize) { 3467 atomic_subtract_long(&bufmallocspace, 3468 bp->b_bufsize); 3469 bufspacewakeup(); 3470 bp->b_bufsize = 0; 3471 } 3472 bp->b_flags &= ~B_MALLOC; 3473 newbsize = round_page(newbsize); 3474 } 3475 vm_hold_load_pages( 3476 bp, 3477 (vm_offset_t) bp->b_data + bp->b_bufsize, 3478 (vm_offset_t) bp->b_data + newbsize); 3479 if (origbuf) { 3480 bcopy(origbuf, bp->b_data, origbufsize); 3481 free(origbuf, M_BIOBUF); 3482 } 3483 } 3484 } else { 3485 int desiredpages; 3486 3487 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3488 desiredpages = (size == 0) ? 0 : 3489 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3490 3491 if (bp->b_flags & B_MALLOC) 3492 panic("allocbuf: VMIO buffer can't be malloced"); 3493 /* 3494 * Set B_CACHE initially if buffer is 0 length or will become 3495 * 0-length. 3496 */ 3497 if (size == 0 || bp->b_bufsize == 0) 3498 bp->b_flags |= B_CACHE; 3499 3500 if (newbsize < bp->b_bufsize) { 3501 /* 3502 * DEV_BSIZE aligned new buffer size is less then the 3503 * DEV_BSIZE aligned existing buffer size. Figure out 3504 * if we have to remove any pages. 3505 */ 3506 if (desiredpages < bp->b_npages) { 3507 vm_page_t m; 3508 3509 if ((bp->b_flags & B_UNMAPPED) == 0) { 3510 BUF_CHECK_MAPPED(bp); 3511 pmap_qremove((vm_offset_t)trunc_page( 3512 (vm_offset_t)bp->b_data) + 3513 (desiredpages << PAGE_SHIFT), 3514 (bp->b_npages - desiredpages)); 3515 } else 3516 BUF_CHECK_UNMAPPED(bp); 3517 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3518 for (i = desiredpages; i < bp->b_npages; i++) { 3519 /* 3520 * the page is not freed here -- it 3521 * is the responsibility of 3522 * vnode_pager_setsize 3523 */ 3524 m = bp->b_pages[i]; 3525 KASSERT(m != bogus_page, 3526 ("allocbuf: bogus page found")); 3527 while (vm_page_sleep_if_busy(m, 3528 "biodep")) 3529 continue; 3530 3531 bp->b_pages[i] = NULL; 3532 vm_page_lock(m); 3533 vm_page_unwire(m, PQ_INACTIVE); 3534 vm_page_unlock(m); 3535 } 3536 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3537 bp->b_npages = desiredpages; 3538 } 3539 } else if (size > bp->b_bcount) { 3540 /* 3541 * We are growing the buffer, possibly in a 3542 * byte-granular fashion. 3543 */ 3544 vm_object_t obj; 3545 vm_offset_t toff; 3546 vm_offset_t tinc; 3547 3548 /* 3549 * Step 1, bring in the VM pages from the object, 3550 * allocating them if necessary. We must clear 3551 * B_CACHE if these pages are not valid for the 3552 * range covered by the buffer. 3553 */ 3554 3555 obj = bp->b_bufobj->bo_object; 3556 3557 VM_OBJECT_WLOCK(obj); 3558 while (bp->b_npages < desiredpages) { 3559 vm_page_t m; 3560 3561 /* 3562 * We must allocate system pages since blocking 3563 * here could interfere with paging I/O, no 3564 * matter which process we are. 3565 * 3566 * Only exclusive busy can be tested here. 3567 * Blocking on shared busy might lead to 3568 * deadlocks once allocbuf() is called after 3569 * pages are vfs_busy_pages(). 3570 */ 3571 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3572 bp->b_npages, VM_ALLOC_NOBUSY | 3573 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3574 VM_ALLOC_IGN_SBUSY | 3575 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3576 if (m->valid == 0) 3577 bp->b_flags &= ~B_CACHE; 3578 bp->b_pages[bp->b_npages] = m; 3579 ++bp->b_npages; 3580 } 3581 3582 /* 3583 * Step 2. We've loaded the pages into the buffer, 3584 * we have to figure out if we can still have B_CACHE 3585 * set. Note that B_CACHE is set according to the 3586 * byte-granular range ( bcount and size ), new the 3587 * aligned range ( newbsize ). 3588 * 3589 * The VM test is against m->valid, which is DEV_BSIZE 3590 * aligned. Needless to say, the validity of the data 3591 * needs to also be DEV_BSIZE aligned. Note that this 3592 * fails with NFS if the server or some other client 3593 * extends the file's EOF. If our buffer is resized, 3594 * B_CACHE may remain set! XXX 3595 */ 3596 3597 toff = bp->b_bcount; 3598 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3599 3600 while ((bp->b_flags & B_CACHE) && toff < size) { 3601 vm_pindex_t pi; 3602 3603 if (tinc > (size - toff)) 3604 tinc = size - toff; 3605 3606 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3607 PAGE_SHIFT; 