1 /*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * this file contains a new buffer I/O scheme implementing a coherent 30 * VM object and buffer cache scheme. Pains have been taken to make 31 * sure that the performance degradation associated with schemes such 32 * as this is not realized. 33 * 34 * Author: John S. Dyson 35 * Significant help during the development and debugging phases 36 * had been provided by David Greenman, also of the FreeBSD core team. 37 * 38 * see man buf(9) for more info. 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/bio.h> 47 #include <sys/conf.h> 48 #include <sys/buf.h> 49 #include <sys/devicestat.h> 50 #include <sys/eventhandler.h> 51 #include <sys/fail.h> 52 #include <sys/limits.h> 53 #include <sys/lock.h> 54 #include <sys/malloc.h> 55 #include <sys/mount.h> 56 #include <sys/mutex.h> 57 #include <sys/kernel.h> 58 #include <sys/kthread.h> 59 #include <sys/proc.h> 60 #include <sys/resourcevar.h> 61 #include <sys/sysctl.h> 62 #include <sys/vmmeter.h> 63 #include <sys/vnode.h> 64 #include <geom/geom.h> 65 #include <vm/vm.h> 66 #include <vm/vm_param.h> 67 #include <vm/vm_kern.h> 68 #include <vm/vm_pageout.h> 69 #include <vm/vm_page.h> 70 #include <vm/vm_object.h> 71 #include <vm/vm_extern.h> 72 #include <vm/vm_map.h> 73 #include "opt_compat.h" 74 #include "opt_directio.h" 75 #include "opt_swap.h" 76 77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 78 79 struct bio_ops bioops; /* I/O operation notification */ 80 81 struct buf_ops buf_ops_bio = { 82 .bop_name = "buf_ops_bio", 83 .bop_write = bufwrite, 84 .bop_strategy = bufstrategy, 85 .bop_sync = bufsync, 86 .bop_bdflush = bufbdflush, 87 }; 88 89 /* 90 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 91 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 92 */ 93 struct buf *buf; /* buffer header pool */ 94 95 static struct proc *bufdaemonproc; 96 97 static int inmem(struct vnode *vp, daddr_t blkno); 98 static void vm_hold_free_pages(struct buf *bp, int newbsize); 99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 100 vm_offset_t to); 101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 103 vm_page_t m); 104 static void vfs_drain_busy_pages(struct buf *bp); 105 static void vfs_clean_pages_dirty_buf(struct buf *bp); 106 static void vfs_setdirty_locked_object(struct buf *bp); 107 static void vfs_vmio_release(struct buf *bp); 108 static int vfs_bio_clcheck(struct vnode *vp, int size, 109 daddr_t lblkno, daddr_t blkno); 110 static int buf_do_flush(struct vnode *vp); 111 static int flushbufqueues(struct vnode *, int, int); 112 static void buf_daemon(void); 113 static void bremfreel(struct buf *bp); 114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 117 #endif 118 119 int vmiodirenable = TRUE; 120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 121 "Use the VM system for directory writes"); 122 long runningbufspace; 123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 124 "Amount of presently outstanding async buffer io"); 125 static long bufspace; 126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 130 #else 131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 132 "Virtual memory used for buffers"); 133 #endif 134 static long maxbufspace; 135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 136 "Maximum allowed value of bufspace (including buf_daemon)"); 137 static long bufmallocspace; 138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 139 "Amount of malloced memory for buffers"); 140 static long maxbufmallocspace; 141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 142 "Maximum amount of malloced memory for buffers"); 143 static long lobufspace; 144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 145 "Minimum amount of buffers we want to have"); 146 long hibufspace; 147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 148 "Maximum allowed value of bufspace (excluding buf_daemon)"); 149 static int bufreusecnt; 150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 151 "Number of times we have reused a buffer"); 152 static int buffreekvacnt; 153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 154 "Number of times we have freed the KVA space from some buffer"); 155 static int bufdefragcnt; 156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 157 "Number of times we have had to repeat buffer allocation to defragment"); 158 static long lorunningspace; 159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 160 "Minimum preferred space used for in-progress I/O"); 161 static long hirunningspace; 162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 163 "Maximum amount of space to use for in-progress I/O"); 164 int dirtybufferflushes; 165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 167 int bdwriteskip; 168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 170 int altbufferflushes; 171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 172 0, "Number of fsync flushes to limit dirty buffers"); 173 static int recursiveflushes; 174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 175 0, "Number of flushes skipped due to being recursive"); 176 static int numdirtybuffers; 177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 178 "Number of buffers that are dirty (has unwritten changes) at the moment"); 179 static int lodirtybuffers; 180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 181 "How many buffers we want to have free before bufdaemon can sleep"); 182 static int hidirtybuffers; 183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 184 "When the number of dirty buffers is considered severe"); 185 int dirtybufthresh; 186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 188 static int numfreebuffers; 189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 190 "Number of free buffers"); 191 static int lofreebuffers; 192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 193 "XXX Unused"); 194 static int hifreebuffers; 195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 196 "XXX Complicatedly unused"); 197 static int getnewbufcalls; 198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 199 "Number of calls to getnewbuf"); 200 static int getnewbufrestarts; 201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 202 "Number of times getnewbuf has had to restart a buffer aquisition"); 203 static int flushbufqtarget = 100; 204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 205 "Amount of work to do in flushbufqueues when helping bufdaemon"); 206 static long notbufdflashes; 207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 208 "Number of dirty buffer flushes done by the bufdaemon helpers"); 209 210 /* 211 * Wakeup point for bufdaemon, as well as indicator of whether it is already 212 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 213 * is idling. 214 */ 215 static int bd_request; 216 217 /* 218 * Request for the buf daemon to write more buffers than is indicated by 219 * lodirtybuf. This may be necessary to push out excess dependencies or 220 * defragment the address space where a simple count of the number of dirty 221 * buffers is insufficient to characterize the demand for flushing them. 222 */ 223 static int bd_speedupreq; 224 225 /* 226 * This lock synchronizes access to bd_request. 227 */ 228 static struct mtx bdlock; 229 230 /* 231 * bogus page -- for I/O to/from partially complete buffers 232 * this is a temporary solution to the problem, but it is not 233 * really that bad. it would be better to split the buffer 234 * for input in the case of buffers partially already in memory, 235 * but the code is intricate enough already. 236 */ 237 vm_page_t bogus_page; 238 239 /* 240 * Synchronization (sleep/wakeup) variable for active buffer space requests. 241 * Set when wait starts, cleared prior to wakeup(). 242 * Used in runningbufwakeup() and waitrunningbufspace(). 243 */ 244 static int runningbufreq; 245 246 /* 247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 248 * waitrunningbufspace(). 249 */ 250 static struct mtx rbreqlock; 251 252 /* 253 * Synchronization (sleep/wakeup) variable for buffer requests. 254 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 255 * by and/or. 256 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 257 * getnewbuf(), and getblk(). 258 */ 259 static int needsbuffer; 260 261 /* 262 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 263 */ 264 static struct mtx nblock; 265 266 /* 267 * Definitions for the buffer free lists. 268 */ 269 #define BUFFER_QUEUES 6 /* number of free buffer queues */ 270 271 #define QUEUE_NONE 0 /* on no queue */ 272 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 273 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 274 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 275 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 276 #define QUEUE_EMPTY 5 /* empty buffer headers */ 277 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 278 279 /* Queues for free buffers with various properties */ 280 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 281 282 /* Lock for the bufqueues */ 283 static struct mtx bqlock; 284 285 /* 286 * Single global constant for BUF_WMESG, to avoid getting multiple references. 287 * buf_wmesg is referred from macros. 288 */ 289 const char *buf_wmesg = BUF_WMESG; 290 291 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 292 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 293 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 294 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 295 296 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 297 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 298 static int 299 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 300 { 301 long lvalue; 302 int ivalue; 303 304 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 305 return (sysctl_handle_long(oidp, arg1, arg2, req)); 306 lvalue = *(long *)arg1; 307 if (lvalue > INT_MAX) 308 /* On overflow, still write out a long to trigger ENOMEM. */ 309 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 310 ivalue = lvalue; 311 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 312 } 313 #endif 314 315 #ifdef DIRECTIO 316 extern void ffs_rawread_setup(void); 317 #endif /* DIRECTIO */ 318 /* 319 * numdirtywakeup: 320 * 321 * If someone is blocked due to there being too many dirty buffers, 322 * and numdirtybuffers is now reasonable, wake them up. 323 */ 324 325 static __inline void 326 numdirtywakeup(int level) 327 { 328 329 if (numdirtybuffers <= level) { 330 mtx_lock(&nblock); 331 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 332 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 333 wakeup(&needsbuffer); 334 } 335 mtx_unlock(&nblock); 336 } 337 } 338 339 /* 340 * bufspacewakeup: 341 * 342 * Called when buffer space is potentially available for recovery. 343 * getnewbuf() will block on this flag when it is unable to free 344 * sufficient buffer space. Buffer space becomes recoverable when 345 * bp's get placed back in the queues. 346 */ 347 348 static __inline void 349 bufspacewakeup(void) 350 { 351 352 /* 353 * If someone is waiting for BUF space, wake them up. Even 354 * though we haven't freed the kva space yet, the waiting 355 * process will be able to now. 356 */ 357 mtx_lock(&nblock); 358 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 359 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 360 wakeup(&needsbuffer); 361 } 362 mtx_unlock(&nblock); 363 } 364 365 /* 366 * runningbufwakeup() - in-progress I/O accounting. 367 * 368 */ 369 void 370 runningbufwakeup(struct buf *bp) 371 { 372 373 if (bp->b_runningbufspace) { 374 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 375 bp->b_runningbufspace = 0; 376 mtx_lock(&rbreqlock); 377 if (runningbufreq && runningbufspace <= lorunningspace) { 378 runningbufreq = 0; 379 wakeup(&runningbufreq); 380 } 381 mtx_unlock(&rbreqlock); 382 } 383 } 384 385 /* 386 * bufcountwakeup: 387 * 388 * Called when a buffer has been added to one of the free queues to 389 * account for the buffer and to wakeup anyone waiting for free buffers. 390 * This typically occurs when large amounts of metadata are being handled 391 * by the buffer cache ( else buffer space runs out first, usually ). 392 */ 393 394 static __inline void 395 bufcountwakeup(struct buf *bp) 396 { 397 int old; 398 399 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 400 ("buf %p already counted as free", bp)); 401 bp->b_vflags |= BV_INFREECNT; 402 old = atomic_fetchadd_int(&numfreebuffers, 1); 403 KASSERT(old >= 0 && old < nbuf, 404 ("numfreebuffers climbed to %d", old + 1)); 405 mtx_lock(&nblock); 406 if (needsbuffer) { 407 needsbuffer &= ~VFS_BIO_NEED_ANY; 408 if (numfreebuffers >= hifreebuffers) 409 needsbuffer &= ~VFS_BIO_NEED_FREE; 410 wakeup(&needsbuffer); 411 } 412 mtx_unlock(&nblock); 413 } 414 415 /* 416 * waitrunningbufspace() 417 * 418 * runningbufspace is a measure of the amount of I/O currently 419 * running. This routine is used in async-write situations to 420 * prevent creating huge backups of pending writes to a device. 421 * Only asynchronous writes are governed by this function. 422 * 423 * Reads will adjust runningbufspace, but will not block based on it. 424 * The read load has a side effect of reducing the allowed write load. 425 * 426 * This does NOT turn an async write into a sync write. It waits 427 * for earlier writes to complete and generally returns before the 428 * caller's write has reached the device. 429 */ 430 void 431 waitrunningbufspace(void) 432 { 433 434 mtx_lock(&rbreqlock); 435 while (runningbufspace > hirunningspace) { 436 ++runningbufreq; 437 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 438 } 439 mtx_unlock(&rbreqlock); 440 } 441 442 443 /* 444 * vfs_buf_test_cache: 445 * 446 * Called when a buffer is extended. This function clears the B_CACHE 447 * bit if the newly extended portion of the buffer does not contain 448 * valid data. 449 */ 450 static __inline 451 void 452 vfs_buf_test_cache(struct buf *bp, 453 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 454 vm_page_t m) 455 { 456 457 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 458 if (bp->b_flags & B_CACHE) { 459 int base = (foff + off) & PAGE_MASK; 460 if (vm_page_is_valid(m, base, size) == 0) 461 bp->b_flags &= ~B_CACHE; 462 } 463 } 464 465 /* Wake up the buffer daemon if necessary */ 466 static __inline 467 void 468 bd_wakeup(int dirtybuflevel) 469 { 470 471 mtx_lock(&bdlock); 472 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 473 bd_request = 1; 474 wakeup(&bd_request); 475 } 476 mtx_unlock(&bdlock); 477 } 478 479 /* 480 * bd_speedup - speedup the buffer cache flushing code 481 */ 482 483 void 484 bd_speedup(void) 485 { 486 int needwake; 487 488 mtx_lock(&bdlock); 489 needwake = 0; 490 if (bd_speedupreq == 0 || bd_request == 0) 491 needwake = 1; 492 bd_speedupreq = 1; 493 bd_request = 1; 494 if (needwake) 495 wakeup(&bd_request); 496 mtx_unlock(&bdlock); 497 } 498 499 /* 500 * Calculating buffer cache scaling values and reserve space for buffer 501 * headers. This is called during low level kernel initialization and 502 * may be called more then once. We CANNOT write to the memory area 503 * being reserved at this time. 504 */ 505 caddr_t 506 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 507 { 508 int tuned_nbuf; 509 long maxbuf; 510 511 /* 512 * physmem_est is in pages. Convert it to kilobytes (assumes 513 * PAGE_SIZE is >= 1K) 514 */ 515 physmem_est = physmem_est * (PAGE_SIZE / 1024); 516 517 /* 518 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 519 * For the first 64MB of ram nominally allocate sufficient buffers to 520 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 521 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 522 * the buffer cache we limit the eventual kva reservation to 523 * maxbcache bytes. 