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/vmmeter.h> 68 #include <sys/vnode.h> 69 #include <geom/geom.h> 70 #include <vm/vm.h> 71 #include <vm/vm_param.h> 72 #include <vm/vm_kern.h> 73 #include <vm/vm_pageout.h> 74 #include <vm/vm_page.h> 75 #include <vm/vm_object.h> 76 #include <vm/vm_extern.h> 77 #include <vm/vm_map.h> 78 #include "opt_compat.h" 79 #include "opt_directio.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 static struct proc *bufdaemonproc; 102 103 static int inmem(struct vnode *vp, daddr_t blkno); 104 static void vm_hold_free_pages(struct buf *bp, int newbsize); 105 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 106 vm_offset_t to); 107 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 108 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 109 vm_page_t m); 110 static void vfs_drain_busy_pages(struct buf *bp); 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_do_flush(struct vnode *vp); 117 static int flushbufqueues(struct vnode *, int, int); 118 static void buf_daemon(void); 119 static void bremfreel(struct buf *bp); 120 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 121 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 122 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 123 #endif 124 125 int vmiodirenable = TRUE; 126 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 127 "Use the VM system for directory writes"); 128 long runningbufspace; 129 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 130 "Amount of presently outstanding async buffer io"); 131 static long bufspace; 132 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 133 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 134 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 135 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 136 #else 137 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 138 "Virtual memory used for buffers"); 139 #endif 140 static long unmapped_bufspace; 141 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD, 142 &unmapped_bufspace, 0, 143 "Amount of unmapped buffers, inclusive in the bufspace"); 144 static long maxbufspace; 145 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 146 "Maximum allowed value of bufspace (including buf_daemon)"); 147 static long bufmallocspace; 148 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 149 "Amount of malloced memory for buffers"); 150 static long maxbufmallocspace; 151 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 152 "Maximum amount of malloced memory for buffers"); 153 static long lobufspace; 154 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 155 "Minimum amount of buffers we want to have"); 156 long hibufspace; 157 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 158 "Maximum allowed value of bufspace (excluding buf_daemon)"); 159 static int bufreusecnt; 160 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 161 "Number of times we have reused a buffer"); 162 static int buffreekvacnt; 163 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 164 "Number of times we have freed the KVA space from some buffer"); 165 static int bufdefragcnt; 166 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 167 "Number of times we have had to repeat buffer allocation to defragment"); 168 static long lorunningspace; 169 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 170 "Minimum preferred space used for in-progress I/O"); 171 static long hirunningspace; 172 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 173 "Maximum amount of space to use for in-progress I/O"); 174 int dirtybufferflushes; 175 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 176 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 177 int bdwriteskip; 178 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 179 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 180 int altbufferflushes; 181 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 182 0, "Number of fsync flushes to limit dirty buffers"); 183 static int recursiveflushes; 184 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 185 0, "Number of flushes skipped due to being recursive"); 186 static int numdirtybuffers; 187 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 188 "Number of buffers that are dirty (has unwritten changes) at the moment"); 189 static int lodirtybuffers; 190 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 191 "How many buffers we want to have free before bufdaemon can sleep"); 192 static int hidirtybuffers; 193 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 194 "When the number of dirty buffers is considered severe"); 195 int dirtybufthresh; 196 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 197 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 198 static int numfreebuffers; 199 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 200 "Number of free buffers"); 201 static int lofreebuffers; 202 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 203 "XXX Unused"); 204 static int hifreebuffers; 205 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 206 "XXX Complicatedly unused"); 207 static int getnewbufcalls; 208 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 209 "Number of calls to getnewbuf"); 210 static int getnewbufrestarts; 211 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 212 "Number of times getnewbuf has had to restart a buffer aquisition"); 213 static int mappingrestarts; 214 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 215 "Number of times getblk has had to restart a buffer mapping for " 216 "unmapped buffer"); 217 static int flushbufqtarget = 100; 218 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 219 "Amount of work to do in flushbufqueues when helping bufdaemon"); 220 static long notbufdflashes; 221 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 222 "Number of dirty buffer flushes done by the bufdaemon helpers"); 223 static long barrierwrites; 224 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 225 "Number of barrier writes"); 226 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 227 &unmapped_buf_allowed, 0, 228 "Permit the use of the unmapped i/o"); 229 230 /* 231 * Wakeup point for bufdaemon, as well as indicator of whether it is already 232 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 233 * is idling. 234 */ 235 static int bd_request; 236 237 /* 238 * Request for the buf daemon to write more buffers than is indicated by 239 * lodirtybuf. This may be necessary to push out excess dependencies or 240 * defragment the address space where a simple count of the number of dirty 241 * buffers is insufficient to characterize the demand for flushing them. 242 */ 243 static int bd_speedupreq; 244 245 /* 246 * This lock synchronizes access to bd_request. 247 */ 248 static struct mtx bdlock; 249 250 /* 251 * bogus page -- for I/O to/from partially complete buffers 252 * this is a temporary solution to the problem, but it is not 253 * really that bad. it would be better to split the buffer 254 * for input in the case of buffers partially already in memory, 255 * but the code is intricate enough already. 256 */ 257 vm_page_t bogus_page; 258 259 /* 260 * Synchronization (sleep/wakeup) variable for active buffer space requests. 261 * Set when wait starts, cleared prior to wakeup(). 262 * Used in runningbufwakeup() and waitrunningbufspace(). 263 */ 264 static int runningbufreq; 265 266 /* 267 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 268 * waitrunningbufspace(). 269 */ 270 static struct mtx rbreqlock; 271 272 /* 273 * Synchronization (sleep/wakeup) variable for buffer requests. 274 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 275 * by and/or. 276 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 277 * getnewbuf(), and getblk(). 278 */ 279 static int needsbuffer; 280 281 /* 282 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 283 */ 284 static struct mtx nblock; 285 286 /* 287 * Definitions for the buffer free lists. 288 */ 289 #define BUFFER_QUEUES 5 /* number of free buffer queues */ 290 291 #define QUEUE_NONE 0 /* on no queue */ 292 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 293 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 294 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 295 #define QUEUE_EMPTY 4 /* empty buffer headers */ 296 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 297 298 /* Queues for free buffers with various properties */ 299 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 300 #ifdef INVARIANTS 301 static int bq_len[BUFFER_QUEUES]; 302 #endif 303 304 /* Lock for the bufqueues */ 305 static struct mtx bqlock; 306 307 /* 308 * Single global constant for BUF_WMESG, to avoid getting multiple references. 309 * buf_wmesg is referred from macros. 310 */ 311 const char *buf_wmesg = BUF_WMESG; 312 313 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 314 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 315 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 316 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 317 318 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 319 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 320 static int 321 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 322 { 323 long lvalue; 324 int ivalue; 325 326 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 327 return (sysctl_handle_long(oidp, arg1, arg2, req)); 328 lvalue = *(long *)arg1; 329 if (lvalue > INT_MAX) 330 /* On overflow, still write out a long to trigger ENOMEM. */ 331 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 332 ivalue = lvalue; 333 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 334 } 335 #endif 336 337 #ifdef DIRECTIO 338 extern void ffs_rawread_setup(void); 339 #endif /* DIRECTIO */ 340 /* 341 * numdirtywakeup: 342 * 343 * If someone is blocked due to there being too many dirty buffers, 344 * and numdirtybuffers is now reasonable, wake them up. 345 */ 346 347 static __inline void 348 numdirtywakeup(int level) 349 { 350 351 if (numdirtybuffers <= level) { 352 mtx_lock(&nblock); 353 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 354 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 355 wakeup(&needsbuffer); 356 } 357 mtx_unlock(&nblock); 358 } 359 } 360 361 /* 362 * bufspacewakeup: 363 * 364 * Called when buffer space is potentially available for recovery. 365 * getnewbuf() will block on this flag when it is unable to free 366 * sufficient buffer space. Buffer space becomes recoverable when 367 * bp's get placed back in the queues. 368 */ 369 370 static __inline void 371 bufspacewakeup(void) 372 { 373 374 /* 375 * If someone is waiting for BUF space, wake them up. Even 376 * though we haven't freed the kva space yet, the waiting 377 * process will be able to now. 378 */ 379 mtx_lock(&nblock); 380 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 381 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 382 wakeup(&needsbuffer); 383 } 384 mtx_unlock(&nblock); 385 } 386 387 /* 388 * runningbufwakeup() - in-progress I/O accounting. 389 * 390 */ 391 void 392 runningbufwakeup(struct buf *bp) 393 { 394 395 if (bp->b_runningbufspace) { 396 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 397 bp->b_runningbufspace = 0; 398 mtx_lock(&rbreqlock); 399 if (runningbufreq && runningbufspace <= lorunningspace) { 400 runningbufreq = 0; 401 wakeup(&runningbufreq); 402 } 403 mtx_unlock(&rbreqlock); 404 } 405 } 406 407 /* 408 * bufcountwakeup: 409 * 410 * Called when a buffer has been added to one of the free queues to 411 * account for the buffer and to wakeup anyone waiting for free buffers. 412 * This typically occurs when large amounts of metadata are being handled 413 * by the buffer cache ( else buffer space runs out first, usually ). 414 */ 415 416 static __inline void 417 bufcountwakeup(struct buf *bp) 418 { 419 int old; 420 421 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 422 ("buf %p already counted as free", bp)); 423 if (bp->b_bufobj != NULL) 424 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 425 bp->b_vflags |= BV_INFREECNT; 426 old = atomic_fetchadd_int(&numfreebuffers, 1); 427 KASSERT(old >= 0 && old < nbuf, 428 ("numfreebuffers climbed to %d", old + 1)); 429 mtx_lock(&nblock); 430 if (needsbuffer) { 431 needsbuffer &= ~VFS_BIO_NEED_ANY; 432 if (numfreebuffers >= hifreebuffers) 433 needsbuffer &= ~VFS_BIO_NEED_FREE; 434 wakeup(&needsbuffer); 435 } 436 mtx_unlock(&nblock); 437 } 438 439 /* 440 * waitrunningbufspace() 441 * 442 * runningbufspace is a measure of the amount of I/O currently 443 * running. This routine is used in async-write situations to 444 * prevent creating huge backups of pending writes to a device. 445 * Only asynchronous writes are governed by this function. 446 * 447 * Reads will adjust runningbufspace, but will not block based on it. 448 * The read load has a side effect of reducing the allowed write load. 449 * 450 * This does NOT turn an async write into a sync write. It waits 451 * for earlier writes to complete and generally returns before the 452 * caller's write has reached the device. 453 */ 454 void 455 waitrunningbufspace(void) 456 { 457 458 mtx_lock(&rbreqlock); 459 while (runningbufspace > hirunningspace) { 460 ++runningbufreq; 461 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 462 } 463 mtx_unlock(&rbreqlock); 464 } 465 466 467 /* 468 * vfs_buf_test_cache: 469 * 470 * Called when a buffer is extended. This function clears the B_CACHE 471 * bit if the newly extended portion of the buffer does not contain 472 * valid data. 473 */ 474 static __inline 475 void 476 vfs_buf_test_cache(struct buf *bp, 477 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 478 vm_page_t m) 479 { 480 481 VM_OBJECT_ASSERT_WLOCKED(m->object); 482 if (bp->b_flags & B_CACHE) { 483 int base = (foff + off) & PAGE_MASK; 484 if (vm_page_is_valid(m, base, size) == 0) 485 bp->b_flags &= ~B_CACHE; 486 } 487 } 488 489 /* Wake up the buffer daemon if necessary */ 490 static __inline 491 void 492 bd_wakeup(int dirtybuflevel) 493 { 494 495 mtx_lock(&bdlock); 496 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 497 bd_request = 1; 498 wakeup(&bd_request); 499 } 500 mtx_unlock(&bdlock); 501 } 502 503 /* 504 * bd_speedup - speedup the buffer cache flushing code 505 */ 506 507 void 508 bd_speedup(void) 509 { 510 int needwake; 511 512 mtx_lock(&bdlock); 513 needwake = 0; 514 if (bd_speedupreq == 0 || bd_request == 0) 515 needwake = 1; 516 bd_speedupreq = 1; 517 bd_request = 1; 518 if (needwake) 519 wakeup(&bd_request); 520 mtx_unlock(&bdlock); 521 } 522 523 #ifdef __i386__ 524 #define TRANSIENT_DENOM 5 525 #else 526 #define TRANSIENT_DENOM 10 527 #endif 528 529 /* 530 * Calculating buffer cache scaling values and reserve space for buffer 531 * headers. This is called during low level kernel initialization and 532 * may be called more then once. We CANNOT write to the memory area 533 * being reserved at this time. 534 */ 535 caddr_t 536 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 537 { 538 int tuned_nbuf; 539 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 540 541 /* 542 * physmem_est is in pages. Convert it to kilobytes (assumes 543 * PAGE_SIZE is >= 1K) 544 */ 545 physmem_est = physmem_est * (PAGE_SIZE / 1024); 546 547 /* 548 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 549 * For the first 64MB of ram nominally allocate sufficient buffers to 550 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 551 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 552 * the buffer cache we limit the eventual kva reservation to 553 * maxbcache bytes. 554 * 555 * factor represents the 1/4 x ram conversion. 556 */ 557 if (nbuf == 0) { 558 int factor = 4 * BKVASIZE / 1024; 559 560 nbuf = 50; 561 if (physmem_est > 4096) 562 nbuf += min((physmem_est - 4096) / factor, 563 65536 / factor); 564 if (physmem_est > 65536) 565 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 566 567 if (maxbcache && nbuf > maxbcache / BKVASIZE) 568 nbuf = maxbcache / BKVASIZE; 569 tuned_nbuf = 1; 570 } else 571 tuned_nbuf = 0; 572 573 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 574 maxbuf = (LONG_MAX / 3) / BKVASIZE; 575 if (nbuf > maxbuf) { 576 if (!tuned_nbuf) 577 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 578 maxbuf); 579 nbuf = maxbuf; 580 } 581 582 /* 583 * Ideal allocation size for the transient bio submap if 10% 584 * of the maximal space buffer map. This roughly corresponds 585 * to the amount of the buffer mapped for typical UFS load. 586 * 587 * Clip the buffer map to reserve space for the transient 588 * BIOs, if its extent is bigger than 90% (80% on i386) of the 589 * maximum buffer map extent on the platform. 