vfs_bio.c revision 251171
168651Skris/*- 268651Skris * Copyright (c) 2004 Poul-Henning Kamp 368651Skris * Copyright (c) 1994,1997 John S. Dyson 468651Skris * Copyright (c) 2013 The FreeBSD Foundation 568651Skris * All rights reserved. 668651Skris * 768651Skris * Portions of this software were developed by Konstantin Belousov 868651Skris * under sponsorship from the FreeBSD Foundation. 968651Skris * 1068651Skris * Redistribution and use in source and binary forms, with or without 1168651Skris * modification, are permitted provided that the following conditions 1268651Skris * are met: 1368651Skris * 1. Redistributions of source code must retain the above copyright 1468651Skris * notice, this list of conditions and the following disclaimer. 1568651Skris * 2. Redistributions in binary form must reproduce the above copyright 1668651Skris * notice, this list of conditions and the following disclaimer in the 1768651Skris * documentation and/or other materials provided with the distribution. 1868651Skris * 1968651Skris * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 2068651Skris * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2168651Skris * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2268651Skris * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 2368651Skris * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2468651Skris * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2568651Skris * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2668651Skris * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2768651Skris * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2868651Skris * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2968651Skris * SUCH DAMAGE. 3068651Skris */ 3168651Skris 3268651Skris/* 3368651Skris * this file contains a new buffer I/O scheme implementing a coherent 3468651Skris * VM object and buffer cache scheme. Pains have been taken to make 3568651Skris * sure that the performance degradation associated with schemes such 3668651Skris * as this is not realized. 3768651Skris * 3868651Skris * Author: John S. Dyson 3968651Skris * Significant help during the development and debugging phases 4076866Skris * had been provided by David Greenman, also of the FreeBSD core team. 4176866Skris * 4276866Skris * see man buf(9) for more info. 4376866Skris */ 4476866Skris 4568651Skris#include <sys/cdefs.h> 4668651Skris__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 251171 2013-05-31 00:43:41Z jeff $"); 4768651Skris 4868651Skris#include <sys/param.h> 4968651Skris#include <sys/systm.h> 5068651Skris#include <sys/bio.h> 5168651Skris#include <sys/conf.h> 5268651Skris#include <sys/buf.h> 5368651Skris#include <sys/devicestat.h> 5468651Skris#include <sys/eventhandler.h> 5568651Skris#include <sys/fail.h> 5668651Skris#include <sys/limits.h> 5768651Skris#include <sys/lock.h> 5868651Skris#include <sys/malloc.h> 5968651Skris#include <sys/mount.h> 6068651Skris#include <sys/mutex.h> 6168651Skris#include <sys/kernel.h> 6272613Skris#include <sys/kthread.h> 6368651Skris#include <sys/proc.h> 6468651Skris#include <sys/resourcevar.h> 6568651Skris#include <sys/rwlock.h> 6668651Skris#include <sys/sysctl.h> 6768651Skris#include <sys/vmmeter.h> 6868651Skris#include <sys/vnode.h> 6968651Skris#include <geom/geom.h> 7068651Skris#include <vm/vm.h> 7168651Skris#include <vm/vm_param.h> 7268651Skris#include <vm/vm_kern.h> 7368651Skris#include <vm/vm_pageout.h> 7468651Skris#include <vm/vm_page.h> 7576866Skris#include <vm/vm_object.h> 7676866Skris#include <vm/vm_extern.h> 7776866Skris#include <vm/vm_map.h> 7868651Skris#include "opt_compat.h" 7968651Skris#include "opt_directio.h" 80#include "opt_swap.h" 81 82static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 83 84struct bio_ops bioops; /* I/O operation notification */ 85 86struct buf_ops buf_ops_bio = { 87 .bop_name = "buf_ops_bio", 88 .bop_write = bufwrite, 89 .bop_strategy = bufstrategy, 90 .bop_sync = bufsync, 91 .bop_bdflush = bufbdflush, 92}; 93 94/* 95 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 96 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 97 */ 98struct buf *buf; /* buffer header pool */ 99caddr_t unmapped_buf; 100 101static struct proc *bufdaemonproc; 102 103static int inmem(struct vnode *vp, daddr_t blkno); 104static void vm_hold_free_pages(struct buf *bp, int newbsize); 105static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 106 vm_offset_t to); 107static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 108static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 109 vm_page_t m); 110static void vfs_drain_busy_pages(struct buf *bp); 111static void vfs_clean_pages_dirty_buf(struct buf *bp); 112static void vfs_setdirty_locked_object(struct buf *bp); 113static void vfs_vmio_release(struct buf *bp); 114static int vfs_bio_clcheck(struct vnode *vp, int size, 115 daddr_t lblkno, daddr_t blkno); 116static int buf_do_flush(struct vnode *vp); 117static int flushbufqueues(struct vnode *, int, int); 118static void buf_daemon(void); 119static void bremfreel(struct buf *bp); 120#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 121 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 122static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 123#endif 124 125int vmiodirenable = TRUE; 126SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 127 "Use the VM system for directory writes"); 128long runningbufspace; 129SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 130 "Amount of presently outstanding async buffer io"); 131static long bufspace; 132#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 133 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 134SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 135 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 136#else 137SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 138 "Virtual memory used for buffers"); 139#endif 140static long unmapped_bufspace; 141SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD, 142 &unmapped_bufspace, 0, 143 "Amount of unmapped buffers, inclusive in the bufspace"); 144static long maxbufspace; 145SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 146 "Maximum allowed value of bufspace (including buf_daemon)"); 147static long bufmallocspace; 148SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 149 "Amount of malloced memory for buffers"); 150static long maxbufmallocspace; 151SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 152 "Maximum amount of malloced memory for buffers"); 153static long lobufspace; 154SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 155 "Minimum amount of buffers we want to have"); 156long hibufspace; 157SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 158 "Maximum allowed value of bufspace (excluding buf_daemon)"); 159static int bufreusecnt; 160SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 161 "Number of times we have reused a buffer"); 162static int buffreekvacnt; 163SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 164 "Number of times we have freed the KVA space from some buffer"); 165static int bufdefragcnt; 166SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 167 "Number of times we have had to repeat buffer allocation to defragment"); 168static long lorunningspace; 169SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 170 "Minimum preferred space used for in-progress I/O"); 171static long hirunningspace; 172SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 173 "Maximum amount of space to use for in-progress I/O"); 174int dirtybufferflushes; 175SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 176 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 177int bdwriteskip; 178SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 179 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 180int altbufferflushes; 181SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 182 0, "Number of fsync flushes to limit dirty buffers"); 183static int recursiveflushes; 184SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 185 0, "Number of flushes skipped due to being recursive"); 186static int numdirtybuffers; 187SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 188 "Number of buffers that are dirty (has unwritten changes) at the moment"); 189static int lodirtybuffers; 190SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 191 "How many buffers we want to have free before bufdaemon can sleep"); 192static int hidirtybuffers; 193SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 194 "When the number of dirty buffers is considered severe"); 195int dirtybufthresh; 196SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 197 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 198static int numfreebuffers; 199SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 200 "Number of free buffers"); 201static int lofreebuffers; 202SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 203 "XXX Unused"); 204static int hifreebuffers; 205SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 206 "XXX Complicatedly unused"); 207static int getnewbufcalls; 208SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 209 "Number of calls to getnewbuf"); 210static int getnewbufrestarts; 211SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 212 "Number of times getnewbuf has had to restart a buffer aquisition"); 213static int mappingrestarts; 214SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 215 "Number of times getblk has had to restart a buffer mapping for " 216 "unmapped buffer"); 217static int flushbufqtarget = 100; 218SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 219 "Amount of work to do in flushbufqueues when helping bufdaemon"); 220static long notbufdflashes; 221SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 222 "Number of dirty buffer flushes done by the bufdaemon helpers"); 223static long barrierwrites; 224SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 225 "Number of barrier writes"); 226SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 227 &unmapped_buf_allowed, 0, 228 "Permit the use of the unmapped i/o"); 229 230/* 231 * Wakeup point for bufdaemon, as well as indicator of whether it is already 232 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 233 * is idling. 