3608 3609 vfs_buf_test_cache( 3610 bp, 3611 bp->b_offset, 3612 toff, 3613 tinc, 3614 bp->b_pages[pi] 3615 ); 3616 toff += tinc; 3617 tinc = PAGE_SIZE; 3618 } 3619 VM_OBJECT_WUNLOCK(obj); 3620 3621 /* 3622 * Step 3, fixup the KVM pmap. 3623 */ 3624 if ((bp->b_flags & B_UNMAPPED) == 0) 3625 bpmap_qenter(bp); 3626 else 3627 BUF_CHECK_UNMAPPED(bp); 3628 } 3629 } 3630 if (newbsize < bp->b_bufsize) 3631 bufspacewakeup(); 3632 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3633 bp->b_bcount = size; /* requested buffer size */ 3634 return 1; 3635 } 3636 3637 extern int inflight_transient_maps; 3638 3639 void 3640 biodone(struct bio *bp) 3641 { 3642 struct mtx *mtxp; 3643 void (*done)(struct bio *); 3644 vm_offset_t start, end; 3645 3646 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3647 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3648 bp->bio_flags |= BIO_UNMAPPED; 3649 start = trunc_page((vm_offset_t)bp->bio_data); 3650 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3651 bp->bio_data = unmapped_buf; 3652 pmap_qremove(start, OFF_TO_IDX(end - start)); 3653 vmem_free(transient_arena, start, end - start); 3654 atomic_add_int(&inflight_transient_maps, -1); 3655 } 3656 done = bp->bio_done; 3657 if (done == NULL) { 3658 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3659 mtx_lock(mtxp); 3660 bp->bio_flags |= BIO_DONE; 3661 wakeup(bp); 3662 mtx_unlock(mtxp); 3663 } else { 3664 bp->bio_flags |= BIO_DONE; 3665 done(bp); 3666 } 3667 } 3668 3669 /* 3670 * Wait for a BIO to finish. 3671 */ 3672 int 3673 biowait(struct bio *bp, const char *wchan) 3674 { 3675 struct mtx *mtxp; 3676 3677 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3678 mtx_lock(mtxp); 3679 while ((bp->bio_flags & BIO_DONE) == 0) 3680 msleep(bp, mtxp, PRIBIO, wchan, 0); 3681 mtx_unlock(mtxp); 3682 if (bp->bio_error != 0) 3683 return (bp->bio_error); 3684 if (!(bp->bio_flags & BIO_ERROR)) 3685 return (0); 3686 return (EIO); 3687 } 3688 3689 void 3690 biofinish(struct bio *bp, struct devstat *stat, int error) 3691 { 3692 3693 if (error) { 3694 bp->bio_error = error; 3695 bp->bio_flags |= BIO_ERROR; 3696 } 3697 if (stat != NULL) 3698 devstat_end_transaction_bio(stat, bp); 3699 biodone(bp); 3700 } 3701 3702 /* 3703 * bufwait: 3704 * 3705 * Wait for buffer I/O completion, returning error status. The buffer 3706 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3707 * error and cleared. 3708 */ 3709 int 3710 bufwait(struct buf *bp) 3711 { 3712 if (bp->b_iocmd == BIO_READ) 3713 bwait(bp, PRIBIO, "biord"); 3714 else 3715 bwait(bp, PRIBIO, "biowr"); 3716 if (bp->b_flags & B_EINTR) { 3717 bp->b_flags &= ~B_EINTR; 3718 return (EINTR); 3719 } 3720 if (bp->b_ioflags & BIO_ERROR) { 3721 return (bp->b_error ? bp->b_error : EIO); 3722 } else { 3723 return (0); 3724 } 3725 } 3726 3727 /* 3728 * Call back function from struct bio back up to struct buf. 3729 */ 3730 static void 3731 bufdonebio(struct bio *bip) 3732 { 3733 struct buf *bp; 3734 3735 bp = bip->bio_caller2; 3736 bp->b_resid = bip->bio_resid; 3737 bp->b_ioflags = bip->bio_flags; 3738 bp->b_error = bip->bio_error; 3739 if (bp->b_error) 3740 bp->b_ioflags |= BIO_ERROR; 3741 bufdone(bp); 3742 g_destroy_bio(bip); 3743 } 3744 3745 void 3746 dev_strategy(struct cdev *dev, struct buf *bp) 3747 { 3748 struct cdevsw *csw; 3749 int ref; 3750 3751 KASSERT(dev->si_refcount > 0, 3752 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3753 devtoname(dev), dev)); 3754 3755 csw = dev_refthread(dev, &ref); 3756 dev_strategy_csw(dev, csw, bp); 3757 dev_relthread(dev, ref); 3758 } 3759 3760 void 3761 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3762 { 3763 struct bio *bip; 3764 3765 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3766 ("b_iocmd botch")); 3767 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3768 dev->si_threadcount > 0, 3769 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3770 dev)); 3771 if (csw == NULL) { 3772 bp->b_error = ENXIO; 3773 bp->b_ioflags = BIO_ERROR; 3774 bufdone(bp); 3775 return; 3776 } 3777 for (;;) { 3778 bip = g_new_bio(); 3779 if (bip != NULL) 3780 break; 3781 /* Try again later */ 3782 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3783 } 3784 bip->bio_cmd = bp->b_iocmd; 3785 bip->bio_offset = bp->b_iooffset; 3786 bip->bio_length = bp->b_bcount; 3787 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3788 bdata2bio(bp, bip); 3789 bip->bio_done = bufdonebio; 3790 bip->bio_caller2 = bp; 3791 bip->bio_dev = dev; 3792 (*csw->d_strategy)(bip); 3793 } 3794 3795 /* 3796 * bufdone: 3797 * 3798 * Finish I/O on a buffer, optionally calling a completion function. 