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 nlock = bqlock(qindex); 998 /* Handle delayed bremfree() processing. */ 999 if (bp->b_flags & B_REMFREE) { 1000 olock = bqlock(bp->b_qindex); 1001 mtx_lock(olock); 1002 bremfreel(bp); 1003 if (olock != nlock) { 1004 mtx_unlock(olock); 1005 mtx_lock(nlock); 1006 } 1007 } else 1008 mtx_lock(nlock); 1009 1010 if (bp->b_qindex != QUEUE_NONE) 1011 panic("binsfree: free buffer onto another queue???"); 1012 1013 bp->b_qindex = qindex; 1014 if (bp->b_flags & B_AGE) 1015 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1016 else 1017 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1018 #ifdef INVARIANTS 1019 bq_len[bp->b_qindex]++; 1020 #endif 1021 mtx_unlock(nlock); 1022 1023 /* 1024 * Something we can maybe free or reuse. 1025 */ 1026 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1027 bufspacewakeup(); 1028 1029 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1030 bufcountadd(bp); 1031 } 1032 1033 /* 1034 * bremfree: 1035 * 1036 * Mark the buffer for removal from the appropriate free list. 1037 * 1038 */ 1039 void 1040 bremfree(struct buf *bp) 1041 { 1042 1043 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1044 KASSERT((bp->b_flags & B_REMFREE) == 0, 1045 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1046 KASSERT(bp->b_qindex != QUEUE_NONE, 1047 ("bremfree: buffer %p not on a queue.", bp)); 1048 BUF_ASSERT_XLOCKED(bp); 1049 1050 bp->b_flags |= B_REMFREE; 1051 bufcountsub(bp); 1052 } 1053 1054 /* 1055 * bremfreef: 1056 * 1057 * Force an immediate removal from a free list. Used only in nfs when 1058 * it abuses the b_freelist pointer. 1059 */ 1060 void 1061 bremfreef(struct buf *bp) 1062 { 1063 struct mtx *qlock; 1064 1065 qlock = bqlock(bp->b_qindex); 1066 mtx_lock(qlock); 1067 bremfreel(bp); 1068 mtx_unlock(qlock); 1069 } 1070 1071 /* 1072 * bremfreel: 1073 * 1074 * Removes a buffer from the free list, must be called with the 1075 * correct qlock held. 1076 */ 1077 static void 1078 bremfreel(struct buf *bp) 1079 { 1080 1081 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1082 bp, bp->b_vp, bp->b_flags); 1083 KASSERT(bp->b_qindex != QUEUE_NONE, 1084 ("bremfreel: buffer %p not on a queue.", bp)); 1085 BUF_ASSERT_XLOCKED(bp); 1086 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1087 1088 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1089 #ifdef INVARIANTS 1090 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1091 bp->b_qindex)); 1092 bq_len[bp->b_qindex]--; 1093 #endif 1094 bp->b_qindex = QUEUE_NONE; 1095 /* 1096 * If this was a delayed bremfree() we only need to remove the buffer 1097 * from the queue and return the stats are already done. 1098 */ 1099 if (bp->b_flags & B_REMFREE) { 1100 bp->b_flags &= ~B_REMFREE; 1101 return; 1102 } 1103 bufcountsub(bp); 1104 } 1105 1106 /* 1107 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1108 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1109 * the buffer is valid and we do not have to do anything. 1110 */ 1111 void 1112 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 1113 int cnt, struct ucred * cred) 1114 { 1115 struct buf *rabp; 1116 int i; 1117 1118 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1119 if (inmem(vp, *rablkno)) 1120 continue; 1121 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1122 1123 if ((rabp->b_flags & B_CACHE) == 0) { 1124 if (!TD_IS_IDLETHREAD(curthread)) 1125 curthread->td_ru.ru_inblock++; 1126 rabp->b_flags |= B_ASYNC; 1127 rabp->b_flags &= ~B_INVAL; 1128 rabp->b_ioflags &= ~BIO_ERROR; 1129 rabp->b_iocmd = BIO_READ; 1130 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1131 rabp->b_rcred = crhold(cred); 1132 vfs_busy_pages(rabp, 0); 1133 BUF_KERNPROC(rabp); 1134 rabp->b_iooffset = dbtob(rabp->b_blkno); 1135 bstrategy(rabp); 1136 } else { 1137 brelse(rabp); 1138 } 1139 } 1140 } 1141 1142 /* 1143 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1144 * 1145 * Get a buffer with the specified data. Look in the cache first. We 1146 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1147 * is set, the buffer is valid and we do not have to do anything, see 1148 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1149 */ 1150 int 1151 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1152 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 1153 { 1154 struct buf *bp; 1155 int rv = 0, readwait = 0; 1156 1157 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1158 /* 1159 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1160 */ 1161 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1162 if (bp == NULL) 1163 return (EBUSY); 1164 1165 /* if not found in cache, do some I/O */ 1166 if ((bp->b_flags & B_CACHE) == 0) { 1167 if (!TD_IS_IDLETHREAD(curthread)) 1168 curthread->td_ru.ru_inblock++; 1169 bp->b_iocmd = BIO_READ; 1170 bp->b_flags &= ~B_INVAL; 1171 bp->b_ioflags &= ~BIO_ERROR; 1172 if (bp->b_rcred == NOCRED && cred != NOCRED) 1173 bp->b_rcred = crhold(cred); 1174 vfs_busy_pages(bp, 0); 1175 bp->b_iooffset = dbtob(bp->b_blkno); 1176 bstrategy(bp); 1177 ++readwait; 1178 } 1179 1180 breada(vp, rablkno, rabsize, cnt, cred); 1181 1182 if (readwait) { 1183 rv = bufwait(bp); 1184 } 1185 return (rv); 1186 } 1187 1188 /* 1189 * Write, release buffer on completion. (Done by iodone 1190 * if async). Do not bother writing anything if the buffer 1191 * is invalid. 1192 * 1193 * Note that we set B_CACHE here, indicating that buffer is 1194 * fully valid and thus cacheable. This is true even of NFS 1195 * now so we set it generally. This could be set either here 1196 * or in biodone() since the I/O is synchronous. We put it 1197 * here. 1198 */ 1199 int 1200 bufwrite(struct buf *bp) 1201 { 1202 int oldflags; 1203 struct vnode *vp; 1204 long space; 1205 int vp_md; 1206 1207 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1208 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 1209 bp->b_flags |= B_INVAL | B_RELBUF; 1210 bp->b_flags &= ~B_CACHE; 1211 brelse(bp); 1212 return (ENXIO); 1213 } 1214 if (bp->b_flags & B_INVAL) { 1215 brelse(bp); 1216 return (0); 1217 } 1218 1219 if (bp->b_flags & B_BARRIER) 1220 barrierwrites++; 1221 1222 oldflags = bp->b_flags; 1223 1224 BUF_ASSERT_HELD(bp); 1225 1226 if (bp->b_pin_count > 0) 1227 bunpin_wait(bp); 1228 1229 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1230 ("FFS background buffer should not get here %p", bp)); 1231 1232 vp = bp->b_vp; 1233 if (vp) 1234 vp_md = vp->v_vflag & VV_MD; 1235 else 1236 vp_md = 0; 1237 1238 /* 1239 * Mark the buffer clean. Increment the bufobj write count 1240 * before bundirty() call, to prevent other thread from seeing 1241 * empty dirty list and zero counter for writes in progress, 1242 * falsely indicating that the bufobj is clean. 1243 */ 1244 bufobj_wref(bp->b_bufobj); 1245 bundirty(bp); 1246 1247 bp->b_flags &= ~B_DONE; 1248 bp->b_ioflags &= ~BIO_ERROR; 1249 bp->b_flags |= B_CACHE; 1250 bp->b_iocmd = BIO_WRITE; 1251 1252 vfs_busy_pages(bp, 1); 1253 1254 /* 1255 * Normal bwrites pipeline writes 1256 */ 1257 bp->b_runningbufspace = bp->b_bufsize; 1258 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1259 1260 if (!TD_IS_IDLETHREAD(curthread)) 1261 curthread->td_ru.ru_oublock++; 1262 if (oldflags & B_ASYNC) 1263 BUF_KERNPROC(bp); 1264 bp->b_iooffset = dbtob(bp->b_blkno); 1265 bstrategy(bp); 1266 1267 if ((oldflags & B_ASYNC) == 0) { 1268 int rtval = bufwait(bp); 1269 brelse(bp); 1270 return (rtval); 1271 } else if (space > hirunningspace) { 1272 /* 1273 * don't allow the async write to saturate the I/O 1274 * system. We will not deadlock here because 1275 * we are blocking waiting for I/O that is already in-progress 1276 * to complete. We do not block here if it is the update 1277 * or syncer daemon trying to clean up as that can lead 1278 * to deadlock. 1279 */ 1280 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1281 waitrunningbufspace(); 1282 } 1283 1284 return (0); 1285 } 1286 1287 void 1288 bufbdflush(struct bufobj *bo, struct buf *bp) 1289 { 1290 struct buf *nbp; 1291 1292 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1293 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1294 altbufferflushes++; 1295 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1296 BO_LOCK(bo); 1297 /* 1298 * Try to find a buffer to flush. 1299 */ 1300 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1301 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1302 BUF_LOCK(nbp, 1303 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1304 continue; 1305 if (bp == nbp) 1306 panic("bdwrite: found ourselves"); 1307 BO_UNLOCK(bo); 1308 /* Don't countdeps with the bo lock held. */ 1309 if (buf_countdeps(nbp, 0)) { 1310 BO_LOCK(bo); 1311 BUF_UNLOCK(nbp); 1312 continue; 1313 } 1314 if (nbp->b_flags & B_CLUSTEROK) { 1315 vfs_bio_awrite(nbp); 1316 } else { 1317 bremfree(nbp); 1318 bawrite(nbp); 1319 } 1320 dirtybufferflushes++; 1321 break; 1322 } 1323 if (nbp == NULL) 1324 BO_UNLOCK(bo); 1325 } 1326 } 1327 1328 /* 1329 * Delayed write. (Buffer is marked dirty). Do not bother writing 1330 * anything if the buffer is marked invalid. 1331 * 1332 * Note that since the buffer must be completely valid, we can safely 1333 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1334 * biodone() in order to prevent getblk from writing the buffer 1335 * out synchronously. 1336 */ 1337 void 1338 bdwrite(struct buf *bp) 1339 { 1340 struct thread *td = curthread; 1341 struct vnode *vp; 1342 struct bufobj *bo; 1343 1344 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1345 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1346 KASSERT((bp->b_flags & B_BARRIER) == 0, 1347 ("Barrier request in delayed write %p", bp)); 1348 BUF_ASSERT_HELD(bp); 1349 1350 if (bp->b_flags & B_INVAL) { 1351 brelse(bp); 1352 return; 1353 } 1354 1355 /* 1356 * If we have too many dirty buffers, don't create any more. 1357 * If we are wildly over our limit, then force a complete 1358 * cleanup. Otherwise, just keep the situation from getting 1359 * out of control. Note that we have to avoid a recursive 1360 * disaster and not try to clean up after our own cleanup! 1361 */ 1362 vp = bp->b_vp; 1363 bo = bp->b_bufobj; 1364 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1365 td->td_pflags |= TDP_INBDFLUSH; 1366 BO_BDFLUSH(bo, bp); 1367 td->td_pflags &= ~TDP_INBDFLUSH; 1368 } else 1369 recursiveflushes++; 1370 1371 bdirty(bp); 1372 /* 1373 * Set B_CACHE, indicating that the buffer is fully valid. This is 1374 * true even of NFS now. 1375 */ 1376 bp->b_flags |= B_CACHE; 1377 1378 /* 1379 * This bmap keeps the system from needing to do the bmap later, 1380 * perhaps when the system is attempting to do a sync. Since it 1381 * is likely that the indirect block -- or whatever other datastructure 1382 * that the filesystem needs is still in memory now, it is a good 1383 * thing to do this. Note also, that if the pageout daemon is 1384 * requesting a sync -- there might not be enough memory to do 1385 * the bmap then... So, this is important to do. 1386 */ 1387 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1388 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1389 } 1390 1391 /* 1392 * Set the *dirty* buffer range based upon the VM system dirty 1393 * pages. 1394 * 1395 * Mark the buffer pages as clean. We need to do this here to 1396 * satisfy the vnode_pager and the pageout daemon, so that it 1397 * thinks that the pages have been "cleaned". Note that since 1398 * the pages are in a delayed write buffer -- the VFS layer 1399 * "will" see that the pages get written out on the next sync, 1400 * or perhaps the cluster will be completed. 1401 */ 1402 vfs_clean_pages_dirty_buf(bp); 1403 bqrelse(bp); 1404 1405 /* 1406 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1407 * due to the softdep code. 1408 */ 1409 } 1410 1411 /* 1412 * bdirty: 1413 * 1414 * Turn buffer into delayed write request. We must clear BIO_READ and 1415 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1416 * itself to properly update it in the dirty/clean lists. We mark it 1417 * B_DONE to ensure that any asynchronization of the buffer properly 1418 * clears B_DONE ( else a panic will occur later ). 1419 * 1420 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1421 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1422 * should only be called if the buffer is known-good. 1423 * 1424 * Since the buffer is not on a queue, we do not update the numfreebuffers 1425 * count. 1426 * 1427 * The buffer must be on QUEUE_NONE. 1428 */ 1429 void 1430 bdirty(struct buf *bp) 1431 { 1432 1433 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1434 bp, bp->b_vp, bp->b_flags); 1435 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1436 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1437 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1438 BUF_ASSERT_HELD(bp); 1439 bp->b_flags &= ~(B_RELBUF); 1440 bp->b_iocmd = BIO_WRITE; 1441 1442 if ((bp->b_flags & B_DELWRI) == 0) { 1443 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1444 reassignbuf(bp); 1445 bdirtyadd(); 1446 } 1447 } 1448 1449 /* 1450 * bundirty: 1451 * 1452 * Clear B_DELWRI for buffer. 1453 * 1454 * Since the buffer is not on a queue, we do not update the numfreebuffers 1455 * count. 