524 * 525 * factor represents the 1/4 x ram conversion. 526 */ 527 if (nbuf == 0) { 528 int factor = 4 * BKVASIZE / 1024; 529 530 nbuf = 50; 531 if (physmem_est > 4096) 532 nbuf += min((physmem_est - 4096) / factor, 533 65536 / factor); 534 if (physmem_est > 65536) 535 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 536 537 if (maxbcache && nbuf > maxbcache / BKVASIZE) 538 nbuf = maxbcache / BKVASIZE; 539 tuned_nbuf = 1; 540 } else 541 tuned_nbuf = 0; 542 543 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 544 maxbuf = (LONG_MAX / 3) / BKVASIZE; 545 if (nbuf > maxbuf) { 546 if (!tuned_nbuf) 547 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 548 maxbuf); 549 nbuf = maxbuf; 550 } 551 552 /* 553 * swbufs are used as temporary holders for I/O, such as paging I/O. 554 * We have no less then 16 and no more then 256. 555 */ 556 nswbuf = max(min(nbuf/4, 256), 16); 557 #ifdef NSWBUF_MIN 558 if (nswbuf < NSWBUF_MIN) 559 nswbuf = NSWBUF_MIN; 560 #endif 561 #ifdef DIRECTIO 562 ffs_rawread_setup(); 563 #endif 564 565 /* 566 * Reserve space for the buffer cache buffers 567 */ 568 swbuf = (void *)v; 569 v = (caddr_t)(swbuf + nswbuf); 570 buf = (void *)v; 571 v = (caddr_t)(buf + nbuf); 572 573 return(v); 574 } 575 576 /* Initialize the buffer subsystem. Called before use of any buffers. */ 577 void 578 bufinit(void) 579 { 580 struct buf *bp; 581 int i; 582 583 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 584 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 585 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 586 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 587 588 /* next, make a null set of free lists */ 589 for (i = 0; i < BUFFER_QUEUES; i++) 590 TAILQ_INIT(&bufqueues[i]); 591 592 /* finally, initialize each buffer header and stick on empty q */ 593 for (i = 0; i < nbuf; i++) { 594 bp = &buf[i]; 595 bzero(bp, sizeof *bp); 596 bp->b_flags = B_INVAL; /* we're just an empty header */ 597 bp->b_rcred = NOCRED; 598 bp->b_wcred = NOCRED; 599 bp->b_qindex = QUEUE_EMPTY; 600 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 601 bp->b_xflags = 0; 602 LIST_INIT(&bp->b_dep); 603 BUF_LOCKINIT(bp); 604 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 605 } 606 607 /* 608 * maxbufspace is the absolute maximum amount of buffer space we are 609 * allowed to reserve in KVM and in real terms. The absolute maximum 610 * is nominally used by buf_daemon. hibufspace is the nominal maximum 611 * used by most other processes. The differential is required to 612 * ensure that buf_daemon is able to run when other processes might 613 * be blocked waiting for buffer space. 614 * 615 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 616 * this may result in KVM fragmentation which is not handled optimally 617 * by the system. 618 */ 619 maxbufspace = (long)nbuf * BKVASIZE; 620 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 621 lobufspace = hibufspace - MAXBSIZE; 622 623 /* 624 * Note: The 16 MB upper limit for hirunningspace was chosen 625 * arbitrarily and may need further tuning. It corresponds to 626 * 128 outstanding write IO requests (if IO size is 128 KiB), 627 * which fits with many RAID controllers' tagged queuing limits. 628 * The lower 1 MB limit is the historical upper limit for 629 * hirunningspace. 630 */ 631 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 632 16 * 1024 * 1024), 1024 * 1024); 633 lorunningspace = roundup(hirunningspace / 2, MAXBSIZE); 634 635 /* 636 * Limit the amount of malloc memory since it is wired permanently into 637 * the kernel space. Even though this is accounted for in the buffer 638 * allocation, we don't want the malloced region to grow uncontrolled. 639 * The malloc scheme improves memory utilization significantly on average 640 * (small) directories. 641 */ 642 maxbufmallocspace = hibufspace / 20; 643 644 /* 645 * Reduce the chance of a deadlock occuring by limiting the number 646 * of delayed-write dirty buffers we allow to stack up. 647 */ 648 hidirtybuffers = nbuf / 4 + 20; 649 dirtybufthresh = hidirtybuffers * 9 / 10; 650 numdirtybuffers = 0; 651 /* 652 * To support extreme low-memory systems, make sure hidirtybuffers cannot 653 * eat up all available buffer space. This occurs when our minimum cannot 654 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 655 * BKVASIZE'd (8K) buffers. 656 */ 657 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 658 hidirtybuffers >>= 1; 659 } 660 lodirtybuffers = hidirtybuffers / 2; 661 662 /* 663 * Try to keep the number of free buffers in the specified range, 664 * and give special processes (e.g. like buf_daemon) access to an 665 * emergency reserve. 666 */ 667 lofreebuffers = nbuf / 18 + 5; 668 hifreebuffers = 2 * lofreebuffers; 669 numfreebuffers = nbuf; 670 671 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 672 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 673 } 674 675 /* 676 * bfreekva() - free the kva allocation for a buffer. 677 * 678 * Since this call frees up buffer space, we call bufspacewakeup(). 679 */ 680 static void 681 bfreekva(struct buf *bp) 682 { 683 684 if (bp->b_kvasize) { 685 atomic_add_int(&buffreekvacnt, 1); 686 atomic_subtract_long(&bufspace, bp->b_kvasize); 687 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 688 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 689 bp->b_kvasize = 0; 690 bufspacewakeup(); 691 } 692 } 693 694 /* 695 * bremfree: 696 * 697 * Mark the buffer for removal from the appropriate free list in brelse. 698 * 699 */ 700 void 701 bremfree(struct buf *bp) 702 { 703 int old; 704 705 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 706 KASSERT((bp->b_flags & B_REMFREE) == 0, 707 ("bremfree: buffer %p already marked for delayed removal.", bp)); 708 KASSERT(bp->b_qindex != QUEUE_NONE, 709 ("bremfree: buffer %p not on a queue.", bp)); 710 BUF_ASSERT_HELD(bp); 711 712 bp->b_flags |= B_REMFREE; 713 /* Fixup numfreebuffers count. */ 714 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 715 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 716 ("buf %p not counted in numfreebuffers", bp)); 717 bp->b_vflags &= ~BV_INFREECNT; 718 old = atomic_fetchadd_int(&numfreebuffers, -1); 719 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 720 } 721 } 722 723 /* 724 * bremfreef: 725 * 726 * Force an immediate removal from a free list. Used only in nfs when 727 * it abuses the b_freelist pointer. 728 */ 729 void 730 bremfreef(struct buf *bp) 731 { 732 mtx_lock(&bqlock); 733 bremfreel(bp); 734 mtx_unlock(&bqlock); 735 } 736 737 /* 738 * bremfreel: 739 * 740 * Removes a buffer from the free list, must be called with the 741 * bqlock held. 742 */ 743 static void 744 bremfreel(struct buf *bp) 745 { 746 int old; 747 748 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 749 bp, bp->b_vp, bp->b_flags); 750 KASSERT(bp->b_qindex != QUEUE_NONE, 751 ("bremfreel: buffer %p not on a queue.", bp)); 752 BUF_ASSERT_HELD(bp); 753 mtx_assert(&bqlock, MA_OWNED); 754 755 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 756 bp->b_qindex = QUEUE_NONE; 757 /* 758 * If this was a delayed bremfree() we only need to remove the buffer 759 * from the queue and return the stats are already done. 760 */ 761 if (bp->b_flags & B_REMFREE) { 762 bp->b_flags &= ~B_REMFREE; 763 return; 764 } 765 /* 766 * Fixup numfreebuffers count. If the buffer is invalid or not 767 * delayed-write, the buffer was free and we must decrement 768 * numfreebuffers. 769 */ 770 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 771 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 772 ("buf %p not counted in numfreebuffers", bp)); 773 bp->b_vflags &= ~BV_INFREECNT; 774 old = atomic_fetchadd_int(&numfreebuffers, -1); 775 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 776 } 777 } 778 779 780 /* 781 * Get a buffer with the specified data. Look in the cache first. We 782 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 783 * is set, the buffer is valid and we do not have to do anything ( see 784 * getblk() ). This is really just a special case of breadn(). 785 */ 786 int 787 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 788 struct buf **bpp) 789 { 790 791 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 792 } 793 794 /* 795 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 796 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 797 * the buffer is valid and we do not have to do anything. 798 */ 799 void 800 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 801 int cnt, struct ucred * cred) 802 { 803 struct buf *rabp; 804 int i; 805 806 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 807 if (inmem(vp, *rablkno)) 808 continue; 809 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 810 811 if ((rabp->b_flags & B_CACHE) == 0) { 812 if (!TD_IS_IDLETHREAD(curthread)) 813 curthread->td_ru.ru_inblock++; 814 rabp->b_flags |= B_ASYNC; 815 rabp->b_flags &= ~B_INVAL; 816 rabp->b_ioflags &= ~BIO_ERROR; 817 rabp->b_iocmd = BIO_READ; 818 if (rabp->b_rcred == NOCRED && cred != NOCRED) 819 rabp->b_rcred = crhold(cred); 820 vfs_busy_pages(rabp, 0); 821 BUF_KERNPROC(rabp); 822 rabp->b_iooffset = dbtob(rabp->b_blkno); 823 bstrategy(rabp); 824 } else { 825 brelse(rabp); 826 } 827 } 828 } 829 830 /* 831 * Operates like bread, but also starts asynchronous I/O on 832 * read-ahead blocks. 833 */ 834 int 835 breadn(struct vnode * vp, daddr_t blkno, int size, 836 daddr_t * rablkno, int *rabsize, 837 int cnt, struct ucred * cred, struct buf **bpp) 838 { 839 struct buf *bp; 840 int rv = 0, readwait = 0; 841 842 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 843 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 844 845 /* if not found in cache, do some I/O */ 846 if ((bp->b_flags & B_CACHE) == 0) { 847 if (!TD_IS_IDLETHREAD(curthread)) 848 curthread->td_ru.ru_inblock++; 849 bp->b_iocmd = BIO_READ; 850 bp->b_flags &= ~B_INVAL; 851 bp->b_ioflags &= ~BIO_ERROR; 852 if (bp->b_rcred == NOCRED && cred != NOCRED) 853 bp->b_rcred = crhold(cred); 854 vfs_busy_pages(bp, 0); 855 bp->b_iooffset = dbtob(bp->b_blkno); 856 bstrategy(bp); 857 ++readwait; 858 } 859 860 breada(vp, rablkno, rabsize, cnt, cred); 861 862 if (readwait) { 863 rv = bufwait(bp); 864 } 865 return (rv); 866 } 867 868 /* 869 * Write, release buffer on completion. (Done by iodone 870 * if async). Do not bother writing anything if the buffer 871 * is invalid. 872 * 873 * Note that we set B_CACHE here, indicating that buffer is 874 * fully valid and thus cacheable. This is true even of NFS 875 * now so we set it generally. This could be set either here 876 * or in biodone() since the I/O is synchronous. We put it 877 * here. 878 */ 879 int 880 bufwrite(struct buf *bp) 881 { 882 int oldflags; 883 struct vnode *vp; 884 int vp_md; 885 886 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 887 if (bp->b_flags & B_INVAL) { 888 brelse(bp); 889 return (0); 890 } 891 892 oldflags = bp->b_flags; 893 894 BUF_ASSERT_HELD(bp); 895 896 if (bp->b_pin_count > 0) 897 bunpin_wait(bp); 898 899 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 900 ("FFS background buffer should not get here %p", bp)); 901 902 vp = bp->b_vp; 903 if (vp) 904 vp_md = vp->v_vflag & VV_MD; 905 else 906 vp_md = 0; 907 908 /* Mark the buffer clean */ 909 bundirty(bp); 910 911 bp->b_flags &= ~B_DONE; 912 bp->b_ioflags &= ~BIO_ERROR; 913 bp->b_flags |= B_CACHE; 914 bp->b_iocmd = BIO_WRITE; 915 916 bufobj_wref(bp->b_bufobj); 917 vfs_busy_pages(bp, 1); 918 919 /* 920 * Normal bwrites pipeline writes 921 */ 922 bp->b_runningbufspace = bp->b_bufsize; 923 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 924 925 if (!TD_IS_IDLETHREAD(curthread)) 926 curthread->td_ru.ru_oublock++; 927 if (oldflags & B_ASYNC) 928 BUF_KERNPROC(bp); 929 bp->b_iooffset = dbtob(bp->b_blkno); 930 bstrategy(bp); 931 932 if ((oldflags & B_ASYNC) == 0) { 933 int rtval = bufwait(bp); 934 brelse(bp); 935 return (rtval); 936 } else { 937 /* 938 * don't allow the async write to saturate the I/O 939 * system. We will not deadlock here because 940 * we are blocking waiting for I/O that is already in-progress 941 * to complete. We do not block here if it is the update 942 * or syncer daemon trying to clean up as that can lead 943 * to deadlock. 944 */ 945 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 946 waitrunningbufspace(); 947 } 948 949 return (0); 950 } 951 952 void 953 bufbdflush(struct bufobj *bo, struct buf *bp) 954 { 955 struct buf *nbp; 956 957 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 958 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 959 altbufferflushes++; 960 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 961 BO_LOCK(bo); 962 /* 963 * Try to find a buffer to flush. 964 */ 965 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 966 if ((nbp->b_vflags & BV_BKGRDINPROG) || 967 BUF_LOCK(nbp, 968 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 969 continue; 970 if (bp == nbp) 971 panic("bdwrite: found ourselves"); 972 BO_UNLOCK(bo); 973 /* Don't countdeps with the bo lock held. */ 974 if (buf_countdeps(nbp, 0)) { 975 BO_LOCK(bo); 976 BUF_UNLOCK(nbp); 977 continue; 978 } 979 if (nbp->b_flags & B_CLUSTEROK) { 980 vfs_bio_awrite(nbp); 981 } else { 982 bremfree(nbp); 983 bawrite(nbp); 984 } 985 dirtybufferflushes++; 986 break; 987 } 988 if (nbp == NULL) 989 BO_UNLOCK(bo); 990 } 991 } 992 993 /* 994 * Delayed write. (Buffer is marked dirty). Do not bother writing 995 * anything if the buffer is marked invalid. 996 * 997 * Note that since the buffer must be completely valid, we can safely 998 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 999 * biodone() in order to prevent getblk from writing the buffer 1000 * out synchronously. 