3799 * This is usually called from an interrupt so process blocking is 3800 * not allowed. 3801 * 3802 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3803 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3804 * assuming B_INVAL is clear. 3805 * 3806 * For the VMIO case, we set B_CACHE if the op was a read and no 3807 * read error occured, or if the op was a write. B_CACHE is never 3808 * set if the buffer is invalid or otherwise uncacheable. 3809 * 3810 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3811 * initiator to leave B_INVAL set to brelse the buffer out of existance 3812 * in the biodone routine. 3813 */ 3814 void 3815 bufdone(struct buf *bp) 3816 { 3817 struct bufobj *dropobj; 3818 void (*biodone)(struct buf *); 3819 3820 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3821 dropobj = NULL; 3822 3823 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3824 BUF_ASSERT_HELD(bp); 3825 3826 runningbufwakeup(bp); 3827 if (bp->b_iocmd == BIO_WRITE) 3828 dropobj = bp->b_bufobj; 3829 /* call optional completion function if requested */ 3830 if (bp->b_iodone != NULL) { 3831 biodone = bp->b_iodone; 3832 bp->b_iodone = NULL; 3833 (*biodone) (bp); 3834 if (dropobj) 3835 bufobj_wdrop(dropobj); 3836 return; 3837 } 3838 3839 bufdone_finish(bp); 3840 3841 if (dropobj) 3842 bufobj_wdrop(dropobj); 3843 } 3844 3845 void 3846 bufdone_finish(struct buf *bp) 3847 { 3848 BUF_ASSERT_HELD(bp); 3849 3850 if (!LIST_EMPTY(&bp->b_dep)) 3851 buf_complete(bp); 3852 3853 if (bp->b_flags & B_VMIO) { 3854 vm_ooffset_t foff; 3855 vm_page_t m; 3856 vm_object_t obj; 3857 struct vnode *vp; 3858 int bogus, i, iosize; 3859 3860 obj = bp->b_bufobj->bo_object; 3861 KASSERT(obj->paging_in_progress >= bp->b_npages, 3862 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3863 obj->paging_in_progress, bp->b_npages)); 3864 3865 vp = bp->b_vp; 3866 KASSERT(vp->v_holdcnt > 0, 3867 ("biodone_finish: vnode %p has zero hold count", vp)); 3868 KASSERT(vp->v_object != NULL, 3869 ("biodone_finish: vnode %p has no vm_object", vp)); 3870 3871 foff = bp->b_offset; 3872 KASSERT(bp->b_offset != NOOFFSET, 3873 ("biodone_finish: bp %p has no buffer offset", bp)); 3874 3875 /* 3876 * Set B_CACHE if the op was a normal read and no error 3877 * occured. B_CACHE is set for writes in the b*write() 3878 * routines. 3879 */ 3880 iosize = bp->b_bcount - bp->b_resid; 3881 if (bp->b_iocmd == BIO_READ && 3882 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3883 !(bp->b_ioflags & BIO_ERROR)) { 3884 bp->b_flags |= B_CACHE; 3885 } 3886 bogus = 0; 3887 VM_OBJECT_WLOCK(obj); 3888 for (i = 0; i < bp->b_npages; i++) { 3889 int bogusflag = 0; 3890 int resid; 3891 3892 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3893 if (resid > iosize) 3894 resid = iosize; 3895 3896 /* 3897 * cleanup bogus pages, restoring the originals 3898 */ 3899 m = bp->b_pages[i]; 3900 if (m == bogus_page) { 3901 bogus = bogusflag = 1; 3902 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3903 if (m == NULL) 3904 panic("biodone: page disappeared!"); 3905 bp->b_pages[i] = m; 3906 } 3907 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3908 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3909 (intmax_t)foff, (uintmax_t)m->pindex)); 3910 3911 /* 3912 * In the write case, the valid and clean bits are 3913 * already changed correctly ( see bdwrite() ), so we 3914 * only need to do this here in the read case. 3915 */ 3916 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3917 KASSERT((m->dirty & vm_page_bits(foff & 3918 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3919 " page %p has unexpected dirty bits", m)); 3920 vfs_page_set_valid(bp, foff, m); 3921 } 3922 3923 vm_page_sunbusy(m); 3924 vm_object_pip_subtract(obj, 1); 3925 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3926 iosize -= resid; 3927 } 3928 vm_object_pip_wakeupn(obj, 0); 3929 VM_OBJECT_WUNLOCK(obj); 3930 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3931 BUF_CHECK_MAPPED(bp); 3932 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3933 bp->b_pages, bp->b_npages); 3934 } 3935 } 3936 3937 /* 3938 * For asynchronous completions, release the buffer now. The brelse 3939 * will do a wakeup there if necessary - so no need to do a wakeup 3940 * here in the async case. The sync case always needs to do a wakeup. 