1456 * 1457 * The buffer must be on QUEUE_NONE. 1458 */ 1459 1460 void 1461 bundirty(struct buf *bp) 1462 { 1463 1464 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1465 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1466 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1467 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1468 BUF_ASSERT_HELD(bp); 1469 1470 if (bp->b_flags & B_DELWRI) { 1471 bp->b_flags &= ~B_DELWRI; 1472 reassignbuf(bp); 1473 bdirtysub(); 1474 } 1475 /* 1476 * Since it is now being written, we can clear its deferred write flag. 1477 */ 1478 bp->b_flags &= ~B_DEFERRED; 1479 } 1480 1481 /* 1482 * bawrite: 1483 * 1484 * Asynchronous write. Start output on a buffer, but do not wait for 1485 * it to complete. The buffer is released when the output completes. 1486 * 1487 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1488 * B_INVAL buffers. Not us. 1489 */ 1490 void 1491 bawrite(struct buf *bp) 1492 { 1493 1494 bp->b_flags |= B_ASYNC; 1495 (void) bwrite(bp); 1496 } 1497 1498 /* 1499 * babarrierwrite: 1500 * 1501 * Asynchronous barrier write. Start output on a buffer, but do not 1502 * wait for it to complete. Place a write barrier after this write so 1503 * that this buffer and all buffers written before it are committed to 1504 * the disk before any buffers written after this write are committed 1505 * to the disk. The buffer is released when the output completes. 1506 */ 1507 void 1508 babarrierwrite(struct buf *bp) 1509 { 1510 1511 bp->b_flags |= B_ASYNC | B_BARRIER; 1512 (void) bwrite(bp); 1513 } 1514 1515 /* 1516 * bbarrierwrite: 1517 * 1518 * Synchronous barrier write. Start output on a buffer and wait for 1519 * it to complete. Place a write barrier after this write so that 1520 * this buffer and all buffers written before it are committed to 1521 * the disk before any buffers written after this write are committed 1522 * to the disk. The buffer is released when the output completes. 1523 */ 1524 int 1525 bbarrierwrite(struct buf *bp) 1526 { 1527 1528 bp->b_flags |= B_BARRIER; 1529 return (bwrite(bp)); 1530 } 1531 1532 /* 1533 * bwillwrite: 1534 * 1535 * Called prior to the locking of any vnodes when we are expecting to 1536 * write. We do not want to starve the buffer cache with too many 1537 * dirty buffers so we block here. By blocking prior to the locking 1538 * of any vnodes we attempt to avoid the situation where a locked vnode 1539 * prevents the various system daemons from flushing related buffers. 1540 */ 1541 void 1542 bwillwrite(void) 1543 { 1544 1545 if (numdirtybuffers >= hidirtybuffers) { 1546 mtx_lock(&bdirtylock); 1547 while (numdirtybuffers >= hidirtybuffers) { 1548 bdirtywait = 1; 1549 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 1550 "flswai", 0); 1551 } 1552 mtx_unlock(&bdirtylock); 1553 } 1554 } 1555 1556 /* 1557 * Return true if we have too many dirty buffers. 1558 */ 1559 int 1560 buf_dirty_count_severe(void) 1561 { 1562 1563 return(numdirtybuffers >= hidirtybuffers); 1564 } 1565 1566 static __noinline int 1567 buf_vm_page_count_severe(void) 1568 { 1569 1570 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1571 1572 return vm_page_count_severe(); 1573 } 1574 1575 /* 1576 * brelse: 1577 * 1578 * Release a busy buffer and, if requested, free its resources. The 1579 * buffer will be stashed in the appropriate bufqueue[] allowing it 1580 * to be accessed later as a cache entity or reused for other purposes. 1581 */ 1582 void 1583 brelse(struct buf *bp) 1584 { 1585 int qindex; 1586 1587 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1588 bp, bp->b_vp, bp->b_flags); 1589 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1590 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1591 1592 if (BUF_LOCKRECURSED(bp)) { 1593 /* 1594 * Do not process, in particular, do not handle the 1595 * B_INVAL/B_RELBUF and do not release to free list. 1596 */ 1597 BUF_UNLOCK(bp); 1598 return; 1599 } 1600 1601 if (bp->b_flags & B_MANAGED) { 1602 bqrelse(bp); 1603 return; 1604 } 1605 1606 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 1607 BO_LOCK(bp->b_bufobj); 1608 bp->b_vflags &= ~BV_BKGRDERR; 1609 BO_UNLOCK(bp->b_bufobj); 1610 bdirty(bp); 1611 } 1612 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1613 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1614 /* 1615 * Failed write, redirty. Must clear BIO_ERROR to prevent 1616 * pages from being scrapped. If the error is anything 1617 * other than an I/O error (EIO), assume that retrying 1618 * is futile. 1619 */ 1620 bp->b_ioflags &= ~BIO_ERROR; 1621 bdirty(bp); 1622 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1623 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1624 /* 1625 * Either a failed I/O or we were asked to free or not 1626 * cache the buffer. 1627 */ 1628 bp->b_flags |= B_INVAL; 1629 if (!LIST_EMPTY(&bp->b_dep)) 1630 buf_deallocate(bp); 1631 if (bp->b_flags & B_DELWRI) 1632 bdirtysub(); 1633 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1634 if ((bp->b_flags & B_VMIO) == 0) { 1635 if (bp->b_bufsize) 1636 allocbuf(bp, 0); 1637 if (bp->b_vp) 1638 brelvp(bp); 1639 } 1640 } 1641 1642 /* 1643 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1644 * is called with B_DELWRI set, the underlying pages may wind up 1645 * getting freed causing a previous write (bdwrite()) to get 'lost' 1646 * because pages associated with a B_DELWRI bp are marked clean. 1647 * 1648 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1649 * if B_DELWRI is set. 1650 * 1651 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1652 * on pages to return pages to the VM page queues. 1653 */ 1654 if (bp->b_flags & B_DELWRI) 1655 bp->b_flags &= ~B_RELBUF; 1656 else if (buf_vm_page_count_severe()) { 1657 /* 1658 * BKGRDINPROG can only be set with the buf and bufobj 1659 * locks both held. We tolerate a race to clear it here. 1660 */ 1661 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1662 bp->b_flags |= B_RELBUF; 1663 } 1664 1665 /* 1666 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1667 * constituted, not even NFS buffers now. Two flags effect this. If 1668 * B_INVAL, the struct buf is invalidated but the VM object is kept 1669 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1670 * 1671 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1672 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1673 * buffer is also B_INVAL because it hits the re-dirtying code above. 1674 * 1675 * Normally we can do this whether a buffer is B_DELWRI or not. If 1676 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1677 * the commit state and we cannot afford to lose the buffer. If the 1678 * buffer has a background write in progress, we need to keep it 1679 * around to prevent it from being reconstituted and starting a second 1680 * background write. 1681 */ 1682 if ((bp->b_flags & B_VMIO) 1683 && !(bp->b_vp->v_mount != NULL && 1684 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1685 !vn_isdisk(bp->b_vp, NULL) && 1686 (bp->b_flags & B_DELWRI)) 1687 ) { 1688 1689 int i, j, resid; 1690 vm_page_t m; 1691 off_t foff; 1692 vm_pindex_t poff; 1693 vm_object_t obj; 1694 1695 obj = bp->b_bufobj->bo_object; 1696 1697 /* 1698 * Get the base offset and length of the buffer. Note that 1699 * in the VMIO case if the buffer block size is not 1700 * page-aligned then b_data pointer may not be page-aligned. 1701 * But our b_pages[] array *IS* page aligned. 1702 * 1703 * block sizes less then DEV_BSIZE (usually 512) are not 1704 * supported due to the page granularity bits (m->valid, 1705 * m->dirty, etc...). 1706 * 1707 * See man buf(9) for more information 1708 */ 1709 resid = bp->b_bufsize; 1710 foff = bp->b_offset; 1711 for (i = 0; i < bp->b_npages; i++) { 1712 int had_bogus = 0; 1713 1714 m = bp->b_pages[i]; 1715 1716 /* 1717 * If we hit a bogus page, fixup *all* the bogus pages 1718 * now. 1719 */ 1720 if (m == bogus_page) { 1721 poff = OFF_TO_IDX(bp->b_offset); 1722 had_bogus = 1; 1723 1724 VM_OBJECT_RLOCK(obj); 1725 for (j = i; j < bp->b_npages; j++) { 1726 vm_page_t mtmp; 1727 mtmp = bp->b_pages[j]; 1728 if (mtmp == bogus_page) { 1729 mtmp = vm_page_lookup(obj, poff + j); 1730 if (!mtmp) { 1731 panic("brelse: page missing\n"); 1732 } 1733 bp->b_pages[j] = mtmp; 1734 } 1735 } 1736 VM_OBJECT_RUNLOCK(obj); 1737 1738 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1739 BUF_CHECK_MAPPED(bp); 1740 pmap_qenter( 1741 trunc_page((vm_offset_t)bp->b_data), 1742 bp->b_pages, bp->b_npages); 1743 } 1744 m = bp->b_pages[i]; 1745 } 1746 if ((bp->b_flags & B_NOCACHE) || 1747 (bp->b_ioflags & BIO_ERROR && 1748 bp->b_iocmd == BIO_READ)) { 1749 int poffset = foff & PAGE_MASK; 1750 int presid = resid > (PAGE_SIZE - poffset) ? 1751 (PAGE_SIZE - poffset) : resid; 1752 1753 KASSERT(presid >= 0, ("brelse: extra page")); 1754 VM_OBJECT_WLOCK(obj); 1755 while (vm_page_xbusied(m)) { 1756 vm_page_lock(m); 1757 VM_OBJECT_WUNLOCK(obj); 1758 vm_page_busy_sleep(m, "mbncsh"); 1759 VM_OBJECT_WLOCK(obj); 1760 } 1761 if (pmap_page_wired_mappings(m) == 0) 1762 vm_page_set_invalid(m, poffset, presid); 1763 VM_OBJECT_WUNLOCK(obj); 1764 if (had_bogus) 1765 printf("avoided corruption bug in bogus_page/brelse code\n"); 1766 } 1767 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1768 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1769 } 1770 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1771 vfs_vmio_release(bp); 1772 1773 } else if (bp->b_flags & B_VMIO) { 1774 1775 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1776 vfs_vmio_release(bp); 1777 } 1778 1779 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1780 if (bp->b_bufsize != 0) 1781 allocbuf(bp, 0); 1782 if (bp->b_vp != NULL) 1783 brelvp(bp); 1784 } 1785 1786 /* 1787 * If the buffer has junk contents signal it and eventually 1788 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1789 * doesn't find it. 1790 */ 1791 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1792 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1793 bp->b_flags |= B_INVAL; 1794 if (bp->b_flags & B_INVAL) { 1795 if (bp->b_flags & B_DELWRI) 1796 bundirty(bp); 1797 if (bp->b_vp) 1798 brelvp(bp); 1799 } 1800 1801 /* buffers with no memory */ 1802 if (bp->b_bufsize == 0) { 1803 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1804 if (bp->b_vflags & BV_BKGRDINPROG) 1805 panic("losing buffer 1"); 1806 if (bp->b_kvasize) 1807 qindex = QUEUE_EMPTYKVA; 1808 else 1809 qindex = QUEUE_EMPTY; 1810 bp->b_flags |= B_AGE; 1811 /* buffers with junk contents */ 1812 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1813 (bp->b_ioflags & BIO_ERROR)) { 1814 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1815 if (bp->b_vflags & BV_BKGRDINPROG) 1816 panic("losing buffer 2"); 1817 qindex = QUEUE_CLEAN; 1818 bp->b_flags |= B_AGE; 1819 /* remaining buffers */ 1820 } else if (bp->b_flags & B_DELWRI) 1821 qindex = QUEUE_DIRTY; 1822 else 1823 qindex = QUEUE_CLEAN; 1824 1825 binsfree(bp, qindex); 1826 1827 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1828 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1829 panic("brelse: not dirty"); 1830 /* unlock */ 1831 BUF_UNLOCK(bp); 1832 } 1833 1834 /* 1835 * Release a buffer back to the appropriate queue but do not try to free 1836 * it. The buffer is expected to be used again soon. 1837 * 1838 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1839 * biodone() to requeue an async I/O on completion. It is also used when 1840 * known good buffers need to be requeued but we think we may need the data 1841 * again soon. 1842 * 1843 * XXX we should be able to leave the B_RELBUF hint set on completion. 1844 */ 1845 void 1846 bqrelse(struct buf *bp) 1847 { 1848 int qindex; 1849 1850 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1851 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1852 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1853 1854 if (BUF_LOCKRECURSED(bp)) { 1855 /* do not release to free list */ 1856 BUF_UNLOCK(bp); 1857 return; 1858 } 1859 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1860 1861 if (bp->b_flags & B_MANAGED) { 1862 if (bp->b_flags & B_REMFREE) 1863 bremfreef(bp); 1864 goto out; 1865 } 1866 1867 /* buffers with stale but valid contents */ 1868 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 1869 BV_BKGRDERR)) == BV_BKGRDERR) { 1870 BO_LOCK(bp->b_bufobj); 1871 bp->b_vflags &= ~BV_BKGRDERR; 1872 BO_UNLOCK(bp->b_bufobj); 1873 qindex = QUEUE_DIRTY; 1874 } else { 1875 if ((bp->b_flags & B_DELWRI) == 0 && 1876 (bp->b_xflags & BX_VNDIRTY)) 1877 panic("bqrelse: not dirty"); 1878 /* 1879 * BKGRDINPROG can only be set with the buf and bufobj 1880 * locks both held. We tolerate a race to clear it here. 1881 */ 1882 if (buf_vm_page_count_severe() && 1883 (bp->b_vflags & BV_BKGRDINPROG) == 0) { 1884 /* 1885 * We are too low on memory, we have to try to free 1886 * the buffer (most importantly: the wired pages 1887 * making up its backing store) *now*. 1888 */ 1889 brelse(bp); 1890 return; 1891 } 1892 qindex = QUEUE_CLEAN; 1893 } 1894 binsfree(bp, qindex); 1895 1896 out: 1897 /* unlock */ 1898 BUF_UNLOCK(bp); 1899 } 1900 1901 /* Give pages used by the bp back to the VM system (where possible) */ 1902 static void 1903 vfs_vmio_release(struct buf *bp) 1904 { 1905 vm_object_t obj; 1906 vm_page_t m; 1907 int i; 1908 1909 if ((bp->b_flags & B_UNMAPPED) == 0) { 1910 BUF_CHECK_MAPPED(bp); 1911 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1912 } else 1913 BUF_CHECK_UNMAPPED(bp); 1914 obj = bp->b_bufobj->bo_object; 1915 if (obj != NULL) 1916 VM_OBJECT_WLOCK(obj); 1917 for (i = 0; i < bp->b_npages; i++) { 1918 m = bp->b_pages[i]; 1919 bp->b_pages[i] = NULL; 1920 /* 1921 * In order to keep page LRU ordering consistent, put 1922 * everything on the inactive queue. 