234 */ 235static int bd_request; 236 237/* 238 * Request for the buf daemon to write more buffers than is indicated by 239 * lodirtybuf. This may be necessary to push out excess dependencies or 240 * defragment the address space where a simple count of the number of dirty 241 * buffers is insufficient to characterize the demand for flushing them. 242 */ 243static int bd_speedupreq; 244 245/* 246 * This lock synchronizes access to bd_request. 247 */ 248static struct mtx bdlock; 249 250/* 251 * bogus page -- for I/O to/from partially complete buffers 252 * this is a temporary solution to the problem, but it is not 253 * really that bad. it would be better to split the buffer 254 * for input in the case of buffers partially already in memory, 255 * but the code is intricate enough already. 256 */ 257vm_page_t bogus_page; 258 259/* 260 * Synchronization (sleep/wakeup) variable for active buffer space requests. 261 * Set when wait starts, cleared prior to wakeup(). 262 * Used in runningbufwakeup() and waitrunningbufspace(). 263 */ 264static int runningbufreq; 265 266/* 267 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 268 * waitrunningbufspace(). 269 */ 270static struct mtx rbreqlock; 271 272/* 273 * Synchronization (sleep/wakeup) variable for buffer requests. 274 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 275 * by and/or. 276 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 277 * getnewbuf(), and getblk(). 278 */ 279static int needsbuffer; 280 281/* 282 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 283 */ 284static struct mtx nblock; 285 286/* 287 * Definitions for the buffer free lists. 288 */ 289#define BUFFER_QUEUES 5 /* number of free buffer queues */ 290 291#define QUEUE_NONE 0 /* on no queue */ 292#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 293#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 294#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 295#define QUEUE_EMPTY 4 /* empty buffer headers */ 296#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 297 298/* Queues for free buffers with various properties */ 299static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 300#ifdef INVARIANTS 301static int bq_len[BUFFER_QUEUES]; 302#endif 303 304/* Lock for the bufqueues */ 305static struct mtx bqlock; 306 307/* 308 * Single global constant for BUF_WMESG, to avoid getting multiple references. 309 * buf_wmesg is referred from macros. 310 */ 311const char *buf_wmesg = BUF_WMESG; 312 313#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 314#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 315#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 316#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 317 318#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 319 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 320static int 321sysctl_bufspace(SYSCTL_HANDLER_ARGS) 322{ 323 long lvalue; 324 int ivalue; 325 326 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 327 return (sysctl_handle_long(oidp, arg1, arg2, req)); 328 lvalue = *(long *)arg1; 329 if (lvalue > INT_MAX) 330 /* On overflow, still write out a long to trigger ENOMEM. */ 331 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 332 ivalue = lvalue; 333 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 334} 335#endif 336 337#ifdef DIRECTIO 338extern void ffs_rawread_setup(void); 339#endif /* DIRECTIO */ 340/* 341 * numdirtywakeup: 342 * 343 * If someone is blocked due to there being too many dirty buffers, 344 * and numdirtybuffers is now reasonable, wake them up. 345 */ 346 347static __inline void 348numdirtywakeup(int level) 349{ 350 351 if (numdirtybuffers <= level) { 352 mtx_lock(&nblock); 353 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 354 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 355 wakeup(&needsbuffer); 356 } 357 mtx_unlock(&nblock); 358 } 359} 360 361/* 362 * bufspacewakeup: 363 * 364 * Called when buffer space is potentially available for recovery. 365 * getnewbuf() will block on this flag when it is unable to free 366 * sufficient buffer space. Buffer space becomes recoverable when 367 * bp's get placed back in the queues. 368 */ 369 370static __inline void 371bufspacewakeup(void) 372{ 373 374 /* 375 * If someone is waiting for BUF space, wake them up. Even 376 * though we haven't freed the kva space yet, the waiting 377 * process will be able to now. 378 */ 379 mtx_lock(&nblock); 380 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 381 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 382 wakeup(&needsbuffer); 383 } 384 mtx_unlock(&nblock); 385} 386 387/* 388 * runningbufwakeup() - in-progress I/O accounting. 389 * 390 */ 391void 392runningbufwakeup(struct buf *bp) 393{ 394 395 if (bp->b_runningbufspace) { 396 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 397 bp->b_runningbufspace = 0; 398 mtx_lock(&rbreqlock); 399 if (runningbufreq && runningbufspace <= lorunningspace) { 400 runningbufreq = 0; 401 wakeup(&runningbufreq); 402 } 403 mtx_unlock(&rbreqlock); 404 } 405} 406 407/* 408 * bufcountwakeup: 409 * 410 * Called when a buffer has been added to one of the free queues to 411 * account for the buffer and to wakeup anyone waiting for free buffers. 412 * This typically occurs when large amounts of metadata are being handled 413 * by the buffer cache ( else buffer space runs out first, usually ). 414 */ 415 416static __inline void 417bufcountwakeup(struct buf *bp) 418{ 419 int old; 420 421 KASSERT((bp->b_flags & B_INFREECNT) == 0, 422 ("buf %p already counted as free", bp)); 423 bp->b_flags |= B_INFREECNT; 424 old = atomic_fetchadd_int(&numfreebuffers, 1); 425 KASSERT(old >= 0 && old < nbuf, 426 ("numfreebuffers climbed to %d", old + 1)); 427 mtx_lock(&nblock); 428 if (needsbuffer) { 429 needsbuffer &= ~VFS_BIO_NEED_ANY; 430 if (numfreebuffers >= hifreebuffers) 431 needsbuffer &= ~VFS_BIO_NEED_FREE; 432 wakeup(&needsbuffer); 433 } 434 mtx_unlock(&nblock); 435} 436 437/* 438 * waitrunningbufspace() 439 * 440 * runningbufspace is a measure of the amount of I/O currently 441 * running. This routine is used in async-write situations to 442 * prevent creating huge backups of pending writes to a device. 443 * Only asynchronous writes are governed by this function. 444 * 445 * Reads will adjust runningbufspace, but will not block based on it. 446 * The read load has a side effect of reducing the allowed write load. 447 * 448 * This does NOT turn an async write into a sync write. It waits 449 * for earlier writes to complete and generally returns before the 450 * caller's write has reached the device. 451 */ 452void 453waitrunningbufspace(void) 454{ 455 456 mtx_lock(&rbreqlock); 457 while (runningbufspace > hirunningspace) { 458 ++runningbufreq; 459 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 460 } 461 mtx_unlock(&rbreqlock); 462} 463 464 465/* 466 * vfs_buf_test_cache: 467 * 468 * Called when a buffer is extended. This function clears the B_CACHE 469 * bit if the newly extended portion of the buffer does not contain 470 * valid data. 471 */ 472static __inline 473void 474vfs_buf_test_cache(struct buf *bp, 475 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 476 vm_page_t m) 477{ 478 479 VM_OBJECT_ASSERT_WLOCKED(m->object); 480 if (bp->b_flags & B_CACHE) { 481 int base = (foff + off) & PAGE_MASK; 482 if (vm_page_is_valid(m, base, size) == 0) 483 bp->b_flags &= ~B_CACHE; 484 } 485} 486 487/* Wake up the buffer daemon if necessary */ 488static __inline 489void 490bd_wakeup(int dirtybuflevel) 491{ 492 493 mtx_lock(&bdlock); 494 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 495 bd_request = 1; 496 wakeup(&bd_request); 497 } 498 mtx_unlock(&bdlock); 499} 500 501/* 502 * bd_speedup - speedup the buffer cache flushing code 503 */ 504 505void 506bd_speedup(void) 507{ 508 int needwake; 509 510 mtx_lock(&bdlock); 511 needwake = 0; 512 if (bd_speedupreq == 0 || bd_request == 0) 513 needwake = 1; 514 bd_speedupreq = 1; 515 bd_request = 1; 516 if (needwake) 517 wakeup(&bd_request); 518 mtx_unlock(&bdlock); 519} 520 521#ifdef __i386__ 522#define TRANSIENT_DENOM 5 523#else 524#define TRANSIENT_DENOM 10 525#endif 526 527/* 528 * Calculating buffer cache scaling values and reserve space for buffer 529 * headers. This is called during low level kernel initialization and 530 * may be called more then once. We CANNOT write to the memory area 531 * being reserved at this time. 532 */ 533caddr_t 534kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 535{ 536 int tuned_nbuf; 537 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 538 539 /* 540 * physmem_est is in pages. Convert it to kilobytes (assumes 541 * PAGE_SIZE is >= 1K) 542 */ 543 physmem_est = physmem_est * (PAGE_SIZE / 1024); 544 545 /* 546 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 547 * For the first 64MB of ram nominally allocate sufficient buffers to 548 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 549 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 550 * the buffer cache we limit the eventual kva reservation to 551 * maxbcache bytes. 