590 * 591 * The fall-back to the maxbuf in case of maxbcache unset, 592 * allows to not trim the buffer KVA for the architectures 593 * with ample KVA space. 594 */ 595 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 596 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 597 buf_sz = (long)nbuf * BKVASIZE; 598 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 599 (TRANSIENT_DENOM - 1)) { 600 /* 601 * There is more KVA than memory. Do not 602 * adjust buffer map size, and assign the rest 603 * of maxbuf to transient map. 604 */ 605 biotmap_sz = maxbuf_sz - buf_sz; 606 } else { 607 /* 608 * Buffer map spans all KVA we could afford on 609 * this platform. Give 10% (20% on i386) of 610 * the buffer map to the transient bio map. 611 */ 612 biotmap_sz = buf_sz / TRANSIENT_DENOM; 613 buf_sz -= biotmap_sz; 614 } 615 if (biotmap_sz / INT_MAX > MAXPHYS) 616 bio_transient_maxcnt = INT_MAX; 617 else 618 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 619 /* 620 * Artifically limit to 1024 simultaneous in-flight I/Os 621 * using the transient mapping. 622 */ 623 if (bio_transient_maxcnt > 1024) 624 bio_transient_maxcnt = 1024; 625 if (tuned_nbuf) 626 nbuf = buf_sz / BKVASIZE; 627 } 628 629 /* 630 * swbufs are used as temporary holders for I/O, such as paging I/O. 631 * We have no less then 16 and no more then 256. 632 */ 633 nswbuf = max(min(nbuf/4, 256), 16); 634 #ifdef NSWBUF_MIN 635 if (nswbuf < NSWBUF_MIN) 636 nswbuf = NSWBUF_MIN; 637 #endif 638 #ifdef DIRECTIO 639 ffs_rawread_setup(); 640 #endif 641 642 /* 643 * Reserve space for the buffer cache buffers 644 */ 645 swbuf = (void *)v; 646 v = (caddr_t)(swbuf + nswbuf); 647 buf = (void *)v; 648 v = (caddr_t)(buf + nbuf); 649 650 return(v); 651 } 652 653 /* Initialize the buffer subsystem. Called before use of any buffers. */ 654 void 655 bufinit(void) 656 { 657 struct buf *bp; 658 int i; 659 660 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 661 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 662 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 663 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 664 665 /* next, make a null set of free lists */ 666 for (i = 0; i < BUFFER_QUEUES; i++) 667 TAILQ_INIT(&bufqueues[i]); 668 669 /* finally, initialize each buffer header and stick on empty q */ 670 for (i = 0; i < nbuf; i++) { 671 bp = &buf[i]; 672 bzero(bp, sizeof *bp); 673 bp->b_flags = B_INVAL; /* we're just an empty header */ 674 bp->b_rcred = NOCRED; 675 bp->b_wcred = NOCRED; 676 bp->b_qindex = QUEUE_EMPTY; 677 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 678 bp->b_xflags = 0; 679 LIST_INIT(&bp->b_dep); 680 BUF_LOCKINIT(bp); 681 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 682 #ifdef INVARIANTS 683 bq_len[QUEUE_EMPTY]++; 684 #endif 685 } 686 687 /* 688 * maxbufspace is the absolute maximum amount of buffer space we are 689 * allowed to reserve in KVM and in real terms. The absolute maximum 690 * is nominally used by buf_daemon. hibufspace is the nominal maximum 691 * used by most other processes. The differential is required to 692 * ensure that buf_daemon is able to run when other processes might 693 * be blocked waiting for buffer space. 694 * 695 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 696 * this may result in KVM fragmentation which is not handled optimally 697 * by the system. 698 */ 699 maxbufspace = (long)nbuf * BKVASIZE; 700 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 701 lobufspace = hibufspace - MAXBSIZE; 702 703 /* 704 * Note: The 16 MiB upper limit for hirunningspace was chosen 705 * arbitrarily and may need further tuning. It corresponds to 706 * 128 outstanding write IO requests (if IO size is 128 KiB), 707 * which fits with many RAID controllers' tagged queuing limits. 708 * The lower 1 MiB limit is the historical upper limit for 709 * hirunningspace. 710 */ 711 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 712 16 * 1024 * 1024), 1024 * 1024); 713 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 714 715 /* 716 * Limit the amount of malloc memory since it is wired permanently into 717 * the kernel space. Even though this is accounted for in the buffer 718 * allocation, we don't want the malloced region to grow uncontrolled. 719 * The malloc scheme improves memory utilization significantly on average 720 * (small) directories. 721 */ 722 maxbufmallocspace = hibufspace / 20; 723 724 /* 725 * Reduce the chance of a deadlock occuring by limiting the number 726 * of delayed-write dirty buffers we allow to stack up. 727 */ 728 hidirtybuffers = nbuf / 4 + 20; 729 dirtybufthresh = hidirtybuffers * 9 / 10; 730 numdirtybuffers = 0; 731 /* 732 * To support extreme low-memory systems, make sure hidirtybuffers cannot 733 * eat up all available buffer space. This occurs when our minimum cannot 734 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 735 * BKVASIZE'd buffers. 736 */ 737 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 738 hidirtybuffers >>= 1; 739 } 740 lodirtybuffers = hidirtybuffers / 2; 741 742 /* 743 * Try to keep the number of free buffers in the specified range, 744 * and give special processes (e.g. like buf_daemon) access to an 745 * emergency reserve. 746 */ 747 lofreebuffers = nbuf / 18 + 5; 748 hifreebuffers = 2 * lofreebuffers; 749 numfreebuffers = nbuf; 750 751 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 752 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 753 unmapped_buf = (caddr_t)kmem_alloc_nofault(kernel_map, MAXPHYS); 754 } 755 756 #ifdef INVARIANTS 757 static inline void 758 vfs_buf_check_mapped(struct buf *bp) 759 { 760 761 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 762 ("mapped buf %p %x", bp, bp->b_flags)); 763 KASSERT(bp->b_kvabase != unmapped_buf, 764 ("mapped buf: b_kvabase was not updated %p", bp)); 765 KASSERT(bp->b_data != unmapped_buf, 766 ("mapped buf: b_data was not updated %p", bp)); 767 } 768 769 static inline void 770 vfs_buf_check_unmapped(struct buf *bp) 771 { 772 773 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 774 ("unmapped buf %p %x", bp, bp->b_flags)); 775 KASSERT(bp->b_kvabase == unmapped_buf, 776 ("unmapped buf: corrupted b_kvabase %p", bp)); 777 KASSERT(bp->b_data == unmapped_buf, 778 ("unmapped buf: corrupted b_data %p", bp)); 779 } 780 781 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 782 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 783 #else 784 #define BUF_CHECK_MAPPED(bp) do {} while (0) 785 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 786 #endif 787 788 static void 789 bpmap_qenter(struct buf *bp) 790 { 791 792 BUF_CHECK_MAPPED(bp); 793 794 /* 795 * bp->b_data is relative to bp->b_offset, but 796 * bp->b_offset may be offset into the first page. 797 */ 798 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 799 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 800 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 801 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 802 } 803 804 /* 805 * bfreekva() - free the kva allocation for a buffer. 806 * 807 * Since this call frees up buffer space, we call bufspacewakeup(). 808 */ 809 static void 810 bfreekva(struct buf *bp) 811 { 812 813 if (bp->b_kvasize == 0) 814 return; 815 816 atomic_add_int(&buffreekvacnt, 1); 817 atomic_subtract_long(&bufspace, bp->b_kvasize); 818 if ((bp->b_flags & B_UNMAPPED) == 0) { 819 BUF_CHECK_MAPPED(bp); 820 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvabase, 821 (vm_offset_t)bp->b_kvabase + bp->b_kvasize); 822 } else { 823 BUF_CHECK_UNMAPPED(bp); 824 if ((bp->b_flags & B_KVAALLOC) != 0) { 825 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvaalloc, 826 (vm_offset_t)bp->b_kvaalloc + bp->b_kvasize); 827 } 828 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 829 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 830 } 831 bp->b_kvasize = 0; 832 bufspacewakeup(); 833 } 834 835 /* 836 * bremfree: 837 * 838 * Mark the buffer for removal from the appropriate free list in brelse. 839 * 840 */ 841 void 842 bremfree(struct buf *bp) 843 { 844 int old; 845 846 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 847 KASSERT((bp->b_flags & B_REMFREE) == 0, 848 ("bremfree: buffer %p already marked for delayed removal.", bp)); 849 KASSERT(bp->b_qindex != QUEUE_NONE, 850 ("bremfree: buffer %p not on a queue.", bp)); 851 BUF_ASSERT_HELD(bp); 852 853 bp->b_flags |= B_REMFREE; 854 /* Fixup numfreebuffers count. */ 855 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 856 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 857 ("buf %p not counted in numfreebuffers", bp)); 858 if (bp->b_bufobj != NULL) 859 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 860 bp->b_vflags &= ~BV_INFREECNT; 861 old = atomic_fetchadd_int(&numfreebuffers, -1); 862 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 863 } 864 } 865 866 /* 867 * bremfreef: 868 * 869 * Force an immediate removal from a free list. Used only in nfs when 870 * it abuses the b_freelist pointer. 871 */ 872 void 873 bremfreef(struct buf *bp) 874 { 875 mtx_lock(&bqlock); 876 bremfreel(bp); 877 mtx_unlock(&bqlock); 878 } 879 880 /* 881 * bremfreel: 882 * 883 * Removes a buffer from the free list, must be called with the 884 * bqlock held. 885 */ 886 static void 887 bremfreel(struct buf *bp) 888 { 889 int old; 890 891 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 892 bp, bp->b_vp, bp->b_flags); 893 KASSERT(bp->b_qindex != QUEUE_NONE, 894 ("bremfreel: buffer %p not on a queue.", bp)); 895 BUF_ASSERT_HELD(bp); 896 mtx_assert(&bqlock, MA_OWNED); 897 898 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 899 #ifdef INVARIANTS 900 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 901 bp->b_qindex)); 902 bq_len[bp->b_qindex]--; 903 #endif 904 bp->b_qindex = QUEUE_NONE; 905 /* 906 * If this was a delayed bremfree() we only need to remove the buffer 907 * from the queue and return the stats are already done. 908 */ 909 if (bp->b_flags & B_REMFREE) { 910 bp->b_flags &= ~B_REMFREE; 911 return; 912 } 913 /* 914 * Fixup numfreebuffers count. If the buffer is invalid or not 915 * delayed-write, the buffer was free and we must decrement 916 * numfreebuffers. 917 */ 918 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 919 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 920 ("buf %p not counted in numfreebuffers", bp)); 921 if (bp->b_bufobj != NULL) 922 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 923 bp->b_vflags &= ~BV_INFREECNT; 924 old = atomic_fetchadd_int(&numfreebuffers, -1); 925 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 926 } 927 } 928 929 /* 930 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 931 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 932 * the buffer is valid and we do not have to do anything. 933 */ 934 void 935 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 936 int cnt, struct ucred * cred) 937 { 938 struct buf *rabp; 939 int i; 940 941 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 942 if (inmem(vp, *rablkno)) 943 continue; 944 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 945 946 if ((rabp->b_flags & B_CACHE) == 0) { 947 if (!TD_IS_IDLETHREAD(curthread)) 948 curthread->td_ru.ru_inblock++; 949 rabp->b_flags |= B_ASYNC; 950 rabp->b_flags &= ~B_INVAL; 951 rabp->b_ioflags &= ~BIO_ERROR; 952 rabp->b_iocmd = BIO_READ; 953 if (rabp->b_rcred == NOCRED && cred != NOCRED) 954 rabp->b_rcred = crhold(cred); 955 vfs_busy_pages(rabp, 0); 956 BUF_KERNPROC(rabp); 957 rabp->b_iooffset = dbtob(rabp->b_blkno); 958 bstrategy(rabp); 959 } else { 960 brelse(rabp); 961 } 962 } 963 } 964 965 /* 966 * Entry point for bread() and breadn() via #defines in sys/buf.h. 967 * 968 * Get a buffer with the specified data. Look in the cache first. We 969 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 970 * is set, the buffer is valid and we do not have to do anything, see 971 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 972 */ 973 int 974 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 975 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 976 { 977 struct buf *bp; 978 int rv = 0, readwait = 0; 979 980 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 981 /* 982 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 983 */ 984 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 985 if (bp == NULL) 986 return (EBUSY); 987 988 /* if not found in cache, do some I/O */ 989 if ((bp->b_flags & B_CACHE) == 0) { 990 if (!TD_IS_IDLETHREAD(curthread)) 991 curthread->td_ru.ru_inblock++; 992 bp->b_iocmd = BIO_READ; 993 bp->b_flags &= ~B_INVAL; 994 bp->b_ioflags &= ~BIO_ERROR; 995 if (bp->b_rcred == NOCRED && cred != NOCRED) 996 bp->b_rcred = crhold(cred); 997 vfs_busy_pages(bp, 0); 998 bp->b_iooffset = dbtob(bp->b_blkno); 999 bstrategy(bp); 1000 ++readwait; 1001 } 1002 1003 breada(vp, rablkno, rabsize, cnt, cred); 1004 1005 if (readwait) { 1006 rv = bufwait(bp); 1007 } 1008 return (rv); 1009 } 1010 1011 /* 1012 * Write, release buffer on completion. (Done by iodone 1013 * if async). Do not bother writing anything if the buffer 1014 * is invalid. 1015 * 1016 * Note that we set B_CACHE here, indicating that buffer is 1017 * fully valid and thus cacheable. This is true even of NFS 1018 * now so we set it generally. This could be set either here 1019 * or in biodone() since the I/O is synchronous. We put it 1020 * here. 1021 */ 1022 int 1023 bufwrite(struct buf *bp) 1024 { 1025 int oldflags; 1026 struct vnode *vp; 1027 int vp_md; 1028 1029 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1030 if (bp->b_flags & B_INVAL) { 1031 brelse(bp); 1032 return (0); 1033 } 1034 1035 if (bp->b_flags & B_BARRIER) 1036 barrierwrites++; 1037 1038 oldflags = bp->b_flags; 1039 1040 BUF_ASSERT_HELD(bp); 1041 1042 if (bp->b_pin_count > 0) 1043 bunpin_wait(bp); 1044 1045 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1046 ("FFS background buffer should not get here %p", bp)); 1047 1048 vp = bp->b_vp; 1049 if (vp) 1050 vp_md = vp->v_vflag & VV_MD; 1051 else 1052 vp_md = 0; 1053 1054 /* 1055 * Mark the buffer clean. Increment the bufobj write count 1056 * before bundirty() call, to prevent other thread from seeing 1057 * empty dirty list and zero counter for writes in progress, 1058 * falsely indicating that the bufobj is clean. 1059 */ 1060 bufobj_wref(bp->b_bufobj); 1061 bundirty(bp); 1062 1063 bp->b_flags &= ~B_DONE; 1064 bp->b_ioflags &= ~BIO_ERROR; 1065 bp->b_flags |= B_CACHE; 1066 bp->b_iocmd = BIO_WRITE; 1067 1068 vfs_busy_pages(bp, 1); 1069 1070 /* 1071 * Normal bwrites pipeline writes 1072 */ 1073 bp->b_runningbufspace = bp->b_bufsize; 1074 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1075 1076 if (!TD_IS_IDLETHREAD(curthread)) 1077 curthread->td_ru.ru_oublock++; 1078 if (oldflags & B_ASYNC) 1079 BUF_KERNPROC(bp); 1080 bp->b_iooffset = dbtob(bp->b_blkno); 1081 bstrategy(bp); 1082 1083 if ((oldflags & B_ASYNC) == 0) { 1084 int rtval = bufwait(bp); 1085 brelse(bp); 1086 return (rtval); 1087 } else { 1088 /* 1089 * don't allow the async write to saturate the I/O 1090 * system. We will not deadlock here because 1091 * we are blocking waiting for I/O that is already in-progress 1092 * to complete. We do not block here if it is the update 1093 * or syncer daemon trying to clean up as that can lead 1094 * to deadlock. 1095 */ 1096 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1097 waitrunningbufspace(); 1098 } 1099 1100 return (0); 1101 } 1102 1103 void 1104 bufbdflush(struct bufobj *bo, struct buf *bp) 1105 { 1106 struct buf *nbp; 1107 1108 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1109 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1110 altbufferflushes++; 1111 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1112 BO_LOCK(bo); 1113 /* 1114 * Try to find a buffer to flush. 1115 */ 1116 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1117 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1118 BUF_LOCK(nbp, 1119 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1120 continue; 1121 if (bp == nbp) 1122 panic("bdwrite: found ourselves"); 1123 BO_UNLOCK(bo); 1124 /* Don't countdeps with the bo lock held. */ 1125 if (buf_countdeps(nbp, 0)) { 1126 BO_LOCK(bo); 1127 BUF_UNLOCK(nbp); 1128 continue; 1129 } 1130 if (nbp->b_flags & B_CLUSTEROK) { 1131 vfs_bio_awrite(nbp); 1132 } else { 1133 bremfree(nbp); 1134 bawrite(nbp); 1135 } 1136 dirtybufferflushes++; 1137 break; 1138 } 1139 if (nbp == NULL) 1140 BO_UNLOCK(bo); 1141 } 1142 } 1143 1144 /* 1145 * Delayed write. (Buffer is marked dirty). Do not bother writing 1146 * anything if the buffer is marked invalid. 