3941 */ 3942 3943 if (bp->b_flags & B_ASYNC) { 3944 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3945 brelse(bp); 3946 else 3947 bqrelse(bp); 3948 } else 3949 bdone(bp); 3950 } 3951 3952 /* 3953 * This routine is called in lieu of iodone in the case of 3954 * incomplete I/O. This keeps the busy status for pages 3955 * consistant. 3956 */ 3957 void 3958 vfs_unbusy_pages(struct buf *bp) 3959 { 3960 int i; 3961 vm_object_t obj; 3962 vm_page_t m; 3963 3964 runningbufwakeup(bp); 3965 if (!(bp->b_flags & B_VMIO)) 3966 return; 3967 3968 obj = bp->b_bufobj->bo_object; 3969 VM_OBJECT_WLOCK(obj); 3970 for (i = 0; i < bp->b_npages; i++) { 3971 m = bp->b_pages[i]; 3972 if (m == bogus_page) { 3973 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3974 if (!m) 3975 panic("vfs_unbusy_pages: page missing\n"); 3976 bp->b_pages[i] = m; 3977 if ((bp->b_flags & B_UNMAPPED) == 0) { 3978 BUF_CHECK_MAPPED(bp); 3979 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3980 bp->b_pages, bp->b_npages); 3981 } else 3982 BUF_CHECK_UNMAPPED(bp); 3983 } 3984 vm_object_pip_subtract(obj, 1); 3985 vm_page_sunbusy(m); 3986 } 3987 vm_object_pip_wakeupn(obj, 0); 3988 VM_OBJECT_WUNLOCK(obj); 3989 } 3990 3991 /* 3992 * vfs_page_set_valid: 3993 * 3994 * Set the valid bits in a page based on the supplied offset. The 3995 * range is restricted to the buffer's size. 3996 * 3997 * This routine is typically called after a read completes. 3998 */ 3999 static void 4000 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4001 { 4002 vm_ooffset_t eoff; 4003 4004 /* 4005 * Compute the end offset, eoff, such that [off, eoff) does not span a 4006 * page boundary and eoff is not greater than the end of the buffer. 4007 * The end of the buffer, in this case, is our file EOF, not the 4008 * allocation size of the buffer. 4009 */ 4010 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4011 if (eoff > bp->b_offset + bp->b_bcount) 4012 eoff = bp->b_offset + bp->b_bcount; 4013 4014 /* 4015 * Set valid range. This is typically the entire buffer and thus the 4016 * entire page. 4017 */ 4018 if (eoff > off) 4019 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4020 } 4021 4022 /* 4023 * vfs_page_set_validclean: 4024 * 4025 * Set the valid bits and clear the dirty bits in a page based on the 4026 * supplied offset. The range is restricted to the buffer's size. 4027 */ 4028 static void 4029 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4030 { 4031 vm_ooffset_t soff, eoff; 4032 4033 /* 4034 * Start and end offsets in buffer. eoff - soff may not cross a 4035 * page boundry or cross the end of the buffer. The end of the 4036 * buffer, in this case, is our file EOF, not the allocation size 4037 * of the buffer. 4038 */ 4039 soff = off; 4040 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4041 if (eoff > bp->b_offset + bp->b_bcount) 4042 eoff = bp->b_offset + bp->b_bcount; 4043 4044 /* 4045 * Set valid range. This is typically the entire buffer and thus the 4046 * entire page. 4047 */ 4048 if (eoff > soff) { 4049 vm_page_set_validclean( 4050 m, 4051 (vm_offset_t) (soff & PAGE_MASK), 4052 (vm_offset_t) (eoff - soff) 4053 ); 4054 } 4055 } 4056 4057 /* 4058 * Ensure that all buffer pages are not exclusive busied. If any page is 4059 * exclusive busy, drain it. 4060 */ 4061 void 4062 vfs_drain_busy_pages(struct buf *bp) 4063 { 4064 vm_page_t m; 4065 int i, last_busied; 4066 4067 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4068 last_busied = 0; 4069 for (i = 0; i < bp->b_npages; i++) { 4070 m = bp->b_pages[i]; 4071 if (vm_page_xbusied(m)) { 4072 for (; last_busied < i; last_busied++) 4073 vm_page_sbusy(bp->b_pages[last_busied]); 4074 while (vm_page_xbusied(m)) { 4075 vm_page_lock(m); 4076 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4077 vm_page_busy_sleep(m, "vbpage"); 4078 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4079 } 4080 } 4081 } 4082 for (i = 0; i < last_busied; i++) 4083 vm_page_sunbusy(bp->b_pages[i]); 4084 } 4085 4086 /* 4087 * This routine is called before a device strategy routine. 