1923 */ 1924 vm_page_lock(m); 1925 vm_page_unwire(m, PQ_INACTIVE); 1926 1927 /* 1928 * Might as well free the page if we can and it has 1929 * no valid data. We also free the page if the 1930 * buffer was used for direct I/O 1931 */ 1932 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1933 if (m->wire_count == 0 && !vm_page_busied(m)) 1934 vm_page_free(m); 1935 } else if (bp->b_flags & B_DIRECT) 1936 vm_page_try_to_free(m); 1937 else if (buf_vm_page_count_severe()) 1938 vm_page_try_to_cache(m); 1939 vm_page_unlock(m); 1940 } 1941 if (obj != NULL) 1942 VM_OBJECT_WUNLOCK(obj); 1943 1944 if (bp->b_bufsize) { 1945 bufspacewakeup(); 1946 bp->b_bufsize = 0; 1947 } 1948 bp->b_npages = 0; 1949 bp->b_flags &= ~B_VMIO; 1950 if (bp->b_vp) 1951 brelvp(bp); 1952 } 1953 1954 /* 1955 * Check to see if a block at a particular lbn is available for a clustered 1956 * write. 1957 */ 1958 static int 1959 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1960 { 1961 struct buf *bpa; 1962 int match; 1963 1964 match = 0; 1965 1966 /* If the buf isn't in core skip it */ 1967 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1968 return (0); 1969 1970 /* If the buf is busy we don't want to wait for it */ 1971 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1972 return (0); 1973 1974 /* Only cluster with valid clusterable delayed write buffers */ 1975 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1976 (B_DELWRI | B_CLUSTEROK)) 1977 goto done; 1978 1979 if (bpa->b_bufsize != size) 1980 goto done; 1981 1982 /* 1983 * Check to see if it is in the expected place on disk and that the 1984 * block has been mapped. 1985 */ 1986 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1987 match = 1; 1988 done: 1989 BUF_UNLOCK(bpa); 1990 return (match); 1991 } 1992 1993 /* 1994 * vfs_bio_awrite: 1995 * 1996 * Implement clustered async writes for clearing out B_DELWRI buffers. 1997 * This is much better then the old way of writing only one buffer at 1998 * a time. Note that we may not be presented with the buffers in the 1999 * correct order, so we search for the cluster in both directions. 2000 */ 2001 int 2002 vfs_bio_awrite(struct buf *bp) 2003 { 2004 struct bufobj *bo; 2005 int i; 2006 int j; 2007 daddr_t lblkno = bp->b_lblkno; 2008 struct vnode *vp = bp->b_vp; 2009 int ncl; 2010 int nwritten; 2011 int size; 2012 int maxcl; 2013 int gbflags; 2014 2015 bo = &vp->v_bufobj; 2016 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 2017 /* 2018 * right now we support clustered writing only to regular files. If 2019 * we find a clusterable block we could be in the middle of a cluster 2020 * rather then at the beginning. 2021 */ 2022 if ((vp->v_type == VREG) && 2023 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 2024 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 2025 2026 size = vp->v_mount->mnt_stat.f_iosize; 2027 maxcl = MAXPHYS / size; 2028 2029 BO_RLOCK(bo); 2030 for (i = 1; i < maxcl; i++) 2031 if (vfs_bio_clcheck(vp, size, lblkno + i, 2032 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 2033 break; 2034 2035 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 2036 if (vfs_bio_clcheck(vp, size, lblkno - j, 2037 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 2038 break; 2039 BO_RUNLOCK(bo); 2040 --j; 2041 ncl = i + j; 2042 /* 2043 * this is a possible cluster write 2044 */ 2045 if (ncl != 1) { 2046 BUF_UNLOCK(bp); 2047 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 2048 gbflags); 2049 return (nwritten); 2050 } 2051 } 2052 bremfree(bp); 2053 bp->b_flags |= B_ASYNC; 2054 /* 2055 * default (old) behavior, writing out only one block 2056 * 2057 * XXX returns b_bufsize instead of b_bcount for nwritten? 2058 */ 2059 nwritten = bp->b_bufsize; 2060 (void) bwrite(bp); 2061 2062 return (nwritten); 2063 } 2064 2065 static void 2066 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 2067 { 2068 2069 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2070 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 2071 if ((gbflags & GB_UNMAPPED) == 0) { 2072 bp->b_kvabase = (caddr_t)addr; 2073 } else if ((gbflags & GB_KVAALLOC) != 0) { 2074 KASSERT((gbflags & GB_UNMAPPED) != 0, 2075 ("GB_KVAALLOC without GB_UNMAPPED")); 2076 bp->b_kvaalloc = (caddr_t)addr; 2077 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2078 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2079 } 2080 bp->b_kvasize = maxsize; 2081 } 2082 2083 /* 2084 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 2085 * needed. 2086 */ 2087 static int 2088 allocbufkva(struct buf *bp, int maxsize, int gbflags) 2089 { 2090 vm_offset_t addr; 2091 2092 bfreekva(bp); 2093 addr = 0; 2094 2095 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) { 2096 /* 2097 * Buffer map is too fragmented. Request the caller 2098 * to defragment the map. 2099 */ 2100 atomic_add_int(&bufdefragcnt, 1); 2101 return (1); 2102 } 2103 setbufkva(bp, addr, maxsize, gbflags); 2104 atomic_add_long(&bufspace, bp->b_kvasize); 2105 return (0); 2106 } 2107 2108 /* 2109 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2110 * locked vnode is supplied. 2111 */ 2112 static void 2113 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2114 int defrag) 2115 { 2116 struct thread *td; 2117 char *waitmsg; 2118 int error, fl, flags, norunbuf; 2119 2120 mtx_assert(&bqclean, MA_OWNED); 2121 2122 if (defrag) { 2123 flags = VFS_BIO_NEED_BUFSPACE; 2124 waitmsg = "nbufkv"; 2125 } else if (bufspace >= hibufspace) { 2126 waitmsg = "nbufbs"; 2127 flags = VFS_BIO_NEED_BUFSPACE; 2128 } else { 2129 waitmsg = "newbuf"; 2130 flags = VFS_BIO_NEED_ANY; 2131 } 2132 atomic_set_int(&needsbuffer, flags); 2133 mtx_unlock(&bqclean); 2134 2135 bd_speedup(); /* heeeelp */ 2136 if ((gbflags & GB_NOWAIT_BD) != 0) 2137 return; 2138 2139 td = curthread; 2140 rw_wlock(&nblock); 2141 while ((needsbuffer & flags) != 0) { 2142 if (vp != NULL && vp->v_type != VCHR && 2143 (td->td_pflags & TDP_BUFNEED) == 0) { 2144 rw_wunlock(&nblock); 2145 /* 2146 * getblk() is called with a vnode locked, and 2147 * some majority of the dirty buffers may as 2148 * well belong to the vnode. Flushing the 2149 * buffers there would make a progress that 2150 * cannot be achieved by the buf_daemon, that 2151 * cannot lock the vnode. 2152 */ 2153 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2154 (td->td_pflags & TDP_NORUNNINGBUF); 2155 2156 /* 2157 * Play bufdaemon. The getnewbuf() function 2158 * may be called while the thread owns lock 2159 * for another dirty buffer for the same 2160 * vnode, which makes it impossible to use 2161 * VOP_FSYNC() there, due to the buffer lock 2162 * recursion. 2163 */ 2164 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2165 fl = buf_flush(vp, flushbufqtarget); 2166 td->td_pflags &= norunbuf; 2167 rw_wlock(&nblock); 2168 if (fl != 0) 2169 continue; 2170 if ((needsbuffer & flags) == 0) 2171 break; 2172 } 2173 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 2174 (PRIBIO + 4) | slpflag, waitmsg, slptimeo); 2175 if (error != 0) 2176 break; 2177 } 2178 rw_wunlock(&nblock); 2179 } 2180 2181 static void 2182 getnewbuf_reuse_bp(struct buf *bp, int qindex) 2183 { 2184 2185 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2186 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2187 bp->b_kvasize, bp->b_bufsize, qindex); 2188 mtx_assert(&bqclean, MA_NOTOWNED); 2189 2190 /* 2191 * Note: we no longer distinguish between VMIO and non-VMIO 2192 * buffers. 2193 */ 2194 KASSERT((bp->b_flags & B_DELWRI) == 0, 2195 ("delwri buffer %p found in queue %d", bp, qindex)); 2196 2197 if (qindex == QUEUE_CLEAN) { 2198 if (bp->b_flags & B_VMIO) { 2199 bp->b_flags &= ~B_ASYNC; 2200 vfs_vmio_release(bp); 2201 } 2202 if (bp->b_vp != NULL) 2203 brelvp(bp); 2204 } 2205 2206 /* 2207 * Get the rest of the buffer freed up. b_kva* is still valid 2208 * after this operation. 2209 */ 2210 2211 if (bp->b_rcred != NOCRED) { 2212 crfree(bp->b_rcred); 2213 bp->b_rcred = NOCRED; 2214 } 2215 if (bp->b_wcred != NOCRED) { 2216 crfree(bp->b_wcred); 2217 bp->b_wcred = NOCRED; 2218 } 2219 if (!LIST_EMPTY(&bp->b_dep)) 2220 buf_deallocate(bp); 2221 if (bp->b_vflags & BV_BKGRDINPROG) 2222 panic("losing buffer 3"); 2223 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2224 bp, bp->b_vp, qindex)); 2225 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2226 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2227 2228 if (bp->b_bufsize) 2229 allocbuf(bp, 0); 2230 2231 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2232 bp->b_ioflags = 0; 2233 bp->b_xflags = 0; 2234 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2235 ("buf %p still counted as free?", bp)); 2236 bp->b_vflags = 0; 2237 bp->b_vp = NULL; 2238 bp->b_blkno = bp->b_lblkno = 0; 2239 bp->b_offset = NOOFFSET; 2240 bp->b_iodone = 0; 2241 bp->b_error = 0; 2242 bp->b_resid = 0; 2243 bp->b_bcount = 0; 2244 bp->b_npages = 0; 2245 bp->b_dirtyoff = bp->b_dirtyend = 0; 2246 bp->b_bufobj = NULL; 2247 bp->b_pin_count = 0; 2248 bp->b_fsprivate1 = NULL; 2249 bp->b_fsprivate2 = NULL; 2250 bp->b_fsprivate3 = NULL; 2251 2252 LIST_INIT(&bp->b_dep); 2253 } 2254 2255 static int flushingbufs; 2256 2257 static struct buf * 2258 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2259 { 2260 struct buf *bp, *nbp; 2261 int nqindex, qindex, pass; 2262 2263 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2264 2265 pass = 1; 2266 restart: 2267 atomic_add_int(&getnewbufrestarts, 1); 2268 2269 /* 2270 * Setup for scan. If we do not have enough free buffers, 2271 * we setup a degenerate case that immediately fails. Note 2272 * that if we are specially marked process, we are allowed to 2273 * dip into our reserves. 2274 * 2275 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2276 * for the allocation of the mapped buffer. For unmapped, the 2277 * easiest is to start with EMPTY outright. 2278 * 2279 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2280 * However, there are a number of cases (defragging, reusing, ...) 2281 * where we cannot backup. 2282 */ 2283 nbp = NULL; 2284 mtx_lock(&bqclean); 2285 if (!defrag && unmapped) { 2286 nqindex = QUEUE_EMPTY; 2287 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2288 } 2289 if (nbp == NULL) { 2290 nqindex = QUEUE_EMPTYKVA; 2291 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2292 } 2293 2294 /* 2295 * If no EMPTYKVA buffers and we are either defragging or 2296 * reusing, locate a CLEAN buffer to free or reuse. If 2297 * bufspace useage is low skip this step so we can allocate a 2298 * new buffer. 2299 */ 2300 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2301 nqindex = QUEUE_CLEAN; 2302 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2303 } 2304 2305 /* 2306 * If we could not find or were not allowed to reuse a CLEAN 2307 * buffer, check to see if it is ok to use an EMPTY buffer. 2308 * We can only use an EMPTY buffer if allocating its KVA would 2309 * not otherwise run us out of buffer space. No KVA is needed 2310 * for the unmapped allocation. 2311 */ 2312 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2313 metadata)) { 2314 nqindex = QUEUE_EMPTY; 2315 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2316 } 2317 2318 /* 2319 * All available buffers might be clean, retry ignoring the 2320 * lobufspace as the last resort. 2321 */ 2322 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2323 nqindex = QUEUE_CLEAN; 2324 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2325 } 2326 2327 /* 2328 * Run scan, possibly freeing data and/or kva mappings on the fly 2329 * depending. 2330 */ 2331 while ((bp = nbp) != NULL) { 2332 qindex = nqindex; 2333 2334 /* 2335 * Calculate next bp (we can only use it if we do not 2336 * block or do other fancy things). 2337 */ 2338 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2339 switch (qindex) { 2340 case QUEUE_EMPTY: 2341 nqindex = QUEUE_EMPTYKVA; 2342 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2343 if (nbp != NULL) 2344 break; 2345 /* FALLTHROUGH */ 2346 case QUEUE_EMPTYKVA: 2347 nqindex = QUEUE_CLEAN; 2348 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2349 if (nbp != NULL) 2350 break; 2351 /* FALLTHROUGH */ 2352 case QUEUE_CLEAN: 2353 if (metadata && pass == 1) { 2354 pass = 2; 2355 nqindex = QUEUE_EMPTY; 2356 nbp = TAILQ_FIRST( 2357 &bufqueues[QUEUE_EMPTY]); 2358 } 2359 /* 2360 * nbp is NULL. 2361 */ 2362 break; 2363 } 2364 } 2365 /* 2366 * If we are defragging then we need a buffer with 2367 * b_kvasize != 0. XXX this situation should no longer 2368 * occur, if defrag is non-zero the buffer's b_kvasize 2369 * should also be non-zero at this point. XXX 2370 */ 2371 if (defrag && bp->b_kvasize == 0) { 2372 printf("Warning: defrag empty buffer %p\n", bp); 2373 continue; 2374 } 2375 2376 /* 2377 * Start freeing the bp. This is somewhat involved. nbp 2378 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2379 */ 2380 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2381 continue; 2382 /* 2383 * BKGRDINPROG can only be set with the buf and bufobj 2384 * locks both held. We tolerate a race to clear it here. 2385 */ 2386 if (bp->b_vflags & BV_BKGRDINPROG) { 2387 BUF_UNLOCK(bp); 2388 continue; 2389 } 2390 2391 /* 2392 * Requeue the background write buffer with error. 