1147 * 1148 * Note that since the buffer must be completely valid, we can safely 1149 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1150 * biodone() in order to prevent getblk from writing the buffer 1151 * out synchronously. 1152 */ 1153 void 1154 bdwrite(struct buf *bp) 1155 { 1156 struct thread *td = curthread; 1157 struct vnode *vp; 1158 struct bufobj *bo; 1159 1160 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1161 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1162 KASSERT((bp->b_flags & B_BARRIER) == 0, 1163 ("Barrier request in delayed write %p", bp)); 1164 BUF_ASSERT_HELD(bp); 1165 1166 if (bp->b_flags & B_INVAL) { 1167 brelse(bp); 1168 return; 1169 } 1170 1171 /* 1172 * If we have too many dirty buffers, don't create any more. 1173 * If we are wildly over our limit, then force a complete 1174 * cleanup. Otherwise, just keep the situation from getting 1175 * out of control. Note that we have to avoid a recursive 1176 * disaster and not try to clean up after our own cleanup! 1177 */ 1178 vp = bp->b_vp; 1179 bo = bp->b_bufobj; 1180 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1181 td->td_pflags |= TDP_INBDFLUSH; 1182 BO_BDFLUSH(bo, bp); 1183 td->td_pflags &= ~TDP_INBDFLUSH; 1184 } else 1185 recursiveflushes++; 1186 1187 bdirty(bp); 1188 /* 1189 * Set B_CACHE, indicating that the buffer is fully valid. This is 1190 * true even of NFS now. 1191 */ 1192 bp->b_flags |= B_CACHE; 1193 1194 /* 1195 * This bmap keeps the system from needing to do the bmap later, 1196 * perhaps when the system is attempting to do a sync. Since it 1197 * is likely that the indirect block -- or whatever other datastructure 1198 * that the filesystem needs is still in memory now, it is a good 1199 * thing to do this. Note also, that if the pageout daemon is 1200 * requesting a sync -- there might not be enough memory to do 1201 * the bmap then... So, this is important to do. 1202 */ 1203 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1204 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1205 } 1206 1207 /* 1208 * Set the *dirty* buffer range based upon the VM system dirty 1209 * pages. 1210 * 1211 * Mark the buffer pages as clean. We need to do this here to 1212 * satisfy the vnode_pager and the pageout daemon, so that it 1213 * thinks that the pages have been "cleaned". Note that since 1214 * the pages are in a delayed write buffer -- the VFS layer 1215 * "will" see that the pages get written out on the next sync, 1216 * or perhaps the cluster will be completed. 1217 */ 1218 vfs_clean_pages_dirty_buf(bp); 1219 bqrelse(bp); 1220 1221 /* 1222 * Wakeup the buffer flushing daemon if we have a lot of dirty 1223 * buffers (midpoint between our recovery point and our stall 1224 * point). 1225 */ 1226 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1227 1228 /* 1229 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1230 * due to the softdep code. 1231 */ 1232 } 1233 1234 /* 1235 * bdirty: 1236 * 1237 * Turn buffer into delayed write request. We must clear BIO_READ and 1238 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1239 * itself to properly update it in the dirty/clean lists. We mark it 1240 * B_DONE to ensure that any asynchronization of the buffer properly 1241 * clears B_DONE ( else a panic will occur later ). 1242 * 1243 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1244 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1245 * should only be called if the buffer is known-good. 1246 * 1247 * Since the buffer is not on a queue, we do not update the numfreebuffers 1248 * count. 1249 * 1250 * The buffer must be on QUEUE_NONE. 1251 */ 1252 void 1253 bdirty(struct buf *bp) 1254 { 1255 1256 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1257 bp, bp->b_vp, bp->b_flags); 1258 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1259 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1260 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1261 BUF_ASSERT_HELD(bp); 1262 bp->b_flags &= ~(B_RELBUF); 1263 bp->b_iocmd = BIO_WRITE; 1264 1265 if ((bp->b_flags & B_DELWRI) == 0) { 1266 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1267 reassignbuf(bp); 1268 atomic_add_int(&numdirtybuffers, 1); 1269 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1270 } 1271 } 1272 1273 /* 1274 * bundirty: 1275 * 1276 * Clear B_DELWRI for buffer. 1277 * 1278 * Since the buffer is not on a queue, we do not update the numfreebuffers 1279 * count. 1280 * 1281 * The buffer must be on QUEUE_NONE. 1282 */ 1283 1284 void 1285 bundirty(struct buf *bp) 1286 { 1287 1288 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1289 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1290 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1291 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1292 BUF_ASSERT_HELD(bp); 1293 1294 if (bp->b_flags & B_DELWRI) { 1295 bp->b_flags &= ~B_DELWRI; 1296 reassignbuf(bp); 1297 atomic_subtract_int(&numdirtybuffers, 1); 1298 numdirtywakeup(lodirtybuffers); 1299 } 1300 /* 1301 * Since it is now being written, we can clear its deferred write flag. 1302 */ 1303 bp->b_flags &= ~B_DEFERRED; 1304 } 1305 1306 /* 1307 * bawrite: 1308 * 1309 * Asynchronous write. Start output on a buffer, but do not wait for 1310 * it to complete. The buffer is released when the output completes. 1311 * 1312 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1313 * B_INVAL buffers. Not us. 1314 */ 1315 void 1316 bawrite(struct buf *bp) 1317 { 1318 1319 bp->b_flags |= B_ASYNC; 1320 (void) bwrite(bp); 1321 } 1322 1323 /* 1324 * babarrierwrite: 1325 * 1326 * Asynchronous barrier write. Start output on a buffer, but do not 1327 * wait for it to complete. Place a write barrier after this write so 1328 * that this buffer and all buffers written before it are committed to 1329 * the disk before any buffers written after this write are committed 1330 * to the disk. The buffer is released when the output completes. 1331 */ 1332 void 1333 babarrierwrite(struct buf *bp) 1334 { 1335 1336 bp->b_flags |= B_ASYNC | B_BARRIER; 1337 (void) bwrite(bp); 1338 } 1339 1340 /* 1341 * bbarrierwrite: 1342 * 1343 * Synchronous barrier write. Start output on a buffer and wait for 1344 * it to complete. Place a write barrier after this write so that 1345 * this buffer and all buffers written before it are committed to 1346 * the disk before any buffers written after this write are committed 1347 * to the disk. The buffer is released when the output completes. 1348 */ 1349 int 1350 bbarrierwrite(struct buf *bp) 1351 { 1352 1353 bp->b_flags |= B_BARRIER; 1354 return (bwrite(bp)); 1355 } 1356 1357 /* 1358 * bwillwrite: 1359 * 1360 * Called prior to the locking of any vnodes when we are expecting to 1361 * write. We do not want to starve the buffer cache with too many 1362 * dirty buffers so we block here. By blocking prior to the locking 1363 * of any vnodes we attempt to avoid the situation where a locked vnode 1364 * prevents the various system daemons from flushing related buffers. 1365 */ 1366 1367 void 1368 bwillwrite(void) 1369 { 1370 1371 if (numdirtybuffers >= hidirtybuffers) { 1372 mtx_lock(&nblock); 1373 while (numdirtybuffers >= hidirtybuffers) { 1374 bd_wakeup(1); 1375 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1376 msleep(&needsbuffer, &nblock, 1377 (PRIBIO + 4), "flswai", 0); 1378 } 1379 mtx_unlock(&nblock); 1380 } 1381 } 1382 1383 /* 1384 * Return true if we have too many dirty buffers. 1385 */ 1386 int 1387 buf_dirty_count_severe(void) 1388 { 1389 1390 return(numdirtybuffers >= hidirtybuffers); 1391 } 1392 1393 static __noinline int 1394 buf_vm_page_count_severe(void) 1395 { 1396 1397 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1398 1399 return vm_page_count_severe(); 1400 } 1401 1402 /* 1403 * brelse: 1404 * 1405 * Release a busy buffer and, if requested, free its resources. The 1406 * buffer will be stashed in the appropriate bufqueue[] allowing it 1407 * to be accessed later as a cache entity or reused for other purposes. 1408 */ 1409 void 1410 brelse(struct buf *bp) 1411 { 1412 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1413 bp, bp->b_vp, bp->b_flags); 1414 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1415 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1416 1417 if (BUF_LOCKRECURSED(bp)) { 1418 /* 1419 * Do not process, in particular, do not handle the 1420 * B_INVAL/B_RELBUF and do not release to free list. 1421 */ 1422 BUF_UNLOCK(bp); 1423 return; 1424 } 1425 1426 if (bp->b_flags & B_MANAGED) { 1427 bqrelse(bp); 1428 return; 1429 } 1430 1431 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1432 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1433 /* 1434 * Failed write, redirty. Must clear BIO_ERROR to prevent 1435 * pages from being scrapped. If the error is anything 1436 * other than an I/O error (EIO), assume that retrying 1437 * is futile. 1438 */ 1439 bp->b_ioflags &= ~BIO_ERROR; 1440 bdirty(bp); 1441 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1442 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1443 /* 1444 * Either a failed I/O or we were asked to free or not 1445 * cache the buffer. 1446 */ 1447 bp->b_flags |= B_INVAL; 1448 if (!LIST_EMPTY(&bp->b_dep)) 1449 buf_deallocate(bp); 1450 if (bp->b_flags & B_DELWRI) { 1451 atomic_subtract_int(&numdirtybuffers, 1); 1452 numdirtywakeup(lodirtybuffers); 1453 } 1454 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1455 if ((bp->b_flags & B_VMIO) == 0) { 1456 if (bp->b_bufsize) 1457 allocbuf(bp, 0); 1458 if (bp->b_vp) 1459 brelvp(bp); 1460 } 1461 } 1462 1463 /* 1464 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1465 * is called with B_DELWRI set, the underlying pages may wind up 1466 * getting freed causing a previous write (bdwrite()) to get 'lost' 1467 * because pages associated with a B_DELWRI bp are marked clean. 1468 * 1469 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1470 * if B_DELWRI is set. 1471 * 1472 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1473 * on pages to return pages to the VM page queues. 1474 */ 1475 if (bp->b_flags & B_DELWRI) 1476 bp->b_flags &= ~B_RELBUF; 1477 else if (buf_vm_page_count_severe()) { 1478 /* 1479 * The locking of the BO_LOCK is not necessary since 1480 * BKGRDINPROG cannot be set while we hold the buf 1481 * lock, it can only be cleared if it is already 1482 * pending. 1483 */ 1484 if (bp->b_vp) { 1485 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1486 bp->b_flags |= B_RELBUF; 1487 } else 1488 bp->b_flags |= B_RELBUF; 1489 } 1490 1491 /* 1492 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1493 * constituted, not even NFS buffers now. Two flags effect this. If 1494 * B_INVAL, the struct buf is invalidated but the VM object is kept 1495 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1496 * 1497 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1498 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1499 * buffer is also B_INVAL because it hits the re-dirtying code above. 1500 * 1501 * Normally we can do this whether a buffer is B_DELWRI or not. If 1502 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1503 * the commit state and we cannot afford to lose the buffer. If the 1504 * buffer has a background write in progress, we need to keep it 1505 * around to prevent it from being reconstituted and starting a second 1506 * background write. 1507 */ 1508 if ((bp->b_flags & B_VMIO) 1509 && !(bp->b_vp->v_mount != NULL && 1510 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1511 !vn_isdisk(bp->b_vp, NULL) && 1512 (bp->b_flags & B_DELWRI)) 1513 ) { 1514 1515 int i, j, resid; 1516 vm_page_t m; 1517 off_t foff; 1518 vm_pindex_t poff; 1519 vm_object_t obj; 1520 1521 obj = bp->b_bufobj->bo_object; 1522 1523 /* 1524 * Get the base offset and length of the buffer. Note that 1525 * in the VMIO case if the buffer block size is not 1526 * page-aligned then b_data pointer may not be page-aligned. 1527 * But our b_pages[] array *IS* page aligned. 1528 * 1529 * block sizes less then DEV_BSIZE (usually 512) are not 1530 * supported due to the page granularity bits (m->valid, 1531 * m->dirty, etc...). 1532 * 1533 * See man buf(9) for more information 1534 */ 1535 resid = bp->b_bufsize; 1536 foff = bp->b_offset; 1537 VM_OBJECT_WLOCK(obj); 1538 for (i = 0; i < bp->b_npages; i++) { 1539 int had_bogus = 0; 1540 1541 m = bp->b_pages[i]; 1542 1543 /* 1544 * If we hit a bogus page, fixup *all* the bogus pages 1545 * now. 1546 */ 1547 if (m == bogus_page) { 1548 poff = OFF_TO_IDX(bp->b_offset); 1549 had_bogus = 1; 1550 1551 for (j = i; j < bp->b_npages; j++) { 1552 vm_page_t mtmp; 1553 mtmp = bp->b_pages[j]; 1554 if (mtmp == bogus_page) { 1555 mtmp = vm_page_lookup(obj, poff + j); 1556 if (!mtmp) { 1557 panic("brelse: page missing\n"); 1558 } 1559 bp->b_pages[j] = mtmp; 1560 } 1561 } 1562 1563 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1564 BUF_CHECK_MAPPED(bp); 1565 pmap_qenter( 1566 trunc_page((vm_offset_t)bp->b_data), 1567 bp->b_pages, bp->b_npages); 1568 } 1569 m = bp->b_pages[i]; 1570 } 1571 if ((bp->b_flags & B_NOCACHE) || 1572 (bp->b_ioflags & BIO_ERROR && 1573 bp->b_iocmd == BIO_READ)) { 1574 int poffset = foff & PAGE_MASK; 1575 int presid = resid > (PAGE_SIZE - poffset) ? 1576 (PAGE_SIZE - poffset) : resid; 1577 1578 KASSERT(presid >= 0, ("brelse: extra page")); 1579 vm_page_set_invalid(m, poffset, presid); 1580 if (had_bogus) 1581 printf("avoided corruption bug in bogus_page/brelse code\n"); 1582 } 1583 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1584 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1585 } 1586 VM_OBJECT_WUNLOCK(obj); 1587 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1588 vfs_vmio_release(bp); 1589 1590 } else if (bp->b_flags & B_VMIO) { 1591 1592 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1593 vfs_vmio_release(bp); 1594 } 1595 1596 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1597 if (bp->b_bufsize != 0) 1598 allocbuf(bp, 0); 1599 if (bp->b_vp != NULL) 1600 brelvp(bp); 1601 } 1602 1603 /* enqueue */ 1604 mtx_lock(&bqlock); 1605 /* Handle delayed bremfree() processing. */ 1606 if (bp->b_flags & B_REMFREE) { 1607 struct bufobj *bo; 1608 1609 bo = bp->b_bufobj; 1610 if (bo != NULL) 1611 BO_LOCK(bo); 1612 bremfreel(bp); 1613 if (bo != NULL) 1614 BO_UNLOCK(bo); 1615 } 1616 if (bp->b_qindex != QUEUE_NONE) 1617 panic("brelse: free buffer onto another queue???"); 1618 1619 /* 1620 * If the buffer has junk contents signal it and eventually 1621 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1622 * doesn't find it. 1623 */ 1624 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1625 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1626 bp->b_flags |= B_INVAL; 1627 if (bp->b_flags & B_INVAL) { 1628 if (bp->b_flags & B_DELWRI) 1629 bundirty(bp); 1630 if (bp->b_vp) 1631 brelvp(bp); 1632 } 1633 1634 /* buffers with no memory */ 1635 if (bp->b_bufsize == 0) { 1636 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1637 if (bp->b_vflags & BV_BKGRDINPROG) 1638 panic("losing buffer 1"); 1639 if (bp->b_kvasize) { 1640 bp->b_qindex = QUEUE_EMPTYKVA; 1641 } else { 1642 bp->b_qindex = QUEUE_EMPTY; 1643 } 1644 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1645 /* buffers with junk contents */ 1646 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1647 (bp->b_ioflags & BIO_ERROR)) { 1648 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1649 if (bp->b_vflags & BV_BKGRDINPROG) 1650 panic("losing buffer 2"); 1651 bp->b_qindex = QUEUE_CLEAN; 1652 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1653 /* remaining buffers */ 1654 } else { 1655 if (bp->b_flags & B_DELWRI) 1656 bp->b_qindex = QUEUE_DIRTY; 1657 else 1658 bp->b_qindex = QUEUE_CLEAN; 1659 if (bp->b_flags & B_AGE) { 1660 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, 1661 b_freelist); 1662 } else { 1663 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, 1664 b_freelist); 1665 } 1666 } 1667 #ifdef INVARIANTS 1668 bq_len[bp->b_qindex]++; 1669 #endif 1670 mtx_unlock(&bqlock); 1671 1672 /* 1673 * Fixup numfreebuffers count. The bp is on an appropriate queue 1674 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1675 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1676 * if B_INVAL is set ). 1677 */ 1678 1679 if (!