4088 * It is used to tell the VM system that paging I/O is in 4089 * progress, and treat the pages associated with the buffer 4090 * almost as being exclusive busy. Also the object paging_in_progress 4091 * flag is handled to make sure that the object doesn't become 4092 * inconsistant. 4093 * 4094 * Since I/O has not been initiated yet, certain buffer flags 4095 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4096 * and should be ignored. 4097 */ 4098 void 4099 vfs_busy_pages(struct buf *bp, int clear_modify) 4100 { 4101 int i, bogus; 4102 vm_object_t obj; 4103 vm_ooffset_t foff; 4104 vm_page_t m; 4105 4106 if (!(bp->b_flags & B_VMIO)) 4107 return; 4108 4109 obj = bp->b_bufobj->bo_object; 4110 foff = bp->b_offset; 4111 KASSERT(bp->b_offset != NOOFFSET, 4112 ("vfs_busy_pages: no buffer offset")); 4113 VM_OBJECT_WLOCK(obj); 4114 vfs_drain_busy_pages(bp); 4115 if (bp->b_bufsize != 0) 4116 vfs_setdirty_locked_object(bp); 4117 bogus = 0; 4118 for (i = 0; i < bp->b_npages; i++) { 4119 m = bp->b_pages[i]; 4120 4121 if ((bp->b_flags & B_CLUSTER) == 0) { 4122 vm_object_pip_add(obj, 1); 4123 vm_page_sbusy(m); 4124 } 4125 /* 4126 * When readying a buffer for a read ( i.e 4127 * clear_modify == 0 ), it is important to do 4128 * bogus_page replacement for valid pages in 4129 * partially instantiated buffers. Partially 4130 * instantiated buffers can, in turn, occur when 4131 * reconstituting a buffer from its VM backing store 4132 * base. We only have to do this if B_CACHE is 4133 * clear ( which causes the I/O to occur in the 4134 * first place ). The replacement prevents the read 4135 * I/O from overwriting potentially dirty VM-backed 4136 * pages. XXX bogus page replacement is, uh, bogus. 4137 * It may not work properly with small-block devices. 4138 * We need to find a better way. 4139 */ 4140 if (clear_modify) { 4141 pmap_remove_write(m); 4142 vfs_page_set_validclean(bp, foff, m); 4143 } else if (m->valid == VM_PAGE_BITS_ALL && 4144 (bp->b_flags & B_CACHE) == 0) { 4145 bp->b_pages[i] = bogus_page; 4146 bogus++; 4147 } 4148 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4149 } 4150 VM_OBJECT_WUNLOCK(obj); 4151 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4152 BUF_CHECK_MAPPED(bp); 4153 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4154 bp->b_pages, bp->b_npages); 4155 } 4156 } 4157 4158 /* 4159 * vfs_bio_set_valid: 4160 * 4161 * Set the range within the buffer to valid. The range is 4162 * relative to the beginning of the buffer, b_offset. Note that 4163 * b_offset itself may be offset from the beginning of the first 4164 * page. 4165 */ 4166 void 4167 vfs_bio_set_valid(struct buf *bp, int base, int size) 4168 { 4169 int i, n; 4170 vm_page_t m; 4171 4172 if (!(bp->b_flags & B_VMIO)) 4173 return; 4174 4175 /* 4176 * Fixup base to be relative to beginning of first page. 4177 * Set initial n to be the maximum number of bytes in the 4178 * first page that can be validated. 4179 */ 4180 base += (bp->b_offset & PAGE_MASK); 4181 n = PAGE_SIZE - (base & PAGE_MASK); 4182 4183 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4184 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4185 m = bp->b_pages[i]; 4186 if (n > size) 4187 n = size; 4188 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4189 base += n; 4190 size -= n; 4191 n = PAGE_SIZE; 4192 } 4193 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4194 } 4195 4196 /* 4197 * vfs_bio_clrbuf: 4198 * 4199 * If the specified buffer is a non-VMIO buffer, clear the entire 4200 * buffer. If the specified buffer is a VMIO buffer, clear and 4201 * validate only the previously invalid portions of the buffer. 4202 * This routine essentially fakes an I/O, so we need to clear 4203 * BIO_ERROR and B_INVAL. 4204 * 4205 * Note that while we only theoretically need to clear through b_bcount, 4206 * we go ahead and clear through b_bufsize. 