2393 */ 2394 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 2395 bremfreel(bp); 2396 mtx_unlock(&bqclean); 2397 bqrelse(bp); 2398 continue; 2399 } 2400 2401 KASSERT(bp->b_qindex == qindex, 2402 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2403 2404 bremfreel(bp); 2405 mtx_unlock(&bqclean); 2406 /* 2407 * NOTE: nbp is now entirely invalid. We can only restart 2408 * the scan from this point on. 2409 */ 2410 2411 getnewbuf_reuse_bp(bp, qindex); 2412 mtx_assert(&bqclean, MA_NOTOWNED); 2413 2414 /* 2415 * If we are defragging then free the buffer. 2416 */ 2417 if (defrag) { 2418 bp->b_flags |= B_INVAL; 2419 bfreekva(bp); 2420 brelse(bp); 2421 defrag = 0; 2422 goto restart; 2423 } 2424 2425 /* 2426 * Notify any waiters for the buffer lock about 2427 * identity change by freeing the buffer. 2428 */ 2429 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2430 bp->b_flags |= B_INVAL; 2431 bfreekva(bp); 2432 brelse(bp); 2433 goto restart; 2434 } 2435 2436 if (metadata) 2437 break; 2438 2439 /* 2440 * If we are overcomitted then recover the buffer and its 2441 * KVM space. This occurs in rare situations when multiple 2442 * processes are blocked in getnewbuf() or allocbuf(). 2443 */ 2444 if (bufspace >= hibufspace) 2445 flushingbufs = 1; 2446 if (flushingbufs && bp->b_kvasize != 0) { 2447 bp->b_flags |= B_INVAL; 2448 bfreekva(bp); 2449 brelse(bp); 2450 goto restart; 2451 } 2452 if (bufspace < lobufspace) 2453 flushingbufs = 0; 2454 break; 2455 } 2456 return (bp); 2457 } 2458 2459 /* 2460 * getnewbuf: 2461 * 2462 * Find and initialize a new buffer header, freeing up existing buffers 2463 * in the bufqueues as necessary. The new buffer is returned locked. 2464 * 2465 * Important: B_INVAL is not set. If the caller wishes to throw the 2466 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2467 * 2468 * We block if: 2469 * We have insufficient buffer headers 2470 * We have insufficient buffer space 2471 * buffer_arena is too fragmented ( space reservation fails ) 2472 * If we have to flush dirty buffers ( but we try to avoid this ) 2473 */ 2474 static struct buf * 2475 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2476 int gbflags) 2477 { 2478 struct buf *bp; 2479 int defrag, metadata; 2480 2481 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2482 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2483 if (!unmapped_buf_allowed) 2484 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2485 2486 defrag = 0; 2487 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2488 vp->v_type == VCHR) 2489 metadata = 1; 2490 else 2491 metadata = 0; 2492 /* 2493 * We can't afford to block since we might be holding a vnode lock, 2494 * which may prevent system daemons from running. We deal with 2495 * low-memory situations by proactively returning memory and running 2496 * async I/O rather then sync I/O. 2497 */ 2498 atomic_add_int(&getnewbufcalls, 1); 2499 atomic_subtract_int(&getnewbufrestarts, 1); 2500 restart: 2501 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2502 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2503 if (bp != NULL) 2504 defrag = 0; 2505 2506 /* 2507 * If we exhausted our list, sleep as appropriate. We may have to 2508 * wakeup various daemons and write out some dirty buffers. 2509 * 2510 * Generally we are sleeping due to insufficient buffer space. 2511 */ 2512 if (bp == NULL) { 2513 mtx_assert(&bqclean, MA_OWNED); 2514 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2515 mtx_assert(&bqclean, MA_NOTOWNED); 2516 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2517 mtx_assert(&bqclean, MA_NOTOWNED); 2518 2519 bfreekva(bp); 2520 bp->b_flags |= B_UNMAPPED; 2521 bp->b_kvabase = bp->b_data = unmapped_buf; 2522 bp->b_kvasize = maxsize; 2523 atomic_add_long(&bufspace, bp->b_kvasize); 2524 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2525 atomic_add_int(&bufreusecnt, 1); 2526 } else { 2527 mtx_assert(&bqclean, MA_NOTOWNED); 2528 2529 /* 2530 * We finally have a valid bp. We aren't quite out of the 2531 * woods, we still have to reserve kva space. In order 2532 * to keep fragmentation sane we only allocate kva in 2533 * BKVASIZE chunks. 2534 */ 2535 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2536 2537 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2538 B_KVAALLOC)) == B_UNMAPPED) { 2539 if (allocbufkva(bp, maxsize, gbflags)) { 2540 defrag = 1; 2541 bp->b_flags |= B_INVAL; 2542 brelse(bp); 2543 goto restart; 2544 } 2545 atomic_add_int(&bufreusecnt, 1); 2546 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2547 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2548 /* 2549 * If the reused buffer has KVA allocated, 2550 * reassign b_kvaalloc to b_kvabase. 2551 */ 2552 bp->b_kvabase = bp->b_kvaalloc; 2553 bp->b_flags &= ~B_KVAALLOC; 2554 atomic_subtract_long(&unmapped_bufspace, 2555 bp->b_kvasize); 2556 atomic_add_int(&bufreusecnt, 1); 2557 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2558 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2559 GB_KVAALLOC)) { 2560 /* 2561 * The case of reused buffer already have KVA 2562 * mapped, but the request is for unmapped 2563 * buffer with KVA allocated. 2564 */ 2565 bp->b_kvaalloc = bp->b_kvabase; 2566 bp->b_data = bp->b_kvabase = unmapped_buf; 2567 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2568 atomic_add_long(&unmapped_bufspace, 2569 bp->b_kvasize); 2570 atomic_add_int(&bufreusecnt, 1); 2571 } 2572 if ((gbflags & GB_UNMAPPED) == 0) { 2573 bp->b_saveaddr = bp->b_kvabase; 2574 bp->b_data = bp->b_saveaddr; 2575 bp->b_flags &= ~B_UNMAPPED; 2576 BUF_CHECK_MAPPED(bp); 2577 } 2578 } 2579 return (bp); 2580 } 2581 2582 /* 2583 * buf_daemon: 2584 * 2585 * buffer flushing daemon. Buffers are normally flushed by the 2586 * update daemon but if it cannot keep up this process starts to 2587 * take the load in an attempt to prevent getnewbuf() from blocking. 2588 */ 2589 2590 static struct kproc_desc buf_kp = { 2591 "bufdaemon", 2592 buf_daemon, 2593 &bufdaemonproc 2594 }; 2595 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2596 2597 static int 2598 buf_flush(struct vnode *vp, int target) 2599 { 2600 int flushed; 2601 2602 flushed = flushbufqueues(vp, target, 0); 2603 if (flushed == 0) { 2604 /* 2605 * Could not find any buffers without rollback 2606 * dependencies, so just write the first one 2607 * in the hopes of eventually making progress. 2608 */ 2609 if (vp != NULL && target > 2) 2610 target /= 2; 2611 flushbufqueues(vp, target, 1); 2612 } 2613 return (flushed); 2614 } 2615 2616 static void 2617 buf_daemon() 2618 { 2619 int lodirty; 2620 2621 /* 2622 * This process needs to be suspended prior to shutdown sync. 2623 */ 2624 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2625 SHUTDOWN_PRI_LAST); 2626 2627 /* 2628 * This process is allowed to take the buffer cache to the limit 2629 */ 2630 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2631 mtx_lock(&bdlock); 2632 for (;;) { 2633 bd_request = 0; 2634 mtx_unlock(&bdlock); 2635 2636 kproc_suspend_check(bufdaemonproc); 2637 lodirty = lodirtybuffers; 2638 if (bd_speedupreq) { 2639 lodirty = numdirtybuffers / 2; 2640 bd_speedupreq = 0; 2641 } 2642 /* 2643 * Do the flush. Limit the amount of in-transit I/O we 2644 * allow to build up, otherwise we would completely saturate 2645 * the I/O system. 2646 */ 2647 while (numdirtybuffers > lodirty) { 2648 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0) 2649 break; 2650 kern_yield(PRI_USER); 2651 } 2652 2653 /* 2654 * Only clear bd_request if we have reached our low water 2655 * mark. The buf_daemon normally waits 1 second and 2656 * then incrementally flushes any dirty buffers that have 2657 * built up, within reason. 2658 * 2659 * If we were unable to hit our low water mark and couldn't 2660 * find any flushable buffers, we sleep for a short period 2661 * to avoid endless loops on unlockable buffers. 2662 */ 2663 mtx_lock(&bdlock); 2664 if (numdirtybuffers <= lodirtybuffers) { 2665 /* 2666 * We reached our low water mark, reset the 2667 * request and sleep until we are needed again. 2668 * The sleep is just so the suspend code works. 2669 */ 2670 bd_request = 0; 2671 /* 2672 * Do an extra wakeup in case dirty threshold 2673 * changed via sysctl and the explicit transition 2674 * out of shortfall was missed. 2675 */ 2676 bdirtywakeup(); 2677 if (runningbufspace <= lorunningspace) 2678 runningwakeup(); 2679 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2680 } else { 2681 /* 2682 * We couldn't find any flushable dirty buffers but 2683 * still have too many dirty buffers, we 2684 * have to sleep and try again. (rare) 2685 */ 2686 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2687 } 2688 } 2689 } 2690 2691 /* 2692 * flushbufqueues: 2693 * 2694 * Try to flush a buffer in the dirty queue. We must be careful to 2695 * free up B_INVAL buffers instead of write them, which NFS is 2696 * particularly sensitive to. 2697 */ 2698 static int flushwithdeps = 0; 2699 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2700 0, "Number of buffers flushed with dependecies that require rollbacks"); 2701 2702 static int 2703 flushbufqueues(struct vnode *lvp, int target, int flushdeps) 2704 { 2705 struct buf *sentinel; 2706 struct vnode *vp; 2707 struct mount *mp; 2708 struct buf *bp; 2709 int hasdeps; 2710 int flushed; 2711 int queue; 2712 int error; 2713 bool unlock; 2714 2715 flushed = 0; 2716 queue = QUEUE_DIRTY; 2717 bp = NULL; 2718 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2719 sentinel->b_qindex = QUEUE_SENTINEL; 2720 mtx_lock(&bqdirty); 2721 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2722 mtx_unlock(&bqdirty); 2723 while (flushed != target) { 2724 maybe_yield(); 2725 mtx_lock(&bqdirty); 2726 bp = TAILQ_NEXT(sentinel, b_freelist); 2727 if (bp != NULL) { 2728 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2729 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2730 b_freelist); 2731 } else { 2732 mtx_unlock(&bqdirty); 2733 break; 2734 } 2735 /* 2736 * Skip sentinels inserted by other invocations of the 2737 * flushbufqueues(), taking care to not reorder them. 2738 * 2739 * Only flush the buffers that belong to the 2740 * vnode locked by the curthread. 2741 */ 2742 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 2743 bp->b_vp != lvp)) { 2744 mtx_unlock(&bqdirty); 2745 continue; 2746 } 2747 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 2748 mtx_unlock(&bqdirty); 2749 if (error != 0) 2750 continue; 2751 if (bp->b_pin_count > 0) { 2752 BUF_UNLOCK(bp); 2753 continue; 2754 } 2755 /* 2756 * BKGRDINPROG can only be set with the buf and bufobj 2757 * locks both held. We tolerate a race to clear it here. 2758 */ 2759 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2760 (bp->b_flags & B_DELWRI) == 0) { 2761 BUF_UNLOCK(bp); 2762 continue; 2763 } 2764 if (bp->b_flags & B_INVAL) { 2765 bremfreef(bp); 2766 brelse(bp); 2767 flushed++; 2768 continue; 2769 } 2770 2771 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2772 if (flushdeps == 0) { 2773 BUF_UNLOCK(bp); 2774 continue; 2775 } 2776 hasdeps = 1; 2777 } else 2778 hasdeps = 0; 2779 /* 2780 * We must hold the lock on a vnode before writing 2781 * one of its buffers. Otherwise we may confuse, or 2782 * in the case of a snapshot vnode, deadlock the 2783 * system. 2784 * 2785 * The lock order here is the reverse of the normal 2786 * of vnode followed by buf lock. This is ok because 2787 * the NOWAIT will prevent deadlock. 2788 */ 2789 vp = bp->b_vp; 2790 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2791 BUF_UNLOCK(bp); 2792 continue; 2793 } 2794 if (lvp == NULL) { 2795 unlock = true; 2796 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 2797 } else { 2798 ASSERT_VOP_LOCKED(vp, "getbuf"); 2799 unlock = false; 2800 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 2801 vn_lock(vp, LK_TRYUPGRADE); 2802 } 2803 if (error == 0) { 2804 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2805 bp, bp->b_vp, bp->b_flags); 2806 if (curproc == bufdaemonproc) { 2807 vfs_bio_awrite(bp); 2808 } else { 2809 bremfree(bp); 2810 bwrite(bp); 2811 notbufdflushes++; 2812 } 2813 vn_finished_write(mp); 2814 if (unlock) 2815 VOP_UNLOCK(vp, 0); 2816 flushwithdeps += hasdeps; 2817 flushed++; 2818 2819 /* 2820 * Sleeping on runningbufspace while holding 2821 * vnode lock leads to deadlock. 2822 */ 2823 if (curproc == bufdaemonproc && 2824 runningbufspace > hirunningspace) 2825 waitrunningbufspace(); 2826 continue; 2827 } 2828 vn_finished_write(mp); 2829 BUF_UNLOCK(bp); 2830 } 2831 mtx_lock(&bqdirty); 2832 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2833 mtx_unlock(&bqdirty); 2834 free(sentinel, M_TEMP); 2835 return (flushed); 2836 } 2837 2838 /* 2839 * Check to see if a block is currently memory resident. 