(bp->b_flags & B_DELWRI)) { 1680 struct bufobj *bo; 1681 1682 bo = bp->b_bufobj; 1683 if (bo != NULL) 1684 BO_LOCK(bo); 1685 bufcountwakeup(bp); 1686 if (bo != NULL) 1687 BO_UNLOCK(bo); 1688 } 1689 1690 /* 1691 * Something we can maybe free or reuse 1692 */ 1693 if (bp->b_bufsize || bp->b_kvasize) 1694 bufspacewakeup(); 1695 1696 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1697 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1698 panic("brelse: not dirty"); 1699 /* unlock */ 1700 BUF_UNLOCK(bp); 1701 } 1702 1703 /* 1704 * Release a buffer back to the appropriate queue but do not try to free 1705 * it. The buffer is expected to be used again soon. 1706 * 1707 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1708 * biodone() to requeue an async I/O on completion. It is also used when 1709 * known good buffers need to be requeued but we think we may need the data 1710 * again soon. 1711 * 1712 * XXX we should be able to leave the B_RELBUF hint set on completion. 1713 */ 1714 void 1715 bqrelse(struct buf *bp) 1716 { 1717 struct bufobj *bo; 1718 1719 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1720 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1721 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1722 1723 if (BUF_LOCKRECURSED(bp)) { 1724 /* do not release to free list */ 1725 BUF_UNLOCK(bp); 1726 return; 1727 } 1728 1729 bo = bp->b_bufobj; 1730 if (bp->b_flags & B_MANAGED) { 1731 if (bp->b_flags & B_REMFREE) { 1732 mtx_lock(&bqlock); 1733 if (bo != NULL) 1734 BO_LOCK(bo); 1735 bremfreel(bp); 1736 if (bo != NULL) 1737 BO_UNLOCK(bo); 1738 mtx_unlock(&bqlock); 1739 } 1740 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1741 BUF_UNLOCK(bp); 1742 return; 1743 } 1744 1745 mtx_lock(&bqlock); 1746 /* Handle delayed bremfree() processing. */ 1747 if (bp->b_flags & B_REMFREE) { 1748 if (bo != NULL) 1749 BO_LOCK(bo); 1750 bremfreel(bp); 1751 if (bo != NULL) 1752 BO_UNLOCK(bo); 1753 } 1754 if (bp->b_qindex != QUEUE_NONE) 1755 panic("bqrelse: free buffer onto another queue???"); 1756 /* buffers with stale but valid contents */ 1757 if (bp->b_flags & B_DELWRI) { 1758 bp->b_qindex = QUEUE_DIRTY; 1759 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1760 #ifdef INVARIANTS 1761 bq_len[bp->b_qindex]++; 1762 #endif 1763 } else { 1764 /* 1765 * The locking of the BO_LOCK for checking of the 1766 * BV_BKGRDINPROG is not necessary since the 1767 * BV_BKGRDINPROG cannot be set while we hold the buf 1768 * lock, it can only be cleared if it is already 1769 * pending. 1770 */ 1771 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1772 bp->b_qindex = QUEUE_CLEAN; 1773 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1774 b_freelist); 1775 #ifdef INVARIANTS 1776 bq_len[QUEUE_CLEAN]++; 1777 #endif 1778 } else { 1779 /* 1780 * We are too low on memory, we have to try to free 1781 * the buffer (most importantly: the wired pages 1782 * making up its backing store) *now*. 1783 */ 1784 mtx_unlock(&bqlock); 1785 brelse(bp); 1786 return; 1787 } 1788 } 1789 mtx_unlock(&bqlock); 1790 1791 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) { 1792 if (bo != NULL) 1793 BO_LOCK(bo); 1794 bufcountwakeup(bp); 1795 if (bo != NULL) 1796 BO_UNLOCK(bo); 1797 } 1798 1799 /* 1800 * Something we can maybe free or reuse. 1801 */ 1802 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1803 bufspacewakeup(); 1804 1805 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1806 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1807 panic("bqrelse: not dirty"); 1808 /* unlock */ 1809 BUF_UNLOCK(bp); 1810 } 1811 1812 /* Give pages used by the bp back to the VM system (where possible) */ 1813 static void 1814 vfs_vmio_release(struct buf *bp) 1815 { 1816 int i; 1817 vm_page_t m; 1818 1819 if ((bp->b_flags & B_UNMAPPED) == 0) { 1820 BUF_CHECK_MAPPED(bp); 1821 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1822 } else 1823 BUF_CHECK_UNMAPPED(bp); 1824 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 1825 for (i = 0; i < bp->b_npages; i++) { 1826 m = bp->b_pages[i]; 1827 bp->b_pages[i] = NULL; 1828 /* 1829 * In order to keep page LRU ordering consistent, put 1830 * everything on the inactive queue. 1831 */ 1832 vm_page_lock(m); 1833 vm_page_unwire(m, 0); 1834 /* 1835 * We don't mess with busy pages, it is 1836 * the responsibility of the process that 1837 * busied the pages to deal with them. 1838 */ 1839 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1840 m->wire_count == 0) { 1841 /* 1842 * Might as well free the page if we can and it has 1843 * no valid data. We also free the page if the 1844 * buffer was used for direct I/O 1845 */ 1846 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1847 vm_page_free(m); 1848 } else if (bp->b_flags & B_DIRECT) { 1849 vm_page_try_to_free(m); 1850 } else if (buf_vm_page_count_severe()) { 1851 vm_page_try_to_cache(m); 1852 } 1853 } 1854 vm_page_unlock(m); 1855 } 1856 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 1857 1858 if (bp->b_bufsize) { 1859 bufspacewakeup(); 1860 bp->b_bufsize = 0; 1861 } 1862 bp->b_npages = 0; 1863 bp->b_flags &= ~B_VMIO; 1864 if (bp->b_vp) 1865 brelvp(bp); 1866 } 1867 1868 /* 1869 * Check to see if a block at a particular lbn is available for a clustered 1870 * write. 1871 */ 1872 static int 1873 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1874 { 1875 struct buf *bpa; 1876 int match; 1877 1878 match = 0; 1879 1880 /* If the buf isn't in core skip it */ 1881 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1882 return (0); 1883 1884 /* If the buf is busy we don't want to wait for it */ 1885 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1886 return (0); 1887 1888 /* Only cluster with valid clusterable delayed write buffers */ 1889 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1890 (B_DELWRI | B_CLUSTEROK)) 1891 goto done; 1892 1893 if (bpa->b_bufsize != size) 1894 goto done; 1895 1896 /* 1897 * Check to see if it is in the expected place on disk and that the 1898 * block has been mapped. 1899 */ 1900 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1901 match = 1; 1902 done: 1903 BUF_UNLOCK(bpa); 1904 return (match); 1905 } 1906 1907 /* 1908 * vfs_bio_awrite: 1909 * 1910 * Implement clustered async writes for clearing out B_DELWRI buffers. 1911 * This is much better then the old way of writing only one buffer at 1912 * a time. Note that we may not be presented with the buffers in the 1913 * correct order, so we search for the cluster in both directions. 1914 */ 1915 int 1916 vfs_bio_awrite(struct buf *bp) 1917 { 1918 struct bufobj *bo; 1919 int i; 1920 int j; 1921 daddr_t lblkno = bp->b_lblkno; 1922 struct vnode *vp = bp->b_vp; 1923 int ncl; 1924 int nwritten; 1925 int size; 1926 int maxcl; 1927 int gbflags; 1928 1929 bo = &vp->v_bufobj; 1930 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1931 /* 1932 * right now we support clustered writing only to regular files. If 1933 * we find a clusterable block we could be in the middle of a cluster 1934 * rather then at the beginning. 1935 */ 1936 if ((vp->v_type == VREG) && 1937 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1938 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1939 1940 size = vp->v_mount->mnt_stat.f_iosize; 1941 maxcl = MAXPHYS / size; 1942 1943 BO_LOCK(bo); 1944 for (i = 1; i < maxcl; i++) 1945 if (vfs_bio_clcheck(vp, size, lblkno + i, 1946 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1947 break; 1948 1949 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1950 if (vfs_bio_clcheck(vp, size, lblkno - j, 1951 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1952 break; 1953 BO_UNLOCK(bo); 1954 --j; 1955 ncl = i + j; 1956 /* 1957 * this is a possible cluster write 1958 */ 1959 if (ncl != 1) { 1960 BUF_UNLOCK(bp); 1961 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1962 gbflags); 1963 return (nwritten); 1964 } 1965 } 1966 bremfree(bp); 1967 bp->b_flags |= B_ASYNC; 1968 /* 1969 * default (old) behavior, writing out only one block 1970 * 1971 * XXX returns b_bufsize instead of b_bcount for nwritten? 1972 */ 1973 nwritten = bp->b_bufsize; 1974 (void) bwrite(bp); 1975 1976 return (nwritten); 1977 } 1978 1979 static void 1980 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 1981 { 1982 1983 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 1984 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 1985 if ((gbflags & GB_UNMAPPED) == 0) { 1986 bp->b_kvabase = (caddr_t)addr; 1987 } else if ((gbflags & GB_KVAALLOC) != 0) { 1988 KASSERT((gbflags & GB_UNMAPPED) != 0, 1989 ("GB_KVAALLOC without GB_UNMAPPED")); 1990 bp->b_kvaalloc = (caddr_t)addr; 1991 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 1992 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 1993 } 1994 bp->b_kvasize = maxsize; 1995 } 1996 1997 /* 1998 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 1999 * needed. 2000 */ 2001 static int 2002 allocbufkva(struct buf *bp, int maxsize, int gbflags) 2003 { 2004 vm_offset_t addr; 2005 int rv; 2006 2007 bfreekva(bp); 2008 addr = 0; 2009 2010 vm_map_lock(buffer_map); 2011 if (vm_map_findspace(buffer_map, vm_map_min(buffer_map), maxsize, 2012 &addr)) { 2013 vm_map_unlock(buffer_map); 2014 /* 2015 * Buffer map is too fragmented. Request the caller 2016 * to defragment the map. 2017 */ 2018 atomic_add_int(&bufdefragcnt, 1); 2019 return (1); 2020 } 2021 rv = vm_map_insert(buffer_map, NULL, 0, addr, addr + maxsize, 2022 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 2023 KASSERT(rv == KERN_SUCCESS, ("vm_map_insert(buffer_map) rv %d", rv)); 2024 vm_map_unlock(buffer_map); 2025 setbufkva(bp, addr, maxsize, gbflags); 2026 atomic_add_long(&bufspace, bp->b_kvasize); 2027 return (0); 2028 } 2029 2030 /* 2031 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2032 * locked vnode is supplied. 2033 */ 2034 static void 2035 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2036 int defrag) 2037 { 2038 struct thread *td; 2039 char *waitmsg; 2040 int fl, flags, norunbuf; 2041 2042 mtx_assert(&bqlock, MA_OWNED); 2043 2044 if (defrag) { 2045 flags = VFS_BIO_NEED_BUFSPACE; 2046 waitmsg = "nbufkv"; 2047 } else if (bufspace >= hibufspace) { 2048 waitmsg = "nbufbs"; 2049 flags = VFS_BIO_NEED_BUFSPACE; 2050 } else { 2051 waitmsg = "newbuf"; 2052 flags = VFS_BIO_NEED_ANY; 2053 } 2054 mtx_lock(&nblock); 2055 needsbuffer |= flags; 2056 mtx_unlock(&nblock); 2057 mtx_unlock(&bqlock); 2058 2059 bd_speedup(); /* heeeelp */ 2060 if ((gbflags & GB_NOWAIT_BD) != 0) 2061 return; 2062 2063 td = curthread; 2064 mtx_lock(&nblock); 2065 while (needsbuffer & flags) { 2066 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2067 mtx_unlock(&nblock); 2068 /* 2069 * getblk() is called with a vnode locked, and 2070 * some majority of the dirty buffers may as 2071 * well belong to the vnode. Flushing the 2072 * buffers there would make a progress that 2073 * cannot be achieved by the buf_daemon, that 2074 * cannot lock the vnode. 2075 */ 2076 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2077 (td->td_pflags & TDP_NORUNNINGBUF); 2078 /* play bufdaemon */ 2079 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2080 fl = buf_do_flush(vp); 2081 td->td_pflags &= norunbuf; 2082 mtx_lock(&nblock); 2083 if (fl != 0) 2084 continue; 2085 if ((needsbuffer & flags) == 0) 2086 break; 2087 } 2088 if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag, 2089 waitmsg, slptimeo)) 2090 break; 2091 } 2092 mtx_unlock(&nblock); 2093 } 2094 2095 static void 2096 getnewbuf_reuse_bp(struct buf *bp, int qindex) 2097 { 2098 2099 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2100 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2101 bp->b_kvasize, bp->b_bufsize, qindex); 2102 mtx_assert(&bqlock, MA_NOTOWNED); 2103 2104 /* 2105 * Note: we no longer distinguish between VMIO and non-VMIO 2106 * buffers. 2107 */ 2108 KASSERT((bp->b_flags & B_DELWRI) == 0, 2109 ("delwri buffer %p found in queue %d", bp, qindex)); 2110 2111 if (qindex == QUEUE_CLEAN) { 2112 if (bp->b_flags & B_VMIO) { 2113 bp->b_flags &= ~B_ASYNC; 2114 vfs_vmio_release(bp); 2115 } 2116 if (bp->b_vp != NULL) 2117 brelvp(bp); 2118 } 2119 2120 /* 2121 * Get the rest of the buffer freed up. b_kva* is still valid 2122 * after this operation. 2123 */ 2124 2125 if (bp->b_rcred != NOCRED) { 2126 crfree(bp->b_rcred); 2127 bp->b_rcred = NOCRED; 2128 } 2129 if (bp->b_wcred != NOCRED) { 2130 crfree(bp->b_wcred); 2131 bp->b_wcred = NOCRED; 2132 } 2133 if (!LIST_EMPTY(&bp->b_dep)) 2134 buf_deallocate(bp); 2135 if (bp->b_vflags & BV_BKGRDINPROG) 2136 panic("losing buffer 3"); 2137 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2138 bp, bp->b_vp, qindex)); 2139 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2140 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2141 2142 if (bp->b_bufsize) 2143 allocbuf(bp, 0); 2144 2145 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2146 bp->b_ioflags = 0; 2147 bp->b_xflags = 0; 2148 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 2149 ("buf %p still counted as free?", bp)); 2150 bp->b_vflags = 0; 2151 bp->b_vp = NULL; 2152 bp->b_blkno = bp->b_lblkno = 0; 2153 bp->b_offset = NOOFFSET; 2154 bp->b_iodone = 0; 2155 bp->b_error = 0; 2156 bp->b_resid = 0; 2157 bp->b_bcount = 0; 2158 bp->b_npages = 0; 2159 bp->b_dirtyoff = bp->b_dirtyend = 0; 2160 bp->b_bufobj = NULL; 2161 bp->b_pin_count = 0; 2162 bp->b_fsprivate1 = NULL; 2163 bp->b_fsprivate2 = NULL; 2164 bp->b_fsprivate3 = NULL; 2165 2166 LIST_INIT(&bp->b_dep); 2167 } 2168 2169 static int flushingbufs; 2170 2171 static struct buf * 2172 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2173 { 2174 struct buf *bp, *nbp; 2175 int nqindex, qindex, pass; 2176 2177 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2178 2179 pass = 1; 2180 restart: 2181 atomic_add_int(&getnewbufrestarts, 1); 2182 2183 /* 2184 * Setup for scan. If we do not have enough free buffers, 2185 * we setup a degenerate case that immediately fails. Note 2186 * that if we are specially marked process, we are allowed to 2187 * dip into our reserves. 2188 * 2189 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2190 * for the allocation of the mapped buffer. For unmapped, the 2191 * easiest is to start with EMPTY outright. 2192 * 2193 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2194 * However, there are a number of cases (defragging, reusing, ...) 2195 * where we cannot backup. 2196 */ 2197 nbp = NULL; 2198 mtx_lock(&bqlock); 2199 if (!defrag && unmapped) { 2200 nqindex = QUEUE_EMPTY; 2201 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2202 } 2203 if (nbp == NULL) { 2204 nqindex = QUEUE_EMPTYKVA; 2205 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2206 } 2207 2208 /* 2209 * If no EMPTYKVA buffers and we are either defragging or 2210 * reusing, locate a CLEAN buffer to free or reuse. If 2211 * bufspace useage is low skip this step so we can allocate a 2212 * new buffer. 2213 */ 2214 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2215 nqindex = QUEUE_CLEAN; 2216 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2217 } 2218 2219 /* 2220 * If we could not find or were not allowed to reuse a CLEAN 2221 * buffer, check to see if it is ok to use an EMPTY buffer. 2222 * We can only use an EMPTY buffer if allocating its KVA would 2223 * not otherwise run us out of buffer space. No KVA is needed 2224 * for the unmapped allocation. 2225 */ 2226 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2227 metadata)) { 2228 nqindex = QUEUE_EMPTY; 2229 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2230 } 2231 2232 /* 2233 * All available buffers might be clean, retry ignoring the 2234 * lobufspace as the last resort. 2235 */ 2236 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2237 nqindex = QUEUE_CLEAN; 2238 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2239 } 2240 2241 /* 2242 * Run scan, possibly freeing data and/or kva mappings on the fly 2243 * depending. 2244 */ 2245 while ((bp = nbp) != NULL) { 2246 qindex = nqindex; 2247 2248 /* 2249 * Calculate next bp (we can only use it if we do not 2250 * block or do other fancy things). 2251 */ 2252 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2253 switch (qindex) { 2254 case QUEUE_EMPTY: 2255 nqindex = QUEUE_EMPTYKVA; 2256 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2257 if (nbp != NULL) 2258 break; 2259 /* FALLTHROUGH */ 2260 case QUEUE_EMPTYKVA: 2261 nqindex = QUEUE_CLEAN; 2262 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2263 if (nbp != NULL) 2264 break; 2265 /* FALLTHROUGH */ 2266 case QUEUE_CLEAN: 2267 if (metadata && pass == 1) { 2268 pass = 2; 2269 nqindex = QUEUE_EMPTY; 2270 nbp = TAILQ_FIRST( 2271 &bufqueues[QUEUE_EMPTY]); 2272 } 2273 /* 2274 * nbp is NULL. 2275 */ 2276 break; 2277 } 2278 } 2279 /* 2280 * If we are defragging then we need a buffer with 2281 * b_kvasize != 0. XXX this situation should no longer 2282 * occur, if defrag is non-zero the buffer's b_kvasize 2283 * should also be non-zero at this point. XXX 2284 */ 2285 if (defrag && bp->b_kvasize == 0) { 2286 printf("Warning: defrag empty buffer %p\n", bp); 2287 continue; 2288 } 2289 2290 /* 2291 * Start freeing the bp. This is somewhat involved. nbp 2292 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2293 */ 2294 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2295 continue; 2296 if (bp->b_vp) { 2297 BO_LOCK(bp->b_bufobj); 2298 if (bp->b_vflags & BV_BKGRDINPROG) { 2299 BO_UNLOCK(bp->b_bufobj); 2300 BUF_UNLOCK(bp); 2301 continue; 2302 } 2303 BO_UNLOCK(bp->b_bufobj); 2304 } 2305 2306 KASSERT(bp->b_qindex == qindex, 2307 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2308 2309 if (bp->b_bufobj != NULL) 2310 BO_LOCK(bp->b_bufobj); 2311 bremfreel(bp); 2312 if (bp->b_bufobj != NULL) 2313 BO_UNLOCK(bp->b_bufobj); 2314 mtx_unlock(&bqlock); 2315 /* 2316 * NOTE: nbp is now entirely invalid. We can only restart 2317 * the scan from this point on. 2318 */ 2319 2320 getnewbuf_reuse_bp(bp, qindex); 2321 mtx_assert(&bqlock, MA_NOTOWNED); 2322 2323 /* 2324 * If we are defragging then free the buffer. 