1001 */ 1002 void 1003 bdwrite(struct buf *bp) 1004 { 1005 struct thread *td = curthread; 1006 struct vnode *vp; 1007 struct bufobj *bo; 1008 1009 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1010 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1011 BUF_ASSERT_HELD(bp); 1012 1013 if (bp->b_flags & B_INVAL) { 1014 brelse(bp); 1015 return; 1016 } 1017 1018 /* 1019 * If we have too many dirty buffers, don't create any more. 1020 * If we are wildly over our limit, then force a complete 1021 * cleanup. Otherwise, just keep the situation from getting 1022 * out of control. Note that we have to avoid a recursive 1023 * disaster and not try to clean up after our own cleanup! 1024 */ 1025 vp = bp->b_vp; 1026 bo = bp->b_bufobj; 1027 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1028 td->td_pflags |= TDP_INBDFLUSH; 1029 BO_BDFLUSH(bo, bp); 1030 td->td_pflags &= ~TDP_INBDFLUSH; 1031 } else 1032 recursiveflushes++; 1033 1034 bdirty(bp); 1035 /* 1036 * Set B_CACHE, indicating that the buffer is fully valid. This is 1037 * true even of NFS now. 1038 */ 1039 bp->b_flags |= B_CACHE; 1040 1041 /* 1042 * This bmap keeps the system from needing to do the bmap later, 1043 * perhaps when the system is attempting to do a sync. Since it 1044 * is likely that the indirect block -- or whatever other datastructure 1045 * that the filesystem needs is still in memory now, it is a good 1046 * thing to do this. Note also, that if the pageout daemon is 1047 * requesting a sync -- there might not be enough memory to do 1048 * the bmap then... So, this is important to do. 1049 */ 1050 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1051 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1052 } 1053 1054 /* 1055 * Set the *dirty* buffer range based upon the VM system dirty 1056 * pages. 1057 * 1058 * Mark the buffer pages as clean. We need to do this here to 1059 * satisfy the vnode_pager and the pageout daemon, so that it 1060 * thinks that the pages have been "cleaned". Note that since 1061 * the pages are in a delayed write buffer -- the VFS layer 1062 * "will" see that the pages get written out on the next sync, 1063 * or perhaps the cluster will be completed. 1064 */ 1065 vfs_clean_pages_dirty_buf(bp); 1066 bqrelse(bp); 1067 1068 /* 1069 * Wakeup the buffer flushing daemon if we have a lot of dirty 1070 * buffers (midpoint between our recovery point and our stall 1071 * point). 1072 */ 1073 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1074 1075 /* 1076 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1077 * due to the softdep code. 1078 */ 1079 } 1080 1081 /* 1082 * bdirty: 1083 * 1084 * Turn buffer into delayed write request. We must clear BIO_READ and 1085 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1086 * itself to properly update it in the dirty/clean lists. We mark it 1087 * B_DONE to ensure that any asynchronization of the buffer properly 1088 * clears B_DONE ( else a panic will occur later ). 1089 * 1090 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1091 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1092 * should only be called if the buffer is known-good. 1093 * 1094 * Since the buffer is not on a queue, we do not update the numfreebuffers 1095 * count. 1096 * 1097 * The buffer must be on QUEUE_NONE. 1098 */ 1099 void 1100 bdirty(struct buf *bp) 1101 { 1102 1103 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1104 bp, bp->b_vp, bp->b_flags); 1105 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1106 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1107 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1108 BUF_ASSERT_HELD(bp); 1109 bp->b_flags &= ~(B_RELBUF); 1110 bp->b_iocmd = BIO_WRITE; 1111 1112 if ((bp->b_flags & B_DELWRI) == 0) { 1113 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1114 reassignbuf(bp); 1115 atomic_add_int(&numdirtybuffers, 1); 1116 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1117 } 1118 } 1119 1120 /* 1121 * bundirty: 1122 * 1123 * Clear B_DELWRI for buffer. 1124 * 1125 * Since the buffer is not on a queue, we do not update the numfreebuffers 1126 * count. 1127 * 1128 * The buffer must be on QUEUE_NONE. 1129 */ 1130 1131 void 1132 bundirty(struct buf *bp) 1133 { 1134 1135 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1136 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1137 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1138 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1139 BUF_ASSERT_HELD(bp); 1140 1141 if (bp->b_flags & B_DELWRI) { 1142 bp->b_flags &= ~B_DELWRI; 1143 reassignbuf(bp); 1144 atomic_subtract_int(&numdirtybuffers, 1); 1145 numdirtywakeup(lodirtybuffers); 1146 } 1147 /* 1148 * Since it is now being written, we can clear its deferred write flag. 1149 */ 1150 bp->b_flags &= ~B_DEFERRED; 1151 } 1152 1153 /* 1154 * bawrite: 1155 * 1156 * Asynchronous write. Start output on a buffer, but do not wait for 1157 * it to complete. The buffer is released when the output completes. 1158 * 1159 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1160 * B_INVAL buffers. Not us. 1161 */ 1162 void 1163 bawrite(struct buf *bp) 1164 { 1165 1166 bp->b_flags |= B_ASYNC; 1167 (void) bwrite(bp); 1168 } 1169 1170 /* 1171 * bwillwrite: 1172 * 1173 * Called prior to the locking of any vnodes when we are expecting to 1174 * write. We do not want to starve the buffer cache with too many 1175 * dirty buffers so we block here. By blocking prior to the locking 1176 * of any vnodes we attempt to avoid the situation where a locked vnode 1177 * prevents the various system daemons from flushing related buffers. 1178 */ 1179 1180 void 1181 bwillwrite(void) 1182 { 1183 1184 if (numdirtybuffers >= hidirtybuffers) { 1185 mtx_lock(&nblock); 1186 while (numdirtybuffers >= hidirtybuffers) { 1187 bd_wakeup(1); 1188 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1189 msleep(&needsbuffer, &nblock, 1190 (PRIBIO + 4), "flswai", 0); 1191 } 1192 mtx_unlock(&nblock); 1193 } 1194 } 1195 1196 /* 1197 * Return true if we have too many dirty buffers. 1198 */ 1199 int 1200 buf_dirty_count_severe(void) 1201 { 1202 1203 return(numdirtybuffers >= hidirtybuffers); 1204 } 1205 1206 static __noinline int 1207 buf_vm_page_count_severe(void) 1208 { 1209 1210 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1211 1212 return vm_page_count_severe(); 1213 } 1214 1215 /* 1216 * brelse: 1217 * 1218 * Release a busy buffer and, if requested, free its resources. The 1219 * buffer will be stashed in the appropriate bufqueue[] allowing it 1220 * to be accessed later as a cache entity or reused for other purposes. 1221 */ 1222 void 1223 brelse(struct buf *bp) 1224 { 1225 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1226 bp, bp->b_vp, bp->b_flags); 1227 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1228 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1229 1230 if (bp->b_flags & B_MANAGED) { 1231 bqrelse(bp); 1232 return; 1233 } 1234 1235 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1236 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1237 /* 1238 * Failed write, redirty. Must clear BIO_ERROR to prevent 1239 * pages from being scrapped. If the error is anything 1240 * other than an I/O error (EIO), assume that retrying 1241 * is futile. 1242 */ 1243 bp->b_ioflags &= ~BIO_ERROR; 1244 bdirty(bp); 1245 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1246 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1247 /* 1248 * Either a failed I/O or we were asked to free or not 1249 * cache the buffer. 1250 */ 1251 bp->b_flags |= B_INVAL; 1252 if (!LIST_EMPTY(&bp->b_dep)) 1253 buf_deallocate(bp); 1254 if (bp->b_flags & B_DELWRI) { 1255 atomic_subtract_int(&numdirtybuffers, 1); 1256 numdirtywakeup(lodirtybuffers); 1257 } 1258 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1259 if ((bp->b_flags & B_VMIO) == 0) { 1260 if (bp->b_bufsize) 1261 allocbuf(bp, 0); 1262 if (bp->b_vp) 1263 brelvp(bp); 1264 } 1265 } 1266 1267 /* 1268 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1269 * is called with B_DELWRI set, the underlying pages may wind up 1270 * getting freed causing a previous write (bdwrite()) to get 'lost' 1271 * because pages associated with a B_DELWRI bp are marked clean. 1272 * 1273 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1274 * if B_DELWRI is set. 1275 * 1276 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1277 * on pages to return pages to the VM page queues. 1278 */ 1279 if (bp->b_flags & B_DELWRI) 1280 bp->b_flags &= ~B_RELBUF; 1281 else if (buf_vm_page_count_severe()) { 1282 /* 1283 * The locking of the BO_LOCK is not necessary since 1284 * BKGRDINPROG cannot be set while we hold the buf 1285 * lock, it can only be cleared if it is already 1286 * pending. 1287 */ 1288 if (bp->b_vp) { 1289 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1290 bp->b_flags |= B_RELBUF; 1291 } else 1292 bp->b_flags |= B_RELBUF; 1293 } 1294 1295 /* 1296 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1297 * constituted, not even NFS buffers now. Two flags effect this. If 1298 * B_INVAL, the struct buf is invalidated but the VM object is kept 1299 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1300 * 1301 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1302 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1303 * buffer is also B_INVAL because it hits the re-dirtying code above. 1304 * 1305 * Normally we can do this whether a buffer is B_DELWRI or not. If 1306 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1307 * the commit state and we cannot afford to lose the buffer. If the 1308 * buffer has a background write in progress, we need to keep it 1309 * around to prevent it from being reconstituted and starting a second 1310 * background write. 1311 */ 1312 if ((bp->b_flags & B_VMIO) 1313 && !(bp->b_vp->v_mount != NULL && 1314 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1315 !vn_isdisk(bp->b_vp, NULL) && 1316 (bp->b_flags & B_DELWRI)) 1317 ) { 1318 1319 int i, j, resid; 1320 vm_page_t m; 1321 off_t foff; 1322 vm_pindex_t poff; 1323 vm_object_t obj; 1324 1325 obj = bp->b_bufobj->bo_object; 1326 1327 /* 1328 * Get the base offset and length of the buffer. Note that 1329 * in the VMIO case if the buffer block size is not 1330 * page-aligned then b_data pointer may not be page-aligned. 1331 * But our b_pages[] array *IS* page aligned. 1332 * 1333 * block sizes less then DEV_BSIZE (usually 512) are not 1334 * supported due to the page granularity bits (m->valid, 1335 * m->dirty, etc...). 1336 * 1337 * See man buf(9) for more information 1338 */ 1339 resid = bp->b_bufsize; 1340 foff = bp->b_offset; 1341 VM_OBJECT_LOCK(obj); 1342 for (i = 0; i < bp->b_npages; i++) { 1343 int had_bogus = 0; 1344 1345 m = bp->b_pages[i]; 1346 1347 /* 1348 * If we hit a bogus page, fixup *all* the bogus pages 1349 * now. 1350 */ 1351 if (m == bogus_page) { 1352 poff = OFF_TO_IDX(bp->b_offset); 1353 had_bogus = 1; 1354 1355 for (j = i; j < bp->b_npages; j++) { 1356 vm_page_t mtmp; 1357 mtmp = bp->b_pages[j]; 1358 if (mtmp == bogus_page) { 1359 mtmp = vm_page_lookup(obj, poff + j); 1360 if (!mtmp) { 1361 panic("brelse: page missing\n"); 1362 } 1363 bp->b_pages[j] = mtmp; 1364 } 1365 } 1366 1367 if ((bp->b_flags & B_INVAL) == 0) { 1368 pmap_qenter( 1369 trunc_page((vm_offset_t)bp->b_data), 1370 bp->b_pages, bp->b_npages); 1371 } 1372 m = bp->b_pages[i]; 1373 } 1374 if ((bp->b_flags & B_NOCACHE) || 1375 (bp->b_ioflags & BIO_ERROR && 1376 bp->b_iocmd == BIO_READ)) { 1377 int poffset = foff & PAGE_MASK; 1378 int presid = resid > (PAGE_SIZE - poffset) ? 1379 (PAGE_SIZE - poffset) : resid; 1380 1381 KASSERT(presid >= 0, ("brelse: extra page")); 1382 vm_page_set_invalid(m, poffset, presid); 1383 if (had_bogus) 1384 printf("avoided corruption bug in bogus_page/brelse code\n"); 1385 } 1386 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1387 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1388 } 1389 VM_OBJECT_UNLOCK(obj); 1390 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1391 vfs_vmio_release(bp); 1392 1393 } else if (bp->b_flags & B_VMIO) { 1394 1395 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1396 vfs_vmio_release(bp); 1397 } 1398 1399 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1400 if (bp->b_bufsize != 0) 1401 allocbuf(bp, 0); 1402 if (bp->b_vp != NULL) 1403 brelvp(bp); 1404 } 1405 1406 if (BUF_LOCKRECURSED(bp)) { 1407 /* do not release to free list */ 1408 BUF_UNLOCK(bp); 1409 return; 1410 } 1411 1412 /* enqueue */ 1413 mtx_lock(&bqlock); 1414 /* Handle delayed bremfree() processing. */ 1415 if (bp->b_flags & B_REMFREE) 1416 bremfreel(bp); 1417 if (bp->b_qindex != QUEUE_NONE) 1418 panic("brelse: free buffer onto another queue???"); 1419 1420 /* 1421 * If the buffer has junk contents signal it and eventually 1422 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1423 * doesn't find it. 1424 */ 1425 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1426 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1427 bp->b_flags |= B_INVAL; 1428 if (bp->b_flags & B_INVAL) { 1429 if (bp->b_flags & B_DELWRI) 1430 bundirty(bp); 1431 if (bp->b_vp) 1432 brelvp(bp); 1433 } 1434 1435 /* buffers with no memory */ 1436 if (bp->b_bufsize == 0) { 1437 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1438 if (bp->b_vflags & BV_BKGRDINPROG) 1439 panic("losing buffer 1"); 1440 if (bp->b_kvasize) { 1441 bp->b_qindex = QUEUE_EMPTYKVA; 1442 } else { 1443 bp->b_qindex = QUEUE_EMPTY; 1444 } 1445 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1446 /* buffers with junk contents */ 1447 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1448 (bp->b_ioflags & BIO_ERROR)) { 1449 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1450 if (bp->b_vflags & BV_BKGRDINPROG) 1451 panic("losing buffer 2"); 1452 bp->b_qindex = QUEUE_CLEAN; 1453 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1454 /* remaining buffers */ 1455 } else { 1456 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1457 (B_DELWRI|B_NEEDSGIANT)) 1458 bp->b_qindex = QUEUE_DIRTY_GIANT; 1459 else if (bp->b_flags & B_DELWRI) 1460 bp->b_qindex = QUEUE_DIRTY; 1461 else 1462 bp->b_qindex = QUEUE_CLEAN; 1463 if (bp->b_flags & B_AGE) 1464 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1465 else 1466 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1467 } 1468 mtx_unlock(&bqlock); 1469 1470 /* 1471 * Fixup numfreebuffers count. The bp is on an appropriate queue 1472 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1473 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1474 * if B_INVAL is set ). 1475 */ 1476 1477 if (!(bp->b_flags & B_DELWRI)) 1478 bufcountwakeup(bp); 1479 1480 /* 1481 * Something we can maybe free or reuse 1482 */ 1483 if (bp->b_bufsize || bp->b_kvasize) 1484 bufspacewakeup(); 1485 1486 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1487 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1488 panic("brelse: not dirty"); 1489 /* unlock */ 1490 BUF_UNLOCK(bp); 1491 } 1492 1493 /* 1494 * Release a buffer back to the appropriate queue but do not try to free 1495 * it. The buffer is expected to be used again soon. 1496 * 1497 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1498 * biodone() to requeue an async I/O on completion. It is also used when 1499 * known good buffers need to be requeued but we think we may need the data 1500 * again soon. 1501 * 1502 * XXX we should be able to leave the B_RELBUF hint set on completion. 