2840 */ 2841 struct buf * 2842 incore(struct bufobj *bo, daddr_t blkno) 2843 { 2844 struct buf *bp; 2845 2846 BO_RLOCK(bo); 2847 bp = gbincore(bo, blkno); 2848 BO_RUNLOCK(bo); 2849 return (bp); 2850 } 2851 2852 /* 2853 * Returns true if no I/O is needed to access the 2854 * associated VM object. This is like incore except 2855 * it also hunts around in the VM system for the data. 2856 */ 2857 2858 static int 2859 inmem(struct vnode * vp, daddr_t blkno) 2860 { 2861 vm_object_t obj; 2862 vm_offset_t toff, tinc, size; 2863 vm_page_t m; 2864 vm_ooffset_t off; 2865 2866 ASSERT_VOP_LOCKED(vp, "inmem"); 2867 2868 if (incore(&vp->v_bufobj, blkno)) 2869 return 1; 2870 if (vp->v_mount == NULL) 2871 return 0; 2872 obj = vp->v_object; 2873 if (obj == NULL) 2874 return (0); 2875 2876 size = PAGE_SIZE; 2877 if (size > vp->v_mount->mnt_stat.f_iosize) 2878 size = vp->v_mount->mnt_stat.f_iosize; 2879 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2880 2881 VM_OBJECT_RLOCK(obj); 2882 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2883 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2884 if (!m) 2885 goto notinmem; 2886 tinc = size; 2887 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2888 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2889 if (vm_page_is_valid(m, 2890 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2891 goto notinmem; 2892 } 2893 VM_OBJECT_RUNLOCK(obj); 2894 return 1; 2895 2896 notinmem: 2897 VM_OBJECT_RUNLOCK(obj); 2898 return (0); 2899 } 2900 2901 /* 2902 * Set the dirty range for a buffer based on the status of the dirty 2903 * bits in the pages comprising the buffer. The range is limited 2904 * to the size of the buffer. 2905 * 2906 * Tell the VM system that the pages associated with this buffer 2907 * are clean. This is used for delayed writes where the data is 2908 * going to go to disk eventually without additional VM intevention. 2909 * 2910 * Note that while we only really need to clean through to b_bcount, we 2911 * just go ahead and clean through to b_bufsize. 2912 */ 2913 static void 2914 vfs_clean_pages_dirty_buf(struct buf *bp) 2915 { 2916 vm_ooffset_t foff, noff, eoff; 2917 vm_page_t m; 2918 int i; 2919 2920 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2921 return; 2922 2923 foff = bp->b_offset; 2924 KASSERT(bp->b_offset != NOOFFSET, 2925 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2926 2927 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2928 vfs_drain_busy_pages(bp); 2929 vfs_setdirty_locked_object(bp); 2930 for (i = 0; i < bp->b_npages; i++) { 2931 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2932 eoff = noff; 2933 if (eoff > bp->b_offset + bp->b_bufsize) 2934 eoff = bp->b_offset + bp->b_bufsize; 2935 m = bp->b_pages[i]; 2936 vfs_page_set_validclean(bp, foff, m); 2937 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2938 foff = noff; 2939 } 2940 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2941 } 2942 2943 static void 2944 vfs_setdirty_locked_object(struct buf *bp) 2945 { 2946 vm_object_t object; 2947 int i; 2948 2949 object = bp->b_bufobj->bo_object; 2950 VM_OBJECT_ASSERT_WLOCKED(object); 2951 2952 /* 2953 * We qualify the scan for modified pages on whether the 2954 * object has been flushed yet. 2955 */ 2956 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2957 vm_offset_t boffset; 2958 vm_offset_t eoffset; 2959 2960 /* 2961 * test the pages to see if they have been modified directly 2962 * by users through the VM system. 2963 */ 2964 for (i = 0; i < bp->b_npages; i++) 2965 vm_page_test_dirty(bp->b_pages[i]); 2966 2967 /* 2968 * Calculate the encompassing dirty range, boffset and eoffset, 2969 * (eoffset - boffset) bytes. 2970 */ 2971 2972 for (i = 0; i < bp->b_npages; i++) { 2973 if (bp->b_pages[i]->dirty) 2974 break; 2975 } 2976 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2977 2978 for (i = bp->b_npages - 1; i >= 0; --i) { 2979 if (bp->b_pages[i]->dirty) { 2980 break; 2981 } 2982 } 2983 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2984 2985 /* 2986 * Fit it to the buffer. 2987 */ 2988 2989 if (eoffset > bp->b_bcount) 2990 eoffset = bp->b_bcount; 2991 2992 /* 2993 * If we have a good dirty range, merge with the existing 2994 * dirty range. 2995 */ 2996 2997 if (boffset < eoffset) { 2998 if (bp->b_dirtyoff > boffset) 2999 bp->b_dirtyoff = boffset; 3000 if (bp->b_dirtyend < eoffset) 3001 bp->b_dirtyend = eoffset; 3002 } 3003 } 3004 } 3005 3006 /* 3007 * Allocate the KVA mapping for an existing buffer. It handles the 3008 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 3009 * KVA which is not mapped (B_KVAALLOC). 3010 */ 3011 static void 3012 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3013 { 3014 struct buf *scratch_bp; 3015 int bsize, maxsize, need_mapping, need_kva; 3016 off_t offset; 3017 3018 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 3019 (gbflags & GB_UNMAPPED) == 0; 3020 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 3021 (gbflags & GB_KVAALLOC) != 0; 3022 if (!need_mapping && !need_kva) 3023 return; 3024 3025 BUF_CHECK_UNMAPPED(bp); 3026 3027 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 3028 /* 3029 * Buffer is not mapped, but the KVA was already 3030 * reserved at the time of the instantiation. Use the 3031 * allocated space. 3032 */ 3033 bp->b_flags &= ~B_KVAALLOC; 3034 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 3035 bp->b_kvabase = bp->b_kvaalloc; 3036 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 3037 goto has_addr; 3038 } 3039 3040 /* 3041 * Calculate the amount of the address space we would reserve 3042 * if the buffer was mapped. 3043 */ 3044 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3045 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3046 offset = blkno * bsize; 3047 maxsize = size + (offset & PAGE_MASK); 3048 maxsize = imax(maxsize, bsize); 3049 3050 mapping_loop: 3051 if (allocbufkva(bp, maxsize, gbflags)) { 3052 /* 3053 * Request defragmentation. getnewbuf() returns us the 3054 * allocated space by the scratch buffer KVA. 3055 */ 3056 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 3057 (GB_UNMAPPED | GB_KVAALLOC)); 3058 if (scratch_bp == NULL) { 3059 if ((gbflags & GB_NOWAIT_BD) != 0) { 3060 /* 3061 * XXXKIB: defragmentation cannot 3062 * succeed, not sure what else to do. 3063 */ 3064 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 3065 } 3066 atomic_add_int(&mappingrestarts, 1); 3067 goto mapping_loop; 3068 } 3069 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 3070 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 3071 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 3072 scratch_bp->b_kvasize, gbflags); 3073 3074 /* Get rid of the scratch buffer. */ 3075 scratch_bp->b_kvasize = 0; 3076 scratch_bp->b_flags |= B_INVAL; 3077 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 3078 brelse(scratch_bp); 3079 } 3080 if (!need_mapping) 3081 return; 3082 3083 has_addr: 3084 bp->b_saveaddr = bp->b_kvabase; 3085 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 3086 bp->b_flags &= ~B_UNMAPPED; 3087 BUF_CHECK_MAPPED(bp); 3088 bpmap_qenter(bp); 3089 } 3090 3091 /* 3092 * getblk: 3093 * 3094 * Get a block given a specified block and offset into a file/device. 3095 * The buffers B_DONE bit will be cleared on return, making it almost 3096 * ready for an I/O initiation. B_INVAL may or may not be set on 3097 * return. The caller should clear B_INVAL prior to initiating a 3098 * READ. 3099 * 3100 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3101 * an existing buffer. 3102 * 3103 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3104 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3105 * and then cleared based on the backing VM. If the previous buffer is 3106 * non-0-sized but invalid, B_CACHE will be cleared. 3107 * 3108 * If getblk() must create a new buffer, the new buffer is returned with 3109 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3110 * case it is returned with B_INVAL clear and B_CACHE set based on the 3111 * backing VM. 3112 * 3113 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3114 * B_CACHE bit is clear. 3115 * 3116 * What this means, basically, is that the caller should use B_CACHE to 3117 * determine whether the buffer is fully valid or not and should clear 3118 * B_INVAL prior to issuing a read. If the caller intends to validate 3119 * the buffer by loading its data area with something, the caller needs 3120 * to clear B_INVAL. If the caller does this without issuing an I/O, 3121 * the caller should set B_CACHE ( as an optimization ), else the caller 3122 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3123 * a write attempt or if it was a successfull read. If the caller 3124 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3125 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3126 */ 3127 struct buf * 3128 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3129 int flags) 3130 { 3131 struct buf *bp; 3132 struct bufobj *bo; 3133 int bsize, error, maxsize, vmio; 3134 off_t offset; 3135 3136 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3137 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3138 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3139 ASSERT_VOP_LOCKED(vp, "getblk"); 3140 if (size > MAXBCACHEBUF) 3141 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size, 3142 MAXBCACHEBUF); 3143 if (!unmapped_buf_allowed) 3144 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3145 3146 bo = &vp->v_bufobj; 3147 loop: 3148 BO_RLOCK(bo); 3149 bp = gbincore(bo, blkno); 3150 if (bp != NULL) { 3151 int lockflags; 3152 /* 3153 * Buffer is in-core. If the buffer is not busy nor managed, 3154 * it must be on a queue. 3155 */ 3156 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3157 3158 if (flags & GB_LOCK_NOWAIT) 3159 lockflags |= LK_NOWAIT; 3160 3161 error = BUF_TIMELOCK(bp, lockflags, 3162 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3163 3164 /* 3165 * If we slept and got the lock we have to restart in case 3166 * the buffer changed identities. 3167 */ 3168 if (error == ENOLCK) 3169 goto loop; 3170 /* We timed out or were interrupted. */ 3171 else if (error) 3172 return (NULL); 3173 /* If recursed, assume caller knows the rules. */ 3174 else if (BUF_LOCKRECURSED(bp)) 3175 goto end; 3176 3177 /* 3178 * The buffer is locked. B_CACHE is cleared if the buffer is 3179 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3180 * and for a VMIO buffer B_CACHE is adjusted according to the 3181 * backing VM cache. 3182 */ 3183 if (bp->b_flags & B_INVAL) 3184 bp->b_flags &= ~B_CACHE; 3185 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3186 bp->b_flags |= B_CACHE; 3187 if (bp->b_flags & B_MANAGED) 3188 MPASS(bp->b_qindex == QUEUE_NONE); 3189 else 3190 bremfree(bp); 3191 3192 /* 3193 * check for size inconsistencies for non-VMIO case. 3194 */ 3195 if (bp->b_bcount != size) { 3196 if ((bp->b_flags & B_VMIO) == 0 || 3197 (size > bp->b_kvasize)) { 3198 if (bp->b_flags & B_DELWRI) { 3199 /* 3200 * If buffer is pinned and caller does 3201 * not want sleep waiting for it to be 3202 * unpinned, bail out 3203 * */ 3204 if (bp->b_pin_count > 0) { 3205 if (flags & GB_LOCK_NOWAIT) { 3206 bqrelse(bp); 3207 return (NULL); 3208 } else { 3209 bunpin_wait(bp); 3210 } 3211 } 3212 bp->b_flags |= B_NOCACHE; 3213 bwrite(bp); 3214 } else { 3215 if (LIST_EMPTY(&bp->b_dep)) { 3216 bp->b_flags |= B_RELBUF; 3217 brelse(bp); 3218 } else { 3219 bp->b_flags |= B_NOCACHE; 3220 bwrite(bp); 3221 } 3222 } 3223 goto loop; 3224 } 3225 } 3226 3227 /* 3228 * Handle the case of unmapped buffer which should 3229 * become mapped, or the buffer for which KVA 3230 * reservation is requested. 3231 */ 3232 bp_unmapped_get_kva(bp, blkno, size, flags); 3233 3234 /* 3235 * If the size is inconsistant in the VMIO case, we can resize 3236 * the buffer. This might lead to B_CACHE getting set or 3237 * cleared. If the size has not changed, B_CACHE remains 3238 * unchanged from its previous state. 3239 */ 3240 if (bp->b_bcount != size) 3241 allocbuf(bp, size); 3242 3243 KASSERT(bp->b_offset != NOOFFSET, 3244 ("getblk: no buffer offset")); 3245 3246 /* 3247 * A buffer with B_DELWRI set and B_CACHE clear must 3248 * be committed before we can return the buffer in 3249 * order to prevent the caller from issuing a read 3250 * ( due to B_CACHE not being set ) and overwriting 3251 * it. 3252 * 3253 * Most callers, including NFS and FFS, need this to 3254 * operate properly either because they assume they 3255 * can issue a read if B_CACHE is not set, or because 3256 * ( for example ) an uncached B_DELWRI might loop due 3257 * to softupdates re-dirtying the buffer. In the latter 3258 * case, B_CACHE is set after the first write completes, 3259 * preventing further loops. 3260 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3261 * above while extending the buffer, we cannot allow the 3262 * buffer to remain with B_CACHE set after the write 3263 * completes or it will represent a corrupt state. To 3264 * deal with this we set B_NOCACHE to scrap the buffer 3265 * after the write. 3266 * 3267 * We might be able to do something fancy, like setting 3268 * B_CACHE in bwrite() except if B_DELWRI is already set, 3269 * so the below call doesn't set B_CACHE, but that gets real 3270 * confusing. This is much easier. 