2325 */ 2326 if (defrag) { 2327 bp->b_flags |= B_INVAL; 2328 bfreekva(bp); 2329 brelse(bp); 2330 defrag = 0; 2331 goto restart; 2332 } 2333 2334 /* 2335 * Notify any waiters for the buffer lock about 2336 * identity change by freeing the buffer. 2337 */ 2338 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2339 bp->b_flags |= B_INVAL; 2340 bfreekva(bp); 2341 brelse(bp); 2342 goto restart; 2343 } 2344 2345 if (metadata) 2346 break; 2347 2348 /* 2349 * If we are overcomitted then recover the buffer and its 2350 * KVM space. This occurs in rare situations when multiple 2351 * processes are blocked in getnewbuf() or allocbuf(). 2352 */ 2353 if (bufspace >= hibufspace) 2354 flushingbufs = 1; 2355 if (flushingbufs && bp->b_kvasize != 0) { 2356 bp->b_flags |= B_INVAL; 2357 bfreekva(bp); 2358 brelse(bp); 2359 goto restart; 2360 } 2361 if (bufspace < lobufspace) 2362 flushingbufs = 0; 2363 break; 2364 } 2365 return (bp); 2366 } 2367 2368 /* 2369 * getnewbuf: 2370 * 2371 * Find and initialize a new buffer header, freeing up existing buffers 2372 * in the bufqueues as necessary. The new buffer is returned locked. 2373 * 2374 * Important: B_INVAL is not set. If the caller wishes to throw the 2375 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2376 * 2377 * We block if: 2378 * We have insufficient buffer headers 2379 * We have insufficient buffer space 2380 * buffer_map is too fragmented ( space reservation fails ) 2381 * If we have to flush dirty buffers ( but we try to avoid this ) 2382 * 2383 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 2384 * Instead we ask the buf daemon to do it for us. We attempt to 2385 * avoid piecemeal wakeups of the pageout daemon. 2386 */ 2387 static struct buf * 2388 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2389 int gbflags) 2390 { 2391 struct buf *bp; 2392 int defrag, metadata; 2393 2394 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2395 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2396 if (!unmapped_buf_allowed) 2397 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2398 2399 defrag = 0; 2400 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2401 vp->v_type == VCHR) 2402 metadata = 1; 2403 else 2404 metadata = 0; 2405 /* 2406 * We can't afford to block since we might be holding a vnode lock, 2407 * which may prevent system daemons from running. We deal with 2408 * low-memory situations by proactively returning memory and running 2409 * async I/O rather then sync I/O. 2410 */ 2411 atomic_add_int(&getnewbufcalls, 1); 2412 atomic_subtract_int(&getnewbufrestarts, 1); 2413 restart: 2414 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2415 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2416 if (bp != NULL) 2417 defrag = 0; 2418 2419 /* 2420 * If we exhausted our list, sleep as appropriate. We may have to 2421 * wakeup various daemons and write out some dirty buffers. 2422 * 2423 * Generally we are sleeping due to insufficient buffer space. 2424 */ 2425 if (bp == NULL) { 2426 mtx_assert(&bqlock, MA_OWNED); 2427 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2428 mtx_assert(&bqlock, MA_NOTOWNED); 2429 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2430 mtx_assert(&bqlock, MA_NOTOWNED); 2431 2432 bfreekva(bp); 2433 bp->b_flags |= B_UNMAPPED; 2434 bp->b_kvabase = bp->b_data = unmapped_buf; 2435 bp->b_kvasize = maxsize; 2436 atomic_add_long(&bufspace, bp->b_kvasize); 2437 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2438 atomic_add_int(&bufreusecnt, 1); 2439 } else { 2440 mtx_assert(&bqlock, MA_NOTOWNED); 2441 2442 /* 2443 * We finally have a valid bp. We aren't quite out of the 2444 * woods, we still have to reserve kva space. In order 2445 * to keep fragmentation sane we only allocate kva in 2446 * BKVASIZE chunks. 2447 */ 2448 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2449 2450 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2451 B_KVAALLOC)) == B_UNMAPPED) { 2452 if (allocbufkva(bp, maxsize, gbflags)) { 2453 defrag = 1; 2454 bp->b_flags |= B_INVAL; 2455 brelse(bp); 2456 goto restart; 2457 } 2458 atomic_add_int(&bufreusecnt, 1); 2459 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2460 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2461 /* 2462 * If the reused buffer has KVA allocated, 2463 * reassign b_kvaalloc to b_kvabase. 2464 */ 2465 bp->b_kvabase = bp->b_kvaalloc; 2466 bp->b_flags &= ~B_KVAALLOC; 2467 atomic_subtract_long(&unmapped_bufspace, 2468 bp->b_kvasize); 2469 atomic_add_int(&bufreusecnt, 1); 2470 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2471 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2472 GB_KVAALLOC)) { 2473 /* 2474 * The case of reused buffer already have KVA 2475 * mapped, but the request is for unmapped 2476 * buffer with KVA allocated. 2477 */ 2478 bp->b_kvaalloc = bp->b_kvabase; 2479 bp->b_data = bp->b_kvabase = unmapped_buf; 2480 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2481 atomic_add_long(&unmapped_bufspace, 2482 bp->b_kvasize); 2483 atomic_add_int(&bufreusecnt, 1); 2484 } 2485 if ((gbflags & GB_UNMAPPED) == 0) { 2486 bp->b_saveaddr = bp->b_kvabase; 2487 bp->b_data = bp->b_saveaddr; 2488 bp->b_flags &= ~B_UNMAPPED; 2489 BUF_CHECK_MAPPED(bp); 2490 } 2491 } 2492 return (bp); 2493 } 2494 2495 /* 2496 * buf_daemon: 2497 * 2498 * buffer flushing daemon. Buffers are normally flushed by the 2499 * update daemon but if it cannot keep up this process starts to 2500 * take the load in an attempt to prevent getnewbuf() from blocking. 2501 */ 2502 2503 static struct kproc_desc buf_kp = { 2504 "bufdaemon", 2505 buf_daemon, 2506 &bufdaemonproc 2507 }; 2508 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2509 2510 static int 2511 buf_do_flush(struct vnode *vp) 2512 { 2513 int flushed; 2514 2515 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2516 if (flushed == 0) { 2517 /* 2518 * Could not find any buffers without rollback 2519 * dependencies, so just write the first one 2520 * in the hopes of eventually making progress. 2521 */ 2522 flushbufqueues(vp, QUEUE_DIRTY, 1); 2523 } 2524 return (flushed); 2525 } 2526 2527 static void 2528 buf_daemon() 2529 { 2530 int lodirtysave; 2531 2532 /* 2533 * This process needs to be suspended prior to shutdown sync. 2534 */ 2535 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2536 SHUTDOWN_PRI_LAST); 2537 2538 /* 2539 * This process is allowed to take the buffer cache to the limit 2540 */ 2541 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2542 mtx_lock(&bdlock); 2543 for (;;) { 2544 bd_request = 0; 2545 mtx_unlock(&bdlock); 2546 2547 kproc_suspend_check(bufdaemonproc); 2548 lodirtysave = lodirtybuffers; 2549 if (bd_speedupreq) { 2550 lodirtybuffers = numdirtybuffers / 2; 2551 bd_speedupreq = 0; 2552 } 2553 /* 2554 * Do the flush. Limit the amount of in-transit I/O we 2555 * allow to build up, otherwise we would completely saturate 2556 * the I/O system. Wakeup any waiting processes before we 2557 * normally would so they can run in parallel with our drain. 2558 */ 2559 while (numdirtybuffers > lodirtybuffers) { 2560 if (buf_do_flush(NULL) == 0) 2561 break; 2562 kern_yield(PRI_USER); 2563 } 2564 lodirtybuffers = lodirtysave; 2565 2566 /* 2567 * Only clear bd_request if we have reached our low water 2568 * mark. The buf_daemon normally waits 1 second and 2569 * then incrementally flushes any dirty buffers that have 2570 * built up, within reason. 2571 * 2572 * If we were unable to hit our low water mark and couldn't 2573 * find any flushable buffers, we sleep half a second. 2574 * Otherwise we loop immediately. 2575 */ 2576 mtx_lock(&bdlock); 2577 if (numdirtybuffers <= lodirtybuffers) { 2578 /* 2579 * We reached our low water mark, reset the 2580 * request and sleep until we are needed again. 2581 * The sleep is just so the suspend code works. 2582 */ 2583 bd_request = 0; 2584 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2585 } else { 2586 /* 2587 * We couldn't find any flushable dirty buffers but 2588 * still have too many dirty buffers, we 2589 * have to sleep and try again. (rare) 2590 */ 2591 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2592 } 2593 } 2594 } 2595 2596 /* 2597 * flushbufqueues: 2598 * 2599 * Try to flush a buffer in the dirty queue. We must be careful to 2600 * free up B_INVAL buffers instead of write them, which NFS is 2601 * particularly sensitive to. 2602 */ 2603 static int flushwithdeps = 0; 2604 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2605 0, "Number of buffers flushed with dependecies that require rollbacks"); 2606 2607 static int 2608 flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2609 { 2610 struct buf *sentinel; 2611 struct vnode *vp; 2612 struct mount *mp; 2613 struct buf *bp; 2614 int hasdeps; 2615 int flushed; 2616 int target; 2617 2618 if (lvp == NULL) { 2619 target = numdirtybuffers - lodirtybuffers; 2620 if (flushdeps && target > 2) 2621 target /= 2; 2622 } else 2623 target = flushbufqtarget; 2624 flushed = 0; 2625 bp = NULL; 2626 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2627 sentinel->b_qindex = QUEUE_SENTINEL; 2628 mtx_lock(&bqlock); 2629 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2630 while (flushed != target) { 2631 bp = TAILQ_NEXT(sentinel, b_freelist); 2632 if (bp != NULL) { 2633 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2634 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2635 b_freelist); 2636 } else 2637 break; 2638 /* 2639 * Skip sentinels inserted by other invocations of the 2640 * flushbufqueues(), taking care to not reorder them. 2641 */ 2642 if (bp->b_qindex == QUEUE_SENTINEL) 2643 continue; 2644 /* 2645 * Only flush the buffers that belong to the 2646 * vnode locked by the curthread. 2647 */ 2648 if (lvp != NULL && bp->b_vp != lvp) 2649 continue; 2650 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2651 continue; 2652 if (bp->b_pin_count > 0) { 2653 BUF_UNLOCK(bp); 2654 continue; 2655 } 2656 BO_LOCK(bp->b_bufobj); 2657 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2658 (bp->b_flags & B_DELWRI) == 0) { 2659 BO_UNLOCK(bp->b_bufobj); 2660 BUF_UNLOCK(bp); 2661 continue; 2662 } 2663 BO_UNLOCK(bp->b_bufobj); 2664 if (bp->b_flags & B_INVAL) { 2665 bremfreel(bp); 2666 mtx_unlock(&bqlock); 2667 brelse(bp); 2668 flushed++; 2669 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2670 mtx_lock(&bqlock); 2671 continue; 2672 } 2673 2674 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2675 if (flushdeps == 0) { 2676 BUF_UNLOCK(bp); 2677 continue; 2678 } 2679 hasdeps = 1; 2680 } else 2681 hasdeps = 0; 2682 /* 2683 * We must hold the lock on a vnode before writing 2684 * one of its buffers. Otherwise we may confuse, or 2685 * in the case of a snapshot vnode, deadlock the 2686 * system. 2687 * 2688 * The lock order here is the reverse of the normal 2689 * of vnode followed by buf lock. This is ok because 2690 * the NOWAIT will prevent deadlock. 2691 */ 2692 vp = bp->b_vp; 2693 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2694 BUF_UNLOCK(bp); 2695 continue; 2696 } 2697 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2698 mtx_unlock(&bqlock); 2699 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2700 bp, bp->b_vp, bp->b_flags); 2701 if (curproc == bufdaemonproc) 2702 vfs_bio_awrite(bp); 2703 else { 2704 bremfree(bp); 2705 bwrite(bp); 2706 notbufdflashes++; 2707 } 2708 vn_finished_write(mp); 2709 VOP_UNLOCK(vp, 0); 2710 flushwithdeps += hasdeps; 2711 flushed++; 2712 2713 /* 2714 * Sleeping on runningbufspace while holding 2715 * vnode lock leads to deadlock. 2716 */ 2717 if (curproc == bufdaemonproc) 2718 waitrunningbufspace(); 2719 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2720 mtx_lock(&bqlock); 2721 continue; 2722 } 2723 vn_finished_write(mp); 2724 BUF_UNLOCK(bp); 2725 } 2726 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2727 mtx_unlock(&bqlock); 2728 free(sentinel, M_TEMP); 2729 return (flushed); 2730 } 2731 2732 /* 2733 * Check to see if a block is currently memory resident. 2734 */ 2735 struct buf * 2736 incore(struct bufobj *bo, daddr_t blkno) 2737 { 2738 struct buf *bp; 2739 2740 BO_LOCK(bo); 2741 bp = gbincore(bo, blkno); 2742 BO_UNLOCK(bo); 2743 return (bp); 2744 } 2745 2746 /* 2747 * Returns true if no I/O is needed to access the 2748 * associated VM object. This is like incore except 2749 * it also hunts around in the VM system for the data. 2750 */ 2751 2752 static int 2753 inmem(struct vnode * vp, daddr_t blkno) 2754 { 2755 vm_object_t obj; 2756 vm_offset_t toff, tinc, size; 2757 vm_page_t m; 2758 vm_ooffset_t off; 2759 2760 ASSERT_VOP_LOCKED(vp, "inmem"); 2761 2762 if (incore(&vp->v_bufobj, blkno)) 2763 return 1; 2764 if (vp->v_mount == NULL) 2765 return 0; 2766 obj = vp->v_object; 2767 if (obj == NULL) 2768 return (0); 2769 2770 size = PAGE_SIZE; 2771 if (size > vp->v_mount->mnt_stat.f_iosize) 2772 size = vp->v_mount->mnt_stat.f_iosize; 2773 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2774 2775 VM_OBJECT_WLOCK(obj); 2776 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2777 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2778 if (!m) 2779 goto notinmem; 2780 tinc = size; 2781 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2782 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2783 if (vm_page_is_valid(m, 2784 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2785 goto notinmem; 2786 } 2787 VM_OBJECT_WUNLOCK(obj); 2788 return 1; 2789 2790 notinmem: 2791 VM_OBJECT_WUNLOCK(obj); 2792 return (0); 2793 } 2794 2795 /* 2796 * Set the dirty range for a buffer based on the status of the dirty 2797 * bits in the pages comprising the buffer. The range is limited 2798 * to the size of the buffer. 2799 * 2800 * Tell the VM system that the pages associated with this buffer 2801 * are clean. This is used for delayed writes where the data is 2802 * going to go to disk eventually without additional VM intevention. 2803 * 2804 * Note that while we only really need to clean through to b_bcount, we 2805 * just go ahead and clean through to b_bufsize. 2806 */ 2807 static void 2808 vfs_clean_pages_dirty_buf(struct buf *bp) 2809 { 2810 vm_ooffset_t foff, noff, eoff; 2811 vm_page_t m; 2812 int i; 2813 2814 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2815 return; 2816 2817 foff = bp->b_offset; 2818 KASSERT(bp->b_offset != NOOFFSET, 2819 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2820 2821 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2822 vfs_drain_busy_pages(bp); 2823 vfs_setdirty_locked_object(bp); 2824 for (i = 0; i < bp->b_npages; i++) { 2825 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2826 eoff = noff; 2827 if (eoff > bp->b_offset + bp->b_bufsize) 2828 eoff = bp->b_offset + bp->b_bufsize; 2829 m = bp->b_pages[i]; 2830 vfs_page_set_validclean(bp, foff, m); 2831 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2832 foff = noff; 2833 } 2834 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2835 } 2836 2837 static void 2838 vfs_setdirty_locked_object(struct buf *bp) 2839 { 2840 vm_object_t object; 2841 int i; 2842 2843 object = bp->b_bufobj->bo_object; 2844 VM_OBJECT_ASSERT_WLOCKED(object); 2845 2846 /* 2847 * We qualify the scan for modified pages on whether the 2848 * object has been flushed yet. 2849 */ 2850 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2851 vm_offset_t boffset; 2852 vm_offset_t eoffset; 2853 2854 /* 2855 * test the pages to see if they have been modified directly 2856 * by users through the VM system. 2857 */ 2858 for (i = 0; i < bp->b_npages; i++) 2859 vm_page_test_dirty(bp->b_pages[i]); 2860 2861 /* 2862 * Calculate the encompassing dirty range, boffset and eoffset, 2863 * (eoffset - boffset) bytes. 