552 * 553 * factor represents the 1/4 x ram conversion. 554 */ 555 if (nbuf == 0) { 556 int factor = 4 * BKVASIZE / 1024; 557 558 nbuf = 50; 559 if (physmem_est > 4096) 560 nbuf += min((physmem_est - 4096) / factor, 561 65536 / factor); 562 if (physmem_est > 65536) 563 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 564 565 if (maxbcache && nbuf > maxbcache / BKVASIZE) 566 nbuf = maxbcache / BKVASIZE; 567 tuned_nbuf = 1; 568 } else 569 tuned_nbuf = 0; 570 571 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 572 maxbuf = (LONG_MAX / 3) / BKVASIZE; 573 if (nbuf > maxbuf) { 574 if (!tuned_nbuf) 575 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 576 maxbuf); 577 nbuf = maxbuf; 578 } 579 580 /* 581 * Ideal allocation size for the transient bio submap if 10% 582 * of the maximal space buffer map. This roughly corresponds 583 * to the amount of the buffer mapped for typical UFS load. 584 * 585 * Clip the buffer map to reserve space for the transient 586 * BIOs, if its extent is bigger than 90% (80% on i386) of the 587 * maximum buffer map extent on the platform. 588 * 589 * The fall-back to the maxbuf in case of maxbcache unset, 590 * allows to not trim the buffer KVA for the architectures 591 * with ample KVA space. 592 */ 593 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 594 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 595 buf_sz = (long)nbuf * BKVASIZE; 596 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 597 (TRANSIENT_DENOM - 1)) { 598 /* 599 * There is more KVA than memory. Do not 600 * adjust buffer map size, and assign the rest 601 * of maxbuf to transient map. 602 */ 603 biotmap_sz = maxbuf_sz - buf_sz; 604 } else { 605 /* 606 * Buffer map spans all KVA we could afford on 607 * this platform. Give 10% (20% on i386) of 608 * the buffer map to the transient bio map. 609 */ 610 biotmap_sz = buf_sz / TRANSIENT_DENOM; 611 buf_sz -= biotmap_sz; 612 } 613 if (biotmap_sz / INT_MAX > MAXPHYS) 614 bio_transient_maxcnt = INT_MAX; 615 else 616 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 617 /* 618 * Artifically limit to 1024 simultaneous in-flight I/Os 619 * using the transient mapping. 620 */ 621 if (bio_transient_maxcnt > 1024) 622 bio_transient_maxcnt = 1024; 623 if (tuned_nbuf) 624 nbuf = buf_sz / BKVASIZE; 625 } 626 627 /* 628 * swbufs are used as temporary holders for I/O, such as paging I/O. 629 * We have no less then 16 and no more then 256. 630 */ 631 nswbuf = max(min(nbuf/4, 256), 16); 632#ifdef NSWBUF_MIN 633 if (nswbuf < NSWBUF_MIN) 634 nswbuf = NSWBUF_MIN; 635#endif 636#ifdef DIRECTIO 637 ffs_rawread_setup(); 638#endif 639 640 /* 641 * Reserve space for the buffer cache buffers 642 */ 643 swbuf = (void *)v; 644 v = (caddr_t)(swbuf + nswbuf); 645 buf = (void *)v; 646 v = (caddr_t)(buf + nbuf); 647 648 return(v); 649} 650 651/* Initialize the buffer subsystem. Called before use of any buffers. */ 652void 653bufinit(void) 654{ 655 struct buf *bp; 656 int i; 657 658 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 659 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 660 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 661 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 662 663 /* next, make a null set of free lists */ 664 for (i = 0; i < BUFFER_QUEUES; i++) 665 TAILQ_INIT(&bufqueues[i]); 666 667 /* finally, initialize each buffer header and stick on empty q */ 668 for (i = 0; i < nbuf; i++) { 669 bp = &buf[i]; 670 bzero(bp, sizeof *bp); 671 bp->b_flags = B_INVAL | B_INFREECNT; 672 bp->b_rcred = NOCRED; 673 bp->b_wcred = NOCRED; 674 bp->b_qindex = QUEUE_EMPTY; 675 bp->b_xflags = 0; 676 LIST_INIT(&bp->b_dep); 677 BUF_LOCKINIT(bp); 678 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 679#ifdef INVARIANTS 680 bq_len[QUEUE_EMPTY]++; 681#endif 682 } 683 684 /* 685 * maxbufspace is the absolute maximum amount of buffer space we are 686 * allowed to reserve in KVM and in real terms. The absolute maximum 687 * is nominally used by buf_daemon. hibufspace is the nominal maximum 688 * used by most other processes. The differential is required to 689 * ensure that buf_daemon is able to run when other processes might 690 * be blocked waiting for buffer space. 691 * 692 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 693 * this may result in KVM fragmentation which is not handled optimally 694 * by the system. 695 */ 696 maxbufspace = (long)nbuf * BKVASIZE; 697 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 698 lobufspace = hibufspace - MAXBSIZE; 699 700 /* 701 * Note: The 16 MiB upper limit for hirunningspace was chosen 702 * arbitrarily and may need further tuning. It corresponds to 703 * 128 outstanding write IO requests (if IO size is 128 KiB), 704 * which fits with many RAID controllers' tagged queuing limits. 705 * The lower 1 MiB limit is the historical upper limit for 706 * hirunningspace. 707 */ 708 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 709 16 * 1024 * 1024), 1024 * 1024); 710 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 711 712/* 713 * Limit the amount of malloc memory since it is wired permanently into 714 * the kernel space. Even though this is accounted for in the buffer 715 * allocation, we don't want the malloced region to grow uncontrolled. 716 * The malloc scheme improves memory utilization significantly on average 717 * (small) directories. 718 */ 719 maxbufmallocspace = hibufspace / 20; 720 721/* 722 * Reduce the chance of a deadlock occuring by limiting the number 723 * of delayed-write dirty buffers we allow to stack up. 724 */ 725 hidirtybuffers = nbuf / 4 + 20; 726 dirtybufthresh = hidirtybuffers * 9 / 10; 727 numdirtybuffers = 0; 728/* 729 * To support extreme low-memory systems, make sure hidirtybuffers cannot 730 * eat up all available buffer space. This occurs when our minimum cannot 731 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 732 * BKVASIZE'd buffers. 733 */ 734 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 735 hidirtybuffers >>= 1; 736 } 737 lodirtybuffers = hidirtybuffers / 2; 738 739/* 740 * Try to keep the number of free buffers in the specified range, 741 * and give special processes (e.g. like buf_daemon) access to an 742 * emergency reserve. 743 */ 744 lofreebuffers = nbuf / 18 + 5; 745 hifreebuffers = 2 * lofreebuffers; 746 numfreebuffers = nbuf; 747 748 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 749 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 750 unmapped_buf = (caddr_t)kmem_alloc_nofault(kernel_map, MAXPHYS); 751} 752 753#ifdef INVARIANTS 754static inline void 755vfs_buf_check_mapped(struct buf *bp) 756{ 757 758 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 759 ("mapped buf %p %x", bp, bp->b_flags)); 760 KASSERT(bp->b_kvabase != unmapped_buf, 761 ("mapped buf: b_kvabase was not updated %p", bp)); 762 KASSERT(bp->b_data != unmapped_buf, 763 ("mapped buf: b_data was not updated %p", bp)); 764} 765 766static inline void 767vfs_buf_check_unmapped(struct buf *bp) 768{ 769 770 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 771 ("unmapped buf %p %x", bp, bp->b_flags)); 772 KASSERT(bp->b_kvabase == unmapped_buf, 773 ("unmapped buf: corrupted b_kvabase %p", bp)); 774 KASSERT(bp->b_data == unmapped_buf, 775 ("unmapped buf: corrupted b_data %p", bp)); 776} 777 778#define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 779#define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 780#else 781#define BUF_CHECK_MAPPED(bp) do {} while (0) 782#define BUF_CHECK_UNMAPPED(bp) do {} while (0) 783#endif 784 785static void 786bpmap_qenter(struct buf *bp) 787{ 788 789 BUF_CHECK_MAPPED(bp); 790 791 /* 792 * bp->b_data is relative to bp->b_offset, but 793 * bp->b_offset may be offset into the first page. 794 */ 795 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 796 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 797 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 798 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 799} 800 801/* 802 * bfreekva() - free the kva allocation for a buffer. 803 * 804 * Since this call frees up buffer space, we call bufspacewakeup(). 