1503 */ 1504 void 1505 bqrelse(struct buf *bp) 1506 { 1507 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1508 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1509 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1510 1511 if (BUF_LOCKRECURSED(bp)) { 1512 /* do not release to free list */ 1513 BUF_UNLOCK(bp); 1514 return; 1515 } 1516 1517 if (bp->b_flags & B_MANAGED) { 1518 if (bp->b_flags & B_REMFREE) { 1519 mtx_lock(&bqlock); 1520 bremfreel(bp); 1521 mtx_unlock(&bqlock); 1522 } 1523 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1524 BUF_UNLOCK(bp); 1525 return; 1526 } 1527 1528 mtx_lock(&bqlock); 1529 /* Handle delayed bremfree() processing. */ 1530 if (bp->b_flags & B_REMFREE) 1531 bremfreel(bp); 1532 if (bp->b_qindex != QUEUE_NONE) 1533 panic("bqrelse: free buffer onto another queue???"); 1534 /* buffers with stale but valid contents */ 1535 if (bp->b_flags & B_DELWRI) { 1536 if (bp->b_flags & B_NEEDSGIANT) 1537 bp->b_qindex = QUEUE_DIRTY_GIANT; 1538 else 1539 bp->b_qindex = QUEUE_DIRTY; 1540 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1541 } else { 1542 /* 1543 * The locking of the BO_LOCK for checking of the 1544 * BV_BKGRDINPROG is not necessary since the 1545 * BV_BKGRDINPROG cannot be set while we hold the buf 1546 * lock, it can only be cleared if it is already 1547 * pending. 1548 */ 1549 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1550 bp->b_qindex = QUEUE_CLEAN; 1551 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1552 b_freelist); 1553 } else { 1554 /* 1555 * We are too low on memory, we have to try to free 1556 * the buffer (most importantly: the wired pages 1557 * making up its backing store) *now*. 1558 */ 1559 mtx_unlock(&bqlock); 1560 brelse(bp); 1561 return; 1562 } 1563 } 1564 mtx_unlock(&bqlock); 1565 1566 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1567 bufcountwakeup(bp); 1568 1569 /* 1570 * Something we can maybe free or reuse. 1571 */ 1572 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1573 bufspacewakeup(); 1574 1575 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1576 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1577 panic("bqrelse: not dirty"); 1578 /* unlock */ 1579 BUF_UNLOCK(bp); 1580 } 1581 1582 /* Give pages used by the bp back to the VM system (where possible) */ 1583 static void 1584 vfs_vmio_release(struct buf *bp) 1585 { 1586 int i; 1587 vm_page_t m; 1588 1589 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1590 for (i = 0; i < bp->b_npages; i++) { 1591 m = bp->b_pages[i]; 1592 bp->b_pages[i] = NULL; 1593 /* 1594 * In order to keep page LRU ordering consistent, put 1595 * everything on the inactive queue. 1596 */ 1597 vm_page_lock(m); 1598 vm_page_unwire(m, 0); 1599 /* 1600 * We don't mess with busy pages, it is 1601 * the responsibility of the process that 1602 * busied the pages to deal with them. 1603 */ 1604 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1605 m->wire_count == 0) { 1606 /* 1607 * Might as well free the page if we can and it has 1608 * no valid data. We also free the page if the 1609 * buffer was used for direct I/O 1610 */ 1611 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1612 m->hold_count == 0) { 1613 vm_page_free(m); 1614 } else if (bp->b_flags & B_DIRECT) { 1615 vm_page_try_to_free(m); 1616 } else if (buf_vm_page_count_severe()) { 1617 vm_page_try_to_cache(m); 1618 } 1619 } 1620 vm_page_unlock(m); 1621 } 1622 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1623 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1624 1625 if (bp->b_bufsize) { 1626 bufspacewakeup(); 1627 bp->b_bufsize = 0; 1628 } 1629 bp->b_npages = 0; 1630 bp->b_flags &= ~B_VMIO; 1631 if (bp->b_vp) 1632 brelvp(bp); 1633 } 1634 1635 /* 1636 * Check to see if a block at a particular lbn is available for a clustered 1637 * write. 1638 */ 1639 static int 1640 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1641 { 1642 struct buf *bpa; 1643 int match; 1644 1645 match = 0; 1646 1647 /* If the buf isn't in core skip it */ 1648 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1649 return (0); 1650 1651 /* If the buf is busy we don't want to wait for it */ 1652 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1653 return (0); 1654 1655 /* Only cluster with valid clusterable delayed write buffers */ 1656 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1657 (B_DELWRI | B_CLUSTEROK)) 1658 goto done; 1659 1660 if (bpa->b_bufsize != size) 1661 goto done; 1662 1663 /* 1664 * Check to see if it is in the expected place on disk and that the 1665 * block has been mapped. 1666 */ 1667 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1668 match = 1; 1669 done: 1670 BUF_UNLOCK(bpa); 1671 return (match); 1672 } 1673 1674 /* 1675 * vfs_bio_awrite: 1676 * 1677 * Implement clustered async writes for clearing out B_DELWRI buffers. 1678 * This is much better then the old way of writing only one buffer at 1679 * a time. Note that we may not be presented with the buffers in the 1680 * correct order, so we search for the cluster in both directions. 1681 */ 1682 int 1683 vfs_bio_awrite(struct buf *bp) 1684 { 1685 struct bufobj *bo; 1686 int i; 1687 int j; 1688 daddr_t lblkno = bp->b_lblkno; 1689 struct vnode *vp = bp->b_vp; 1690 int ncl; 1691 int nwritten; 1692 int size; 1693 int maxcl; 1694 1695 bo = &vp->v_bufobj; 1696 /* 1697 * right now we support clustered writing only to regular files. If 1698 * we find a clusterable block we could be in the middle of a cluster 1699 * rather then at the beginning. 1700 */ 1701 if ((vp->v_type == VREG) && 1702 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1703 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1704 1705 size = vp->v_mount->mnt_stat.f_iosize; 1706 maxcl = MAXPHYS / size; 1707 1708 BO_LOCK(bo); 1709 for (i = 1; i < maxcl; i++) 1710 if (vfs_bio_clcheck(vp, size, lblkno + i, 1711 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1712 break; 1713 1714 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1715 if (vfs_bio_clcheck(vp, size, lblkno - j, 1716 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1717 break; 1718 BO_UNLOCK(bo); 1719 --j; 1720 ncl = i + j; 1721 /* 1722 * this is a possible cluster write 1723 */ 1724 if (ncl != 1) { 1725 BUF_UNLOCK(bp); 1726 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1727 return nwritten; 1728 } 1729 } 1730 bremfree(bp); 1731 bp->b_flags |= B_ASYNC; 1732 /* 1733 * default (old) behavior, writing out only one block 1734 * 1735 * XXX returns b_bufsize instead of b_bcount for nwritten? 1736 */ 1737 nwritten = bp->b_bufsize; 1738 (void) bwrite(bp); 1739 1740 return nwritten; 1741 } 1742 1743 /* 1744 * getnewbuf: 1745 * 1746 * Find and initialize a new buffer header, freeing up existing buffers 1747 * in the bufqueues as necessary. The new buffer is returned locked. 1748 * 1749 * Important: B_INVAL is not set. If the caller wishes to throw the 1750 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1751 * 1752 * We block if: 1753 * We have insufficient buffer headers 1754 * We have insufficient buffer space 1755 * buffer_map is too fragmented ( space reservation fails ) 1756 * If we have to flush dirty buffers ( but we try to avoid this ) 1757 * 1758 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1759 * Instead we ask the buf daemon to do it for us. We attempt to 1760 * avoid piecemeal wakeups of the pageout daemon. 1761 */ 1762 1763 static struct buf * 1764 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1765 int gbflags) 1766 { 1767 struct thread *td; 1768 struct buf *bp; 1769 struct buf *nbp; 1770 int defrag = 0; 1771 int nqindex; 1772 static int flushingbufs; 1773 1774 td = curthread; 1775 /* 1776 * We can't afford to block since we might be holding a vnode lock, 1777 * which may prevent system daemons from running. We deal with 1778 * low-memory situations by proactively returning memory and running 1779 * async I/O rather then sync I/O. 1780 */ 1781 atomic_add_int(&getnewbufcalls, 1); 1782 atomic_subtract_int(&getnewbufrestarts, 1); 1783 restart: 1784 atomic_add_int(&getnewbufrestarts, 1); 1785 1786 /* 1787 * Setup for scan. If we do not have enough free buffers, 1788 * we setup a degenerate case that immediately fails. Note 1789 * that if we are specially marked process, we are allowed to 1790 * dip into our reserves. 1791 * 1792 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1793 * 1794 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1795 * However, there are a number of cases (defragging, reusing, ...) 1796 * where we cannot backup. 1797 */ 1798 mtx_lock(&bqlock); 1799 nqindex = QUEUE_EMPTYKVA; 1800 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1801 1802 if (nbp == NULL) { 1803 /* 1804 * If no EMPTYKVA buffers and we are either 1805 * defragging or reusing, locate a CLEAN buffer 1806 * to free or reuse. If bufspace useage is low 1807 * skip this step so we can allocate a new buffer. 1808 */ 1809 if (defrag || bufspace >= lobufspace) { 1810 nqindex = QUEUE_CLEAN; 1811 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1812 } 1813 1814 /* 1815 * If we could not find or were not allowed to reuse a 1816 * CLEAN buffer, check to see if it is ok to use an EMPTY 1817 * buffer. We can only use an EMPTY buffer if allocating 1818 * its KVA would not otherwise run us out of buffer space. 1819 */ 1820 if (nbp == NULL && defrag == 0 && 1821 bufspace + maxsize < hibufspace) { 1822 nqindex = QUEUE_EMPTY; 1823 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1824 } 1825 } 1826 1827 /* 1828 * Run scan, possibly freeing data and/or kva mappings on the fly 1829 * depending. 1830 */ 1831 1832 while ((bp = nbp) != NULL) { 1833 int qindex = nqindex; 1834 1835 /* 1836 * Calculate next bp ( we can only use it if we do not block 1837 * or do other fancy things ). 1838 */ 1839 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1840 switch(qindex) { 1841 case QUEUE_EMPTY: 1842 nqindex = QUEUE_EMPTYKVA; 1843 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1844 break; 1845 /* FALLTHROUGH */ 1846 case QUEUE_EMPTYKVA: 1847 nqindex = QUEUE_CLEAN; 1848 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1849 break; 1850 /* FALLTHROUGH */ 1851 case QUEUE_CLEAN: 1852 /* 1853 * nbp is NULL. 1854 */ 1855 break; 1856 } 1857 } 1858 /* 1859 * If we are defragging then we need a buffer with 1860 * b_kvasize != 0. XXX this situation should no longer 1861 * occur, if defrag is non-zero the buffer's b_kvasize 1862 * should also be non-zero at this point. XXX 1863 */ 1864 if (defrag && bp->b_kvasize == 0) { 1865 printf("Warning: defrag empty buffer %p\n", bp); 1866 continue; 1867 } 1868 1869 /* 1870 * Start freeing the bp. This is somewhat involved. nbp 1871 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1872 */ 1873 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1874 continue; 1875 if (bp->b_vp) { 1876 BO_LOCK(bp->b_bufobj); 1877 if (bp->b_vflags & BV_BKGRDINPROG) { 1878 BO_UNLOCK(bp->b_bufobj); 1879 BUF_UNLOCK(bp); 1880 continue; 1881 } 1882 BO_UNLOCK(bp->b_bufobj); 1883 } 1884 CTR6(KTR_BUF, 1885 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1886 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1887 bp->b_kvasize, bp->b_bufsize, qindex); 1888 1889 /* 1890 * Sanity Checks 1891 */ 1892 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1893 1894 /* 1895 * Note: we no longer distinguish between VMIO and non-VMIO 1896 * buffers. 1897 */ 1898 1899 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1900 1901 bremfreel(bp); 1902 mtx_unlock(&bqlock); 1903 1904 if (qindex == QUEUE_CLEAN) { 1905 if (bp->b_flags & B_VMIO) { 1906 bp->b_flags &= ~B_ASYNC; 1907 vfs_vmio_release(bp); 1908 } 1909 if (bp->b_vp) 1910 brelvp(bp); 1911 } 1912 1913 /* 1914 * NOTE: nbp is now entirely invalid. We can only restart 1915 * the scan from this point on. 1916 * 1917 * Get the rest of the buffer freed up. b_kva* is still 1918 * valid after this operation. 1919 */ 1920 1921 if (bp->b_rcred != NOCRED) { 1922 crfree(bp->b_rcred); 1923 bp->b_rcred = NOCRED; 1924 } 1925 if (bp->b_wcred != NOCRED) { 1926 crfree(bp->b_wcred); 1927 bp->b_wcred = NOCRED; 1928 } 1929 if (!LIST_EMPTY(&bp->b_dep)) 1930 buf_deallocate(bp); 1931 if (bp->b_vflags & BV_BKGRDINPROG) 1932 panic("losing buffer 3"); 1933 KASSERT(bp->b_vp == NULL, 1934 ("bp: %p still has vnode %p. qindex: %d", 1935 bp, bp->b_vp, qindex)); 1936 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1937 ("bp: %p still on a buffer list. xflags %X", 1938 bp, bp->b_xflags)); 1939 1940 if (bp->b_bufsize) 1941 allocbuf(bp, 0); 1942 1943 bp->b_flags = 0; 1944 bp->b_ioflags = 0; 1945 bp->b_xflags = 0; 1946 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 1947 ("buf %p still counted as free?", bp)); 1948 bp->b_vflags = 0; 1949 bp->b_vp = NULL; 1950 bp->b_blkno = bp->b_lblkno = 0; 1951 bp->b_offset = NOOFFSET; 1952 bp->b_iodone = 0; 1953 bp->b_error = 0; 1954 bp->b_resid = 0; 1955 bp->b_bcount = 0; 1956 bp->b_npages = 0; 1957 bp->b_dirtyoff = bp->b_dirtyend = 0; 1958 bp->b_bufobj = NULL; 1959 bp->b_pin_count = 0; 1960 bp->b_fsprivate1 = NULL; 1961 bp->b_fsprivate2 = NULL; 1962 bp->b_fsprivate3 = NULL; 1963 1964 LIST_INIT(&bp->b_dep); 1965 1966 /* 1967 * If we are defragging then free the buffer. 1968 */ 1969 if (defrag) { 1970 bp->b_flags |= B_INVAL; 1971 bfreekva(bp); 1972 brelse(bp); 1973 defrag = 0; 1974 goto restart; 1975 } 1976 1977 /* 1978 * Notify any waiters for the buffer lock about 1979 * identity change by freeing the buffer. 1980 */ 1981 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1982 bp->b_flags |= B_INVAL; 1983 bfreekva(bp); 1984 brelse(bp); 1985 goto restart; 1986 } 1987 1988 /* 1989 * If we are overcomitted then recover the buffer and its 1990 * KVM space. This occurs in rare situations when multiple 1991 * processes are blocked in getnewbuf() or allocbuf(). 1992 */ 1993 if (bufspace >= hibufspace) 1994 flushingbufs = 1; 1995 if (flushingbufs && bp->b_kvasize != 0) { 1996 bp->b_flags |= B_INVAL; 1997 bfreekva(bp); 1998 brelse(bp); 1999 goto restart; 2000 } 2001 if (bufspace < lobufspace) 2002 flushingbufs = 0; 2003 break; 2004 } 2005 2006 /* 2007 * If we exhausted our list, sleep as appropriate. We may have to 2008 * wakeup various daemons and write out some dirty buffers. 2009 * 2010 * Generally we are sleeping due to insufficient buffer space. 