2864 */ 2865 2866 for (i = 0; i < bp->b_npages; i++) { 2867 if (bp->b_pages[i]->dirty) 2868 break; 2869 } 2870 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2871 2872 for (i = bp->b_npages - 1; i >= 0; --i) { 2873 if (bp->b_pages[i]->dirty) { 2874 break; 2875 } 2876 } 2877 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2878 2879 /* 2880 * Fit it to the buffer. 2881 */ 2882 2883 if (eoffset > bp->b_bcount) 2884 eoffset = bp->b_bcount; 2885 2886 /* 2887 * If we have a good dirty range, merge with the existing 2888 * dirty range. 2889 */ 2890 2891 if (boffset < eoffset) { 2892 if (bp->b_dirtyoff > boffset) 2893 bp->b_dirtyoff = boffset; 2894 if (bp->b_dirtyend < eoffset) 2895 bp->b_dirtyend = eoffset; 2896 } 2897 } 2898 } 2899 2900 /* 2901 * Allocate the KVA mapping for an existing buffer. It handles the 2902 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2903 * KVA which is not mapped (B_KVAALLOC). 2904 */ 2905 static void 2906 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2907 { 2908 struct buf *scratch_bp; 2909 int bsize, maxsize, need_mapping, need_kva; 2910 off_t offset; 2911 2912 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2913 (gbflags & GB_UNMAPPED) == 0; 2914 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2915 (gbflags & GB_KVAALLOC) != 0; 2916 if (!need_mapping && !need_kva) 2917 return; 2918 2919 BUF_CHECK_UNMAPPED(bp); 2920 2921 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2922 /* 2923 * Buffer is not mapped, but the KVA was already 2924 * reserved at the time of the instantiation. Use the 2925 * allocated space. 2926 */ 2927 bp->b_flags &= ~B_KVAALLOC; 2928 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2929 bp->b_kvabase = bp->b_kvaalloc; 2930 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2931 goto has_addr; 2932 } 2933 2934 /* 2935 * Calculate the amount of the address space we would reserve 2936 * if the buffer was mapped. 2937 */ 2938 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2939 offset = blkno * bsize; 2940 maxsize = size + (offset & PAGE_MASK); 2941 maxsize = imax(maxsize, bsize); 2942 2943 mapping_loop: 2944 if (allocbufkva(bp, maxsize, gbflags)) { 2945 /* 2946 * Request defragmentation. getnewbuf() returns us the 2947 * allocated space by the scratch buffer KVA. 2948 */ 2949 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2950 (GB_UNMAPPED | GB_KVAALLOC)); 2951 if (scratch_bp == NULL) { 2952 if ((gbflags & GB_NOWAIT_BD) != 0) { 2953 /* 2954 * XXXKIB: defragmentation cannot 2955 * succeed, not sure what else to do. 2956 */ 2957 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 2958 } 2959 atomic_add_int(&mappingrestarts, 1); 2960 goto mapping_loop; 2961 } 2962 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 2963 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 2964 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 2965 scratch_bp->b_kvasize, gbflags); 2966 2967 /* Get rid of the scratch buffer. */ 2968 scratch_bp->b_kvasize = 0; 2969 scratch_bp->b_flags |= B_INVAL; 2970 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 2971 brelse(scratch_bp); 2972 } 2973 if (!need_mapping) 2974 return; 2975 2976 has_addr: 2977 bp->b_saveaddr = bp->b_kvabase; 2978 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 2979 bp->b_flags &= ~B_UNMAPPED; 2980 BUF_CHECK_MAPPED(bp); 2981 bpmap_qenter(bp); 2982 } 2983 2984 /* 2985 * getblk: 2986 * 2987 * Get a block given a specified block and offset into a file/device. 2988 * The buffers B_DONE bit will be cleared on return, making it almost 2989 * ready for an I/O initiation. B_INVAL may or may not be set on 2990 * return. The caller should clear B_INVAL prior to initiating a 2991 * READ. 2992 * 2993 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2994 * an existing buffer. 2995 * 2996 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2997 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2998 * and then cleared based on the backing VM. If the previous buffer is 2999 * non-0-sized but invalid, B_CACHE will be cleared. 3000 * 3001 * If getblk() must create a new buffer, the new buffer is returned with 3002 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3003 * case it is returned with B_INVAL clear and B_CACHE set based on the 3004 * backing VM. 3005 * 3006 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3007 * B_CACHE bit is clear. 3008 * 3009 * What this means, basically, is that the caller should use B_CACHE to 3010 * determine whether the buffer is fully valid or not and should clear 3011 * B_INVAL prior to issuing a read. If the caller intends to validate 3012 * the buffer by loading its data area with something, the caller needs 3013 * to clear B_INVAL. If the caller does this without issuing an I/O, 3014 * the caller should set B_CACHE ( as an optimization ), else the caller 3015 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3016 * a write attempt or if it was a successfull read. If the caller 3017 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3018 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3019 */ 3020 struct buf * 3021 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3022 int flags) 3023 { 3024 struct buf *bp; 3025 struct bufobj *bo; 3026 int bsize, error, maxsize, vmio; 3027 off_t offset; 3028 3029 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3030 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3031 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3032 ASSERT_VOP_LOCKED(vp, "getblk"); 3033 if (size > MAXBSIZE) 3034 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 3035 if (!unmapped_buf_allowed) 3036 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3037 3038 bo = &vp->v_bufobj; 3039 loop: 3040 /* 3041 * Block if we are low on buffers. Certain processes are allowed 3042 * to completely exhaust the buffer cache. 3043 * 3044 * If this check ever becomes a bottleneck it may be better to 3045 * move it into the else, when gbincore() fails. At the moment 3046 * it isn't a problem. 3047 */ 3048 if (numfreebuffers == 0) { 3049 if (TD_IS_IDLETHREAD(curthread)) 3050 return NULL; 3051 mtx_lock(&nblock); 3052 needsbuffer |= VFS_BIO_NEED_ANY; 3053 mtx_unlock(&nblock); 3054 } 3055 3056 BO_LOCK(bo); 3057 bp = gbincore(bo, blkno); 3058 if (bp != NULL) { 3059 int lockflags; 3060 /* 3061 * Buffer is in-core. If the buffer is not busy nor managed, 3062 * it must be on a queue. 3063 */ 3064 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3065 3066 if (flags & GB_LOCK_NOWAIT) 3067 lockflags |= LK_NOWAIT; 3068 3069 error = BUF_TIMELOCK(bp, lockflags, 3070 BO_MTX(bo), "getblk", slpflag, slptimeo); 3071 3072 /* 3073 * If we slept and got the lock we have to restart in case 3074 * the buffer changed identities. 3075 */ 3076 if (error == ENOLCK) 3077 goto loop; 3078 /* We timed out or were interrupted. */ 3079 else if (error) 3080 return (NULL); 3081 /* If recursed, assume caller knows the rules. */ 3082 else if (BUF_LOCKRECURSED(bp)) 3083 goto end; 3084 3085 /* 3086 * The buffer is locked. B_CACHE is cleared if the buffer is 3087 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3088 * and for a VMIO buffer B_CACHE is adjusted according to the 3089 * backing VM cache. 3090 */ 3091 if (bp->b_flags & B_INVAL) 3092 bp->b_flags &= ~B_CACHE; 3093 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3094 bp->b_flags |= B_CACHE; 3095 if (bp->b_flags & B_MANAGED) 3096 MPASS(bp->b_qindex == QUEUE_NONE); 3097 else { 3098 BO_LOCK(bo); 3099 bremfree(bp); 3100 BO_UNLOCK(bo); 3101 } 3102 3103 /* 3104 * check for size inconsistencies for non-VMIO case. 3105 */ 3106 if (bp->b_bcount != size) { 3107 if ((bp->b_flags & B_VMIO) == 0 || 3108 (size > bp->b_kvasize)) { 3109 if (bp->b_flags & B_DELWRI) { 3110 /* 3111 * If buffer is pinned and caller does 3112 * not want sleep waiting for it to be 3113 * unpinned, bail out 3114 * */ 3115 if (bp->b_pin_count > 0) { 3116 if (flags & GB_LOCK_NOWAIT) { 3117 bqrelse(bp); 3118 return (NULL); 3119 } else { 3120 bunpin_wait(bp); 3121 } 3122 } 3123 bp->b_flags |= B_NOCACHE; 3124 bwrite(bp); 3125 } else { 3126 if (LIST_EMPTY(&bp->b_dep)) { 3127 bp->b_flags |= B_RELBUF; 3128 brelse(bp); 3129 } else { 3130 bp->b_flags |= B_NOCACHE; 3131 bwrite(bp); 3132 } 3133 } 3134 goto loop; 3135 } 3136 } 3137 3138 /* 3139 * Handle the case of unmapped buffer which should 3140 * become mapped, or the buffer for which KVA 3141 * reservation is requested. 3142 */ 3143 bp_unmapped_get_kva(bp, blkno, size, flags); 3144 3145 /* 3146 * If the size is inconsistant in the VMIO case, we can resize 3147 * the buffer. This might lead to B_CACHE getting set or 3148 * cleared. If the size has not changed, B_CACHE remains 3149 * unchanged from its previous state. 3150 */ 3151 if (bp->b_bcount != size) 3152 allocbuf(bp, size); 3153 3154 KASSERT(bp->b_offset != NOOFFSET, 3155 ("getblk: no buffer offset")); 3156 3157 /* 3158 * A buffer with B_DELWRI set and B_CACHE clear must 3159 * be committed before we can return the buffer in 3160 * order to prevent the caller from issuing a read 3161 * ( due to B_CACHE not being set ) and overwriting 3162 * it. 3163 * 3164 * Most callers, including NFS and FFS, need this to 3165 * operate properly either because they assume they 3166 * can issue a read if B_CACHE is not set, or because 3167 * ( for example ) an uncached B_DELWRI might loop due 3168 * to softupdates re-dirtying the buffer. In the latter 3169 * case, B_CACHE is set after the first write completes, 3170 * preventing further loops. 3171 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3172 * above while extending the buffer, we cannot allow the 3173 * buffer to remain with B_CACHE set after the write 3174 * completes or it will represent a corrupt state. To 3175 * deal with this we set B_NOCACHE to scrap the buffer 3176 * after the write. 3177 * 3178 * We might be able to do something fancy, like setting 3179 * B_CACHE in bwrite() except if B_DELWRI is already set, 3180 * so the below call doesn't set B_CACHE, but that gets real 3181 * confusing. This is much easier. 3182 */ 3183 3184 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3185 bp->b_flags |= B_NOCACHE; 3186 bwrite(bp); 3187 goto loop; 3188 } 3189 bp->b_flags &= ~B_DONE; 3190 } else { 3191 /* 3192 * Buffer is not in-core, create new buffer. The buffer 3193 * returned by getnewbuf() is locked. Note that the returned 3194 * buffer is also considered valid (not marked B_INVAL). 3195 */ 3196 BO_UNLOCK(bo); 3197 /* 3198 * If the user does not want us to create the buffer, bail out 3199 * here. 3200 */ 3201 if (flags & GB_NOCREAT) 3202 return NULL; 3203 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3204 offset = blkno * bsize; 3205 vmio = vp->v_object != NULL; 3206 if (vmio) { 3207 maxsize = size + (offset & PAGE_MASK); 3208 } else { 3209 maxsize = size; 3210 /* Do not allow non-VMIO notmapped buffers. */ 3211 flags &= ~GB_UNMAPPED; 3212 } 3213 maxsize = imax(maxsize, bsize); 3214 3215 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3216 if (bp == NULL) { 3217 if (slpflag || slptimeo) 3218 return NULL; 3219 goto loop; 3220 } 3221 3222 /* 3223 * This code is used to make sure that a buffer is not 3224 * created while the getnewbuf routine is blocked. 3225 * This can be a problem whether the vnode is locked or not. 3226 * If the buffer is created out from under us, we have to 3227 * throw away the one we just created. 3228 * 3229 * Note: this must occur before we associate the buffer 3230 * with the vp especially considering limitations in 3231 * the splay tree implementation when dealing with duplicate 3232 * lblkno's. 3233 */ 3234 BO_LOCK(bo); 3235 if (gbincore(bo, blkno)) { 3236 BO_UNLOCK(bo); 3237 bp->b_flags |= B_INVAL; 3238 brelse(bp); 3239 goto loop; 3240 } 3241 3242 /* 3243 * Insert the buffer into the hash, so that it can 3244 * be found by incore. 3245 */ 3246 bp->b_blkno = bp->b_lblkno = blkno; 3247 bp->b_offset = offset; 3248 bgetvp(vp, bp); 3249 BO_UNLOCK(bo); 3250 3251 /* 3252 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3253 * buffer size starts out as 0, B_CACHE will be set by 3254 * allocbuf() for the VMIO case prior to it testing the 3255 * backing store for validity. 3256 */ 3257 3258 if (vmio) { 3259 bp->b_flags |= B_VMIO; 3260 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3261 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3262 bp, vp->v_object, bp->b_bufobj->bo_object)); 3263 } else { 3264 bp->b_flags &= ~B_VMIO; 3265 KASSERT(bp->b_bufobj->bo_object == NULL, 3266 ("ARGH! has b_bufobj->bo_object %p %p\n", 3267 bp, bp->b_bufobj->bo_object)); 3268 BUF_CHECK_MAPPED(bp); 3269 } 3270 3271 allocbuf(bp, size); 3272 bp->b_flags &= ~B_DONE; 3273 } 3274 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3275 BUF_ASSERT_HELD(bp); 3276 end: 3277 KASSERT(bp->b_bufobj == bo, 3278 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3279 return (bp); 3280 } 3281 3282 /* 3283 * Get an empty, disassociated buffer of given size. The buffer is initially 3284 * set to B_INVAL. 3285 */ 3286 struct buf * 3287 geteblk(int size, int flags) 3288 { 3289 struct buf *bp; 3290 int maxsize; 3291 3292 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3293 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3294 if ((flags & GB_NOWAIT_BD) && 3295 (curthread->td_pflags & TDP_BUFNEED) != 0) 3296 return (NULL); 3297 } 3298 allocbuf(bp, size); 3299 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3300 BUF_ASSERT_HELD(bp); 3301 return (bp); 3302 } 3303 3304 3305 /* 3306 * This code constitutes the buffer memory from either anonymous system 3307 * memory (in the case of non-VMIO operations) or from an associated 3308 * VM object (in the case of VMIO operations). This code is able to 3309 * resize a buffer up or down. 3310 * 3311 * Note that this code is tricky, and has many complications to resolve 3312 * deadlock or inconsistant data situations. Tread lightly!!! 3313 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3314 * the caller. Calling this code willy nilly can result in the loss of data. 3315 * 3316 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3317 * B_CACHE for the non-VMIO case. 3318 */ 3319 3320 int 3321 allocbuf(struct buf *bp, int size) 3322 { 3323 int newbsize, mbsize; 3324 int i; 3325 3326 BUF_ASSERT_HELD(bp); 3327 3328 if (bp->b_kvasize < size) 3329 panic("allocbuf: buffer too small"); 3330 3331 if ((bp->b_flags & B_VMIO) == 0) { 3332 caddr_t origbuf; 3333 int origbufsize; 3334 /* 3335 * Just get anonymous memory from the kernel. Don't 3336 * mess with B_CACHE. 3337 */ 3338 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3339 if (bp->b_flags & B_MALLOC) 3340 newbsize = mbsize; 3341 else 3342 newbsize = round_page(size); 3343 3344 if (newbsize < bp->b_bufsize) { 3345 /* 3346 * malloced buffers are not shrunk 3347 */ 3348 if (bp->b_flags & B_MALLOC) { 3349 if (newbsize) { 3350 bp->b_bcount = size; 3351 } else { 3352 free(bp->b_data, M_BIOBUF); 3353 if (bp->b_bufsize) { 3354 atomic_subtract_long( 3355 &bufmallocspace, 3356 bp->b_bufsize); 3357 bufspacewakeup(); 3358 bp->b_bufsize = 0; 3359 } 3360 bp->b_saveaddr = bp->b_kvabase; 3361 bp->b_data = bp->b_saveaddr; 3362 bp->b_bcount = 0; 3363 bp->b_flags &= ~B_MALLOC; 3364 } 3365 return 1; 3366 } 3367 vm_hold_free_pages(bp, newbsize); 3368 } else if (newbsize > bp->b_bufsize) { 3369 /* 3370 * We only use malloced memory on the first allocation. 3371 * and revert to page-allocated memory when the buffer 3372 * grows. 3373 */ 3374 /* 3375 * There is a potential smp race here that could lead 3376 * to bufmallocspace slightly passing the max. It 3377 * is probably extremely rare and not worth worrying 3378 * over. 3379 */ 3380 if ( (bufmallocspace < maxbufmallocspace) && 3381 (bp->b_bufsize == 0) && 3382 (mbsize <= PAGE_SIZE/2)) { 3383 3384 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3385 bp->b_bufsize = mbsize; 3386 bp->b_bcount = size; 3387 bp->b_flags |= B_MALLOC; 3388 atomic_add_long(&bufmallocspace, mbsize); 3389 return 1; 3390 } 3391 origbuf = NULL; 3392 origbufsize = 0; 3393 /* 3394 * If the buffer is growing on its other-than-first allocation, 3395 * then we revert to the page-allocation scheme. 3396 */ 3397 if (bp->b_flags & B_MALLOC) { 3398 origbuf = bp->b_data; 3399 origbufsize = bp->b_bufsize; 3400 bp->b_data = bp->b_kvabase; 3401 if (bp->b_bufsize) { 3402 atomic_subtract_long(&bufmallocspace, 3403 bp->b_bufsize); 3404 bufspacewakeup(); 3405 bp->b_bufsize = 0; 3406 } 3407 bp->b_flags &= ~B_MALLOC; 3408 newbsize = round_page(newbsize); 3409 } 3410 vm_hold_load_pages( 3411 bp, 3412 (vm_offset_t) bp->b_data + bp->b_bufsize, 3413 (vm_offset_t) bp->b_data + newbsize); 3414 if (origbuf) { 3415 bcopy(origbuf, bp->b_data, origbufsize); 3416 free(origbuf, M_BIOBUF); 3417 } 3418 } 3419 } else { 3420 int desiredpages; 3421 3422 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3423 desiredpages = (size == 0) ? 