805 */ 806static void 807bfreekva(struct buf *bp) 808{ 809 810 if (bp->b_kvasize == 0) 811 return; 812 813 atomic_add_int(&buffreekvacnt, 1); 814 atomic_subtract_long(&bufspace, bp->b_kvasize); 815 if ((bp->b_flags & B_UNMAPPED) == 0) { 816 BUF_CHECK_MAPPED(bp); 817 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvabase, 818 (vm_offset_t)bp->b_kvabase + bp->b_kvasize); 819 } else { 820 BUF_CHECK_UNMAPPED(bp); 821 if ((bp->b_flags & B_KVAALLOC) != 0) { 822 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvaalloc, 823 (vm_offset_t)bp->b_kvaalloc + bp->b_kvasize); 824 } 825 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 826 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 827 } 828 bp->b_kvasize = 0; 829 bufspacewakeup(); 830} 831 832/* 833 * bremfree: 834 * 835 * Mark the buffer for removal from the appropriate free list in brelse. 836 * 837 */ 838void 839bremfree(struct buf *bp) 840{ 841 int old; 842 843 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 844 KASSERT((bp->b_flags & B_REMFREE) == 0, 845 ("bremfree: buffer %p already marked for delayed removal.", bp)); 846 KASSERT(bp->b_qindex != QUEUE_NONE, 847 ("bremfree: buffer %p not on a queue.", bp)); 848 BUF_ASSERT_XLOCKED(bp); 849 850 bp->b_flags |= B_REMFREE; 851 /* Fixup numfreebuffers count. */ 852 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 853 KASSERT((bp->b_flags & B_INFREECNT) != 0, 854 ("buf %p not counted in numfreebuffers", bp)); 855 bp->b_flags &= ~B_INFREECNT; 856 old = atomic_fetchadd_int(&numfreebuffers, -1); 857 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 858 } 859} 860 861/* 862 * bremfreef: 863 * 864 * Force an immediate removal from a free list. Used only in nfs when 865 * it abuses the b_freelist pointer. 866 */ 867void 868bremfreef(struct buf *bp) 869{ 870 mtx_lock(&bqlock); 871 bremfreel(bp); 872 mtx_unlock(&bqlock); 873} 874 875/* 876 * bremfreel: 877 * 878 * Removes a buffer from the free list, must be called with the 879 * bqlock held. 880 */ 881static void 882bremfreel(struct buf *bp) 883{ 884 int old; 885 886 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 887 bp, bp->b_vp, bp->b_flags); 888 KASSERT(bp->b_qindex != QUEUE_NONE, 889 ("bremfreel: buffer %p not on a queue.", bp)); 890 BUF_ASSERT_XLOCKED(bp); 891 mtx_assert(&bqlock, MA_OWNED); 892 893 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 894#ifdef INVARIANTS 895 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 896 bp->b_qindex)); 897 bq_len[bp->b_qindex]--; 898#endif 899 bp->b_qindex = QUEUE_NONE; 900 /* 901 * If this was a delayed bremfree() we only need to remove the buffer 902 * from the queue and return the stats are already done. 903 */ 904 if (bp->b_flags & B_REMFREE) { 905 bp->b_flags &= ~B_REMFREE; 906 return; 907 } 908 /* 909 * Fixup numfreebuffers count. If the buffer is invalid or not 910 * delayed-write, the buffer was free and we must decrement 911 * numfreebuffers. 912 */ 913 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 914 KASSERT((bp->b_flags & B_INFREECNT) != 0, 915 ("buf %p not counted in numfreebuffers", bp)); 916 bp->b_flags &= ~B_INFREECNT; 917 old = atomic_fetchadd_int(&numfreebuffers, -1); 918 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 919 } 920} 921 922/* 923 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 924 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 925 * the buffer is valid and we do not have to do anything. 926 */ 927void 928breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 929 int cnt, struct ucred * cred) 930{ 931 struct buf *rabp; 932 int i; 933 934 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 935 if (inmem(vp, *rablkno)) 936 continue; 937 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 938 939 if ((rabp->b_flags & B_CACHE) == 0) { 940 if (!TD_IS_IDLETHREAD(curthread)) 941 curthread->td_ru.ru_inblock++; 942 rabp->b_flags |= B_ASYNC; 943 rabp->b_flags &= ~B_INVAL; 944 rabp->b_ioflags &= ~BIO_ERROR; 945 rabp->b_iocmd = BIO_READ; 946 if (rabp->b_rcred == NOCRED && cred != NOCRED) 947 rabp->b_rcred = crhold(cred); 948 vfs_busy_pages(rabp, 0); 949 BUF_KERNPROC(rabp); 950 rabp->b_iooffset = dbtob(rabp->b_blkno); 951 bstrategy(rabp); 952 } else { 953 brelse(rabp); 954 } 955 } 956} 957 958/* 959 * Entry point for bread() and breadn() via #defines in sys/buf.h. 960 * 961 * Get a buffer with the specified data. Look in the cache first. We 962 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 963 * is set, the buffer is valid and we do not have to do anything, see 964 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 965 */ 966int 967breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 968 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 969{ 970 struct buf *bp; 971 int rv = 0, readwait = 0; 972 973 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 974 /* 975 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 976 */ 977 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 978 if (bp == NULL) 979 return (EBUSY); 980 981 /* if not found in cache, do some I/O */ 982 if ((bp->b_flags & B_CACHE) == 0) { 983 if (!TD_IS_IDLETHREAD(curthread)) 984 curthread->td_ru.ru_inblock++; 985 bp->b_iocmd = BIO_READ; 986 bp->b_flags &= ~B_INVAL; 987 bp->b_ioflags &= ~BIO_ERROR; 988 if (bp->b_rcred == NOCRED && cred != NOCRED) 989 bp->b_rcred = crhold(cred); 990 vfs_busy_pages(bp, 0); 991 bp->b_iooffset = dbtob(bp->b_blkno); 992 bstrategy(bp); 993 ++readwait; 994 } 995 996 breada(vp, rablkno, rabsize, cnt, cred); 997 998 if (readwait) { 999 rv = bufwait(bp); 1000 } 1001 return (rv); 1002} 1003 1004/* 1005 * Write, release buffer on completion. (Done by iodone 1006 * if async). Do not bother writing anything if the buffer 1007 * is invalid. 1008 * 1009 * Note that we set B_CACHE here, indicating that buffer is 1010 * fully valid and thus cacheable. This is true even of NFS 1011 * now so we set it generally. This could be set either here 1012 * or in biodone() since the I/O is synchronous. We put it 1013 * here. 1014 */ 1015int 1016bufwrite(struct buf *bp) 1017{ 1018 int oldflags; 1019 struct vnode *vp; 1020 int vp_md; 1021 1022 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1023 if (bp->b_flags & B_INVAL) { 1024 brelse(bp); 1025 return (0); 1026 } 1027 1028 if (bp->b_flags & B_BARRIER) 1029 barrierwrites++; 1030 1031 oldflags = bp->b_flags; 1032 1033 BUF_ASSERT_HELD(bp); 1034 1035 if (bp->b_pin_count > 0) 1036 bunpin_wait(bp); 1037 1038 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1039 ("FFS background buffer should not get here %p", bp)); 1040 1041 vp = bp->b_vp; 1042 if (vp) 1043 vp_md = vp->v_vflag & VV_MD; 1044 else 1045 vp_md = 0; 1046 1047 /* 1048 * Mark the buffer clean. Increment the bufobj write count 1049 * before bundirty() call, to prevent other thread from seeing 1050 * empty dirty list and zero counter for writes in progress, 1051 * falsely indicating that the bufobj is clean. 1052 */ 1053 bufobj_wref(bp->b_bufobj); 1054 bundirty(bp); 1055 1056 bp->b_flags &= ~B_DONE; 1057 bp->b_ioflags &= ~BIO_ERROR; 1058 bp->b_flags |= B_CACHE; 1059 bp->b_iocmd = BIO_WRITE; 1060 1061 vfs_busy_pages(bp, 1); 1062 1063 /* 1064 * Normal bwrites pipeline writes 1065 */ 1066 bp->b_runningbufspace = bp->b_bufsize; 1067 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1068 1069 if (!TD_IS_IDLETHREAD(curthread)) 1070 curthread->td_ru.ru_oublock++; 1071 if (oldflags & B_ASYNC) 1072 BUF_KERNPROC(bp); 1073 bp->b_iooffset = dbtob(bp->b_blkno); 1074 bstrategy(bp); 1075 1076 if ((oldflags & B_ASYNC) == 0) { 1077 int rtval = bufwait(bp); 1078 brelse(bp); 1079 return (rtval); 1080 } else { 1081 /* 1082 * don't allow the async write to saturate the I/O 1083 * system. We will not deadlock here because 1084 * we are blocking waiting for I/O that is already in-progress 1085 * to complete. We do not block here if it is the update 1086 * or syncer daemon trying to clean up as that can lead 1087 * to deadlock. 1088 */ 1089 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1090 waitrunningbufspace(); 1091 } 1092 1093 return (0); 1094} 1095 1096void 1097bufbdflush(struct bufobj *bo, struct buf *bp) 1098{ 1099 struct buf *nbp; 1100 1101 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1102 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1103 altbufferflushes++; 1104 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1105 BO_LOCK(bo); 1106 /* 1107 * Try to find a buffer to flush. 1108 */ 1109 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1110 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1111 BUF_LOCK(nbp, 1112 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1113 continue; 1114 if (bp == nbp) 1115 panic("bdwrite: found ourselves"); 1116 BO_UNLOCK(bo); 1117 /* Don't countdeps with the bo lock held. */ 1118 if (buf_countdeps(nbp, 0)) { 1119 BO_LOCK(bo); 1120 BUF_UNLOCK(nbp); 1121 continue; 1122 } 1123 if (nbp->b_flags & B_CLUSTEROK) { 1124 vfs_bio_awrite(nbp); 1125 } else { 1126 bremfree(nbp); 1127 bawrite(nbp); 1128 } 1129 dirtybufferflushes++; 1130 break; 1131 } 1132 if (nbp == NULL) 1133 BO_UNLOCK(bo); 1134 } 1135} 1136 1137/* 1138 * Delayed write. (Buffer is marked dirty). Do not bother writing 1139 * anything if the buffer is marked invalid. 1140 * 1141 * Note that since the buffer must be completely valid, we can safely 1142 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1143 * biodone() in order to prevent getblk from writing the buffer 1144 * out synchronously. 