3271 */ 3272 3273 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3274 bp->b_flags |= B_NOCACHE; 3275 bwrite(bp); 3276 goto loop; 3277 } 3278 bp->b_flags &= ~B_DONE; 3279 } else { 3280 /* 3281 * Buffer is not in-core, create new buffer. The buffer 3282 * returned by getnewbuf() is locked. Note that the returned 3283 * buffer is also considered valid (not marked B_INVAL). 3284 */ 3285 BO_RUNLOCK(bo); 3286 /* 3287 * If the user does not want us to create the buffer, bail out 3288 * here. 3289 */ 3290 if (flags & GB_NOCREAT) 3291 return NULL; 3292 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3293 return NULL; 3294 3295 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3296 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3297 offset = blkno * bsize; 3298 vmio = vp->v_object != NULL; 3299 if (vmio) { 3300 maxsize = size + (offset & PAGE_MASK); 3301 } else { 3302 maxsize = size; 3303 /* Do not allow non-VMIO notmapped buffers. */ 3304 flags &= ~GB_UNMAPPED; 3305 } 3306 maxsize = imax(maxsize, bsize); 3307 3308 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3309 if (bp == NULL) { 3310 if (slpflag || slptimeo) 3311 return NULL; 3312 goto loop; 3313 } 3314 3315 /* 3316 * This code is used to make sure that a buffer is not 3317 * created while the getnewbuf routine is blocked. 3318 * This can be a problem whether the vnode is locked or not. 3319 * If the buffer is created out from under us, we have to 3320 * throw away the one we just created. 3321 * 3322 * Note: this must occur before we associate the buffer 3323 * with the vp especially considering limitations in 3324 * the splay tree implementation when dealing with duplicate 3325 * lblkno's. 3326 */ 3327 BO_LOCK(bo); 3328 if (gbincore(bo, blkno)) { 3329 BO_UNLOCK(bo); 3330 bp->b_flags |= B_INVAL; 3331 brelse(bp); 3332 goto loop; 3333 } 3334 3335 /* 3336 * Insert the buffer into the hash, so that it can 3337 * be found by incore. 3338 */ 3339 bp->b_blkno = bp->b_lblkno = blkno; 3340 bp->b_offset = offset; 3341 bgetvp(vp, bp); 3342 BO_UNLOCK(bo); 3343 3344 /* 3345 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3346 * buffer size starts out as 0, B_CACHE will be set by 3347 * allocbuf() for the VMIO case prior to it testing the 3348 * backing store for validity. 3349 */ 3350 3351 if (vmio) { 3352 bp->b_flags |= B_VMIO; 3353 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3354 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3355 bp, vp->v_object, bp->b_bufobj->bo_object)); 3356 } else { 3357 bp->b_flags &= ~B_VMIO; 3358 KASSERT(bp->b_bufobj->bo_object == NULL, 3359 ("ARGH! has b_bufobj->bo_object %p %p\n", 3360 bp, bp->b_bufobj->bo_object)); 3361 BUF_CHECK_MAPPED(bp); 3362 } 3363 3364 allocbuf(bp, size); 3365 bp->b_flags &= ~B_DONE; 3366 } 3367 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3368 BUF_ASSERT_HELD(bp); 3369 end: 3370 KASSERT(bp->b_bufobj == bo, 3371 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3372 return (bp); 3373 } 3374 3375 /* 3376 * Get an empty, disassociated buffer of given size. The buffer is initially 3377 * set to B_INVAL. 3378 */ 3379 struct buf * 3380 geteblk(int size, int flags) 3381 { 3382 struct buf *bp; 3383 int maxsize; 3384 3385 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3386 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3387 if ((flags & GB_NOWAIT_BD) && 3388 (curthread->td_pflags & TDP_BUFNEED) != 0) 3389 return (NULL); 3390 } 3391 allocbuf(bp, size); 3392 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3393 BUF_ASSERT_HELD(bp); 3394 return (bp); 3395 } 3396 3397 3398 /* 3399 * This code constitutes the buffer memory from either anonymous system 3400 * memory (in the case of non-VMIO operations) or from an associated 3401 * VM object (in the case of VMIO operations). This code is able to 3402 * resize a buffer up or down. 3403 * 3404 * Note that this code is tricky, and has many complications to resolve 3405 * deadlock or inconsistant data situations. Tread lightly!!! 3406 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3407 * the caller. Calling this code willy nilly can result in the loss of data. 3408 * 3409 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3410 * B_CACHE for the non-VMIO case. 3411 */ 3412 3413 int 3414 allocbuf(struct buf *bp, int size) 3415 { 3416 int newbsize, mbsize; 3417 int i; 3418 3419 BUF_ASSERT_HELD(bp); 3420 3421 if (bp->b_kvasize < size) 3422 panic("allocbuf: buffer too small"); 3423 3424 if ((bp->b_flags & B_VMIO) == 0) { 3425 caddr_t origbuf; 3426 int origbufsize; 3427 /* 3428 * Just get anonymous memory from the kernel. Don't 3429 * mess with B_CACHE. 3430 */ 3431 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3432 if (bp->b_flags & B_MALLOC) 3433 newbsize = mbsize; 3434 else 3435 newbsize = round_page(size); 3436 3437 if (newbsize < bp->b_bufsize) { 3438 /* 3439 * malloced buffers are not shrunk 3440 */ 3441 if (bp->b_flags & B_MALLOC) { 3442 if (newbsize) { 3443 bp->b_bcount = size; 3444 } else { 3445 free(bp->b_data, M_BIOBUF); 3446 if (bp->b_bufsize) { 3447 atomic_subtract_long( 3448 &bufmallocspace, 3449 bp->b_bufsize); 3450 bufspacewakeup(); 3451 bp->b_bufsize = 0; 3452 } 3453 bp->b_saveaddr = bp->b_kvabase; 3454 bp->b_data = bp->b_saveaddr; 3455 bp->b_bcount = 0; 3456 bp->b_flags &= ~B_MALLOC; 3457 } 3458 return 1; 3459 } 3460 vm_hold_free_pages(bp, newbsize); 3461 } else if (newbsize > bp->b_bufsize) { 3462 /* 3463 * We only use malloced memory on the first allocation. 3464 * and revert to page-allocated memory when the buffer 3465 * grows. 3466 */ 3467 /* 3468 * There is a potential smp race here that could lead 3469 * to bufmallocspace slightly passing the max. It 3470 * is probably extremely rare and not worth worrying 3471 * over. 3472 */ 3473 if ( (bufmallocspace < maxbufmallocspace) && 3474 (bp->b_bufsize == 0) && 3475 (mbsize <= PAGE_SIZE/2)) { 3476 3477 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3478 bp->b_bufsize = mbsize; 3479 bp->b_bcount = size; 3480 bp->b_flags |= B_MALLOC; 3481 atomic_add_long(&bufmallocspace, mbsize); 3482 return 1; 3483 } 3484 origbuf = NULL; 3485 origbufsize = 0; 3486 /* 3487 * If the buffer is growing on its other-than-first allocation, 3488 * then we revert to the page-allocation scheme. 3489 */ 3490 if (bp->b_flags & B_MALLOC) { 3491 origbuf = bp->b_data; 3492 origbufsize = bp->b_bufsize; 3493 bp->b_data = bp->b_kvabase; 3494 if (bp->b_bufsize) { 3495 atomic_subtract_long(&bufmallocspace, 3496 bp->b_bufsize); 3497 bufspacewakeup(); 3498 bp->b_bufsize = 0; 3499 } 3500 bp->b_flags &= ~B_MALLOC; 3501 newbsize = round_page(newbsize); 3502 } 3503 vm_hold_load_pages( 3504 bp, 3505 (vm_offset_t) bp->b_data + bp->b_bufsize, 3506 (vm_offset_t) bp->b_data + newbsize); 3507 if (origbuf) { 3508 bcopy(origbuf, bp->b_data, origbufsize); 3509 free(origbuf, M_BIOBUF); 3510 } 3511 } 3512 } else { 3513 int desiredpages; 3514 3515 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3516 desiredpages = (size == 0) ? 0 : 3517 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3518 3519 if (bp->b_flags & B_MALLOC) 3520 panic("allocbuf: VMIO buffer can't be malloced"); 3521 /* 3522 * Set B_CACHE initially if buffer is 0 length or will become 3523 * 0-length. 3524 */ 3525 if (size == 0 || bp->b_bufsize == 0) 3526 bp->b_flags |= B_CACHE; 3527 3528 if (newbsize < bp->b_bufsize) { 3529 /* 3530 * DEV_BSIZE aligned new buffer size is less then the 3531 * DEV_BSIZE aligned existing buffer size. Figure out 3532 * if we have to remove any pages. 3533 */ 3534 if (desiredpages < bp->b_npages) { 3535 vm_page_t m; 3536 3537 if ((bp->b_flags & B_UNMAPPED) == 0) { 3538 BUF_CHECK_MAPPED(bp); 3539 pmap_qremove((vm_offset_t)trunc_page( 3540 (vm_offset_t)bp->b_data) + 3541 (desiredpages << PAGE_SHIFT), 3542 (bp->b_npages - desiredpages)); 3543 } else 3544 BUF_CHECK_UNMAPPED(bp); 3545 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3546 for (i = desiredpages; i < bp->b_npages; i++) { 3547 /* 3548 * the page is not freed here -- it 3549 * is the responsibility of 3550 * vnode_pager_setsize 3551 */ 3552 m = bp->b_pages[i]; 3553 KASSERT(m != bogus_page, 3554 ("allocbuf: bogus page found")); 3555 while (vm_page_sleep_if_busy(m, 3556 "biodep")) 3557 continue; 3558 3559 bp->b_pages[i] = NULL; 3560 vm_page_lock(m); 3561 vm_page_unwire(m, PQ_INACTIVE); 3562 vm_page_unlock(m); 3563 } 3564 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3565 bp->b_npages = desiredpages; 3566 } 3567 } else if (size > bp->b_bcount) { 3568 /* 3569 * We are growing the buffer, possibly in a 3570 * byte-granular fashion. 3571 */ 3572 vm_object_t obj; 3573 vm_offset_t toff; 3574 vm_offset_t tinc; 3575 3576 /* 3577 * Step 1, bring in the VM pages from the object, 3578 * allocating them if necessary. We must clear 3579 * B_CACHE if these pages are not valid for the 3580 * range covered by the buffer. 3581 */ 3582 3583 obj = bp->b_bufobj->bo_object; 3584 3585 VM_OBJECT_WLOCK(obj); 3586 while (bp->b_npages < desiredpages) { 3587 vm_page_t m; 3588 3589 /* 3590 * We must allocate system pages since blocking 3591 * here could interfere with paging I/O, no 3592 * matter which process we are. 3593 * 3594 * Only exclusive busy can be tested here. 3595 * Blocking on shared busy might lead to 3596 * deadlocks once allocbuf() is called after 3597 * pages are vfs_busy_pages(). 3598 */ 3599 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3600 bp->b_npages, VM_ALLOC_NOBUSY | 3601 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3602 VM_ALLOC_IGN_SBUSY | 3603 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3604 if (m->valid == 0) 3605 bp->b_flags &= ~B_CACHE; 3606 bp->b_pages[bp->b_npages] = m; 3607 ++bp->b_npages; 3608 } 3609 3610 /* 3611 * Step 2. We've loaded the pages into the buffer, 3612 * we have to figure out if we can still have B_CACHE 3613 * set. Note that B_CACHE is set according to the 3614 * byte-granular range ( bcount and size ), new the 3615 * aligned range ( newbsize ). 3616 * 3617 * The VM test is against m->valid, which is DEV_BSIZE 3618 * aligned. Needless to say, the validity of the data 3619 * needs to also be DEV_BSIZE aligned. Note that this 3620 * fails with NFS if the server or some other client 3621 * extends the file's EOF. If our buffer is resized, 3622 * B_CACHE may remain set! XXX 3623 */ 3624 3625 toff = bp->b_bcount; 3626 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3627 3628 while ((bp->b_flags & B_CACHE) && toff < size) { 3629 vm_pindex_t pi; 3630 3631 if (tinc > (size - toff)) 3632 tinc = size - toff; 3633 3634 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3635 PAGE_SHIFT; 3636 3637 vfs_buf_test_cache( 3638 bp, 3639 bp->b_offset, 3640 toff, 3641 tinc, 3642 bp->b_pages[pi] 3643 ); 3644 toff += tinc; 3645 tinc = PAGE_SIZE; 3646 } 3647 VM_OBJECT_WUNLOCK(obj); 3648 3649 /* 3650 * Step 3, fixup the KVM pmap. 3651 */ 3652 if ((bp->b_flags & B_UNMAPPED) == 0) 3653 bpmap_qenter(bp); 3654 else 3655 BUF_CHECK_UNMAPPED(bp); 3656 } 3657 } 3658 if (newbsize < bp->b_bufsize) 3659 bufspacewakeup(); 3660 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3661 bp->b_bcount = size; /* requested buffer size */ 3662 return 1; 3663 } 3664 3665 extern int inflight_transient_maps; 3666 3667 void 3668 biodone(struct bio *bp) 3669 { 3670 struct mtx *mtxp; 3671 void (*done)(struct bio *); 3672 vm_offset_t start, end; 3673 3674 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3675 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3676 bp->bio_flags |= BIO_UNMAPPED; 3677 start = trunc_page((vm_offset_t)bp->bio_data); 3678 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3679 bp->bio_data = unmapped_buf; 3680 pmap_qremove(start, OFF_TO_IDX(end - start)); 3681 vmem_free(transient_arena, start, end - start); 3682 atomic_add_int(&inflight_transient_maps, -1); 3683 } 3684 done = bp->bio_done; 3685 if (done == NULL) { 3686 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3687 mtx_lock(mtxp); 3688 bp->bio_flags |= BIO_DONE; 3689 wakeup(bp); 3690 mtx_unlock(mtxp); 3691 } else { 3692 bp->bio_flags |= BIO_DONE; 3693 done(bp); 3694 } 3695 } 3696 3697 /* 3698 * Wait for a BIO to finish. 3699 */ 3700 int 3701 biowait(struct bio *bp, const char *wchan) 3702 { 3703 struct mtx *mtxp; 3704 3705 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3706 mtx_lock(mtxp); 3707 while ((bp->bio_flags & BIO_DONE) == 0) 3708 msleep(bp, mtxp, PRIBIO, wchan, 0); 3709 mtx_unlock(mtxp); 3710 if (bp->bio_error != 0) 3711 return (bp->bio_error); 3712 if (!(bp->bio_flags & BIO_ERROR)) 3713 return (0); 3714 return (EIO); 3715 } 3716 3717 void 3718 biofinish(struct bio *bp, struct devstat *stat, int error) 3719 { 3720 3721 if (error) { 3722 bp->bio_error = error; 3723 bp->bio_flags |= BIO_ERROR; 3724 } 3725 if (stat != NULL) 3726 devstat_end_transaction_bio(stat, bp); 3727 biodone(bp); 3728 } 3729 3730 /* 3731 * bufwait: 3732 * 3733 * Wait for buffer I/O completion, returning error status. The buffer 3734 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3735 * error and cleared. 