4207 */ 4208 void 4209 vfs_bio_clrbuf(struct buf *bp) 4210 { 4211 int i, j, mask, sa, ea, slide; 4212 4213 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4214 clrbuf(bp); 4215 return; 4216 } 4217 bp->b_flags &= ~B_INVAL; 4218 bp->b_ioflags &= ~BIO_ERROR; 4219 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4220 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4221 (bp->b_offset & PAGE_MASK) == 0) { 4222 if (bp->b_pages[0] == bogus_page) 4223 goto unlock; 4224 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4225 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4226 if ((bp->b_pages[0]->valid & mask) == mask) 4227 goto unlock; 4228 if ((bp->b_pages[0]->valid & mask) == 0) { 4229 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4230 bp->b_pages[0]->valid |= mask; 4231 goto unlock; 4232 } 4233 } 4234 sa = bp->b_offset & PAGE_MASK; 4235 slide = 0; 4236 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4237 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4238 ea = slide & PAGE_MASK; 4239 if (ea == 0) 4240 ea = PAGE_SIZE; 4241 if (bp->b_pages[i] == bogus_page) 4242 continue; 4243 j = sa / DEV_BSIZE; 4244 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4245 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4246 if ((bp->b_pages[i]->valid & mask) == mask) 4247 continue; 4248 if ((bp->b_pages[i]->valid & mask) == 0) 4249 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4250 else { 4251 for (; sa < ea; sa += DEV_BSIZE, j++) { 4252 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4253 pmap_zero_page_area(bp->b_pages[i], 4254 sa, DEV_BSIZE); 4255 } 4256 } 4257 } 4258 bp->b_pages[i]->valid |= mask; 4259 } 4260 unlock: 4261 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4262 bp->b_resid = 0; 4263 } 4264 4265 void 4266 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4267 { 4268 vm_page_t m; 4269 int i, n; 4270 4271 if ((bp->b_flags & B_UNMAPPED) == 0) { 4272 BUF_CHECK_MAPPED(bp); 4273 bzero(bp->b_data + base, size); 4274 } else { 4275 BUF_CHECK_UNMAPPED(bp); 4276 n = PAGE_SIZE - (base & PAGE_MASK); 4277 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4278 m = bp->b_pages[i]; 4279 if (n > size) 4280 n = size; 4281 pmap_zero_page_area(m, base & PAGE_MASK, n); 4282 base += n; 4283 size -= n; 4284 n = PAGE_SIZE; 4285 } 4286 } 4287 } 4288 4289 /* 4290 * vm_hold_load_pages and vm_hold_free_pages get pages into 4291 * a buffers address space. The pages are anonymous and are 4292 * not associated with a file object. 4293 */ 4294 static void 4295 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4296 { 4297 vm_offset_t pg; 4298 vm_page_t p; 4299 int index; 4300 4301 BUF_CHECK_MAPPED(bp); 4302 4303 to = round_page(to); 4304 from = round_page(from); 4305 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4306 4307 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4308 tryagain: 4309 /* 4310 * note: must allocate system pages since blocking here 4311 * could interfere with paging I/O, no matter which 4312 * process we are. 4313 */ 4314 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4315 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4316 if (p == NULL) { 4317 VM_WAIT; 4318 goto tryagain; 4319 } 4320 pmap_qenter(pg, &p, 1); 4321 bp->b_pages[index] = p; 4322 } 4323 bp->b_npages = index; 4324 } 4325 4326 /* Return pages associated with this buf to the vm system */ 4327 static void 4328 vm_hold_free_pages(struct buf *bp, int newbsize) 4329 { 4330 vm_offset_t from; 4331 vm_page_t p; 4332 int index, newnpages; 4333 4334 BUF_CHECK_MAPPED(bp); 4335 4336 from = round_page((vm_offset_t)bp->b_data + newbsize); 4337 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4338 if (bp->b_npages > newnpages) 4339 pmap_qremove(from, bp->b_npages - newnpages); 4340 for (index = newnpages; index < bp->b_npages; index++) { 4341 p = bp->b_pages[index]; 4342 bp->b_pages[index] = NULL; 4343 if (vm_page_sbusied(p)) 4344 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4345 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4346 p->wire_count--; 4347 vm_page_free(p); 4348 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 4349 } 4350 bp->b_npages = newnpages; 4351 } 4352 4353 /* 4354 * Map an IO request into kernel virtual address space. 