2011 */ 2012 2013 if (bp == NULL) { 2014 int flags, norunbuf; 2015 char *waitmsg; 2016 int fl; 2017 2018 if (defrag) { 2019 flags = VFS_BIO_NEED_BUFSPACE; 2020 waitmsg = "nbufkv"; 2021 } else if (bufspace >= hibufspace) { 2022 waitmsg = "nbufbs"; 2023 flags = VFS_BIO_NEED_BUFSPACE; 2024 } else { 2025 waitmsg = "newbuf"; 2026 flags = VFS_BIO_NEED_ANY; 2027 } 2028 mtx_lock(&nblock); 2029 needsbuffer |= flags; 2030 mtx_unlock(&nblock); 2031 mtx_unlock(&bqlock); 2032 2033 bd_speedup(); /* heeeelp */ 2034 if (gbflags & GB_NOWAIT_BD) 2035 return (NULL); 2036 2037 mtx_lock(&nblock); 2038 while (needsbuffer & flags) { 2039 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2040 mtx_unlock(&nblock); 2041 /* 2042 * getblk() is called with a vnode 2043 * locked, and some majority of the 2044 * dirty buffers may as well belong to 2045 * the vnode. Flushing the buffers 2046 * there would make a progress that 2047 * cannot be achieved by the 2048 * buf_daemon, that cannot lock the 2049 * vnode. 2050 */ 2051 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2052 (td->td_pflags & TDP_NORUNNINGBUF); 2053 /* play bufdaemon */ 2054 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2055 fl = buf_do_flush(vp); 2056 td->td_pflags &= norunbuf; 2057 mtx_lock(&nblock); 2058 if (fl != 0) 2059 continue; 2060 if ((needsbuffer & flags) == 0) 2061 break; 2062 } 2063 if (msleep(&needsbuffer, &nblock, 2064 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2065 mtx_unlock(&nblock); 2066 return (NULL); 2067 } 2068 } 2069 mtx_unlock(&nblock); 2070 } else { 2071 /* 2072 * We finally have a valid bp. We aren't quite out of the 2073 * woods, we still have to reserve kva space. In order 2074 * to keep fragmentation sane we only allocate kva in 2075 * BKVASIZE chunks. 2076 */ 2077 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2078 2079 if (maxsize != bp->b_kvasize) { 2080 vm_offset_t addr = 0; 2081 2082 bfreekva(bp); 2083 2084 vm_map_lock(buffer_map); 2085 if (vm_map_findspace(buffer_map, 2086 vm_map_min(buffer_map), maxsize, &addr)) { 2087 /* 2088 * Uh oh. Buffer map is to fragmented. We 2089 * must defragment the map. 2090 */ 2091 atomic_add_int(&bufdefragcnt, 1); 2092 vm_map_unlock(buffer_map); 2093 defrag = 1; 2094 bp->b_flags |= B_INVAL; 2095 brelse(bp); 2096 goto restart; 2097 } 2098 if (addr) { 2099 vm_map_insert(buffer_map, NULL, 0, 2100 addr, addr + maxsize, 2101 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2102 2103 bp->b_kvabase = (caddr_t) addr; 2104 bp->b_kvasize = maxsize; 2105 atomic_add_long(&bufspace, bp->b_kvasize); 2106 atomic_add_int(&bufreusecnt, 1); 2107 } 2108 vm_map_unlock(buffer_map); 2109 } 2110 bp->b_saveaddr = bp->b_kvabase; 2111 bp->b_data = bp->b_saveaddr; 2112 } 2113 return(bp); 2114 } 2115 2116 /* 2117 * buf_daemon: 2118 * 2119 * buffer flushing daemon. Buffers are normally flushed by the 2120 * update daemon but if it cannot keep up this process starts to 2121 * take the load in an attempt to prevent getnewbuf() from blocking. 2122 */ 2123 2124 static struct kproc_desc buf_kp = { 2125 "bufdaemon", 2126 buf_daemon, 2127 &bufdaemonproc 2128 }; 2129 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2130 2131 static int 2132 buf_do_flush(struct vnode *vp) 2133 { 2134 int flushed; 2135 2136 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2137 /* The list empty check here is slightly racy */ 2138 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2139 mtx_lock(&Giant); 2140 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2141 mtx_unlock(&Giant); 2142 } 2143 if (flushed == 0) { 2144 /* 2145 * Could not find any buffers without rollback 2146 * dependencies, so just write the first one 2147 * in the hopes of eventually making progress. 2148 */ 2149 flushbufqueues(vp, QUEUE_DIRTY, 1); 2150 if (!TAILQ_EMPTY( 2151 &bufqueues[QUEUE_DIRTY_GIANT])) { 2152 mtx_lock(&Giant); 2153 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2154 mtx_unlock(&Giant); 2155 } 2156 } 2157 return (flushed); 2158 } 2159 2160 static void 2161 buf_daemon() 2162 { 2163 int lodirtysave; 2164 2165 /* 2166 * This process needs to be suspended prior to shutdown sync. 2167 */ 2168 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2169 SHUTDOWN_PRI_LAST); 2170 2171 /* 2172 * This process is allowed to take the buffer cache to the limit 2173 */ 2174 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2175 mtx_lock(&bdlock); 2176 for (;;) { 2177 bd_request = 0; 2178 mtx_unlock(&bdlock); 2179 2180 kproc_suspend_check(bufdaemonproc); 2181 lodirtysave = lodirtybuffers; 2182 if (bd_speedupreq) { 2183 lodirtybuffers = numdirtybuffers / 2; 2184 bd_speedupreq = 0; 2185 } 2186 /* 2187 * Do the flush. Limit the amount of in-transit I/O we 2188 * allow to build up, otherwise we would completely saturate 2189 * the I/O system. Wakeup any waiting processes before we 2190 * normally would so they can run in parallel with our drain. 2191 */ 2192 while (numdirtybuffers > lodirtybuffers) { 2193 if (buf_do_flush(NULL) == 0) 2194 break; 2195 uio_yield(); 2196 } 2197 lodirtybuffers = lodirtysave; 2198 2199 /* 2200 * Only clear bd_request if we have reached our low water 2201 * mark. The buf_daemon normally waits 1 second and 2202 * then incrementally flushes any dirty buffers that have 2203 * built up, within reason. 2204 * 2205 * If we were unable to hit our low water mark and couldn't 2206 * find any flushable buffers, we sleep half a second. 2207 * Otherwise we loop immediately. 2208 */ 2209 mtx_lock(&bdlock); 2210 if (numdirtybuffers <= lodirtybuffers) { 2211 /* 2212 * We reached our low water mark, reset the 2213 * request and sleep until we are needed again. 2214 * The sleep is just so the suspend code works. 2215 */ 2216 bd_request = 0; 2217 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2218 } else { 2219 /* 2220 * We couldn't find any flushable dirty buffers but 2221 * still have too many dirty buffers, we 2222 * have to sleep and try again. (rare) 2223 */ 2224 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2225 } 2226 } 2227 } 2228 2229 /* 2230 * flushbufqueues: 2231 * 2232 * Try to flush a buffer in the dirty queue. We must be careful to 2233 * free up B_INVAL buffers instead of write them, which NFS is 2234 * particularly sensitive to. 2235 */ 2236 static int flushwithdeps = 0; 2237 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2238 0, "Number of buffers flushed with dependecies that require rollbacks"); 2239 2240 static int 2241 flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2242 { 2243 struct buf *sentinel; 2244 struct vnode *vp; 2245 struct mount *mp; 2246 struct buf *bp; 2247 int hasdeps; 2248 int flushed; 2249 int target; 2250 2251 if (lvp == NULL) { 2252 target = numdirtybuffers - lodirtybuffers; 2253 if (flushdeps && target > 2) 2254 target /= 2; 2255 } else 2256 target = flushbufqtarget; 2257 flushed = 0; 2258 bp = NULL; 2259 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2260 sentinel->b_qindex = QUEUE_SENTINEL; 2261 mtx_lock(&bqlock); 2262 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2263 while (flushed != target) { 2264 bp = TAILQ_NEXT(sentinel, b_freelist); 2265 if (bp != NULL) { 2266 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2267 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2268 b_freelist); 2269 } else 2270 break; 2271 /* 2272 * Skip sentinels inserted by other invocations of the 2273 * flushbufqueues(), taking care to not reorder them. 2274 */ 2275 if (bp->b_qindex == QUEUE_SENTINEL) 2276 continue; 2277 /* 2278 * Only flush the buffers that belong to the 2279 * vnode locked by the curthread. 2280 */ 2281 if (lvp != NULL && bp->b_vp != lvp) 2282 continue; 2283 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2284 continue; 2285 if (bp->b_pin_count > 0) { 2286 BUF_UNLOCK(bp); 2287 continue; 2288 } 2289 BO_LOCK(bp->b_bufobj); 2290 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2291 (bp->b_flags & B_DELWRI) == 0) { 2292 BO_UNLOCK(bp->b_bufobj); 2293 BUF_UNLOCK(bp); 2294 continue; 2295 } 2296 BO_UNLOCK(bp->b_bufobj); 2297 if (bp->b_flags & B_INVAL) { 2298 bremfreel(bp); 2299 mtx_unlock(&bqlock); 2300 brelse(bp); 2301 flushed++; 2302 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2303 mtx_lock(&bqlock); 2304 continue; 2305 } 2306 2307 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2308 if (flushdeps == 0) { 2309 BUF_UNLOCK(bp); 2310 continue; 2311 } 2312 hasdeps = 1; 2313 } else 2314 hasdeps = 0; 2315 /* 2316 * We must hold the lock on a vnode before writing 2317 * one of its buffers. Otherwise we may confuse, or 2318 * in the case of a snapshot vnode, deadlock the 2319 * system. 2320 * 2321 * The lock order here is the reverse of the normal 2322 * of vnode followed by buf lock. This is ok because 2323 * the NOWAIT will prevent deadlock. 2324 */ 2325 vp = bp->b_vp; 2326 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2327 BUF_UNLOCK(bp); 2328 continue; 2329 } 2330 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2331 mtx_unlock(&bqlock); 2332 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2333 bp, bp->b_vp, bp->b_flags); 2334 if (curproc == bufdaemonproc) 2335 vfs_bio_awrite(bp); 2336 else { 2337 bremfree(bp); 2338 bwrite(bp); 2339 notbufdflashes++; 2340 } 2341 vn_finished_write(mp); 2342 VOP_UNLOCK(vp, 0); 2343 flushwithdeps += hasdeps; 2344 flushed++; 2345 2346 /* 2347 * Sleeping on runningbufspace while holding 2348 * vnode lock leads to deadlock. 2349 */ 2350 if (curproc == bufdaemonproc) 2351 waitrunningbufspace(); 2352 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2353 mtx_lock(&bqlock); 2354 continue; 2355 } 2356 vn_finished_write(mp); 2357 BUF_UNLOCK(bp); 2358 } 2359 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2360 mtx_unlock(&bqlock); 2361 free(sentinel, M_TEMP); 2362 return (flushed); 2363 } 2364 2365 /* 2366 * Check to see if a block is currently memory resident. 2367 */ 2368 struct buf * 2369 incore(struct bufobj *bo, daddr_t blkno) 2370 { 2371 struct buf *bp; 2372 2373 BO_LOCK(bo); 2374 bp = gbincore(bo, blkno); 2375 BO_UNLOCK(bo); 2376 return (bp); 2377 } 2378 2379 /* 2380 * Returns true if no I/O is needed to access the 2381 * associated VM object. This is like incore except 2382 * it also hunts around in the VM system for the data. 2383 */ 2384 2385 static int 2386 inmem(struct vnode * vp, daddr_t blkno) 2387 { 2388 vm_object_t obj; 2389 vm_offset_t toff, tinc, size; 2390 vm_page_t m; 2391 vm_ooffset_t off; 2392 2393 ASSERT_VOP_LOCKED(vp, "inmem"); 2394 2395 if (incore(&vp->v_bufobj, blkno)) 2396 return 1; 2397 if (vp->v_mount == NULL) 2398 return 0; 2399 obj = vp->v_object; 2400 if (obj == NULL) 2401 return (0); 2402 2403 size = PAGE_SIZE; 2404 if (size > vp->v_mount->mnt_stat.f_iosize) 2405 size = vp->v_mount->mnt_stat.f_iosize; 2406 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2407 2408 VM_OBJECT_LOCK(obj); 2409 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2410 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2411 if (!m) 2412 goto notinmem; 2413 tinc = size; 2414 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2415 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2416 if (vm_page_is_valid(m, 2417 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2418 goto notinmem; 2419 } 2420 VM_OBJECT_UNLOCK(obj); 2421 return 1; 2422 2423 notinmem: 2424 VM_OBJECT_UNLOCK(obj); 2425 return (0); 2426 } 2427 2428 /* 2429 * Set the dirty range for a buffer based on the status of the dirty 2430 * bits in the pages comprising the buffer. The range is limited 2431 * to the size of the buffer. 2432 * 2433 * Tell the VM system that the pages associated with this buffer 2434 * are clean. This is used for delayed writes where the data is 2435 * going to go to disk eventually without additional VM intevention. 2436 * 2437 * Note that while we only really need to clean through to b_bcount, we 2438 * just go ahead and clean through to b_bufsize. 2439 */ 2440 static void 2441 vfs_clean_pages_dirty_buf(struct buf *bp) 2442 { 2443 vm_ooffset_t foff, noff, eoff; 2444 vm_page_t m; 2445 int i; 2446 2447 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2448 return; 2449 2450 foff = bp->b_offset; 2451 KASSERT(bp->b_offset != NOOFFSET, 2452 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2453 2454 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2455 vfs_drain_busy_pages(bp); 2456 vfs_setdirty_locked_object(bp); 2457 for (i = 0; i < bp->b_npages; i++) { 2458 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2459 eoff = noff; 2460 if (eoff > bp->b_offset + bp->b_bufsize) 2461 eoff = bp->b_offset + bp->b_bufsize; 2462 m = bp->b_pages[i]; 2463 vfs_page_set_validclean(bp, foff, m); 2464 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2465 foff = noff; 2466 } 2467 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2468 } 2469 2470 static void 2471 vfs_setdirty_locked_object(struct buf *bp) 2472 { 2473 vm_object_t object; 2474 int i; 2475 2476 object = bp->b_bufobj->bo_object; 2477 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2478 2479 /* 2480 * We qualify the scan for modified pages on whether the 2481 * object has been flushed yet. 2482 */ 2483 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2484 vm_offset_t boffset; 2485 vm_offset_t eoffset; 2486 2487 /* 2488 * test the pages to see if they have been modified directly 2489 * by users through the VM system. 2490 */ 2491 for (i = 0; i < bp->b_npages; i++) 2492 vm_page_test_dirty(bp->b_pages[i]); 2493 2494 /* 2495 * Calculate the encompassing dirty range, boffset and eoffset, 2496 * (eoffset - boffset) bytes. 2497 */ 2498 2499 for (i = 0; i < bp->b_npages; i++) { 2500 if (bp->b_pages[i]->dirty) 2501 break; 2502 } 2503 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2504 2505 for (i = bp->b_npages - 1; i >= 0; --i) { 2506 if (bp->b_pages[i]->dirty) { 2507 break; 2508 } 2509 } 2510 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2511 2512 /* 2513 * Fit it to the buffer. 2514 */ 2515 2516 if (eoffset > bp->b_bcount) 2517 eoffset = bp->b_bcount; 2518 2519 /* 2520 * If we have a good dirty range, merge with the existing 2521 * dirty range. 2522 */ 2523 2524 if (boffset < eoffset) { 2525 if (bp->b_dirtyoff > boffset) 2526 bp->b_dirtyoff = boffset; 2527 if (bp->b_dirtyend < eoffset) 2528 bp->b_dirtyend = eoffset; 2529 } 2530 } 2531 } 2532 2533 /* 2534 * getblk: 2535 * 2536 * Get a block given a specified block and offset into a file/device. 2537 * The buffers B_DONE bit will be cleared on return, making it almost 2538 * ready for an I/O initiation. B_INVAL may or may not be set on 2539 * return. The caller should clear B_INVAL prior to initiating a 2540 * READ. 2541 * 2542 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2543 * an existing buffer. 2544 * 2545 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2546 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2547 * and then cleared based on the backing VM. If the previous buffer is 2548 * non-0-sized but invalid, B_CACHE will be cleared. 