0 : 3424 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3425 3426 if (bp->b_flags & B_MALLOC) 3427 panic("allocbuf: VMIO buffer can't be malloced"); 3428 /* 3429 * Set B_CACHE initially if buffer is 0 length or will become 3430 * 0-length. 3431 */ 3432 if (size == 0 || bp->b_bufsize == 0) 3433 bp->b_flags |= B_CACHE; 3434 3435 if (newbsize < bp->b_bufsize) { 3436 /* 3437 * DEV_BSIZE aligned new buffer size is less then the 3438 * DEV_BSIZE aligned existing buffer size. Figure out 3439 * if we have to remove any pages. 3440 */ 3441 if (desiredpages < bp->b_npages) { 3442 vm_page_t m; 3443 3444 if ((bp->b_flags & B_UNMAPPED) == 0) { 3445 BUF_CHECK_MAPPED(bp); 3446 pmap_qremove((vm_offset_t)trunc_page( 3447 (vm_offset_t)bp->b_data) + 3448 (desiredpages << PAGE_SHIFT), 3449 (bp->b_npages - desiredpages)); 3450 } else 3451 BUF_CHECK_UNMAPPED(bp); 3452 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3453 for (i = desiredpages; i < bp->b_npages; i++) { 3454 /* 3455 * the page is not freed here -- it 3456 * is the responsibility of 3457 * vnode_pager_setsize 3458 */ 3459 m = bp->b_pages[i]; 3460 KASSERT(m != bogus_page, 3461 ("allocbuf: bogus page found")); 3462 while (vm_page_sleep_if_busy(m, TRUE, 3463 "biodep")) 3464 continue; 3465 3466 bp->b_pages[i] = NULL; 3467 vm_page_lock(m); 3468 vm_page_unwire(m, 0); 3469 vm_page_unlock(m); 3470 } 3471 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3472 bp->b_npages = desiredpages; 3473 } 3474 } else if (size > bp->b_bcount) { 3475 /* 3476 * We are growing the buffer, possibly in a 3477 * byte-granular fashion. 3478 */ 3479 vm_object_t obj; 3480 vm_offset_t toff; 3481 vm_offset_t tinc; 3482 3483 /* 3484 * Step 1, bring in the VM pages from the object, 3485 * allocating them if necessary. We must clear 3486 * B_CACHE if these pages are not valid for the 3487 * range covered by the buffer. 3488 */ 3489 3490 obj = bp->b_bufobj->bo_object; 3491 3492 VM_OBJECT_WLOCK(obj); 3493 while (bp->b_npages < desiredpages) { 3494 vm_page_t m; 3495 3496 /* 3497 * We must allocate system pages since blocking 3498 * here could interfere with paging I/O, no 3499 * matter which process we are. 3500 * 3501 * We can only test VPO_BUSY here. Blocking on 3502 * m->busy might lead to a deadlock: 3503 * vm_fault->getpages->cluster_read->allocbuf 3504 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3505 */ 3506 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3507 bp->b_npages, VM_ALLOC_NOBUSY | 3508 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3509 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3510 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3511 if (m->valid == 0) 3512 bp->b_flags &= ~B_CACHE; 3513 bp->b_pages[bp->b_npages] = m; 3514 ++bp->b_npages; 3515 } 3516 3517 /* 3518 * Step 2. We've loaded the pages into the buffer, 3519 * we have to figure out if we can still have B_CACHE 3520 * set. Note that B_CACHE is set according to the 3521 * byte-granular range ( bcount and size ), new the 3522 * aligned range ( newbsize ). 3523 * 3524 * The VM test is against m->valid, which is DEV_BSIZE 3525 * aligned. Needless to say, the validity of the data 3526 * needs to also be DEV_BSIZE aligned. Note that this 3527 * fails with NFS if the server or some other client 3528 * extends the file's EOF. If our buffer is resized, 3529 * B_CACHE may remain set! XXX 3530 */ 3531 3532 toff = bp->b_bcount; 3533 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3534 3535 while ((bp->b_flags & B_CACHE) && toff < size) { 3536 vm_pindex_t pi; 3537 3538 if (tinc > (size - toff)) 3539 tinc = size - toff; 3540 3541 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3542 PAGE_SHIFT; 3543 3544 vfs_buf_test_cache( 3545 bp, 3546 bp->b_offset, 3547 toff, 3548 tinc, 3549 bp->b_pages[pi] 3550 ); 3551 toff += tinc; 3552 tinc = PAGE_SIZE; 3553 } 3554 VM_OBJECT_WUNLOCK(obj); 3555 3556 /* 3557 * Step 3, fixup the KVM pmap. 3558 */ 3559 if ((bp->b_flags & B_UNMAPPED) == 0) 3560 bpmap_qenter(bp); 3561 else 3562 BUF_CHECK_UNMAPPED(bp); 3563 } 3564 } 3565 if (newbsize < bp->b_bufsize) 3566 bufspacewakeup(); 3567 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3568 bp->b_bcount = size; /* requested buffer size */ 3569 return 1; 3570 } 3571 3572 extern int inflight_transient_maps; 3573 3574 void 3575 biodone(struct bio *bp) 3576 { 3577 struct mtx *mtxp; 3578 void (*done)(struct bio *); 3579 vm_offset_t start, end; 3580 int transient; 3581 3582 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3583 mtx_lock(mtxp); 3584 bp->bio_flags |= BIO_DONE; 3585 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3586 start = trunc_page((vm_offset_t)bp->bio_data); 3587 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3588 transient = 1; 3589 } else { 3590 transient = 0; 3591 start = end = 0; 3592 } 3593 done = bp->bio_done; 3594 if (done == NULL) 3595 wakeup(bp); 3596 mtx_unlock(mtxp); 3597 if (done != NULL) 3598 done(bp); 3599 if (transient) { 3600 pmap_qremove(start, OFF_TO_IDX(end - start)); 3601 vm_map_remove(bio_transient_map, start, end); 3602 atomic_add_int(&inflight_transient_maps, -1); 3603 } 3604 } 3605 3606 /* 3607 * Wait for a BIO to finish. 3608 * 3609 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3610 * case is not yet clear. 3611 */ 3612 int 3613 biowait(struct bio *bp, const char *wchan) 3614 { 3615 struct mtx *mtxp; 3616 3617 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3618 mtx_lock(mtxp); 3619 while ((bp->bio_flags & BIO_DONE) == 0) 3620 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3621 mtx_unlock(mtxp); 3622 if (bp->bio_error != 0) 3623 return (bp->bio_error); 3624 if (!(bp->bio_flags & BIO_ERROR)) 3625 return (0); 3626 return (EIO); 3627 } 3628 3629 void 3630 biofinish(struct bio *bp, struct devstat *stat, int error) 3631 { 3632 3633 if (error) { 3634 bp->bio_error = error; 3635 bp->bio_flags |= BIO_ERROR; 3636 } 3637 if (stat != NULL) 3638 devstat_end_transaction_bio(stat, bp); 3639 biodone(bp); 3640 } 3641 3642 /* 3643 * bufwait: 3644 * 3645 * Wait for buffer I/O completion, returning error status. The buffer 3646 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3647 * error and cleared. 3648 */ 3649 int 3650 bufwait(struct buf *bp) 3651 { 3652 if (bp->b_iocmd == BIO_READ) 3653 bwait(bp, PRIBIO, "biord"); 3654 else 3655 bwait(bp, PRIBIO, "biowr"); 3656 if (bp->b_flags & B_EINTR) { 3657 bp->b_flags &= ~B_EINTR; 3658 return (EINTR); 3659 } 3660 if (bp->b_ioflags & BIO_ERROR) { 3661 return (bp->b_error ? bp->b_error : EIO); 3662 } else { 3663 return (0); 3664 } 3665 } 3666 3667 /* 3668 * Call back function from struct bio back up to struct buf. 3669 */ 3670 static void 3671 bufdonebio(struct bio *bip) 3672 { 3673 struct buf *bp; 3674 3675 bp = bip->bio_caller2; 3676 bp->b_resid = bp->b_bcount - bip->bio_completed; 3677 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3678 bp->b_ioflags = bip->bio_flags; 3679 bp->b_error = bip->bio_error; 3680 if (bp->b_error) 3681 bp->b_ioflags |= BIO_ERROR; 3682 bufdone(bp); 3683 g_destroy_bio(bip); 3684 } 3685 3686 void 3687 dev_strategy(struct cdev *dev, struct buf *bp) 3688 { 3689 struct cdevsw *csw; 3690 int ref; 3691 3692 KASSERT(dev->si_refcount > 0, 3693 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3694 devtoname(dev), dev)); 3695 3696 csw = dev_refthread(dev, &ref); 3697 dev_strategy_csw(dev, csw, bp); 3698 dev_relthread(dev, ref); 3699 } 3700 3701 void 3702 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3703 { 3704 struct bio *bip; 3705 3706 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3707 ("b_iocmd botch")); 3708 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3709 dev->si_threadcount > 0, 3710 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3711 dev)); 3712 if (csw == NULL) { 3713 bp->b_error = ENXIO; 3714 bp->b_ioflags = BIO_ERROR; 3715 bufdone(bp); 3716 return; 3717 } 3718 for (;;) { 3719 bip = g_new_bio(); 3720 if (bip != NULL) 3721 break; 3722 /* Try again later */ 3723 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3724 } 3725 bip->bio_cmd = bp->b_iocmd; 3726 bip->bio_offset = bp->b_iooffset; 3727 bip->bio_length = bp->b_bcount; 3728 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3729 bdata2bio(bp, bip); 3730 bip->bio_done = bufdonebio; 3731 bip->bio_caller2 = bp; 3732 bip->bio_dev = dev; 3733 (*csw->d_strategy)(bip); 3734 } 3735 3736 /* 3737 * bufdone: 3738 * 3739 * Finish I/O on a buffer, optionally calling a completion function. 3740 * This is usually called from an interrupt so process blocking is 3741 * not allowed. 3742 * 3743 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3744 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3745 * assuming B_INVAL is clear. 3746 * 3747 * For the VMIO case, we set B_CACHE if the op was a read and no 3748 * read error occured, or if the op was a write. B_CACHE is never 3749 * set if the buffer is invalid or otherwise uncacheable. 3750 * 3751 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3752 * initiator to leave B_INVAL set to brelse the buffer out of existance 3753 * in the biodone routine. 3754 */ 3755 void 3756 bufdone(struct buf *bp) 3757 { 3758 struct bufobj *dropobj; 3759 void (*biodone)(struct buf *); 3760 3761 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3762 dropobj = NULL; 3763 3764 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3765 BUF_ASSERT_HELD(bp); 3766 3767 runningbufwakeup(bp); 3768 if (bp->b_iocmd == BIO_WRITE) 3769 dropobj = bp->b_bufobj; 3770 /* call optional completion function if requested */ 3771 if (bp->b_iodone != NULL) { 3772 biodone = bp->b_iodone; 3773 bp->b_iodone = NULL; 3774 (*biodone) (bp); 3775 if (dropobj) 3776 bufobj_wdrop(dropobj); 3777 return; 3778 } 3779 3780 bufdone_finish(bp); 3781 3782 if (dropobj) 3783 bufobj_wdrop(dropobj); 3784 } 3785 3786 void 3787 bufdone_finish(struct buf *bp) 3788 { 3789 BUF_ASSERT_HELD(bp); 3790 3791 if (!LIST_EMPTY(&bp->b_dep)) 3792 buf_complete(bp); 3793 3794 if (bp->b_flags & B_VMIO) { 3795 vm_ooffset_t foff; 3796 vm_page_t m; 3797 vm_object_t obj; 3798 struct vnode *vp; 3799 int bogus, i, iosize; 3800 3801 obj = bp->b_bufobj->bo_object; 3802 KASSERT(obj->paging_in_progress >= bp->b_npages, 3803 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3804 obj->paging_in_progress, bp->b_npages)); 3805 3806 vp = bp->b_vp; 3807 KASSERT(vp->v_holdcnt > 0, 3808 ("biodone_finish: vnode %p has zero hold count", vp)); 3809 KASSERT(vp->v_object != NULL, 3810 ("biodone_finish: vnode %p has no vm_object", vp)); 3811 3812 foff = bp->b_offset; 3813 KASSERT(bp->b_offset != NOOFFSET, 3814 ("biodone_finish: bp %p has no buffer offset", bp)); 3815 3816 /* 3817 * Set B_CACHE if the op was a normal read and no error 3818 * occured. B_CACHE is set for writes in the b*write() 3819 * routines. 3820 */ 3821 iosize = bp->b_bcount - bp->b_resid; 3822 if (bp->b_iocmd == BIO_READ && 3823 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3824 !(bp->b_ioflags & BIO_ERROR)) { 3825 bp->b_flags |= B_CACHE; 3826 } 3827 bogus = 0; 3828 VM_OBJECT_WLOCK(obj); 3829 for (i = 0; i < bp->b_npages; i++) { 3830 int bogusflag = 0; 3831 int resid; 3832 3833 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3834 if (resid > iosize) 3835 resid = iosize; 3836 3837 /* 3838 * cleanup bogus pages, restoring the originals 3839 */ 3840 m = bp->b_pages[i]; 3841 if (m == bogus_page) { 3842 bogus = bogusflag = 1; 3843 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3844 if (m == NULL) 3845 panic("biodone: page disappeared!"); 3846 bp->b_pages[i] = m; 3847 } 3848 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3849 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3850 (intmax_t)foff, (uintmax_t)m->pindex)); 3851 3852 /* 3853 * In the write case, the valid and clean bits are 3854 * already changed correctly ( see bdwrite() ), so we 3855 * only need to do this here in the read case. 3856 */ 3857 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3858 KASSERT((m->dirty & vm_page_bits(foff & 3859 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3860 " page %p has unexpected dirty bits", m)); 3861 vfs_page_set_valid(bp, foff, m); 3862 } 3863 3864 vm_page_io_finish(m); 3865 vm_object_pip_subtract(obj, 1); 3866 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3867 iosize -= resid; 3868 } 3869 vm_object_pip_wakeupn(obj, 0); 3870 VM_OBJECT_WUNLOCK(obj); 3871 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3872 BUF_CHECK_MAPPED(bp); 3873 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3874 bp->b_pages, bp->b_npages); 3875 } 3876 } 3877 3878 /* 3879 * For asynchronous completions, release the buffer now. The brelse 3880 * will do a wakeup there if necessary - so no need to do a wakeup 3881 * here in the async case. The sync case always needs to do a wakeup. 3882 */ 3883 3884 if (bp->b_flags & B_ASYNC) { 3885 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3886 brelse(bp); 3887 else 3888 bqrelse(bp); 3889 } else 3890 bdone(bp); 3891 } 3892 3893 /* 3894 * This routine is called in lieu of iodone in the case of 3895 * incomplete I/O. This keeps the busy status for pages 3896 * consistant. 3897 */ 3898 void 3899 vfs_unbusy_pages(struct buf *bp) 3900 { 3901 int i; 3902 vm_object_t obj; 3903 vm_page_t m; 3904 3905 runningbufwakeup(bp); 3906 if (!(bp->b_flags & B_VMIO)) 3907 return; 3908 3909 obj = bp->b_bufobj->bo_object; 3910 VM_OBJECT_WLOCK(obj); 3911 for (i = 0; i < bp->b_npages; i++) { 3912 m = bp->b_pages[i]; 3913 if (m == bogus_page) { 3914 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3915 if (!m) 3916 panic("vfs_unbusy_pages: page missing\n"); 3917 bp->b_pages[i] = m; 3918 if ((bp->b_flags & B_UNMAPPED) == 0) { 3919 BUF_CHECK_MAPPED(bp); 3920 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3921 bp->b_pages, bp->b_npages); 3922 } else 3923 BUF_CHECK_UNMAPPED(bp); 3924 } 3925 vm_object_pip_subtract(obj, 1); 3926 vm_page_io_finish(m); 3927 } 3928 vm_object_pip_wakeupn(obj, 0); 3929 VM_OBJECT_WUNLOCK(obj); 3930 } 3931 3932 /* 3933 * vfs_page_set_valid: 3934 * 3935 * Set the valid bits in a page based on the supplied offset. The 3936 * range is restricted to the buffer's size. 3937 * 3938 * This routine is typically called after a read completes. 3939 */ 3940 static void 3941 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3942 { 3943 vm_ooffset_t eoff; 3944 3945 /* 3946 * Compute the end offset, eoff, such that [off, eoff) does not span a 3947 * page boundary and eoff is not greater than the end of the buffer. 3948 * The end of the buffer, in this case, is our file EOF, not the 3949 * allocation size of the buffer. 3950 */ 3951 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3952 if (eoff > bp->b_offset + bp->b_bcount) 3953 eoff = bp->b_offset + bp->b_bcount; 3954 3955 /* 3956 * Set valid range. This is typically the entire buffer and thus the 3957 * entire page. 3958 */ 3959 if (eoff > off) 3960 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3961 } 3962 3963 /* 3964 * vfs_page_set_validclean: 3965 * 3966 * Set the valid bits and clear the dirty bits in a page based on the 3967 * supplied offset. The range is restricted to the buffer's size. 3968 */ 3969 static void 3970 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3971 { 3972 vm_ooffset_t soff, eoff; 3973 3974 /* 3975 * Start and end offsets in buffer. eoff - soff may not cross a 3976 * page boundry or cross the end of the buffer. The end of the 3977 * buffer, in this case, is our file EOF, not the allocation size 3978 * of the buffer. 3979 */ 3980 soff = off; 3981 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3982 if (eoff > bp->b_offset + bp->b_bcount) 3983 eoff = bp->b_offset + bp->b_bcount; 3984 3985 /* 3986 * Set valid range. This is typically the entire buffer and thus the 3987 * entire page. 3988 */ 3989 if (eoff > soff) { 3990 vm_page_set_validclean( 3991 m, 3992 (vm_offset_t) (soff & PAGE_MASK), 3993 (vm_offset_t) (eoff - soff) 3994 ); 3995 } 3996 } 3997 3998 /* 3999 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 4000 * any page is busy, drain the flag. 4001 */ 4002 static void 4003 vfs_drain_busy_pages(struct buf *bp) 4004 { 4005 vm_page_t m; 4006 int i, last_busied; 4007 4008 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4009 last_busied = 0; 4010 for (i = 0; i < bp->b_npages; i++) { 4011 m = bp->b_pages[i]; 4012 if ((m->oflags & VPO_BUSY) != 0) { 4013 for (; last_busied < i; last_busied++) 4014 vm_page_busy(bp->b_pages[last_busied]); 4015 while ((m->oflags & VPO_BUSY) != 0) 4016 vm_page_sleep(m, "vbpage"); 4017 } 4018 } 4019 for (i = 0; i < last_busied; i++) 4020 vm_page_wakeup(bp->b_pages[i]); 4021 } 4022 4023 /* 4024 * This routine is called before a device strategy routine. 