4355 * 4356 * All requests are (re)mapped into kernel VA space. 4357 * Notice that we use b_bufsize for the size of the buffer 4358 * to be mapped. b_bcount might be modified by the driver. 4359 * 4360 * Note that even if the caller determines that the address space should 4361 * be valid, a race or a smaller-file mapped into a larger space may 4362 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4363 * check the return value. 4364 */ 4365 int 4366 vmapbuf(struct buf *bp, int mapbuf) 4367 { 4368 caddr_t kva; 4369 vm_prot_t prot; 4370 int pidx; 4371 4372 if (bp->b_bufsize < 0) 4373 return (-1); 4374 prot = VM_PROT_READ; 4375 if (bp->b_iocmd == BIO_READ) 4376 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4377 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4378 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4379 btoc(MAXPHYS))) < 0) 4380 return (-1); 4381 bp->b_npages = pidx; 4382 if (mapbuf || !unmapped_buf_allowed) { 4383 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4384 kva = bp->b_saveaddr; 4385 bp->b_saveaddr = bp->b_data; 4386 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4387 bp->b_flags &= ~B_UNMAPPED; 4388 } else { 4389 bp->b_flags |= B_UNMAPPED; 4390 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4391 bp->b_saveaddr = bp->b_data; 4392 bp->b_data = unmapped_buf; 4393 } 4394 return(0); 4395 } 4396 4397 /* 4398 * Free the io map PTEs associated with this IO operation. 4399 * We also invalidate the TLB entries and restore the original b_addr. 4400 */ 4401 void 4402 vunmapbuf(struct buf *bp) 4403 { 4404 int npages; 4405 4406 npages = bp->b_npages; 4407 if (bp->b_flags & B_UNMAPPED) 4408 bp->b_flags &= ~B_UNMAPPED; 4409 else 4410 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4411 vm_page_unhold_pages(bp->b_pages, npages); 4412 4413 bp->b_data = bp->b_saveaddr; 4414 } 4415 4416 void 4417 bdone(struct buf *bp) 4418 { 4419 struct mtx *mtxp; 4420 4421 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4422 mtx_lock(mtxp); 4423 bp->b_flags |= B_DONE; 4424 wakeup(bp); 4425 mtx_unlock(mtxp); 4426 } 4427 4428 void 4429 bwait(struct buf *bp, u_char pri, const char *wchan) 4430 { 4431 struct mtx *mtxp; 4432 4433 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4434 mtx_lock(mtxp); 4435 while ((bp->b_flags & B_DONE) == 0) 4436 msleep(bp, mtxp, pri, wchan, 0); 4437 mtx_unlock(mtxp); 4438 } 4439 4440 int 4441 bufsync(struct bufobj *bo, int waitfor) 4442 { 4443 4444 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4445 } 4446 4447 void 4448 bufstrategy(struct bufobj *bo, struct buf *bp) 4449 { 4450 int i = 0; 4451 struct vnode *vp; 4452 4453 vp = bp->b_vp; 4454 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4455 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4456 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4457 i = VOP_STRATEGY(vp, bp); 4458 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4459 } 4460 4461 void 4462 bufobj_wrefl(struct bufobj *bo) 4463 { 4464 4465 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4466 ASSERT_BO_WLOCKED(bo); 4467 bo->bo_numoutput++; 4468 } 4469 4470 void 4471 bufobj_wref(struct bufobj *bo) 4472 { 4473 4474 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4475 BO_LOCK(bo); 4476 bo->bo_numoutput++; 4477 BO_UNLOCK(bo); 4478 } 4479 4480 void 4481 bufobj_wdrop(struct bufobj *bo) 4482 { 4483 4484 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4485 BO_LOCK(bo); 4486 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4487 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4488 bo->bo_flag &= ~BO_WWAIT; 4489 wakeup(&bo->bo_numoutput); 4490 } 4491 BO_UNLOCK(bo); 4492 } 4493 4494 int 4495 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4496 { 4497 int error; 4498 4499 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4500 ASSERT_BO_WLOCKED(bo); 4501 error = 0; 4502 while (bo->bo_numoutput) { 4503 bo->bo_flag |= BO_WWAIT; 4504 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4505 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4506 if (error) 4507 break; 4508 } 4509 return (error); 4510 } 