3736 */ 3737 int 3738 bufwait(struct buf *bp) 3739 { 3740 if (bp->b_iocmd == BIO_READ) 3741 bwait(bp, PRIBIO, "biord"); 3742 else 3743 bwait(bp, PRIBIO, "biowr"); 3744 if (bp->b_flags & B_EINTR) { 3745 bp->b_flags &= ~B_EINTR; 3746 return (EINTR); 3747 } 3748 if (bp->b_ioflags & BIO_ERROR) { 3749 return (bp->b_error ? bp->b_error : EIO); 3750 } else { 3751 return (0); 3752 } 3753 } 3754 3755 /* 3756 * Call back function from struct bio back up to struct buf. 3757 */ 3758 static void 3759 bufdonebio(struct bio *bip) 3760 { 3761 struct buf *bp; 3762 3763 bp = bip->bio_caller2; 3764 bp->b_resid = bip->bio_resid; 3765 bp->b_ioflags = bip->bio_flags; 3766 bp->b_error = bip->bio_error; 3767 if (bp->b_error) 3768 bp->b_ioflags |= BIO_ERROR; 3769 bufdone(bp); 3770 g_destroy_bio(bip); 3771 } 3772 3773 void 3774 dev_strategy(struct cdev *dev, struct buf *bp) 3775 { 3776 struct cdevsw *csw; 3777 int ref; 3778 3779 KASSERT(dev->si_refcount > 0, 3780 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3781 devtoname(dev), dev)); 3782 3783 csw = dev_refthread(dev, &ref); 3784 dev_strategy_csw(dev, csw, bp); 3785 dev_relthread(dev, ref); 3786 } 3787 3788 void 3789 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3790 { 3791 struct bio *bip; 3792 3793 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3794 ("b_iocmd botch")); 3795 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3796 dev->si_threadcount > 0, 3797 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3798 dev)); 3799 if (csw == NULL) { 3800 bp->b_error = ENXIO; 3801 bp->b_ioflags = BIO_ERROR; 3802 bufdone(bp); 3803 return; 3804 } 3805 for (;;) { 3806 bip = g_new_bio(); 3807 if (bip != NULL) 3808 break; 3809 /* Try again later */ 3810 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3811 } 3812 bip->bio_cmd = bp->b_iocmd; 3813 bip->bio_offset = bp->b_iooffset; 3814 bip->bio_length = bp->b_bcount; 3815 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3816 bdata2bio(bp, bip); 3817 bip->bio_done = bufdonebio; 3818 bip->bio_caller2 = bp; 3819 bip->bio_dev = dev; 3820 (*csw->d_strategy)(bip); 3821 } 3822 3823 /* 3824 * bufdone: 3825 * 3826 * Finish I/O on a buffer, optionally calling a completion function. 3827 * This is usually called from an interrupt so process blocking is 3828 * not allowed. 3829 * 3830 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3831 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3832 * assuming B_INVAL is clear. 3833 * 3834 * For the VMIO case, we set B_CACHE if the op was a read and no 3835 * read error occured, or if the op was a write. B_CACHE is never 3836 * set if the buffer is invalid or otherwise uncacheable. 3837 * 3838 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3839 * initiator to leave B_INVAL set to brelse the buffer out of existance 3840 * in the biodone routine. 3841 */ 3842 void 3843 bufdone(struct buf *bp) 3844 { 3845 struct bufobj *dropobj; 3846 void (*biodone)(struct buf *); 3847 3848 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3849 dropobj = NULL; 3850 3851 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3852 BUF_ASSERT_HELD(bp); 3853 3854 runningbufwakeup(bp); 3855 if (bp->b_iocmd == BIO_WRITE) 3856 dropobj = bp->b_bufobj; 3857 /* call optional completion function if requested */ 3858 if (bp->b_iodone != NULL) { 3859 biodone = bp->b_iodone; 3860 bp->b_iodone = NULL; 3861 (*biodone) (bp); 3862 if (dropobj) 3863 bufobj_wdrop(dropobj); 3864 return; 3865 } 3866 3867 bufdone_finish(bp); 3868 3869 if (dropobj) 3870 bufobj_wdrop(dropobj); 3871 } 3872 3873 void 3874 bufdone_finish(struct buf *bp) 3875 { 3876 BUF_ASSERT_HELD(bp); 3877 3878 if (!LIST_EMPTY(&bp->b_dep)) 3879 buf_complete(bp); 3880 3881 if (bp->b_flags & B_VMIO) { 3882 vm_ooffset_t foff; 3883 vm_page_t m; 3884 vm_object_t obj; 3885 struct vnode *vp; 3886 int bogus, i, iosize; 3887 3888 obj = bp->b_bufobj->bo_object; 3889 KASSERT(obj->paging_in_progress >= bp->b_npages, 3890 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3891 obj->paging_in_progress, bp->b_npages)); 3892 3893 vp = bp->b_vp; 3894 KASSERT(vp->v_holdcnt > 0, 3895 ("biodone_finish: vnode %p has zero hold count", vp)); 3896 KASSERT(vp->v_object != NULL, 3897 ("biodone_finish: vnode %p has no vm_object", vp)); 3898 3899 foff = bp->b_offset; 3900 KASSERT(bp->b_offset != NOOFFSET, 3901 ("biodone_finish: bp %p has no buffer offset", bp)); 3902 3903 /* 3904 * Set B_CACHE if the op was a normal read and no error 3905 * occured. B_CACHE is set for writes in the b*write() 3906 * routines. 3907 */ 3908 iosize = bp->b_bcount - bp->b_resid; 3909 if (bp->b_iocmd == BIO_READ && 3910 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3911 !(bp->b_ioflags & BIO_ERROR)) { 3912 bp->b_flags |= B_CACHE; 3913 } 3914 bogus = 0; 3915 VM_OBJECT_WLOCK(obj); 3916 for (i = 0; i < bp->b_npages; i++) { 3917 int bogusflag = 0; 3918 int resid; 3919 3920 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3921 if (resid > iosize) 3922 resid = iosize; 3923 3924 /* 3925 * cleanup bogus pages, restoring the originals 3926 */ 3927 m = bp->b_pages[i]; 3928 if (m == bogus_page) { 3929 bogus = bogusflag = 1; 3930 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3931 if (m == NULL) 3932 panic("biodone: page disappeared!"); 3933 bp->b_pages[i] = m; 3934 } 3935 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3936 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3937 (intmax_t)foff, (uintmax_t)m->pindex)); 3938 3939 /* 3940 * In the write case, the valid and clean bits are 3941 * already changed correctly ( see bdwrite() ), so we 3942 * only need to do this here in the read case. 3943 */ 3944 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3945 KASSERT((m->dirty & vm_page_bits(foff & 3946 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3947 " page %p has unexpected dirty bits", m)); 3948 vfs_page_set_valid(bp, foff, m); 3949 } 3950 3951 vm_page_sunbusy(m); 3952 vm_object_pip_subtract(obj, 1); 3953 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3954 iosize -= resid; 3955 } 3956 vm_object_pip_wakeupn(obj, 0); 3957 VM_OBJECT_WUNLOCK(obj); 3958 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3959 BUF_CHECK_MAPPED(bp); 3960 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3961 bp->b_pages, bp->b_npages); 3962 } 3963 } 3964 3965 /* 3966 * For asynchronous completions, release the buffer now. The brelse 3967 * will do a wakeup there if necessary - so no need to do a wakeup 3968 * here in the async case. The sync case always needs to do a wakeup. 3969 */ 3970 3971 if (bp->b_flags & B_ASYNC) { 3972 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3973 brelse(bp); 3974 else 3975 bqrelse(bp); 3976 } else 3977 bdone(bp); 3978 } 3979 3980 /* 3981 * This routine is called in lieu of iodone in the case of 3982 * incomplete I/O. This keeps the busy status for pages 3983 * consistant. 3984 */ 3985 void 3986 vfs_unbusy_pages(struct buf *bp) 3987 { 3988 int i; 3989 vm_object_t obj; 3990 vm_page_t m; 3991 3992 runningbufwakeup(bp); 3993 if (!(bp->b_flags & B_VMIO)) 3994 return; 3995 3996 obj = bp->b_bufobj->bo_object; 3997 VM_OBJECT_WLOCK(obj); 3998 for (i = 0; i < bp->b_npages; i++) { 3999 m = bp->b_pages[i]; 4000 if (m == bogus_page) { 4001 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4002 if (!m) 4003 panic("vfs_unbusy_pages: page missing\n"); 4004 bp->b_pages[i] = m; 4005 if ((bp->b_flags & B_UNMAPPED) == 0) { 4006 BUF_CHECK_MAPPED(bp); 4007 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4008 bp->b_pages, bp->b_npages); 4009 } else 4010 BUF_CHECK_UNMAPPED(bp); 4011 } 4012 vm_object_pip_subtract(obj, 1); 4013 vm_page_sunbusy(m); 4014 } 4015 vm_object_pip_wakeupn(obj, 0); 4016 VM_OBJECT_WUNLOCK(obj); 4017 } 4018 4019 /* 4020 * vfs_page_set_valid: 4021 * 4022 * Set the valid bits in a page based on the supplied offset. The 4023 * range is restricted to the buffer's size. 4024 * 4025 * This routine is typically called after a read completes. 4026 */ 4027 static void 4028 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4029 { 4030 vm_ooffset_t eoff; 4031 4032 /* 4033 * Compute the end offset, eoff, such that [off, eoff) does not span a 4034 * page boundary and eoff is not greater than the end of the buffer. 4035 * The end of the buffer, in this case, is our file EOF, not the 4036 * allocation size of the buffer. 4037 */ 4038 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4039 if (eoff > bp->b_offset + bp->b_bcount) 4040 eoff = bp->b_offset + bp->b_bcount; 4041 4042 /* 4043 * Set valid range. This is typically the entire buffer and thus the 4044 * entire page. 4045 */ 4046 if (eoff > off) 4047 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4048 } 4049 4050 /* 4051 * vfs_page_set_validclean: 4052 * 4053 * Set the valid bits and clear the dirty bits in a page based on the 4054 * supplied offset. The range is restricted to the buffer's size. 4055 */ 4056 static void 4057 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4058 { 4059 vm_ooffset_t soff, eoff; 4060 4061 /* 4062 * Start and end offsets in buffer. eoff - soff may not cross a 4063 * page boundry or cross the end of the buffer. The end of the 4064 * buffer, in this case, is our file EOF, not the allocation size 4065 * of the buffer. 4066 */ 4067 soff = off; 4068 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4069 if (eoff > bp->b_offset + bp->b_bcount) 4070 eoff = bp->b_offset + bp->b_bcount; 4071 4072 /* 4073 * Set valid range. This is typically the entire buffer and thus the 4074 * entire page. 4075 */ 4076 if (eoff > soff) { 4077 vm_page_set_validclean( 4078 m, 4079 (vm_offset_t) (soff & PAGE_MASK), 4080 (vm_offset_t) (eoff - soff) 4081 ); 4082 } 4083 } 4084 4085 /* 4086 * Ensure that all buffer pages are not exclusive busied. If any page is 4087 * exclusive busy, drain it. 4088 */ 4089 void 4090 vfs_drain_busy_pages(struct buf *bp) 4091 { 4092 vm_page_t m; 4093 int i, last_busied; 4094 4095 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4096 last_busied = 0; 4097 for (i = 0; i < bp->b_npages; i++) { 4098 m = bp->b_pages[i]; 4099 if (vm_page_xbusied(m)) { 4100 for (; last_busied < i; last_busied++) 4101 vm_page_sbusy(bp->b_pages[last_busied]); 4102 while (vm_page_xbusied(m)) { 4103 vm_page_lock(m); 4104 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4105 vm_page_busy_sleep(m, "vbpage"); 4106 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4107 } 4108 } 4109 } 4110 for (i = 0; i < last_busied; i++) 4111 vm_page_sunbusy(bp->b_pages[i]); 4112 } 4113 4114 /* 4115 * This routine is called before a device strategy routine. 4116 * It is used to tell the VM system that paging I/O is in 4117 * progress, and treat the pages associated with the buffer 4118 * almost as being exclusive busy. Also the object paging_in_progress 4119 * flag is handled to make sure that the object doesn't become 4120 * inconsistant. 4121 * 4122 * Since I/O has not been initiated yet, certain buffer flags 4123 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4124 * and should be ignored. 4125 */ 4126 void 4127 vfs_busy_pages(struct buf *bp, int clear_modify) 4128 { 4129 int i, bogus; 4130 vm_object_t obj; 4131 vm_ooffset_t foff; 4132 vm_page_t m; 4133 4134 if (!(bp->b_flags & B_VMIO)) 4135 return; 4136 4137 obj = bp->b_bufobj->bo_object; 4138 foff = bp->b_offset; 4139 KASSERT(bp->b_offset != NOOFFSET, 4140 ("vfs_busy_pages: no buffer offset")); 4141 VM_OBJECT_WLOCK(obj); 4142 vfs_drain_busy_pages(bp); 4143 if (bp->b_bufsize != 0) 4144 vfs_setdirty_locked_object(bp); 4145 bogus = 0; 4146 for (i = 0; i < bp->b_npages; i++) { 4147 m = bp->b_pages[i]; 4148 4149 if ((bp->b_flags & B_CLUSTER) == 0) { 4150 vm_object_pip_add(obj, 1); 4151 vm_page_sbusy(m); 4152 } 4153 /* 4154 * When readying a buffer for a read ( i.e 4155 * clear_modify == 0 ), it is important to do 4156 * bogus_page replacement for valid pages in 4157 * partially instantiated buffers. Partially 4158 * instantiated buffers can, in turn, occur when 4159 * reconstituting a buffer from its VM backing store 4160 * base. We only have to do this if B_CACHE is 4161 * clear ( which causes the I/O to occur in the 4162 * first place ). The replacement prevents the read 4163 * I/O from overwriting potentially dirty VM-backed 4164 * pages. XXX bogus page replacement is, uh, bogus. 4165 * It may not work properly with small-block devices. 4166 * We need to find a better way. 4167 */ 4168 if (clear_modify) { 4169 pmap_remove_write(m); 4170 vfs_page_set_validclean(bp, foff, m); 4171 } else if (m->valid == VM_PAGE_BITS_ALL && 4172 (bp->b_flags & B_CACHE) == 0) { 4173 bp->b_pages[i] = bogus_page; 4174 bogus++; 4175 } 4176 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4177 } 4178 VM_OBJECT_WUNLOCK(obj); 4179 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4180 BUF_CHECK_MAPPED(bp); 4181 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4182 bp->b_pages, bp->b_npages); 4183 } 4184 } 4185 4186 /* 4187 * vfs_bio_set_valid: 4188 * 4189 * Set the range within the buffer to valid. The range is 4190 * relative to the beginning of the buffer, b_offset. Note that 4191 * b_offset itself may be offset from the beginning of the first 4192 * page. 4193 */ 4194 void 4195 vfs_bio_set_valid(struct buf *bp, int base, int size) 4196 { 4197 int i, n; 4198 vm_page_t m; 4199 4200 if (!(bp->b_flags & B_VMIO)) 4201 return; 4202 4203 /* 4204 * Fixup base to be relative to beginning of first page. 4205 * Set initial n to be the maximum number of bytes in the 4206 * first page that can be validated. 