1145 */ 1146void 1147bdwrite(struct buf *bp) 1148{ 1149 struct thread *td = curthread; 1150 struct vnode *vp; 1151 struct bufobj *bo; 1152 1153 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1154 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1155 KASSERT((bp->b_flags & B_BARRIER) == 0, 1156 ("Barrier request in delayed write %p", bp)); 1157 BUF_ASSERT_HELD(bp); 1158 1159 if (bp->b_flags & B_INVAL) { 1160 brelse(bp); 1161 return; 1162 } 1163 1164 /* 1165 * If we have too many dirty buffers, don't create any more. 1166 * If we are wildly over our limit, then force a complete 1167 * cleanup. Otherwise, just keep the situation from getting 1168 * out of control. Note that we have to avoid a recursive 1169 * disaster and not try to clean up after our own cleanup! 1170 */ 1171 vp = bp->b_vp; 1172 bo = bp->b_bufobj; 1173 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1174 td->td_pflags |= TDP_INBDFLUSH; 1175 BO_BDFLUSH(bo, bp); 1176 td->td_pflags &= ~TDP_INBDFLUSH; 1177 } else 1178 recursiveflushes++; 1179 1180 bdirty(bp); 1181 /* 1182 * Set B_CACHE, indicating that the buffer is fully valid. This is 1183 * true even of NFS now. 1184 */ 1185 bp->b_flags |= B_CACHE; 1186 1187 /* 1188 * This bmap keeps the system from needing to do the bmap later, 1189 * perhaps when the system is attempting to do a sync. Since it 1190 * is likely that the indirect block -- or whatever other datastructure 1191 * that the filesystem needs is still in memory now, it is a good 1192 * thing to do this. Note also, that if the pageout daemon is 1193 * requesting a sync -- there might not be enough memory to do 1194 * the bmap then... So, this is important to do. 1195 */ 1196 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1197 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1198 } 1199 1200 /* 1201 * Set the *dirty* buffer range based upon the VM system dirty 1202 * pages. 1203 * 1204 * Mark the buffer pages as clean. We need to do this here to 1205 * satisfy the vnode_pager and the pageout daemon, so that it 1206 * thinks that the pages have been "cleaned". Note that since 1207 * the pages are in a delayed write buffer -- the VFS layer 1208 * "will" see that the pages get written out on the next sync, 1209 * or perhaps the cluster will be completed. 1210 */ 1211 vfs_clean_pages_dirty_buf(bp); 1212 bqrelse(bp); 1213 1214 /* 1215 * Wakeup the buffer flushing daemon if we have a lot of dirty 1216 * buffers (midpoint between our recovery point and our stall 1217 * point). 1218 */ 1219 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1220 1221 /* 1222 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1223 * due to the softdep code. 1224 */ 1225} 1226 1227/* 1228 * bdirty: 1229 * 1230 * Turn buffer into delayed write request. We must clear BIO_READ and 1231 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1232 * itself to properly update it in the dirty/clean lists. We mark it 1233 * B_DONE to ensure that any asynchronization of the buffer properly 1234 * clears B_DONE ( else a panic will occur later ). 1235 * 1236 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1237 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1238 * should only be called if the buffer is known-good. 1239 * 1240 * Since the buffer is not on a queue, we do not update the numfreebuffers 1241 * count. 1242 * 1243 * The buffer must be on QUEUE_NONE. 1244 */ 1245void 1246bdirty(struct buf *bp) 1247{ 1248 1249 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1250 bp, bp->b_vp, bp->b_flags); 1251 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1252 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1253 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1254 BUF_ASSERT_HELD(bp); 1255 bp->b_flags &= ~(B_RELBUF); 1256 bp->b_iocmd = BIO_WRITE; 1257 1258 if ((bp->b_flags & B_DELWRI) == 0) { 1259 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1260 reassignbuf(bp); 1261 atomic_add_int(&numdirtybuffers, 1); 1262 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1263 } 1264} 1265 1266/* 1267 * bundirty: 1268 * 1269 * Clear B_DELWRI for buffer. 1270 * 1271 * Since the buffer is not on a queue, we do not update the numfreebuffers 1272 * count. 1273 * 1274 * The buffer must be on QUEUE_NONE. 1275 */ 1276 1277void 1278bundirty(struct buf *bp) 1279{ 1280 1281 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1282 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1283 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1284 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1285 BUF_ASSERT_HELD(bp); 1286 1287 if (bp->b_flags & B_DELWRI) { 1288 bp->b_flags &= ~B_DELWRI; 1289 reassignbuf(bp); 1290 atomic_subtract_int(&numdirtybuffers, 1); 1291 numdirtywakeup(lodirtybuffers); 1292 } 1293 /* 1294 * Since it is now being written, we can clear its deferred write flag. 1295 */ 1296 bp->b_flags &= ~B_DEFERRED; 1297} 1298 1299/* 1300 * bawrite: 1301 * 1302 * Asynchronous write. Start output on a buffer, but do not wait for 1303 * it to complete. The buffer is released when the output completes. 1304 * 1305 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1306 * B_INVAL buffers. Not us. 1307 */ 1308void 1309bawrite(struct buf *bp) 1310{ 1311 1312 bp->b_flags |= B_ASYNC; 1313 (void) bwrite(bp); 1314} 1315 1316/* 1317 * babarrierwrite: 1318 * 1319 * Asynchronous barrier write. Start output on a buffer, but do not 1320 * wait for it to complete. Place a write barrier after this write so 1321 * that this buffer and all buffers written before it are committed to 1322 * the disk before any buffers written after this write are committed 1323 * to the disk. The buffer is released when the output completes. 1324 */ 1325void 1326babarrierwrite(struct buf *bp) 1327{ 1328 1329 bp->b_flags |= B_ASYNC | B_BARRIER; 1330 (void) bwrite(bp); 1331} 1332 1333/* 1334 * bbarrierwrite: 1335 * 1336 * Synchronous barrier write. Start output on a buffer and wait for 1337 * it to complete. Place a write barrier after this write so that 1338 * this buffer and all buffers written before it are committed to 1339 * the disk before any buffers written after this write are committed 1340 * to the disk. The buffer is released when the output completes. 1341 */ 1342int 1343bbarrierwrite(struct buf *bp) 1344{ 1345 1346 bp->b_flags |= B_BARRIER; 1347 return (bwrite(bp)); 1348} 1349 1350/* 1351 * bwillwrite: 1352 * 1353 * Called prior to the locking of any vnodes when we are expecting to 1354 * write. We do not want to starve the buffer cache with too many 1355 * dirty buffers so we block here. By blocking prior to the locking 1356 * of any vnodes we attempt to avoid the situation where a locked vnode 1357 * prevents the various system daemons from flushing related buffers. 1358 */ 1359 1360void 1361bwillwrite(void) 1362{ 1363 1364 if (numdirtybuffers >= hidirtybuffers) { 1365 mtx_lock(&nblock); 1366 while (numdirtybuffers >= hidirtybuffers) { 1367 bd_wakeup(1); 1368 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1369 msleep(&needsbuffer, &nblock, 1370 (PRIBIO + 4), "flswai", 0); 1371 } 1372 mtx_unlock(&nblock); 1373 } 1374} 1375 1376/* 1377 * Return true if we have too many dirty buffers. 1378 */ 1379int 1380buf_dirty_count_severe(void) 1381{ 1382 1383 return(numdirtybuffers >= hidirtybuffers); 1384} 1385 1386static __noinline int 1387buf_vm_page_count_severe(void) 1388{ 1389 1390 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1391 1392 return vm_page_count_severe(); 1393} 1394 1395/* 1396 * brelse: 1397 * 1398 * Release a busy buffer and, if requested, free its resources. The 1399 * buffer will be stashed in the appropriate bufqueue[] allowing it 1400 * to be accessed later as a cache entity or reused for other purposes. 1401 */ 1402void 1403brelse(struct buf *bp) 1404{ 1405 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1406 bp, bp->b_vp, bp->b_flags); 1407 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1408 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1409 1410 if (BUF_LOCKRECURSED(bp)) { 1411 /* 1412 * Do not process, in particular, do not handle the 1413 * B_INVAL/B_RELBUF and do not release to free list. 1414 */ 1415 BUF_UNLOCK(bp); 1416 return; 1417 } 1418 1419 if (bp->b_flags & B_MANAGED) { 1420 bqrelse(bp); 1421 return; 1422 } 1423 1424 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1425 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1426 /* 1427 * Failed write, redirty. Must clear BIO_ERROR to prevent 1428 * pages from being scrapped. If the error is anything 1429 * other than an I/O error (EIO), assume that retrying 1430 * is futile. 1431 */ 1432 bp->b_ioflags &= ~BIO_ERROR; 1433 bdirty(bp); 1434 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1435 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1436 /* 1437 * Either a failed I/O or we were asked to free or not 1438 * cache the buffer. 1439 */ 1440 bp->b_flags |= B_INVAL; 1441 if (!LIST_EMPTY(&bp->b_dep)) 1442 buf_deallocate(bp); 1443 if (bp->b_flags & B_DELWRI) { 1444 atomic_subtract_int(&numdirtybuffers, 1); 1445 numdirtywakeup(lodirtybuffers); 1446 } 1447 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1448 if ((bp->b_flags & B_VMIO) == 0) { 1449 if (bp->b_bufsize) 1450 allocbuf(bp, 0); 1451 if (bp->b_vp) 1452 brelvp(bp); 1453 } 1454 } 1455 1456 /* 1457 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1458 * is called with B_DELWRI set, the underlying pages may wind up 1459 * getting freed causing a previous write (bdwrite()) to get 'lost' 1460 * because pages associated with a B_DELWRI bp are marked clean. 