2549 * 2550 * If getblk() must create a new buffer, the new buffer is returned with 2551 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2552 * case it is returned with B_INVAL clear and B_CACHE set based on the 2553 * backing VM. 2554 * 2555 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2556 * B_CACHE bit is clear. 2557 * 2558 * What this means, basically, is that the caller should use B_CACHE to 2559 * determine whether the buffer is fully valid or not and should clear 2560 * B_INVAL prior to issuing a read. If the caller intends to validate 2561 * the buffer by loading its data area with something, the caller needs 2562 * to clear B_INVAL. If the caller does this without issuing an I/O, 2563 * the caller should set B_CACHE ( as an optimization ), else the caller 2564 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2565 * a write attempt or if it was a successfull read. If the caller 2566 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2567 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2568 */ 2569 struct buf * 2570 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2571 int flags) 2572 { 2573 struct buf *bp; 2574 struct bufobj *bo; 2575 int error; 2576 2577 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2578 ASSERT_VOP_LOCKED(vp, "getblk"); 2579 if (size > MAXBSIZE) 2580 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2581 2582 bo = &vp->v_bufobj; 2583 loop: 2584 /* 2585 * Block if we are low on buffers. Certain processes are allowed 2586 * to completely exhaust the buffer cache. 2587 * 2588 * If this check ever becomes a bottleneck it may be better to 2589 * move it into the else, when gbincore() fails. At the moment 2590 * it isn't a problem. 2591 * 2592 * XXX remove if 0 sections (clean this up after its proven) 2593 */ 2594 if (numfreebuffers == 0) { 2595 if (TD_IS_IDLETHREAD(curthread)) 2596 return NULL; 2597 mtx_lock(&nblock); 2598 needsbuffer |= VFS_BIO_NEED_ANY; 2599 mtx_unlock(&nblock); 2600 } 2601 2602 BO_LOCK(bo); 2603 bp = gbincore(bo, blkno); 2604 if (bp != NULL) { 2605 int lockflags; 2606 /* 2607 * Buffer is in-core. If the buffer is not busy, it must 2608 * be on a queue. 2609 */ 2610 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2611 2612 if (flags & GB_LOCK_NOWAIT) 2613 lockflags |= LK_NOWAIT; 2614 2615 error = BUF_TIMELOCK(bp, lockflags, 2616 BO_MTX(bo), "getblk", slpflag, slptimeo); 2617 2618 /* 2619 * If we slept and got the lock we have to restart in case 2620 * the buffer changed identities. 2621 */ 2622 if (error == ENOLCK) 2623 goto loop; 2624 /* We timed out or were interrupted. */ 2625 else if (error) 2626 return (NULL); 2627 2628 /* 2629 * The buffer is locked. B_CACHE is cleared if the buffer is 2630 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2631 * and for a VMIO buffer B_CACHE is adjusted according to the 2632 * backing VM cache. 2633 */ 2634 if (bp->b_flags & B_INVAL) 2635 bp->b_flags &= ~B_CACHE; 2636 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2637 bp->b_flags |= B_CACHE; 2638 bremfree(bp); 2639 2640 /* 2641 * check for size inconsistancies for non-VMIO case. 2642 */ 2643 2644 if (bp->b_bcount != size) { 2645 if ((bp->b_flags & B_VMIO) == 0 || 2646 (size > bp->b_kvasize)) { 2647 if (bp->b_flags & B_DELWRI) { 2648 /* 2649 * If buffer is pinned and caller does 2650 * not want sleep waiting for it to be 2651 * unpinned, bail out 2652 * */ 2653 if (bp->b_pin_count > 0) { 2654 if (flags & GB_LOCK_NOWAIT) { 2655 bqrelse(bp); 2656 return (NULL); 2657 } else { 2658 bunpin_wait(bp); 2659 } 2660 } 2661 bp->b_flags |= B_NOCACHE; 2662 bwrite(bp); 2663 } else { 2664 if (LIST_EMPTY(&bp->b_dep)) { 2665 bp->b_flags |= B_RELBUF; 2666 brelse(bp); 2667 } else { 2668 bp->b_flags |= B_NOCACHE; 2669 bwrite(bp); 2670 } 2671 } 2672 goto loop; 2673 } 2674 } 2675 2676 /* 2677 * If the size is inconsistant in the VMIO case, we can resize 2678 * the buffer. This might lead to B_CACHE getting set or 2679 * cleared. If the size has not changed, B_CACHE remains 2680 * unchanged from its previous state. 2681 */ 2682 2683 if (bp->b_bcount != size) 2684 allocbuf(bp, size); 2685 2686 KASSERT(bp->b_offset != NOOFFSET, 2687 ("getblk: no buffer offset")); 2688 2689 /* 2690 * A buffer with B_DELWRI set and B_CACHE clear must 2691 * be committed before we can return the buffer in 2692 * order to prevent the caller from issuing a read 2693 * ( due to B_CACHE not being set ) and overwriting 2694 * it. 2695 * 2696 * Most callers, including NFS and FFS, need this to 2697 * operate properly either because they assume they 2698 * can issue a read if B_CACHE is not set, or because 2699 * ( for example ) an uncached B_DELWRI might loop due 2700 * to softupdates re-dirtying the buffer. In the latter 2701 * case, B_CACHE is set after the first write completes, 2702 * preventing further loops. 2703 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2704 * above while extending the buffer, we cannot allow the 2705 * buffer to remain with B_CACHE set after the write 2706 * completes or it will represent a corrupt state. To 2707 * deal with this we set B_NOCACHE to scrap the buffer 2708 * after the write. 2709 * 2710 * We might be able to do something fancy, like setting 2711 * B_CACHE in bwrite() except if B_DELWRI is already set, 2712 * so the below call doesn't set B_CACHE, but that gets real 2713 * confusing. This is much easier. 2714 */ 2715 2716 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2717 bp->b_flags |= B_NOCACHE; 2718 bwrite(bp); 2719 goto loop; 2720 } 2721 bp->b_flags &= ~B_DONE; 2722 } else { 2723 int bsize, maxsize, vmio; 2724 off_t offset; 2725 2726 /* 2727 * Buffer is not in-core, create new buffer. The buffer 2728 * returned by getnewbuf() is locked. Note that the returned 2729 * buffer is also considered valid (not marked B_INVAL). 2730 */ 2731 BO_UNLOCK(bo); 2732 /* 2733 * If the user does not want us to create the buffer, bail out 2734 * here. 2735 */ 2736 if (flags & GB_NOCREAT) 2737 return NULL; 2738 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2739 offset = blkno * bsize; 2740 vmio = vp->v_object != NULL; 2741 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2742 maxsize = imax(maxsize, bsize); 2743 2744 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2745 if (bp == NULL) { 2746 if (slpflag || slptimeo) 2747 return NULL; 2748 goto loop; 2749 } 2750 2751 /* 2752 * This code is used to make sure that a buffer is not 2753 * created while the getnewbuf routine is blocked. 2754 * This can be a problem whether the vnode is locked or not. 2755 * If the buffer is created out from under us, we have to 2756 * throw away the one we just created. 2757 * 2758 * Note: this must occur before we associate the buffer 2759 * with the vp especially considering limitations in 2760 * the splay tree implementation when dealing with duplicate 2761 * lblkno's. 2762 */ 2763 BO_LOCK(bo); 2764 if (gbincore(bo, blkno)) { 2765 BO_UNLOCK(bo); 2766 bp->b_flags |= B_INVAL; 2767 brelse(bp); 2768 goto loop; 2769 } 2770 2771 /* 2772 * Insert the buffer into the hash, so that it can 2773 * be found by incore. 2774 */ 2775 bp->b_blkno = bp->b_lblkno = blkno; 2776 bp->b_offset = offset; 2777 bgetvp(vp, bp); 2778 BO_UNLOCK(bo); 2779 2780 /* 2781 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2782 * buffer size starts out as 0, B_CACHE will be set by 2783 * allocbuf() for the VMIO case prior to it testing the 2784 * backing store for validity. 2785 */ 2786 2787 if (vmio) { 2788 bp->b_flags |= B_VMIO; 2789 #if defined(VFS_BIO_DEBUG) 2790 if (vn_canvmio(vp) != TRUE) 2791 printf("getblk: VMIO on vnode type %d\n", 2792 vp->v_type); 2793 #endif 2794 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2795 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2796 bp, vp->v_object, bp->b_bufobj->bo_object)); 2797 } else { 2798 bp->b_flags &= ~B_VMIO; 2799 KASSERT(bp->b_bufobj->bo_object == NULL, 2800 ("ARGH! has b_bufobj->bo_object %p %p\n", 2801 bp, bp->b_bufobj->bo_object)); 2802 } 2803 2804 allocbuf(bp, size); 2805 bp->b_flags &= ~B_DONE; 2806 } 2807 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2808 BUF_ASSERT_HELD(bp); 2809 KASSERT(bp->b_bufobj == bo, 2810 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2811 return (bp); 2812 } 2813 2814 /* 2815 * Get an empty, disassociated buffer of given size. The buffer is initially 2816 * set to B_INVAL. 2817 */ 2818 struct buf * 2819 geteblk(int size, int flags) 2820 { 2821 struct buf *bp; 2822 int maxsize; 2823 2824 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2825 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2826 if ((flags & GB_NOWAIT_BD) && 2827 (curthread->td_pflags & TDP_BUFNEED) != 0) 2828 return (NULL); 2829 } 2830 allocbuf(bp, size); 2831 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2832 BUF_ASSERT_HELD(bp); 2833 return (bp); 2834 } 2835 2836 2837 /* 2838 * This code constitutes the buffer memory from either anonymous system 2839 * memory (in the case of non-VMIO operations) or from an associated 2840 * VM object (in the case of VMIO operations). This code is able to 2841 * resize a buffer up or down. 2842 * 2843 * Note that this code is tricky, and has many complications to resolve 2844 * deadlock or inconsistant data situations. Tread lightly!!! 2845 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2846 * the caller. Calling this code willy nilly can result in the loss of data. 2847 * 2848 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2849 * B_CACHE for the non-VMIO case. 2850 */ 2851 2852 int 2853 allocbuf(struct buf *bp, int size) 2854 { 2855 int newbsize, mbsize; 2856 int i; 2857 2858 BUF_ASSERT_HELD(bp); 2859 2860 if (bp->b_kvasize < size) 2861 panic("allocbuf: buffer too small"); 2862 2863 if ((bp->b_flags & B_VMIO) == 0) { 2864 caddr_t origbuf; 2865 int origbufsize; 2866 /* 2867 * Just get anonymous memory from the kernel. Don't 2868 * mess with B_CACHE. 2869 */ 2870 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2871 if (bp->b_flags & B_MALLOC) 2872 newbsize = mbsize; 2873 else 2874 newbsize = round_page(size); 2875 2876 if (newbsize < bp->b_bufsize) { 2877 /* 2878 * malloced buffers are not shrunk 2879 */ 2880 if (bp->b_flags & B_MALLOC) { 2881 if (newbsize) { 2882 bp->b_bcount = size; 2883 } else { 2884 free(bp->b_data, M_BIOBUF); 2885 if (bp->b_bufsize) { 2886 atomic_subtract_long( 2887 &bufmallocspace, 2888 bp->b_bufsize); 2889 bufspacewakeup(); 2890 bp->b_bufsize = 0; 2891 } 2892 bp->b_saveaddr = bp->b_kvabase; 2893 bp->b_data = bp->b_saveaddr; 2894 bp->b_bcount = 0; 2895 bp->b_flags &= ~B_MALLOC; 2896 } 2897 return 1; 2898 } 2899 vm_hold_free_pages(bp, newbsize); 2900 } else if (newbsize > bp->b_bufsize) { 2901 /* 2902 * We only use malloced memory on the first allocation. 2903 * and revert to page-allocated memory when the buffer 2904 * grows. 2905 */ 2906 /* 2907 * There is a potential smp race here that could lead 2908 * to bufmallocspace slightly passing the max. It 2909 * is probably extremely rare and not worth worrying 2910 * over. 2911 */ 2912 if ( (bufmallocspace < maxbufmallocspace) && 2913 (bp->b_bufsize == 0) && 2914 (mbsize <= PAGE_SIZE/2)) { 2915 2916 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2917 bp->b_bufsize = mbsize; 2918 bp->b_bcount = size; 2919 bp->b_flags |= B_MALLOC; 2920 atomic_add_long(&bufmallocspace, mbsize); 2921 return 1; 2922 } 2923 origbuf = NULL; 2924 origbufsize = 0; 2925 /* 2926 * If the buffer is growing on its other-than-first allocation, 2927 * then we revert to the page-allocation scheme. 2928 */ 2929 if (bp->b_flags & B_MALLOC) { 2930 origbuf = bp->b_data; 2931 origbufsize = bp->b_bufsize; 2932 bp->b_data = bp->b_kvabase; 2933 if (bp->b_bufsize) { 2934 atomic_subtract_long(&bufmallocspace, 2935 bp->b_bufsize); 2936 bufspacewakeup(); 2937 bp->b_bufsize = 0; 2938 } 2939 bp->b_flags &= ~B_MALLOC; 2940 newbsize = round_page(newbsize); 2941 } 2942 vm_hold_load_pages( 2943 bp, 2944 (vm_offset_t) bp->b_data + bp->b_bufsize, 2945 (vm_offset_t) bp->b_data + newbsize); 2946 if (origbuf) { 2947 bcopy(origbuf, bp->b_data, origbufsize); 2948 free(origbuf, M_BIOBUF); 2949 } 2950 } 2951 } else { 2952 int desiredpages; 2953 2954 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2955 desiredpages = (size == 0) ? 0 : 2956 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2957 2958 if (bp->b_flags & B_MALLOC) 2959 panic("allocbuf: VMIO buffer can't be malloced"); 2960 /* 2961 * Set B_CACHE initially if buffer is 0 length or will become 2962 * 0-length. 2963 */ 2964 if (size == 0 || bp->b_bufsize == 0) 2965 bp->b_flags |= B_CACHE; 2966 2967 if (newbsize < bp->b_bufsize) { 2968 /* 2969 * DEV_BSIZE aligned new buffer size is less then the 2970 * DEV_BSIZE aligned existing buffer size. Figure out 2971 * if we have to remove any pages. 2972 */ 2973 if (desiredpages < bp->b_npages) { 2974 vm_page_t m; 2975 2976 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2977 for (i = desiredpages; i < bp->b_npages; i++) { 2978 /* 2979 * the page is not freed here -- it 2980 * is the responsibility of 2981 * vnode_pager_setsize 2982 */ 2983 m = bp->b_pages[i]; 2984 KASSERT(m != bogus_page, 2985 ("allocbuf: bogus page found")); 2986 while (vm_page_sleep_if_busy(m, TRUE, 2987 "biodep")) 2988 continue; 2989 2990 bp->b_pages[i] = NULL; 2991 vm_page_lock(m); 2992 vm_page_unwire(m, 0); 2993 vm_page_unlock(m); 2994 } 2995 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2996 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2997 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2998 bp->b_npages = desiredpages; 2999 } 3000 } else if (size > bp->b_bcount) { 3001 /* 3002 * We are growing the buffer, possibly in a 3003 * byte-granular fashion. 3004 */ 3005 vm_object_t obj; 3006 vm_offset_t toff; 3007 vm_offset_t tinc; 3008 3009 /* 3010 * Step 1, bring in the VM pages from the object, 3011 * allocating them if necessary. We must clear 3012 * B_CACHE if these pages are not valid for the 3013 * range covered by the buffer. 3014 */ 3015 3016 obj = bp->b_bufobj->bo_object; 3017 3018 VM_OBJECT_LOCK(obj); 3019 while (bp->b_npages < desiredpages) { 3020 vm_page_t m; 3021 3022 /* 3023 * We must allocate system pages since blocking 3024 * here could intefere with paging I/O, no 3025 * matter which process we are. 3026 * 3027 * We can only test VPO_BUSY here. Blocking on 3028 * m->busy might lead to a deadlock: 3029 * vm_fault->getpages->cluster_read->allocbuf 3030 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3031 */ 3032 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3033 bp->b_npages, VM_ALLOC_NOBUSY | 3034 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3035 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3036 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3037 if (m->valid == 0) 3038 bp->b_flags &= ~B_CACHE; 3039 bp->b_pages[bp->b_npages] = m; 3040 ++bp->b_npages; 3041 } 3042 3043 /* 3044 * Step 2. We've loaded the pages into the buffer, 3045 * we have to figure out if we can still have B_CACHE 3046 * set. Note that B_CACHE is set according to the 3047 * byte-granular range ( bcount and size ), new the 3048 * aligned range ( newbsize ). 