4511 4512 void 4513 bpin(struct buf *bp) 4514 { 4515 struct mtx *mtxp; 4516 4517 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4518 mtx_lock(mtxp); 4519 bp->b_pin_count++; 4520 mtx_unlock(mtxp); 4521 } 4522 4523 void 4524 bunpin(struct buf *bp) 4525 { 4526 struct mtx *mtxp; 4527 4528 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4529 mtx_lock(mtxp); 4530 if (--bp->b_pin_count == 0) 4531 wakeup(bp); 4532 mtx_unlock(mtxp); 4533 } 4534 4535 void 4536 bunpin_wait(struct buf *bp) 4537 { 4538 struct mtx *mtxp; 4539 4540 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4541 mtx_lock(mtxp); 4542 while (bp->b_pin_count > 0) 4543 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4544 mtx_unlock(mtxp); 4545 } 4546 4547 /* 4548 * Set bio_data or bio_ma for struct bio from the struct buf. 4549 */ 4550 void 4551 bdata2bio(struct buf *bp, struct bio *bip) 4552 { 4553 4554 if ((bp->b_flags & B_UNMAPPED) != 0) { 4555 KASSERT(unmapped_buf_allowed, ("unmapped")); 4556 bip->bio_ma = bp->b_pages; 4557 bip->bio_ma_n = bp->b_npages; 4558 bip->bio_data = unmapped_buf; 4559 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4560 bip->bio_flags |= BIO_UNMAPPED; 4561 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4562 PAGE_SIZE == bp->b_npages, 4563 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4564 (long long)bip->bio_length, bip->bio_ma_n)); 4565 } else { 4566 bip->bio_data = bp->b_data; 4567 bip->bio_ma = NULL; 4568 } 4569 } 4570 4571 #include "opt_ddb.h" 4572 #ifdef DDB 4573 #include <ddb/ddb.h> 4574 4575 /* DDB command to show buffer data */ 4576 DB_SHOW_COMMAND(buffer, db_show_buffer) 4577 { 4578 /* get args */ 4579 struct buf *bp = (struct buf *)addr; 4580 4581 if (!have_addr) { 4582 db_printf("usage: show buffer <addr>\n"); 4583 return; 4584 } 4585 4586 db_printf("buf at %p\n", bp); 4587 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4588 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4589 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4590 db_printf( 4591 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4592 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4593 "b_dep = %p\n", 4594 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4595 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4596 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4597 if (bp->b_npages) { 4598 int i; 4599 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4600 for (i = 0; i < bp->b_npages; i++) { 4601 vm_page_t m; 4602 m = bp->b_pages[i]; 4603 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4604 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4605 if ((i + 1) < bp->b_npages) 4606 db_printf(","); 4607 } 4608 db_printf("\n"); 4609 } 4610 db_printf(" "); 4611 BUF_LOCKPRINTINFO(bp); 4612 } 4613 4614 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4615 { 4616 struct buf *bp; 4617 int i; 4618 4619 for (i = 0; i < nbuf; i++) { 4620 bp = &buf[i]; 4621 if (BUF_ISLOCKED(bp)) { 4622 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4623 db_printf("\n"); 4624 } 4625 } 4626 } 4627 4628 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4629 { 4630 struct vnode *vp; 4631 struct buf *bp; 4632 4633 if (!have_addr) { 4634 db_printf("usage: show vnodebufs <addr>\n"); 4635 return; 4636 } 4637 vp = (struct vnode *)addr; 4638 db_printf("Clean buffers:\n"); 4639 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4640 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4641 db_printf("\n"); 4642 } 4643 db_printf("Dirty buffers:\n"); 4644 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4645 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4646 db_printf("\n"); 4647 } 4648 } 4649 4650 DB_COMMAND(countfreebufs, db_coundfreebufs) 4651 { 4652 struct buf *bp; 4653 int i, used = 0, nfree = 0; 4654 4655 if (have_addr) { 4656 db_printf("usage: countfreebufs\n"); 4657 return; 4658 } 4659 4660 for (i = 0; i < nbuf; i++) { 4661 bp = &buf[i]; 4662 if ((bp->b_flags & B_INFREECNT) != 0) 4663 nfree++; 4664 else 4665 used++; 4666 } 4667 4668 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4669 nfree + used); 4670 db_printf("numfreebuffers is %d\n", numfreebuffers); 4671 } 4672 #endif /* DDB */ 4673