1461 * 1462 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1463 * if B_DELWRI is set. 1464 * 1465 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1466 * on pages to return pages to the VM page queues. 1467 */ 1468 if (bp->b_flags & B_DELWRI) 1469 bp->b_flags &= ~B_RELBUF; 1470 else if (buf_vm_page_count_severe()) { 1471 /* 1472 * BKGRDINPROG can only be set with the buf and bufobj 1473 * locks both held. We tolerate a race to clear it here. 1474 */ 1475 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1476 bp->b_flags |= B_RELBUF; 1477 } 1478 1479 /* 1480 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1481 * constituted, not even NFS buffers now. Two flags effect this. If 1482 * B_INVAL, the struct buf is invalidated but the VM object is kept 1483 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1484 * 1485 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1486 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1487 * buffer is also B_INVAL because it hits the re-dirtying code above. 1488 * 1489 * Normally we can do this whether a buffer is B_DELWRI or not. If 1490 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1491 * the commit state and we cannot afford to lose the buffer. If the 1492 * buffer has a background write in progress, we need to keep it 1493 * around to prevent it from being reconstituted and starting a second 1494 * background write. 1495 */ 1496 if ((bp->b_flags & B_VMIO) 1497 && !(bp->b_vp->v_mount != NULL && 1498 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1499 !vn_isdisk(bp->b_vp, NULL) && 1500 (bp->b_flags & B_DELWRI)) 1501 ) { 1502 1503 int i, j, resid; 1504 vm_page_t m; 1505 off_t foff; 1506 vm_pindex_t poff; 1507 vm_object_t obj; 1508 1509 obj = bp->b_bufobj->bo_object; 1510 1511 /* 1512 * Get the base offset and length of the buffer. Note that 1513 * in the VMIO case if the buffer block size is not 1514 * page-aligned then b_data pointer may not be page-aligned. 1515 * But our b_pages[] array *IS* page aligned. 1516 * 1517 * block sizes less then DEV_BSIZE (usually 512) are not 1518 * supported due to the page granularity bits (m->valid, 1519 * m->dirty, etc...). 1520 * 1521 * See man buf(9) for more information 1522 */ 1523 resid = bp->b_bufsize; 1524 foff = bp->b_offset; 1525 VM_OBJECT_WLOCK(obj); 1526 for (i = 0; i < bp->b_npages; i++) { 1527 int had_bogus = 0; 1528 1529 m = bp->b_pages[i]; 1530 1531 /* 1532 * If we hit a bogus page, fixup *all* the bogus pages 1533 * now. 1534 */ 1535 if (m == bogus_page) { 1536 poff = OFF_TO_IDX(bp->b_offset); 1537 had_bogus = 1; 1538 1539 for (j = i; j < bp->b_npages; j++) { 1540 vm_page_t mtmp; 1541 mtmp = bp->b_pages[j]; 1542 if (mtmp == bogus_page) { 1543 mtmp = vm_page_lookup(obj, poff + j); 1544 if (!mtmp) { 1545 panic("brelse: page missing\n"); 1546 } 1547 bp->b_pages[j] = mtmp; 1548 } 1549 } 1550 1551 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1552 BUF_CHECK_MAPPED(bp); 1553 pmap_qenter( 1554 trunc_page((vm_offset_t)bp->b_data), 1555 bp->b_pages, bp->b_npages); 1556 } 1557 m = bp->b_pages[i]; 1558 } 1559 if ((bp->b_flags & B_NOCACHE) || 1560 (bp->b_ioflags & BIO_ERROR && 1561 bp->b_iocmd == BIO_READ)) { 1562 int poffset = foff & PAGE_MASK; 1563 int presid = resid > (PAGE_SIZE - poffset) ? 1564 (PAGE_SIZE - poffset) : resid; 1565 1566 KASSERT(presid >= 0, ("brelse: extra page")); 1567 vm_page_set_invalid(m, poffset, presid); 1568 if (had_bogus) 1569 printf("avoided corruption bug in bogus_page/brelse code\n"); 1570 } 1571 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1572 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1573 } 1574 VM_OBJECT_WUNLOCK(obj); 1575 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1576 vfs_vmio_release(bp); 1577 1578 } else if (bp->b_flags & B_VMIO) { 1579 1580 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1581 vfs_vmio_release(bp); 1582 } 1583 1584 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1585 if (bp->b_bufsize != 0) 1586 allocbuf(bp, 0); 1587 if (bp->b_vp != NULL) 1588 brelvp(bp); 1589 } 1590 1591 /* enqueue */ 1592 mtx_lock(&bqlock); 1593 /* Handle delayed bremfree() processing. */ 1594 if (bp->b_flags & B_REMFREE) 1595 bremfreel(bp); 1596 1597 if (bp->b_qindex != QUEUE_NONE) 1598 panic("brelse: free buffer onto another queue???"); 1599 1600 /* 1601 * If the buffer has junk contents signal it and eventually 1602 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1603 * doesn't find it. 1604 */ 1605 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1606 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1607 bp->b_flags |= B_INVAL; 1608 if (bp->b_flags & B_INVAL) { 1609 if (bp->b_flags & B_DELWRI) 1610 bundirty(bp); 1611 if (bp->b_vp) 1612 brelvp(bp); 1613 } 1614 1615 /* buffers with no memory */ 1616 if (bp->b_bufsize == 0) { 1617 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1618 if (bp->b_vflags & BV_BKGRDINPROG) 1619 panic("losing buffer 1"); 1620 if (bp->b_kvasize) { 1621 bp->b_qindex = QUEUE_EMPTYKVA; 1622 } else { 1623 bp->b_qindex = QUEUE_EMPTY; 1624 } 1625 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1626 /* buffers with junk contents */ 1627 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1628 (bp->b_ioflags & BIO_ERROR)) { 1629 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1630 if (bp->b_vflags & BV_BKGRDINPROG) 1631 panic("losing buffer 2"); 1632 bp->b_qindex = QUEUE_CLEAN; 1633 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1634 /* remaining buffers */ 1635 } else { 1636 if (bp->b_flags & B_DELWRI) 1637 bp->b_qindex = QUEUE_DIRTY; 1638 else 1639 bp->b_qindex = QUEUE_CLEAN; 1640 if (bp->b_flags & B_AGE) { 1641 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, 1642 b_freelist); 1643 } else { 1644 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, 1645 b_freelist); 1646 } 1647 } 1648#ifdef INVARIANTS 1649 bq_len[bp->b_qindex]++; 1650#endif 1651 mtx_unlock(&bqlock); 1652 1653 /* 1654 * Fixup numfreebuffers count. The bp is on an appropriate queue 1655 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1656 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1657 * if B_INVAL is set ). 1658 */ 1659 1660 if (!(bp->b_flags & B_DELWRI)) 1661 bufcountwakeup(bp); 1662 1663 /* 1664 * Something we can maybe free or reuse 1665 */ 1666 if (bp->b_bufsize || bp->b_kvasize) 1667 bufspacewakeup(); 1668 1669 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1670 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1671 panic("brelse: not dirty"); 1672 /* unlock */ 1673 BUF_UNLOCK(bp); 1674} 1675 1676/* 1677 * Release a buffer back to the appropriate queue but do not try to free 1678 * it. The buffer is expected to be used again soon. 1679 * 1680 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1681 * biodone() to requeue an async I/O on completion. It is also used when 1682 * known good buffers need to be requeued but we think we may need the data 1683 * again soon. 1684 * 1685 * XXX we should be able to leave the B_RELBUF hint set on completion. 1686 */ 1687void 1688bqrelse(struct buf *bp) 1689{ 1690 struct bufobj *bo; 1691 1692 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1693 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1694 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1695 1696 if (BUF_LOCKRECURSED(bp)) { 1697 /* do not release to free list */ 1698 BUF_UNLOCK(bp); 1699 return; 1700 } 1701 1702 bo = bp->b_bufobj; 1703 if (bp->b_flags & B_MANAGED) { 1704 if (bp->b_flags & B_REMFREE) { 1705 mtx_lock(&bqlock); 1706 bremfreel(bp); 1707 mtx_unlock(&bqlock); 1708 } 1709 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1710 BUF_UNLOCK(bp); 1711 return; 1712 } 1713 1714 mtx_lock(&bqlock); 1715 /* Handle delayed bremfree() processing. */ 1716 if (bp->b_flags & B_REMFREE) 1717 bremfreel(bp); 1718 1719 if (bp->b_qindex != QUEUE_NONE) 1720 panic("bqrelse: free buffer onto another queue???"); 1721 /* buffers with stale but valid contents */ 1722 if (bp->b_flags & B_DELWRI) { 1723 bp->b_qindex = QUEUE_DIRTY; 1724 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1725#ifdef INVARIANTS 1726 bq_len[bp->b_qindex]++; 1727#endif 1728 } else { 1729 /* 1730 * BKGRDINPROG can only be set with the buf and bufobj 1731 * locks both held. We tolerate a race to clear it here. 1732 */ 1733 if (!buf_vm_page_count_severe() || 1734 (bp->b_vflags & BV_BKGRDINPROG)) { 1735 bp->b_qindex = QUEUE_CLEAN; 1736 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1737 b_freelist); 1738#ifdef INVARIANTS 1739 bq_len[QUEUE_CLEAN]++; 1740#endif 1741 } else { 1742 /* 1743 * We are too low on memory, we have to try to free 1744 * the buffer (most importantly: the wired pages 1745 * making up its backing store) *now*. 1746 */ 1747 mtx_unlock(&bqlock); 1748 brelse(bp); 1749 return; 1750 } 1751 } 1752 mtx_unlock(&bqlock); 1753 1754 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1755 bufcountwakeup(bp); 1756 1757 /* 1758 * Something we can maybe free or reuse. 