3049 * 3050 * The VM test is against m->valid, which is DEV_BSIZE 3051 * aligned. Needless to say, the validity of the data 3052 * needs to also be DEV_BSIZE aligned. Note that this 3053 * fails with NFS if the server or some other client 3054 * extends the file's EOF. If our buffer is resized, 3055 * B_CACHE may remain set! XXX 3056 */ 3057 3058 toff = bp->b_bcount; 3059 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3060 3061 while ((bp->b_flags & B_CACHE) && toff < size) { 3062 vm_pindex_t pi; 3063 3064 if (tinc > (size - toff)) 3065 tinc = size - toff; 3066 3067 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3068 PAGE_SHIFT; 3069 3070 vfs_buf_test_cache( 3071 bp, 3072 bp->b_offset, 3073 toff, 3074 tinc, 3075 bp->b_pages[pi] 3076 ); 3077 toff += tinc; 3078 tinc = PAGE_SIZE; 3079 } 3080 VM_OBJECT_UNLOCK(obj); 3081 3082 /* 3083 * Step 3, fixup the KVM pmap. Remember that 3084 * bp->b_data is relative to bp->b_offset, but 3085 * bp->b_offset may be offset into the first page. 3086 */ 3087 3088 bp->b_data = (caddr_t) 3089 trunc_page((vm_offset_t)bp->b_data); 3090 pmap_qenter( 3091 (vm_offset_t)bp->b_data, 3092 bp->b_pages, 3093 bp->b_npages 3094 ); 3095 3096 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3097 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3098 } 3099 } 3100 if (newbsize < bp->b_bufsize) 3101 bufspacewakeup(); 3102 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3103 bp->b_bcount = size; /* requested buffer size */ 3104 return 1; 3105 } 3106 3107 void 3108 biodone(struct bio *bp) 3109 { 3110 struct mtx *mtxp; 3111 void (*done)(struct bio *); 3112 3113 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3114 mtx_lock(mtxp); 3115 bp->bio_flags |= BIO_DONE; 3116 done = bp->bio_done; 3117 if (done == NULL) 3118 wakeup(bp); 3119 mtx_unlock(mtxp); 3120 if (done != NULL) 3121 done(bp); 3122 } 3123 3124 /* 3125 * Wait for a BIO to finish. 3126 * 3127 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3128 * case is not yet clear. 3129 */ 3130 int 3131 biowait(struct bio *bp, const char *wchan) 3132 { 3133 struct mtx *mtxp; 3134 3135 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3136 mtx_lock(mtxp); 3137 while ((bp->bio_flags & BIO_DONE) == 0) 3138 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3139 mtx_unlock(mtxp); 3140 if (bp->bio_error != 0) 3141 return (bp->bio_error); 3142 if (!(bp->bio_flags & BIO_ERROR)) 3143 return (0); 3144 return (EIO); 3145 } 3146 3147 void 3148 biofinish(struct bio *bp, struct devstat *stat, int error) 3149 { 3150 3151 if (error) { 3152 bp->bio_error = error; 3153 bp->bio_flags |= BIO_ERROR; 3154 } 3155 if (stat != NULL) 3156 devstat_end_transaction_bio(stat, bp); 3157 biodone(bp); 3158 } 3159 3160 /* 3161 * bufwait: 3162 * 3163 * Wait for buffer I/O completion, returning error status. The buffer 3164 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3165 * error and cleared. 3166 */ 3167 int 3168 bufwait(struct buf *bp) 3169 { 3170 if (bp->b_iocmd == BIO_READ) 3171 bwait(bp, PRIBIO, "biord"); 3172 else 3173 bwait(bp, PRIBIO, "biowr"); 3174 if (bp->b_flags & B_EINTR) { 3175 bp->b_flags &= ~B_EINTR; 3176 return (EINTR); 3177 } 3178 if (bp->b_ioflags & BIO_ERROR) { 3179 return (bp->b_error ? bp->b_error : EIO); 3180 } else { 3181 return (0); 3182 } 3183 } 3184 3185 /* 3186 * Call back function from struct bio back up to struct buf. 3187 */ 3188 static void 3189 bufdonebio(struct bio *bip) 3190 { 3191 struct buf *bp; 3192 3193 bp = bip->bio_caller2; 3194 bp->b_resid = bp->b_bcount - bip->bio_completed; 3195 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3196 bp->b_ioflags = bip->bio_flags; 3197 bp->b_error = bip->bio_error; 3198 if (bp->b_error) 3199 bp->b_ioflags |= BIO_ERROR; 3200 bufdone(bp); 3201 g_destroy_bio(bip); 3202 } 3203 3204 void 3205 dev_strategy(struct cdev *dev, struct buf *bp) 3206 { 3207 struct cdevsw *csw; 3208 struct bio *bip; 3209 int ref; 3210 3211 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3212 panic("b_iocmd botch"); 3213 for (;;) { 3214 bip = g_new_bio(); 3215 if (bip != NULL) 3216 break; 3217 /* Try again later */ 3218 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3219 } 3220 bip->bio_cmd = bp->b_iocmd; 3221 bip->bio_offset = bp->b_iooffset; 3222 bip->bio_length = bp->b_bcount; 3223 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3224 bip->bio_data = bp->b_data; 3225 bip->bio_done = bufdonebio; 3226 bip->bio_caller2 = bp; 3227 bip->bio_dev = dev; 3228 KASSERT(dev->si_refcount > 0, 3229 ("dev_strategy on un-referenced struct cdev *(%s)", 3230 devtoname(dev))); 3231 csw = dev_refthread(dev, &ref); 3232 if (csw == NULL) { 3233 g_destroy_bio(bip); 3234 bp->b_error = ENXIO; 3235 bp->b_ioflags = BIO_ERROR; 3236 bufdone(bp); 3237 return; 3238 } 3239 (*csw->d_strategy)(bip); 3240 dev_relthread(dev, ref); 3241 } 3242 3243 /* 3244 * bufdone: 3245 * 3246 * Finish I/O on a buffer, optionally calling a completion function. 3247 * This is usually called from an interrupt so process blocking is 3248 * not allowed. 3249 * 3250 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3251 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3252 * assuming B_INVAL is clear. 3253 * 3254 * For the VMIO case, we set B_CACHE if the op was a read and no 3255 * read error occured, or if the op was a write. B_CACHE is never 3256 * set if the buffer is invalid or otherwise uncacheable. 3257 * 3258 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3259 * initiator to leave B_INVAL set to brelse the buffer out of existance 3260 * in the biodone routine. 3261 */ 3262 void 3263 bufdone(struct buf *bp) 3264 { 3265 struct bufobj *dropobj; 3266 void (*biodone)(struct buf *); 3267 3268 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3269 dropobj = NULL; 3270 3271 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3272 BUF_ASSERT_HELD(bp); 3273 3274 runningbufwakeup(bp); 3275 if (bp->b_iocmd == BIO_WRITE) 3276 dropobj = bp->b_bufobj; 3277 /* call optional completion function if requested */ 3278 if (bp->b_iodone != NULL) { 3279 biodone = bp->b_iodone; 3280 bp->b_iodone = NULL; 3281 (*biodone) (bp); 3282 if (dropobj) 3283 bufobj_wdrop(dropobj); 3284 return; 3285 } 3286 3287 bufdone_finish(bp); 3288 3289 if (dropobj) 3290 bufobj_wdrop(dropobj); 3291 } 3292 3293 void 3294 bufdone_finish(struct buf *bp) 3295 { 3296 BUF_ASSERT_HELD(bp); 3297 3298 if (!LIST_EMPTY(&bp->b_dep)) 3299 buf_complete(bp); 3300 3301 if (bp->b_flags & B_VMIO) { 3302 int i; 3303 vm_ooffset_t foff; 3304 vm_page_t m; 3305 vm_object_t obj; 3306 int bogus, iosize; 3307 struct vnode *vp = bp->b_vp; 3308 3309 obj = bp->b_bufobj->bo_object; 3310 3311 #if defined(VFS_BIO_DEBUG) 3312 mp_fixme("usecount and vflag accessed without locks."); 3313 if (vp->v_usecount == 0) { 3314 panic("biodone: zero vnode ref count"); 3315 } 3316 3317 KASSERT(vp->v_object != NULL, 3318 ("biodone: vnode %p has no vm_object", vp)); 3319 #endif 3320 3321 foff = bp->b_offset; 3322 KASSERT(bp->b_offset != NOOFFSET, 3323 ("biodone: no buffer offset")); 3324 3325 VM_OBJECT_LOCK(obj); 3326 #if defined(VFS_BIO_DEBUG) 3327 if (obj->paging_in_progress < bp->b_npages) { 3328 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3329 obj->paging_in_progress, bp->b_npages); 3330 } 3331 #endif 3332 3333 /* 3334 * Set B_CACHE if the op was a normal read and no error 3335 * occured. B_CACHE is set for writes in the b*write() 3336 * routines. 3337 */ 3338 iosize = bp->b_bcount - bp->b_resid; 3339 if (bp->b_iocmd == BIO_READ && 3340 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3341 !(bp->b_ioflags & BIO_ERROR)) { 3342 bp->b_flags |= B_CACHE; 3343 } 3344 bogus = 0; 3345 for (i = 0; i < bp->b_npages; i++) { 3346 int bogusflag = 0; 3347 int resid; 3348 3349 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3350 if (resid > iosize) 3351 resid = iosize; 3352 3353 /* 3354 * cleanup bogus pages, restoring the originals 3355 */ 3356 m = bp->b_pages[i]; 3357 if (m == bogus_page) { 3358 bogus = bogusflag = 1; 3359 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3360 if (m == NULL) 3361 panic("biodone: page disappeared!"); 3362 bp->b_pages[i] = m; 3363 } 3364 #if defined(VFS_BIO_DEBUG) 3365 if (OFF_TO_IDX(foff) != m->pindex) { 3366 printf( 3367 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3368 (intmax_t)foff, (uintmax_t)m->pindex); 3369 } 3370 #endif 3371 3372 /* 3373 * In the write case, the valid and clean bits are 3374 * already changed correctly ( see bdwrite() ), so we 3375 * only need to do this here in the read case. 3376 */ 3377 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3378 KASSERT((m->dirty & vm_page_bits(foff & 3379 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3380 " page %p has unexpected dirty bits", m)); 3381 vfs_page_set_valid(bp, foff, m); 3382 } 3383 3384 /* 3385 * when debugging new filesystems or buffer I/O methods, this 3386 * is the most common error that pops up. if you see this, you 3387 * have not set the page busy flag correctly!!! 3388 */ 3389 if (m->busy == 0) { 3390 printf("biodone: page busy < 0, " 3391 "pindex: %d, foff: 0x(%x,%x), " 3392 "resid: %d, index: %d\n", 3393 (int) m->pindex, (int)(foff >> 32), 3394 (int) foff & 0xffffffff, resid, i); 3395 if (!vn_isdisk(vp, NULL)) 3396 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3397 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3398 (intmax_t) bp->b_lblkno, 3399 bp->b_flags, bp->b_npages); 3400 else 3401 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3402 (intmax_t) bp->b_lblkno, 3403 bp->b_flags, bp->b_npages); 3404 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3405 (u_long)m->valid, (u_long)m->dirty, 3406 m->wire_count); 3407 panic("biodone: page busy < 0\n"); 3408 } 3409 vm_page_io_finish(m); 3410 vm_object_pip_subtract(obj, 1); 3411 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3412 iosize -= resid; 3413 } 3414 vm_object_pip_wakeupn(obj, 0); 3415 VM_OBJECT_UNLOCK(obj); 3416 if (bogus) 3417 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3418 bp->b_pages, bp->b_npages); 3419 } 3420 3421 /* 3422 * For asynchronous completions, release the buffer now. The brelse 3423 * will do a wakeup there if necessary - so no need to do a wakeup 3424 * here in the async case. The sync case always needs to do a wakeup. 3425 */ 3426 3427 if (bp->b_flags & B_ASYNC) { 3428 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3429 brelse(bp); 3430 else 3431 bqrelse(bp); 3432 } else 3433 bdone(bp); 3434 } 3435 3436 /* 3437 * This routine is called in lieu of iodone in the case of 3438 * incomplete I/O. This keeps the busy status for pages 3439 * consistant. 3440 */ 3441 void 3442 vfs_unbusy_pages(struct buf *bp) 3443 { 3444 int i; 3445 vm_object_t obj; 3446 vm_page_t m; 3447 3448 runningbufwakeup(bp); 3449 if (!(bp->b_flags & B_VMIO)) 3450 return; 3451 3452 obj = bp->b_bufobj->bo_object; 3453 VM_OBJECT_LOCK(obj); 3454 for (i = 0; i < bp->b_npages; i++) { 3455 m = bp->b_pages[i]; 3456 if (m == bogus_page) { 3457 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3458 if (!m) 3459 panic("vfs_unbusy_pages: page missing\n"); 3460 bp->b_pages[i] = m; 3461 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3462 bp->b_pages, bp->b_npages); 3463 } 3464 vm_object_pip_subtract(obj, 1); 3465 vm_page_io_finish(m); 3466 } 3467 vm_object_pip_wakeupn(obj, 0); 3468 VM_OBJECT_UNLOCK(obj); 3469 } 3470 3471 /* 3472 * vfs_page_set_valid: 3473 * 3474 * Set the valid bits in a page based on the supplied offset. The 3475 * range is restricted to the buffer's size. 3476 * 3477 * This routine is typically called after a read completes. 3478 */ 3479 static void 3480 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3481 { 3482 vm_ooffset_t eoff; 3483 3484 /* 3485 * Compute the end offset, eoff, such that [off, eoff) does not span a 3486 * page boundary and eoff is not greater than the end of the buffer. 3487 * The end of the buffer, in this case, is our file EOF, not the 3488 * allocation size of the buffer. 3489 */ 3490 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3491 if (eoff > bp->b_offset + bp->b_bcount) 3492 eoff = bp->b_offset + bp->b_bcount; 3493 3494 /* 3495 * Set valid range. This is typically the entire buffer and thus the 3496 * entire page. 3497 */ 3498 if (eoff > off) 3499 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3500 } 3501 3502 /* 3503 * vfs_page_set_validclean: 3504 * 3505 * Set the valid bits and clear the dirty bits in a page based on the 3506 * supplied offset. The range is restricted to the buffer's size. 3507 */ 3508 static void 3509 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3510 { 3511 vm_ooffset_t soff, eoff; 3512 3513 /* 3514 * Start and end offsets in buffer. eoff - soff may not cross a 3515 * page boundry or cross the end of the buffer. The end of the 3516 * buffer, in this case, is our file EOF, not the allocation size 3517 * of the buffer. 3518 */ 3519 soff = off; 3520 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3521 if (eoff > bp->b_offset + bp->b_bcount) 3522 eoff = bp->b_offset + bp->b_bcount; 3523 3524 /* 3525 * Set valid range. This is typically the entire buffer and thus the 3526 * entire page. 3527 */ 3528 if (eoff > soff) { 3529 vm_page_set_validclean( 3530 m, 3531 (vm_offset_t) (soff & PAGE_MASK), 3532 (vm_offset_t) (eoff - soff) 3533 ); 3534 } 3535 } 3536 3537 /* 3538 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3539 * any page is busy, drain the flag. 3540 */ 3541 static void 3542 vfs_drain_busy_pages(struct buf *bp) 3543 { 3544 vm_page_t m; 3545 int i, last_busied; 3546 3547 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED); 3548 last_busied = 0; 3549 for (i = 0; i < bp->b_npages; i++) { 3550 m = bp->b_pages[i]; 3551 if ((m->oflags & VPO_BUSY) != 0) { 3552 for (; last_busied < i; last_busied++) 3553 vm_page_busy(bp->b_pages[last_busied]); 3554 while ((m->oflags & VPO_BUSY) != 0) 3555 vm_page_sleep(m, "vbpage"); 3556 } 3557 } 3558 for (i = 0; i < last_busied; i++) 3559 vm_page_wakeup(bp->b_pages[i]); 3560 } 3561 3562 /* 3563 * This routine is called before a device strategy routine. 3564 * It is used to tell the VM system that paging I/O is in 3565 * progress, and treat the pages associated with the buffer 3566 * almost as being VPO_BUSY. Also the object paging_in_progress 3567 * flag is handled to make sure that the object doesn't become 3568 * inconsistant. 3569 * 3570 * Since I/O has not been initiated yet, certain buffer flags 3571 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3572 * and should be ignored. 3573 */ 3574 void 3575 vfs_busy_pages(struct buf *bp, int clear_modify) 3576 { 3577 int i, bogus; 3578 vm_object_t obj; 3579 vm_ooffset_t foff; 3580 vm_page_t m; 3581 3582 if (!