4025 * It is used to tell the VM system that paging I/O is in 4026 * progress, and treat the pages associated with the buffer 4027 * almost as being VPO_BUSY. Also the object paging_in_progress 4028 * flag is handled to make sure that the object doesn't become 4029 * inconsistant. 4030 * 4031 * Since I/O has not been initiated yet, certain buffer flags 4032 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4033 * and should be ignored. 4034 */ 4035 void 4036 vfs_busy_pages(struct buf *bp, int clear_modify) 4037 { 4038 int i, bogus; 4039 vm_object_t obj; 4040 vm_ooffset_t foff; 4041 vm_page_t m; 4042 4043 if (!(bp->b_flags & B_VMIO)) 4044 return; 4045 4046 obj = bp->b_bufobj->bo_object; 4047 foff = bp->b_offset; 4048 KASSERT(bp->b_offset != NOOFFSET, 4049 ("vfs_busy_pages: no buffer offset")); 4050 VM_OBJECT_WLOCK(obj); 4051 vfs_drain_busy_pages(bp); 4052 if (bp->b_bufsize != 0) 4053 vfs_setdirty_locked_object(bp); 4054 bogus = 0; 4055 for (i = 0; i < bp->b_npages; i++) { 4056 m = bp->b_pages[i]; 4057 4058 if ((bp->b_flags & B_CLUSTER) == 0) { 4059 vm_object_pip_add(obj, 1); 4060 vm_page_io_start(m); 4061 } 4062 /* 4063 * When readying a buffer for a read ( i.e 4064 * clear_modify == 0 ), it is important to do 4065 * bogus_page replacement for valid pages in 4066 * partially instantiated buffers. Partially 4067 * instantiated buffers can, in turn, occur when 4068 * reconstituting a buffer from its VM backing store 4069 * base. We only have to do this if B_CACHE is 4070 * clear ( which causes the I/O to occur in the 4071 * first place ). The replacement prevents the read 4072 * I/O from overwriting potentially dirty VM-backed 4073 * pages. XXX bogus page replacement is, uh, bogus. 4074 * It may not work properly with small-block devices. 4075 * We need to find a better way. 4076 */ 4077 if (clear_modify) { 4078 pmap_remove_write(m); 4079 vfs_page_set_validclean(bp, foff, m); 4080 } else if (m->valid == VM_PAGE_BITS_ALL && 4081 (bp->b_flags & B_CACHE) == 0) { 4082 bp->b_pages[i] = bogus_page; 4083 bogus++; 4084 } 4085 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4086 } 4087 VM_OBJECT_WUNLOCK(obj); 4088 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4089 BUF_CHECK_MAPPED(bp); 4090 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4091 bp->b_pages, bp->b_npages); 4092 } 4093 } 4094 4095 /* 4096 * vfs_bio_set_valid: 4097 * 4098 * Set the range within the buffer to valid. The range is 4099 * relative to the beginning of the buffer, b_offset. Note that 4100 * b_offset itself may be offset from the beginning of the first 4101 * page. 4102 */ 4103 void 4104 vfs_bio_set_valid(struct buf *bp, int base, int size) 4105 { 4106 int i, n; 4107 vm_page_t m; 4108 4109 if (!(bp->b_flags & B_VMIO)) 4110 return; 4111 4112 /* 4113 * Fixup base to be relative to beginning of first page. 4114 * Set initial n to be the maximum number of bytes in the 4115 * first page that can be validated. 4116 */ 4117 base += (bp->b_offset & PAGE_MASK); 4118 n = PAGE_SIZE - (base & PAGE_MASK); 4119 4120 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4121 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4122 m = bp->b_pages[i]; 4123 if (n > size) 4124 n = size; 4125 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4126 base += n; 4127 size -= n; 4128 n = PAGE_SIZE; 4129 } 4130 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4131 } 4132 4133 /* 4134 * vfs_bio_clrbuf: 4135 * 4136 * If the specified buffer is a non-VMIO buffer, clear the entire 4137 * buffer. If the specified buffer is a VMIO buffer, clear and 4138 * validate only the previously invalid portions of the buffer. 4139 * This routine essentially fakes an I/O, so we need to clear 4140 * BIO_ERROR and B_INVAL. 4141 * 4142 * Note that while we only theoretically need to clear through b_bcount, 4143 * we go ahead and clear through b_bufsize. 4144 */ 4145 void 4146 vfs_bio_clrbuf(struct buf *bp) 4147 { 4148 int i, j, mask, sa, ea, slide; 4149 4150 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4151 clrbuf(bp); 4152 return; 4153 } 4154 bp->b_flags &= ~B_INVAL; 4155 bp->b_ioflags &= ~BIO_ERROR; 4156 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4157 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4158 (bp->b_offset & PAGE_MASK) == 0) { 4159 if (bp->b_pages[0] == bogus_page) 4160 goto unlock; 4161 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4162 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4163 if ((bp->b_pages[0]->valid & mask) == mask) 4164 goto unlock; 4165 if ((bp->b_pages[0]->valid & mask) == 0) { 4166 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4167 bp->b_pages[0]->valid |= mask; 4168 goto unlock; 4169 } 4170 } 4171 sa = bp->b_offset & PAGE_MASK; 4172 slide = 0; 4173 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4174 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4175 ea = slide & PAGE_MASK; 4176 if (ea == 0) 4177 ea = PAGE_SIZE; 4178 if (bp->b_pages[i] == bogus_page) 4179 continue; 4180 j = sa / DEV_BSIZE; 4181 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4182 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4183 if ((bp->b_pages[i]->valid & mask) == mask) 4184 continue; 4185 if ((bp->b_pages[i]->valid & mask) == 0) 4186 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4187 else { 4188 for (; sa < ea; sa += DEV_BSIZE, j++) { 4189 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4190 pmap_zero_page_area(bp->b_pages[i], 4191 sa, DEV_BSIZE); 4192 } 4193 } 4194 } 4195 bp->b_pages[i]->valid |= mask; 4196 } 4197 unlock: 4198 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4199 bp->b_resid = 0; 4200 } 4201 4202 void 4203 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4204 { 4205 vm_page_t m; 4206 int i, n; 4207 4208 if ((bp->b_flags & B_UNMAPPED) == 0) { 4209 BUF_CHECK_MAPPED(bp); 4210 bzero(bp->b_data + base, size); 4211 } else { 4212 BUF_CHECK_UNMAPPED(bp); 4213 n = PAGE_SIZE - (base & PAGE_MASK); 4214 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4215 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4216 m = bp->b_pages[i]; 4217 if (n > size) 4218 n = size; 4219 pmap_zero_page_area(m, base & PAGE_MASK, n); 4220 base += n; 4221 size -= n; 4222 n = PAGE_SIZE; 4223 } 4224 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4225 } 4226 } 4227 4228 /* 4229 * vm_hold_load_pages and vm_hold_free_pages get pages into 4230 * a buffers address space. The pages are anonymous and are 4231 * not associated with a file object. 4232 */ 4233 static void 4234 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4235 { 4236 vm_offset_t pg; 4237 vm_page_t p; 4238 int index; 4239 4240 BUF_CHECK_MAPPED(bp); 4241 4242 to = round_page(to); 4243 from = round_page(from); 4244 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4245 4246 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4247 tryagain: 4248 /* 4249 * note: must allocate system pages since blocking here 4250 * could interfere with paging I/O, no matter which 4251 * process we are. 4252 */ 4253 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4254 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4255 if (p == NULL) { 4256 VM_WAIT; 4257 goto tryagain; 4258 } 4259 pmap_qenter(pg, &p, 1); 4260 bp->b_pages[index] = p; 4261 } 4262 bp->b_npages = index; 4263 } 4264 4265 /* Return pages associated with this buf to the vm system */ 4266 static void 4267 vm_hold_free_pages(struct buf *bp, int newbsize) 4268 { 4269 vm_offset_t from; 4270 vm_page_t p; 4271 int index, newnpages; 4272 4273 BUF_CHECK_MAPPED(bp); 4274 4275 from = round_page((vm_offset_t)bp->b_data + newbsize); 4276 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4277 if (bp->b_npages > newnpages) 4278 pmap_qremove(from, bp->b_npages - newnpages); 4279 for (index = newnpages; index < bp->b_npages; index++) { 4280 p = bp->b_pages[index]; 4281 bp->b_pages[index] = NULL; 4282 if (p->busy != 0) 4283 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4284 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4285 p->wire_count--; 4286 vm_page_free(p); 4287 atomic_subtract_int(&cnt.v_wire_count, 1); 4288 } 4289 bp->b_npages = newnpages; 4290 } 4291 4292 /* 4293 * Map an IO request into kernel virtual address space. 4294 * 4295 * All requests are (re)mapped into kernel VA space. 4296 * Notice that we use b_bufsize for the size of the buffer 4297 * to be mapped. b_bcount might be modified by the driver. 4298 * 4299 * Note that even if the caller determines that the address space should 4300 * be valid, a race or a smaller-file mapped into a larger space may 4301 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4302 * check the return value. 4303 */ 4304 int 4305 vmapbuf(struct buf *bp, int mapbuf) 4306 { 4307 caddr_t kva; 4308 vm_prot_t prot; 4309 int pidx; 4310 4311 if (bp->b_bufsize < 0) 4312 return (-1); 4313 prot = VM_PROT_READ; 4314 if (bp->b_iocmd == BIO_READ) 4315 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4316 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4317 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4318 btoc(MAXPHYS))) < 0) 4319 return (-1); 4320 bp->b_npages = pidx; 4321 if (mapbuf || !unmapped_buf_allowed) { 4322 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4323 kva = bp->b_saveaddr; 4324 bp->b_saveaddr = bp->b_data; 4325 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4326 bp->b_flags &= ~B_UNMAPPED; 4327 } else { 4328 bp->b_flags |= B_UNMAPPED; 4329 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4330 bp->b_saveaddr = bp->b_data; 4331 bp->b_data = unmapped_buf; 4332 } 4333 return(0); 4334 } 4335 4336 /* 4337 * Free the io map PTEs associated with this IO operation. 4338 * We also invalidate the TLB entries and restore the original b_addr. 4339 */ 4340 void 4341 vunmapbuf(struct buf *bp) 4342 { 4343 int npages; 4344 4345 npages = bp->b_npages; 4346 if (bp->b_flags & B_UNMAPPED) 4347 bp->b_flags &= ~B_UNMAPPED; 4348 else 4349 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4350 vm_page_unhold_pages(bp->b_pages, npages); 4351 4352 bp->b_data = bp->b_saveaddr; 4353 } 4354 4355 void 4356 bdone(struct buf *bp) 4357 { 4358 struct mtx *mtxp; 4359 4360 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4361 mtx_lock(mtxp); 4362 bp->b_flags |= B_DONE; 4363 wakeup(bp); 4364 mtx_unlock(mtxp); 4365 } 4366 4367 void 4368 bwait(struct buf *bp, u_char pri, const char *wchan) 4369 { 4370 struct mtx *mtxp; 4371 4372 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4373 mtx_lock(mtxp); 4374 while ((bp->b_flags & B_DONE) == 0) 4375 msleep(bp, mtxp, pri, wchan, 0); 4376 mtx_unlock(mtxp); 4377 } 4378 4379 int 4380 bufsync(struct bufobj *bo, int waitfor) 4381 { 4382 4383 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4384 } 4385 4386 void 4387 bufstrategy(struct bufobj *bo, struct buf *bp) 4388 { 4389 int i = 0; 4390 struct vnode *vp; 4391 4392 vp = bp->b_vp; 4393 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4394 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4395 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4396 i = VOP_STRATEGY(vp, bp); 4397 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4398 } 4399 4400 void 4401 bufobj_wrefl(struct bufobj *bo) 4402 { 4403 4404 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4405 ASSERT_BO_LOCKED(bo); 4406 bo->bo_numoutput++; 4407 } 4408 4409 void 4410 bufobj_wref(struct bufobj *bo) 4411 { 4412 4413 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4414 BO_LOCK(bo); 4415 bo->bo_numoutput++; 4416 BO_UNLOCK(bo); 4417 } 4418 4419 void 4420 bufobj_wdrop(struct bufobj *bo) 4421 { 4422 4423 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4424 BO_LOCK(bo); 4425 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4426 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4427 bo->bo_flag &= ~BO_WWAIT; 4428 wakeup(&bo->bo_numoutput); 4429 } 4430 BO_UNLOCK(bo); 4431 } 4432 4433 int 4434 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4435 { 4436 int error; 4437 4438 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4439 ASSERT_BO_LOCKED(bo); 4440 error = 0; 4441 while (bo->bo_numoutput) { 4442 bo->bo_flag |= BO_WWAIT; 4443 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 4444 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4445 if (error) 4446 break; 4447 } 4448 return (error); 4449 } 4450 4451 void 4452 bpin(struct buf *bp) 4453 { 4454 struct mtx *mtxp; 4455 4456 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4457 mtx_lock(mtxp); 4458 bp->b_pin_count++; 4459 mtx_unlock(mtxp); 4460 } 4461 4462 void 4463 bunpin(struct buf *bp) 4464 { 4465 struct mtx *mtxp; 4466 4467 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4468 mtx_lock(mtxp); 4469 if (--bp->b_pin_count == 0) 4470 wakeup(bp); 4471 mtx_unlock(mtxp); 4472 } 4473 4474 void 4475 bunpin_wait(struct buf *bp) 4476 { 4477 struct mtx *mtxp; 4478 4479 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4480 mtx_lock(mtxp); 4481 while (bp->b_pin_count > 0) 4482 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4483 mtx_unlock(mtxp); 4484 } 4485 4486 /* 4487 * Set bio_data or bio_ma for struct bio from the struct buf. 4488 */ 4489 void 4490 bdata2bio(struct buf *bp, struct bio *bip) 4491 { 4492 4493 if ((bp->b_flags & B_UNMAPPED) != 0) { 4494 KASSERT(unmapped_buf_allowed, ("unmapped")); 4495 bip->bio_ma = bp->b_pages; 4496 bip->bio_ma_n = bp->b_npages; 4497 bip->bio_data = unmapped_buf; 4498 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4499 bip->bio_flags |= BIO_UNMAPPED; 4500 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4501 PAGE_SIZE == bp->b_npages, 4502 ("Buffer %p too short: %d %d %d", bp, bip->bio_ma_offset, 4503 bip->bio_length, bip->bio_ma_n)); 4504 } else { 4505 bip->bio_data = bp->b_data; 4506 bip->bio_ma = NULL; 4507 } 4508 } 4509 4510 #include "opt_ddb.h" 4511 #ifdef DDB 4512 #include <ddb/ddb.h> 4513 4514 /* DDB command to show buffer data */ 4515 DB_SHOW_COMMAND(buffer, db_show_buffer) 4516 { 4517 /* get args */ 4518 struct buf *bp = (struct buf *)addr; 4519 4520 if (!have_addr) { 4521 db_printf("usage: show buffer <addr>\n"); 4522 return; 4523 } 4524 4525 db_printf("buf at %p\n", bp); 4526 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4527 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4528 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4529 db_printf( 4530 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4531 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4532 "b_dep = %p\n", 4533 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4534 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4535 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4536 if (bp->b_npages) { 4537 int i; 4538 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4539 for (i = 0; i < bp->b_npages; i++) { 4540 vm_page_t m; 4541 m = bp->b_pages[i]; 4542 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4543 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4544 if ((i + 1) < bp->b_npages) 4545 db_printf(","); 4546 } 4547 db_printf("\n"); 4548 } 4549 db_printf(" "); 4550 BUF_LOCKPRINTINFO(bp); 4551 } 4552 4553 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4554 { 4555 struct buf *bp; 4556 int i; 4557 4558 for (i = 0; i < nbuf; i++) { 4559 bp = &buf[i]; 4560 if (BUF_ISLOCKED(bp)) { 4561 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4562 db_printf("\n"); 4563 } 4564 } 4565 } 4566 4567 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4568 { 4569 struct vnode *vp; 4570 struct buf *bp; 4571 4572 if (!have_addr) { 4573 db_printf("usage: show vnodebufs <addr>\n"); 4574 return; 4575 } 4576 vp = (struct vnode *)addr; 4577 db_printf("Clean buffers:\n"); 4578 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4579 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4580 db_printf("\n"); 4581 } 4582 db_printf("Dirty buffers:\n"); 4583 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4584 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4585 db_printf("\n"); 4586 } 4587 } 4588 4589 DB_COMMAND(countfreebufs, db_coundfreebufs) 4590 { 4591 struct buf *bp; 4592 int i, used = 0, nfree = 0; 4593 4594 if (have_addr) { 4595 db_printf("usage: countfreebufs\n"); 4596 return; 4597 } 4598 4599 for (i = 0; i < nbuf; i++) { 4600 bp = &buf[i]; 4601 if ((bp->b_vflags & BV_INFREECNT) != 0) 4602 nfree++; 4603 else 4604 used++; 4605 } 4606 4607 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4608 nfree + used); 4609 db_printf("numfreebuffers is %d\n", numfreebuffers); 4610 } 4611 #endif /* DDB */ 4612