1759 */ 1760 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1761 bufspacewakeup(); 1762 1763 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1764 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1765 panic("bqrelse: not dirty"); 1766 /* unlock */ 1767 BUF_UNLOCK(bp); 1768} 1769 1770/* Give pages used by the bp back to the VM system (where possible) */ 1771static void 1772vfs_vmio_release(struct buf *bp) 1773{ 1774 int i; 1775 vm_page_t m; 1776 1777 if ((bp->b_flags & B_UNMAPPED) == 0) { 1778 BUF_CHECK_MAPPED(bp); 1779 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1780 } else 1781 BUF_CHECK_UNMAPPED(bp); 1782 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 1783 for (i = 0; i < bp->b_npages; i++) { 1784 m = bp->b_pages[i]; 1785 bp->b_pages[i] = NULL; 1786 /* 1787 * In order to keep page LRU ordering consistent, put 1788 * everything on the inactive queue. 1789 */ 1790 vm_page_lock(m); 1791 vm_page_unwire(m, 0); 1792 /* 1793 * We don't mess with busy pages, it is 1794 * the responsibility of the process that 1795 * busied the pages to deal with them. 1796 */ 1797 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1798 m->wire_count == 0) { 1799 /* 1800 * Might as well free the page if we can and it has 1801 * no valid data. We also free the page if the 1802 * buffer was used for direct I/O 1803 */ 1804 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1805 vm_page_free(m); 1806 } else if (bp->b_flags & B_DIRECT) { 1807 vm_page_try_to_free(m); 1808 } else if (buf_vm_page_count_severe()) { 1809 vm_page_try_to_cache(m); 1810 } 1811 } 1812 vm_page_unlock(m); 1813 } 1814 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 1815 1816 if (bp->b_bufsize) { 1817 bufspacewakeup(); 1818 bp->b_bufsize = 0; 1819 } 1820 bp->b_npages = 0; 1821 bp->b_flags &= ~B_VMIO; 1822 if (bp->b_vp) 1823 brelvp(bp); 1824} 1825 1826/* 1827 * Check to see if a block at a particular lbn is available for a clustered 1828 * write. 1829 */ 1830static int 1831vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1832{ 1833 struct buf *bpa; 1834 int match; 1835 1836 match = 0; 1837 1838 /* If the buf isn't in core skip it */ 1839 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1840 return (0); 1841 1842 /* If the buf is busy we don't want to wait for it */ 1843 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1844 return (0); 1845 1846 /* Only cluster with valid clusterable delayed write buffers */ 1847 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1848 (B_DELWRI | B_CLUSTEROK)) 1849 goto done; 1850 1851 if (bpa->b_bufsize != size) 1852 goto done; 1853 1854 /* 1855 * Check to see if it is in the expected place on disk and that the 1856 * block has been mapped. 1857 */ 1858 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1859 match = 1; 1860done: 1861 BUF_UNLOCK(bpa); 1862 return (match); 1863} 1864 1865/* 1866 * vfs_bio_awrite: 1867 * 1868 * Implement clustered async writes for clearing out B_DELWRI buffers. 1869 * This is much better then the old way of writing only one buffer at 1870 * a time. Note that we may not be presented with the buffers in the 1871 * correct order, so we search for the cluster in both directions. 1872 */ 1873int 1874vfs_bio_awrite(struct buf *bp) 1875{ 1876 struct bufobj *bo; 1877 int i; 1878 int j; 1879 daddr_t lblkno = bp->b_lblkno; 1880 struct vnode *vp = bp->b_vp; 1881 int ncl; 1882 int nwritten; 1883 int size; 1884 int maxcl; 1885 int gbflags; 1886 1887 bo = &vp->v_bufobj; 1888 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1889 /* 1890 * right now we support clustered writing only to regular files. If 1891 * we find a clusterable block we could be in the middle of a cluster 1892 * rather then at the beginning. 1893 */ 1894 if ((vp->v_type == VREG) && 1895 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1896 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1897 1898 size = vp->v_mount->mnt_stat.f_iosize; 1899 maxcl = MAXPHYS / size; 1900 1901 BO_RLOCK(bo); 1902 for (i = 1; i < maxcl; i++) 1903 if (vfs_bio_clcheck(vp, size, lblkno + i, 1904 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1905 break; 1906 1907 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1908 if (vfs_bio_clcheck(vp, size, lblkno - j, 1909 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1910 break; 1911 BO_RUNLOCK(bo); 1912 --j; 1913 ncl = i + j; 1914 /* 1915 * this is a possible cluster write 1916 */ 1917 if (ncl != 1) { 1918 BUF_UNLOCK(bp); 1919 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1920 gbflags); 1921 return (nwritten); 1922 } 1923 } 1924 bremfree(bp); 1925 bp->b_flags |= B_ASYNC; 1926 /* 1927 * default (old) behavior, writing out only one block 1928 * 1929 * XXX returns b_bufsize instead of b_bcount for nwritten? 1930 */ 1931 nwritten = bp->b_bufsize; 1932 (void) bwrite(bp); 1933 1934 return (nwritten); 1935} 1936 1937static void 1938setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 1939{ 1940 1941 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 1942 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 1943 if ((gbflags & GB_UNMAPPED) == 0) { 1944 bp->b_kvabase = (caddr_t)addr; 1945 } else if ((gbflags & GB_KVAALLOC) != 0) { 1946 KASSERT((gbflags & GB_UNMAPPED) != 0, 1947 ("GB_KVAALLOC without GB_UNMAPPED")); 1948 bp->b_kvaalloc = (caddr_t)addr; 1949 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 1950 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 1951 } 1952 bp->b_kvasize = maxsize; 1953} 1954 1955/* 1956 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 1957 * needed. 1958 */ 1959static int 1960allocbufkva(struct buf *bp, int maxsize, int gbflags) 1961{ 1962 vm_offset_t addr; 1963 int rv; 1964 1965 bfreekva(bp); 1966 addr = 0; 1967 1968 vm_map_lock(buffer_map); 1969 if (vm_map_findspace(buffer_map, vm_map_min(buffer_map), maxsize, 1970 &addr)) { 1971 vm_map_unlock(buffer_map); 1972 /* 1973 * Buffer map is too fragmented. Request the caller 1974 * to defragment the map. 1975 */ 1976 atomic_add_int(&bufdefragcnt, 1); 1977 return (1); 1978 } 1979 rv = vm_map_insert(buffer_map, NULL, 0, addr, addr + maxsize, 1980 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 1981 KASSERT(rv == KERN_SUCCESS, ("vm_map_insert(buffer_map) rv %d", rv)); 1982 vm_map_unlock(buffer_map); 1983 setbufkva(bp, addr, maxsize, gbflags); 1984 atomic_add_long(&bufspace, bp->b_kvasize); 1985 return (0); 1986} 1987 1988/* 1989 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 1990 * locked vnode is supplied. 1991 */ 1992static void 1993getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 1994 int defrag) 1995{ 1996 struct thread *td; 1997 char *waitmsg; 1998 int fl, flags, norunbuf; 1999 2000 mtx_assert(&bqlock, MA_OWNED); 2001 2002 if (defrag) { 2003 flags = VFS_BIO_NEED_BUFSPACE; 2004 waitmsg = "nbufkv"; 2005 } else if (bufspace >= hibufspace) { 2006 waitmsg = "nbufbs"; 2007 flags = VFS_BIO_NEED_BUFSPACE; 2008 } else { 2009 waitmsg = "newbuf"; 2010 flags = VFS_BIO_NEED_ANY; 2011 } 2012 mtx_lock(&nblock); 2013 needsbuffer |= flags; 2014 mtx_unlock(&nblock); 2015 mtx_unlock(&bqlock); 2016 2017 bd_speedup(); /* heeeelp */ 2018 if ((gbflags & GB_NOWAIT_BD) != 0) 2019 return; 2020 2021 td = curthread; 2022 mtx_lock(&nblock); 2023 while (needsbuffer & flags) { 2024 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2025 mtx_unlock(&nblock); 2026 /* 2027 * getblk() is called with a vnode locked, and 2028 * some majority of the dirty buffers may as 2029 * well belong to the vnode. Flushing the 2030 * buffers there would make a progress that 2031 * cannot be achieved by the buf_daemon, that 2032 * cannot lock the vnode. 2033 */ 2034 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2035 (td->td_pflags & TDP_NORUNNINGBUF); 2036 /* play bufdaemon */ 2037 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2038 fl = buf_do_flush(vp); 2039 td->td_pflags &= norunbuf; 2040 mtx_lock(&nblock); 2041 if (fl != 0) 2042 continue; 2043 if ((needsbuffer & flags) == 0) 2044 break; 2045 } 2046 if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag, 2047 waitmsg, slptimeo)) 2048 break; 2049 } 2050 mtx_unlock(&nblock); 2051} 2052 2053static void 2054getnewbuf_reuse_bp(struct buf *bp, int qindex) 2055{ 2056 2057 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2058 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2059 bp->b_kvasize, bp->b_bufsize, qindex); 2060 mtx_assert(&bqlock, MA_NOTOWNED); 2061 2062 /* 2063 * Note: we no longer distinguish between VMIO and non-VMIO 2064 * buffers. 2065 */ 2066 KASSERT((bp->b_flags & B_DELWRI) == 0, 2067 ("delwri buffer %p found in queue %d", bp, qindex)); 2068 2069 if (qindex == QUEUE_CLEAN) { 2070 if (bp->b_flags & B_VMIO) { 2071 bp->b_flags &= ~B_ASYNC; 2072 vfs_vmio_release(bp); 2073 } 2074 if (bp->b_vp != NULL) 2075 brelvp(bp); 2076 } 2077 2078 /* 2079 * Get the rest of the buffer freed up. b_kva* is still valid 2080 * after this operation. 