(bp->b_flags & B_VMIO)) 3583 return; 3584 3585 obj = bp->b_bufobj->bo_object; 3586 foff = bp->b_offset; 3587 KASSERT(bp->b_offset != NOOFFSET, 3588 ("vfs_busy_pages: no buffer offset")); 3589 VM_OBJECT_LOCK(obj); 3590 vfs_drain_busy_pages(bp); 3591 if (bp->b_bufsize != 0) 3592 vfs_setdirty_locked_object(bp); 3593 bogus = 0; 3594 for (i = 0; i < bp->b_npages; i++) { 3595 m = bp->b_pages[i]; 3596 3597 if ((bp->b_flags & B_CLUSTER) == 0) { 3598 vm_object_pip_add(obj, 1); 3599 vm_page_io_start(m); 3600 } 3601 /* 3602 * When readying a buffer for a read ( i.e 3603 * clear_modify == 0 ), it is important to do 3604 * bogus_page replacement for valid pages in 3605 * partially instantiated buffers. Partially 3606 * instantiated buffers can, in turn, occur when 3607 * reconstituting a buffer from its VM backing store 3608 * base. We only have to do this if B_CACHE is 3609 * clear ( which causes the I/O to occur in the 3610 * first place ). The replacement prevents the read 3611 * I/O from overwriting potentially dirty VM-backed 3612 * pages. XXX bogus page replacement is, uh, bogus. 3613 * It may not work properly with small-block devices. 3614 * We need to find a better way. 3615 */ 3616 if (clear_modify) { 3617 pmap_remove_write(m); 3618 vfs_page_set_validclean(bp, foff, m); 3619 } else if (m->valid == VM_PAGE_BITS_ALL && 3620 (bp->b_flags & B_CACHE) == 0) { 3621 bp->b_pages[i] = bogus_page; 3622 bogus++; 3623 } 3624 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3625 } 3626 VM_OBJECT_UNLOCK(obj); 3627 if (bogus) 3628 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3629 bp->b_pages, bp->b_npages); 3630 } 3631 3632 /* 3633 * vfs_bio_set_valid: 3634 * 3635 * Set the range within the buffer to valid. The range is 3636 * relative to the beginning of the buffer, b_offset. Note that 3637 * b_offset itself may be offset from the beginning of the first 3638 * page. 3639 */ 3640 void 3641 vfs_bio_set_valid(struct buf *bp, int base, int size) 3642 { 3643 int i, n; 3644 vm_page_t m; 3645 3646 if (!(bp->b_flags & B_VMIO)) 3647 return; 3648 3649 /* 3650 * Fixup base to be relative to beginning of first page. 3651 * Set initial n to be the maximum number of bytes in the 3652 * first page that can be validated. 3653 */ 3654 base += (bp->b_offset & PAGE_MASK); 3655 n = PAGE_SIZE - (base & PAGE_MASK); 3656 3657 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3658 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3659 m = bp->b_pages[i]; 3660 if (n > size) 3661 n = size; 3662 vm_page_set_valid(m, base & PAGE_MASK, n); 3663 base += n; 3664 size -= n; 3665 n = PAGE_SIZE; 3666 } 3667 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3668 } 3669 3670 /* 3671 * vfs_bio_clrbuf: 3672 * 3673 * If the specified buffer is a non-VMIO buffer, clear the entire 3674 * buffer. If the specified buffer is a VMIO buffer, clear and 3675 * validate only the previously invalid portions of the buffer. 3676 * This routine essentially fakes an I/O, so we need to clear 3677 * BIO_ERROR and B_INVAL. 3678 * 3679 * Note that while we only theoretically need to clear through b_bcount, 3680 * we go ahead and clear through b_bufsize. 3681 */ 3682 void 3683 vfs_bio_clrbuf(struct buf *bp) 3684 { 3685 int i, j, mask; 3686 caddr_t sa, ea; 3687 3688 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3689 clrbuf(bp); 3690 return; 3691 } 3692 bp->b_flags &= ~B_INVAL; 3693 bp->b_ioflags &= ~BIO_ERROR; 3694 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3695 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3696 (bp->b_offset & PAGE_MASK) == 0) { 3697 if (bp->b_pages[0] == bogus_page) 3698 goto unlock; 3699 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3700 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3701 if ((bp->b_pages[0]->valid & mask) == mask) 3702 goto unlock; 3703 if ((bp->b_pages[0]->valid & mask) == 0) { 3704 bzero(bp->b_data, bp->b_bufsize); 3705 bp->b_pages[0]->valid |= mask; 3706 goto unlock; 3707 } 3708 } 3709 ea = sa = bp->b_data; 3710 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3711 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3712 ea = (caddr_t)(vm_offset_t)ulmin( 3713 (u_long)(vm_offset_t)ea, 3714 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3715 if (bp->b_pages[i] == bogus_page) 3716 continue; 3717 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3718 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3719 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3720 if ((bp->b_pages[i]->valid & mask) == mask) 3721 continue; 3722 if ((bp->b_pages[i]->valid & mask) == 0) 3723 bzero(sa, ea - sa); 3724 else { 3725 for (; sa < ea; sa += DEV_BSIZE, j++) { 3726 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3727 bzero(sa, DEV_BSIZE); 3728 } 3729 } 3730 bp->b_pages[i]->valid |= mask; 3731 } 3732 unlock: 3733 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3734 bp->b_resid = 0; 3735 } 3736 3737 /* 3738 * vm_hold_load_pages and vm_hold_free_pages get pages into 3739 * a buffers address space. The pages are anonymous and are 3740 * not associated with a file object. 3741 */ 3742 static void 3743 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3744 { 3745 vm_offset_t pg; 3746 vm_page_t p; 3747 int index; 3748 3749 to = round_page(to); 3750 from = round_page(from); 3751 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3752 3753 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3754 tryagain: 3755 /* 3756 * note: must allocate system pages since blocking here 3757 * could interfere with paging I/O, no matter which 3758 * process we are. 3759 */ 3760 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ | 3761 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3762 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 3763 if (!p) { 3764 VM_WAIT; 3765 goto tryagain; 3766 } 3767 pmap_qenter(pg, &p, 1); 3768 bp->b_pages[index] = p; 3769 } 3770 bp->b_npages = index; 3771 } 3772 3773 /* Return pages associated with this buf to the vm system */ 3774 static void 3775 vm_hold_free_pages(struct buf *bp, int newbsize) 3776 { 3777 vm_offset_t from; 3778 vm_page_t p; 3779 int index, newnpages; 3780 3781 from = round_page((vm_offset_t)bp->b_data + newbsize); 3782 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3783 if (bp->b_npages > newnpages) 3784 pmap_qremove(from, bp->b_npages - newnpages); 3785 for (index = newnpages; index < bp->b_npages; index++) { 3786 p = bp->b_pages[index]; 3787 bp->b_pages[index] = NULL; 3788 if (p->busy != 0) 3789 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3790 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 3791 p->wire_count--; 3792 vm_page_free(p); 3793 atomic_subtract_int(&cnt.v_wire_count, 1); 3794 } 3795 bp->b_npages = newnpages; 3796 } 3797 3798 /* 3799 * Map an IO request into kernel virtual address space. 3800 * 3801 * All requests are (re)mapped into kernel VA space. 3802 * Notice that we use b_bufsize for the size of the buffer 3803 * to be mapped. b_bcount might be modified by the driver. 3804 * 3805 * Note that even if the caller determines that the address space should 3806 * be valid, a race or a smaller-file mapped into a larger space may 3807 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3808 * check the return value. 3809 */ 3810 int 3811 vmapbuf(struct buf *bp) 3812 { 3813 caddr_t addr, kva; 3814 vm_prot_t prot; 3815 int pidx, i; 3816 struct vm_page *m; 3817 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3818 3819 if (bp->b_bufsize < 0) 3820 return (-1); 3821 prot = VM_PROT_READ; 3822 if (bp->b_iocmd == BIO_READ) 3823 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3824 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3825 addr < bp->b_data + bp->b_bufsize; 3826 addr += PAGE_SIZE, pidx++) { 3827 /* 3828 * Do the vm_fault if needed; do the copy-on-write thing 3829 * when reading stuff off device into memory. 3830 * 3831 * NOTE! Must use pmap_extract() because addr may be in 3832 * the userland address space, and kextract is only guarenteed 3833 * to work for the kernland address space (see: sparc64 port). 3834 */ 3835 retry: 3836 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3837 prot) < 0) { 3838 for (i = 0; i < pidx; ++i) { 3839 vm_page_lock(bp->b_pages[i]); 3840 vm_page_unhold(bp->b_pages[i]); 3841 vm_page_unlock(bp->b_pages[i]); 3842 bp->b_pages[i] = NULL; 3843 } 3844 return(-1); 3845 } 3846 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3847 if (m == NULL) 3848 goto retry; 3849 bp->b_pages[pidx] = m; 3850 } 3851 if (pidx > btoc(MAXPHYS)) 3852 panic("vmapbuf: mapped more than MAXPHYS"); 3853 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3854 3855 kva = bp->b_saveaddr; 3856 bp->b_npages = pidx; 3857 bp->b_saveaddr = bp->b_data; 3858 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3859 return(0); 3860 } 3861 3862 /* 3863 * Free the io map PTEs associated with this IO operation. 3864 * We also invalidate the TLB entries and restore the original b_addr. 3865 */ 3866 void 3867 vunmapbuf(struct buf *bp) 3868 { 3869 int pidx; 3870 int npages; 3871 3872 npages = bp->b_npages; 3873 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3874 for (pidx = 0; pidx < npages; pidx++) { 3875 vm_page_lock(bp->b_pages[pidx]); 3876 vm_page_unhold(bp->b_pages[pidx]); 3877 vm_page_unlock(bp->b_pages[pidx]); 3878 } 3879 3880 bp->b_data = bp->b_saveaddr; 3881 } 3882 3883 void 3884 bdone(struct buf *bp) 3885 { 3886 struct mtx *mtxp; 3887 3888 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3889 mtx_lock(mtxp); 3890 bp->b_flags |= B_DONE; 3891 wakeup(bp); 3892 mtx_unlock(mtxp); 3893 } 3894 3895 void 3896 bwait(struct buf *bp, u_char pri, const char *wchan) 3897 { 3898 struct mtx *mtxp; 3899 3900 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3901 mtx_lock(mtxp); 3902 while ((bp->b_flags & B_DONE) == 0) 3903 msleep(bp, mtxp, pri, wchan, 0); 3904 mtx_unlock(mtxp); 3905 } 3906 3907 int 3908 bufsync(struct bufobj *bo, int waitfor) 3909 { 3910 3911 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3912 } 3913 3914 void 3915 bufstrategy(struct bufobj *bo, struct buf *bp) 3916 { 3917 int i = 0; 3918 struct vnode *vp; 3919 3920 vp = bp->b_vp; 3921 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3922 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3923 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3924 i = VOP_STRATEGY(vp, bp); 3925 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3926 } 3927 3928 void 3929 bufobj_wrefl(struct bufobj *bo) 3930 { 3931 3932 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3933 ASSERT_BO_LOCKED(bo); 3934 bo->bo_numoutput++; 3935 } 3936 3937 void 3938 bufobj_wref(struct bufobj *bo) 3939 { 3940 3941 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3942 BO_LOCK(bo); 3943 bo->bo_numoutput++; 3944 BO_UNLOCK(bo); 3945 } 3946 3947 void 3948 bufobj_wdrop(struct bufobj *bo) 3949 { 3950 3951 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3952 BO_LOCK(bo); 3953 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3954 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3955 bo->bo_flag &= ~BO_WWAIT; 3956 wakeup(&bo->bo_numoutput); 3957 } 3958 BO_UNLOCK(bo); 3959 } 3960 3961 int 3962 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3963 { 3964 int error; 3965 3966 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3967 ASSERT_BO_LOCKED(bo); 3968 error = 0; 3969 while (bo->bo_numoutput) { 3970 bo->bo_flag |= BO_WWAIT; 3971 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3972 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3973 if (error) 3974 break; 3975 } 3976 return (error); 3977 } 3978 3979 void 3980 bpin(struct buf *bp) 3981 { 3982 struct mtx *mtxp; 3983 3984 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3985 mtx_lock(mtxp); 3986 bp->b_pin_count++; 3987 mtx_unlock(mtxp); 3988 } 3989 3990 void 3991 bunpin(struct buf *bp) 3992 { 3993 struct mtx *mtxp; 3994 3995 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3996 mtx_lock(mtxp); 3997 if (--bp->b_pin_count == 0) 3998 wakeup(bp); 3999 mtx_unlock(mtxp); 4000 } 4001 4002 void 4003 bunpin_wait(struct buf *bp) 4004 { 4005 struct mtx *mtxp; 4006 4007 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4008 mtx_lock(mtxp); 4009 while (bp->b_pin_count > 0) 4010 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4011 mtx_unlock(mtxp); 4012 } 4013 4014 #include "opt_ddb.h" 4015 #ifdef DDB 4016 #include <ddb/ddb.h> 4017 4018 /* DDB command to show buffer data */ 4019 DB_SHOW_COMMAND(buffer, db_show_buffer) 4020 { 4021 /* get args */ 4022 struct buf *bp = (struct buf *)addr; 4023 4024 if (!have_addr) { 4025 db_printf("usage: show buffer <addr>\n"); 4026 return; 4027 } 4028 4029 db_printf("buf at %p\n", bp); 4030 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4031 db_printf( 4032 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4033 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4034 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4035 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4036 bp->b_dep.lh_first); 4037 if (bp->b_npages) { 4038 int i; 4039 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4040 for (i = 0; i < bp->b_npages; i++) { 4041 vm_page_t m; 4042 m = bp->b_pages[i]; 4043 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4044 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4045 if ((i + 1) < bp->b_npages) 4046 db_printf(","); 4047 } 4048 db_printf("\n"); 4049 } 4050 db_printf(" "); 4051 lockmgr_printinfo(&bp->b_lock); 4052 } 4053 4054 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4055 { 4056 struct buf *bp; 4057 int i; 4058 4059 for (i = 0; i < nbuf; i++) { 4060 bp = &buf[i]; 4061 if (BUF_ISLOCKED(bp)) { 4062 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4063 db_printf("\n"); 4064 } 4065 } 4066 } 4067 4068 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4069 { 4070 struct vnode *vp; 4071 struct buf *bp; 4072 4073 if (!have_addr) { 4074 db_printf("usage: show vnodebufs <addr>\n"); 4075 return; 4076 } 4077 vp = (struct vnode *)addr; 4078 db_printf("Clean buffers:\n"); 4079 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4080 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4081 db_printf("\n"); 4082 } 4083 db_printf("Dirty buffers:\n"); 4084 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4085 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4086 db_printf("\n"); 4087 } 4088 } 4089 4090 DB_COMMAND(countfreebufs, db_coundfreebufs) 4091 { 4092 struct buf *bp; 4093 int i, used = 0, nfree = 0; 4094 4095 if (have_addr) { 4096 db_printf("usage: countfreebufs\n"); 4097 return; 4098 } 4099 4100 for (i = 0; i < nbuf; i++) { 4101 bp = &buf[i]; 4102 if ((bp->b_vflags & BV_INFREECNT) != 0) 4103 nfree++; 4104 else 4105 used++; 4106 } 4107 4108 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4109 nfree + used); 4110 db_printf("numfreebuffers is %d\n", numfreebuffers); 4111 } 4112 #endif /* DDB */ 4113