4207 */ 4208 base += (bp->b_offset & PAGE_MASK); 4209 n = PAGE_SIZE - (base & PAGE_MASK); 4210 4211 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4212 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4213 m = bp->b_pages[i]; 4214 if (n > size) 4215 n = size; 4216 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4217 base += n; 4218 size -= n; 4219 n = PAGE_SIZE; 4220 } 4221 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4222 } 4223 4224 /* 4225 * vfs_bio_clrbuf: 4226 * 4227 * If the specified buffer is a non-VMIO buffer, clear the entire 4228 * buffer. If the specified buffer is a VMIO buffer, clear and 4229 * validate only the previously invalid portions of the buffer. 4230 * This routine essentially fakes an I/O, so we need to clear 4231 * BIO_ERROR and B_INVAL. 4232 * 4233 * Note that while we only theoretically need to clear through b_bcount, 4234 * we go ahead and clear through b_bufsize. 4235 */ 4236 void 4237 vfs_bio_clrbuf(struct buf *bp) 4238 { 4239 int i, j, mask, sa, ea, slide; 4240 4241 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4242 clrbuf(bp); 4243 return; 4244 } 4245 bp->b_flags &= ~B_INVAL; 4246 bp->b_ioflags &= ~BIO_ERROR; 4247 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4248 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4249 (bp->b_offset & PAGE_MASK) == 0) { 4250 if (bp->b_pages[0] == bogus_page) 4251 goto unlock; 4252 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4253 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4254 if ((bp->b_pages[0]->valid & mask) == mask) 4255 goto unlock; 4256 if ((bp->b_pages[0]->valid & mask) == 0) { 4257 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4258 bp->b_pages[0]->valid |= mask; 4259 goto unlock; 4260 } 4261 } 4262 sa = bp->b_offset & PAGE_MASK; 4263 slide = 0; 4264 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4265 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4266 ea = slide & PAGE_MASK; 4267 if (ea == 0) 4268 ea = PAGE_SIZE; 4269 if (bp->b_pages[i] == bogus_page) 4270 continue; 4271 j = sa / DEV_BSIZE; 4272 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4273 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4274 if ((bp->b_pages[i]->valid & mask) == mask) 4275 continue; 4276 if ((bp->b_pages[i]->valid & mask) == 0) 4277 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4278 else { 4279 for (; sa < ea; sa += DEV_BSIZE, j++) { 4280 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4281 pmap_zero_page_area(bp->b_pages[i], 4282 sa, DEV_BSIZE); 4283 } 4284 } 4285 } 4286 bp->b_pages[i]->valid |= mask; 4287 } 4288 unlock: 4289 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4290 bp->b_resid = 0; 4291 } 4292 4293 void 4294 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4295 { 4296 vm_page_t m; 4297 int i, n; 4298 4299 if ((bp->b_flags & B_UNMAPPED) == 0) { 4300 BUF_CHECK_MAPPED(bp); 4301 bzero(bp->b_data + base, size); 4302 } else { 4303 BUF_CHECK_UNMAPPED(bp); 4304 n = PAGE_SIZE - (base & PAGE_MASK); 4305 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4306 m = bp->b_pages[i]; 4307 if (n > size) 4308 n = size; 4309 pmap_zero_page_area(m, base & PAGE_MASK, n); 4310 base += n; 4311 size -= n; 4312 n = PAGE_SIZE; 4313 } 4314 } 4315 } 4316 4317 /* 4318 * vm_hold_load_pages and vm_hold_free_pages get pages into 4319 * a buffers address space. The pages are anonymous and are 4320 * not associated with a file object. 4321 */ 4322 static void 4323 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4324 { 4325 vm_offset_t pg; 4326 vm_page_t p; 4327 int index; 4328 4329 BUF_CHECK_MAPPED(bp); 4330 4331 to = round_page(to); 4332 from = round_page(from); 4333 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4334 4335 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4336 tryagain: 4337 /* 4338 * note: must allocate system pages since blocking here 4339 * could interfere with paging I/O, no matter which 4340 * process we are. 4341 */ 4342 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4343 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4344 if (p == NULL) { 4345 VM_WAIT; 4346 goto tryagain; 4347 } 4348 pmap_qenter(pg, &p, 1); 4349 bp->b_pages[index] = p; 4350 } 4351 bp->b_npages = index; 4352 } 4353 4354 /* Return pages associated with this buf to the vm system */ 4355 static void 4356 vm_hold_free_pages(struct buf *bp, int newbsize) 4357 { 4358 vm_offset_t from; 4359 vm_page_t p; 4360 int index, newnpages; 4361 4362 BUF_CHECK_MAPPED(bp); 4363 4364 from = round_page((vm_offset_t)bp->b_data + newbsize); 4365 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4366 if (bp->b_npages > newnpages) 4367 pmap_qremove(from, bp->b_npages - newnpages); 4368 for (index = newnpages; index < bp->b_npages; index++) { 4369 p = bp->b_pages[index]; 4370 bp->b_pages[index] = NULL; 4371 if (vm_page_sbusied(p)) 4372 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4373 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4374 p->wire_count--; 4375 vm_page_free(p); 4376 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 4377 } 4378 bp->b_npages = newnpages; 4379 } 4380 4381 /* 4382 * Map an IO request into kernel virtual address space. 4383 * 4384 * All requests are (re)mapped into kernel VA space. 4385 * Notice that we use b_bufsize for the size of the buffer 4386 * to be mapped. b_bcount might be modified by the driver. 4387 * 4388 * Note that even if the caller determines that the address space should 4389 * be valid, a race or a smaller-file mapped into a larger space may 4390 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4391 * check the return value. 4392 */ 4393 int 4394 vmapbuf(struct buf *bp, int mapbuf) 4395 { 4396 caddr_t kva; 4397 vm_prot_t prot; 4398 int pidx; 4399 4400 if (bp->b_bufsize < 0) 4401 return (-1); 4402 prot = VM_PROT_READ; 4403 if (bp->b_iocmd == BIO_READ) 4404 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4405 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4406 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4407 btoc(MAXPHYS))) < 0) 4408 return (-1); 4409 bp->b_npages = pidx; 4410 if (mapbuf || !unmapped_buf_allowed) { 4411 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4412 kva = bp->b_saveaddr; 4413 bp->b_saveaddr = bp->b_data; 4414 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4415 bp->b_flags &= ~B_UNMAPPED; 4416 } else { 4417 bp->b_flags |= B_UNMAPPED; 4418 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4419 bp->b_saveaddr = bp->b_data; 4420 bp->b_data = unmapped_buf; 4421 } 4422 return(0); 4423 } 4424 4425 /* 4426 * Free the io map PTEs associated with this IO operation. 4427 * We also invalidate the TLB entries and restore the original b_addr. 4428 */ 4429 void 4430 vunmapbuf(struct buf *bp) 4431 { 4432 int npages; 4433 4434 npages = bp->b_npages; 4435 if (bp->b_flags & B_UNMAPPED) 4436 bp->b_flags &= ~B_UNMAPPED; 4437 else 4438 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4439 vm_page_unhold_pages(bp->b_pages, npages); 4440 4441 bp->b_data = bp->b_saveaddr; 4442 } 4443 4444 void 4445 bdone(struct buf *bp) 4446 { 4447 struct mtx *mtxp; 4448 4449 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4450 mtx_lock(mtxp); 4451 bp->b_flags |= B_DONE; 4452 wakeup(bp); 4453 mtx_unlock(mtxp); 4454 } 4455 4456 void 4457 bwait(struct buf *bp, u_char pri, const char *wchan) 4458 { 4459 struct mtx *mtxp; 4460 4461 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4462 mtx_lock(mtxp); 4463 while ((bp->b_flags & B_DONE) == 0) 4464 msleep(bp, mtxp, pri, wchan, 0); 4465 mtx_unlock(mtxp); 4466 } 4467 4468 int 4469 bufsync(struct bufobj *bo, int waitfor) 4470 { 4471 4472 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4473 } 4474 4475 void 4476 bufstrategy(struct bufobj *bo, struct buf *bp) 4477 { 4478 int i = 0; 4479 struct vnode *vp; 4480 4481 vp = bp->b_vp; 4482 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4483 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4484 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4485 i = VOP_STRATEGY(vp, bp); 4486 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4487 } 4488 4489 void 4490 bufobj_wrefl(struct bufobj *bo) 4491 { 4492 4493 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4494 ASSERT_BO_WLOCKED(bo); 4495 bo->bo_numoutput++; 4496 } 4497 4498 void 4499 bufobj_wref(struct bufobj *bo) 4500 { 4501 4502 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4503 BO_LOCK(bo); 4504 bo->bo_numoutput++; 4505 BO_UNLOCK(bo); 4506 } 4507 4508 void 4509 bufobj_wdrop(struct bufobj *bo) 4510 { 4511 4512 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4513 BO_LOCK(bo); 4514 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4515 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4516 bo->bo_flag &= ~BO_WWAIT; 4517 wakeup(&bo->bo_numoutput); 4518 } 4519 BO_UNLOCK(bo); 4520 } 4521 4522 int 4523 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4524 { 4525 int error; 4526 4527 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4528 ASSERT_BO_WLOCKED(bo); 4529 error = 0; 4530 while (bo->bo_numoutput) { 4531 bo->bo_flag |= BO_WWAIT; 4532 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4533 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4534 if (error) 4535 break; 4536 } 4537 return (error); 4538 } 4539 4540 void 4541 bpin(struct buf *bp) 4542 { 4543 struct mtx *mtxp; 4544 4545 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4546 mtx_lock(mtxp); 4547 bp->b_pin_count++; 4548 mtx_unlock(mtxp); 4549 } 4550 4551 void 4552 bunpin(struct buf *bp) 4553 { 4554 struct mtx *mtxp; 4555 4556 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4557 mtx_lock(mtxp); 4558 if (--bp->b_pin_count == 0) 4559 wakeup(bp); 4560 mtx_unlock(mtxp); 4561 } 4562 4563 void 4564 bunpin_wait(struct buf *bp) 4565 { 4566 struct mtx *mtxp; 4567 4568 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4569 mtx_lock(mtxp); 4570 while (bp->b_pin_count > 0) 4571 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4572 mtx_unlock(mtxp); 4573 } 4574 4575 /* 4576 * Set bio_data or bio_ma for struct bio from the struct buf. 4577 */ 4578 void 4579 bdata2bio(struct buf *bp, struct bio *bip) 4580 { 4581 4582 if ((bp->b_flags & B_UNMAPPED) != 0) { 4583 KASSERT(unmapped_buf_allowed, ("unmapped")); 4584 bip->bio_ma = bp->b_pages; 4585 bip->bio_ma_n = bp->b_npages; 4586 bip->bio_data = unmapped_buf; 4587 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4588 bip->bio_flags |= BIO_UNMAPPED; 4589 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4590 PAGE_SIZE == bp->b_npages, 4591 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4592 (long long)bip->bio_length, bip->bio_ma_n)); 4593 } else { 4594 bip->bio_data = bp->b_data; 4595 bip->bio_ma = NULL; 4596 } 4597 } 4598 4599 #include "opt_ddb.h" 4600 #ifdef DDB 4601 #include <ddb/ddb.h> 4602 4603 /* DDB command to show buffer data */ 4604 DB_SHOW_COMMAND(buffer, db_show_buffer) 4605 { 4606 /* get args */ 4607 struct buf *bp = (struct buf *)addr; 4608 4609 if (!have_addr) { 4610 db_printf("usage: show buffer <addr>\n"); 4611 return; 4612 } 4613 4614 db_printf("buf at %p\n", bp); 4615 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4616 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4617 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4618 db_printf( 4619 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4620 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4621 "b_dep = %p\n", 4622 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4623 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4624 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4625 if (bp->b_npages) { 4626 int i; 4627 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4628 for (i = 0; i < bp->b_npages; i++) { 4629 vm_page_t m; 4630 m = bp->b_pages[i]; 4631 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4632 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4633 if ((i + 1) < bp->b_npages) 4634 db_printf(","); 4635 } 4636 db_printf("\n"); 4637 } 4638 db_printf(" "); 4639 BUF_LOCKPRINTINFO(bp); 4640 } 4641 4642 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4643 { 4644 struct buf *bp; 4645 int i; 4646 4647 for (i = 0; i < nbuf; i++) { 4648 bp = &buf[i]; 4649 if (BUF_ISLOCKED(bp)) { 4650 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4651 db_printf("\n"); 4652 } 4653 } 4654 } 4655 4656 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4657 { 4658 struct vnode *vp; 4659 struct buf *bp; 4660 4661 if (!have_addr) { 4662 db_printf("usage: show vnodebufs <addr>\n"); 4663 return; 4664 } 4665 vp = (struct vnode *)addr; 4666 db_printf("Clean buffers:\n"); 4667 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4668 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4669 db_printf("\n"); 4670 } 4671 db_printf("Dirty buffers:\n"); 4672 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4673 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4674 db_printf("\n"); 4675 } 4676 } 4677 4678 DB_COMMAND(countfreebufs, db_coundfreebufs) 4679 { 4680 struct buf *bp; 4681 int i, used = 0, nfree = 0; 4682 4683 if (have_addr) { 4684 db_printf("usage: countfreebufs\n"); 4685 return; 4686 } 4687 4688 for (i = 0; i < nbuf; i++) { 4689 bp = &buf[i]; 4690 if ((bp->b_flags & B_INFREECNT) != 0) 4691 nfree++; 4692 else 4693 used++; 4694 } 4695 4696 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4697 nfree + used); 4698 db_printf("numfreebuffers is %d\n", numfreebuffers); 4699 } 4700 #endif /* DDB */ 4701