2081 */ 2082 2083 if (bp->b_rcred != NOCRED) { 2084 crfree(bp->b_rcred); 2085 bp->b_rcred = NOCRED; 2086 } 2087 if (bp->b_wcred != NOCRED) { 2088 crfree(bp->b_wcred); 2089 bp->b_wcred = NOCRED; 2090 } 2091 if (!LIST_EMPTY(&bp->b_dep)) 2092 buf_deallocate(bp); 2093 if (bp->b_vflags & BV_BKGRDINPROG) 2094 panic("losing buffer 3"); 2095 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2096 bp, bp->b_vp, qindex)); 2097 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2098 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2099 2100 if (bp->b_bufsize) 2101 allocbuf(bp, 0); 2102 2103 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2104 bp->b_ioflags = 0; 2105 bp->b_xflags = 0; 2106 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2107 ("buf %p still counted as free?", bp)); 2108 bp->b_vflags = 0; 2109 bp->b_vp = NULL; 2110 bp->b_blkno = bp->b_lblkno = 0; 2111 bp->b_offset = NOOFFSET; 2112 bp->b_iodone = 0; 2113 bp->b_error = 0; 2114 bp->b_resid = 0; 2115 bp->b_bcount = 0; 2116 bp->b_npages = 0; 2117 bp->b_dirtyoff = bp->b_dirtyend = 0; 2118 bp->b_bufobj = NULL; 2119 bp->b_pin_count = 0; 2120 bp->b_fsprivate1 = NULL; 2121 bp->b_fsprivate2 = NULL; 2122 bp->b_fsprivate3 = NULL; 2123 2124 LIST_INIT(&bp->b_dep); 2125} 2126 2127static int flushingbufs; 2128 2129static struct buf * 2130getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2131{ 2132 struct buf *bp, *nbp; 2133 int nqindex, qindex, pass; 2134 2135 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2136 2137 pass = 1; 2138restart: 2139 atomic_add_int(&getnewbufrestarts, 1); 2140 2141 /* 2142 * Setup for scan. If we do not have enough free buffers, 2143 * we setup a degenerate case that immediately fails. Note 2144 * that if we are specially marked process, we are allowed to 2145 * dip into our reserves. 2146 * 2147 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2148 * for the allocation of the mapped buffer. For unmapped, the 2149 * easiest is to start with EMPTY outright. 2150 * 2151 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2152 * However, there are a number of cases (defragging, reusing, ...) 2153 * where we cannot backup. 2154 */ 2155 nbp = NULL; 2156 mtx_lock(&bqlock); 2157 if (!defrag && unmapped) { 2158 nqindex = QUEUE_EMPTY; 2159 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2160 } 2161 if (nbp == NULL) { 2162 nqindex = QUEUE_EMPTYKVA; 2163 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2164 } 2165 2166 /* 2167 * If no EMPTYKVA buffers and we are either defragging or 2168 * reusing, locate a CLEAN buffer to free or reuse. If 2169 * bufspace useage is low skip this step so we can allocate a 2170 * new buffer. 2171 */ 2172 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2173 nqindex = QUEUE_CLEAN; 2174 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2175 } 2176 2177 /* 2178 * If we could not find or were not allowed to reuse a CLEAN 2179 * buffer, check to see if it is ok to use an EMPTY buffer. 2180 * We can only use an EMPTY buffer if allocating its KVA would 2181 * not otherwise run us out of buffer space. No KVA is needed 2182 * for the unmapped allocation. 2183 */ 2184 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2185 metadata)) { 2186 nqindex = QUEUE_EMPTY; 2187 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2188 } 2189 2190 /* 2191 * All available buffers might be clean, retry ignoring the 2192 * lobufspace as the last resort. 2193 */ 2194 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2195 nqindex = QUEUE_CLEAN; 2196 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2197 } 2198 2199 /* 2200 * Run scan, possibly freeing data and/or kva mappings on the fly 2201 * depending. 2202 */ 2203 while ((bp = nbp) != NULL) { 2204 qindex = nqindex; 2205 2206 /* 2207 * Calculate next bp (we can only use it if we do not 2208 * block or do other fancy things). 2209 */ 2210 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2211 switch (qindex) { 2212 case QUEUE_EMPTY: 2213 nqindex = QUEUE_EMPTYKVA; 2214 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2215 if (nbp != NULL) 2216 break; 2217 /* FALLTHROUGH */ 2218 case QUEUE_EMPTYKVA: 2219 nqindex = QUEUE_CLEAN; 2220 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2221 if (nbp != NULL) 2222 break; 2223 /* FALLTHROUGH */ 2224 case QUEUE_CLEAN: 2225 if (metadata && pass == 1) { 2226 pass = 2; 2227 nqindex = QUEUE_EMPTY; 2228 nbp = TAILQ_FIRST( 2229 &bufqueues[QUEUE_EMPTY]); 2230 } 2231 /* 2232 * nbp is NULL. 2233 */ 2234 break; 2235 } 2236 } 2237 /* 2238 * If we are defragging then we need a buffer with 2239 * b_kvasize != 0. XXX this situation should no longer 2240 * occur, if defrag is non-zero the buffer's b_kvasize 2241 * should also be non-zero at this point. XXX 2242 */ 2243 if (defrag && bp->b_kvasize == 0) { 2244 printf("Warning: defrag empty buffer %p\n", bp); 2245 continue; 2246 } 2247 2248 /* 2249 * Start freeing the bp. This is somewhat involved. nbp 2250 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2251 */ 2252 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2253 continue; 2254 /* 2255 * BKGRDINPROG can only be set with the buf and bufobj 2256 * locks both held. We tolerate a race to clear it here. 2257 */ 2258 if (bp->b_vflags & BV_BKGRDINPROG) { 2259 BUF_UNLOCK(bp); 2260 continue; 2261 } 2262 2263 KASSERT(bp->b_qindex == qindex, 2264 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2265 2266 bremfreel(bp); 2267 mtx_unlock(&bqlock); 2268 /* 2269 * NOTE: nbp is now entirely invalid. We can only restart 2270 * the scan from this point on. 2271 */ 2272 2273 getnewbuf_reuse_bp(bp, qindex); 2274 mtx_assert(&bqlock, MA_NOTOWNED); 2275 2276 /* 2277 * If we are defragging then free the buffer. 2278 */ 2279 if (defrag) { 2280 bp->b_flags |= B_INVAL; 2281 bfreekva(bp); 2282 brelse(bp); 2283 defrag = 0; 2284 goto restart; 2285 } 2286 2287 /* 2288 * Notify any waiters for the buffer lock about 2289 * identity change by freeing the buffer. 2290 */ 2291 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2292 bp->b_flags |= B_INVAL; 2293 bfreekva(bp); 2294 brelse(bp); 2295 goto restart; 2296 } 2297 2298 if (metadata) 2299 break; 2300 2301 /* 2302 * If we are overcomitted then recover the buffer and its 2303 * KVM space. This occurs in rare situations when multiple 2304 * processes are blocked in getnewbuf() or allocbuf(). 2305 */ 2306 if (bufspace >= hibufspace) 2307 flushingbufs = 1; 2308 if (flushingbufs && bp->b_kvasize != 0) { 2309 bp->b_flags |= B_INVAL; 2310 bfreekva(bp); 2311 brelse(bp); 2312 goto restart; 2313 } 2314 if (bufspace < lobufspace) 2315 flushingbufs = 0; 2316 break; 2317 } 2318 return (bp); 2319} 2320 2321/* 2322 * getnewbuf: 2323 * 2324 * Find and initialize a new buffer header, freeing up existing buffers 2325 * in the bufqueues as necessary. The new buffer is returned locked. 2326 * 2327 * Important: B_INVAL is not set. If the caller wishes to throw the 2328 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2329 * 2330 * We block if: 2331 * We have insufficient buffer headers 2332 * We have insufficient buffer space 2333 * buffer_map is too fragmented ( space reservation fails ) 2334 * If we have to flush dirty buffers ( but we try to avoid this ) 2335 * 2336 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 2337 * Instead we ask the buf daemon to do it for us. We attempt to 2338 * avoid piecemeal wakeups of the pageout daemon. 2339 */ 2340static struct buf * 2341getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2342 int gbflags) 2343{ 2344 struct buf *bp; 2345 int defrag, metadata; 2346 2347 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2348 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2349 if (!unmapped_buf_allowed) 2350 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2351 2352 defrag = 0; 2353 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2354 vp->v_type == VCHR) 2355 metadata = 1; 2356 else 2357 metadata = 0; 2358 /* 2359 * We can't afford to block since we might be holding a vnode lock, 2360 * which may prevent system daemons from running. We deal with 2361 * low-memory situations by proactively returning memory and running 2362 * async I/O rather then sync I/O. 2363 */ 2364 atomic_add_int(&getnewbufcalls, 1); 2365 atomic_subtract_int(&getnewbufrestarts, 1); 2366restart: 2367 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2368 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2369 if (bp != NULL) 2370 defrag = 0; 2371 2372 /* 2373 * If we exhausted our list, sleep as appropriate. We may have to 2374 * wakeup various daemons and write out some dirty buffers. 2375 * 2376 * Generally we are sleeping due to insufficient buffer space. 2377 */ 2378 if (bp == NULL) { 2379 mtx_assert(&bqlock, MA_OWNED); 2380 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2381 mtx_assert(&bqlock, MA_NOTOWNED); 2382 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2383 mtx_assert(&bqlock, MA_NOTOWNED); 2384 2385 bfreekva(bp); 2386 bp->b_flags |= B_UNMAPPED; 2387 bp->b_kvabase = bp->b_data = unmapped_buf; 2388 bp->b_kvasize = maxsize; 2389 atomic_add_long(&bufspace, bp->b_kvasize); 2390 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2391 atomic_add_int(&bufreusecnt, 1); 2392 } else { 2393 mtx_assert(&bqlock, MA_NOTOWNED); 2394 2395 /* 2396 * We finally have a valid bp. We aren't quite out of the 2397 * woods, we still have to reserve kva space. In order 2398 * to keep fragmentation sane we only allocate kva in 2399 * BKVASIZE chunks. 2400 */ 2401 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2402 2403 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2404 B_KVAALLOC)) == B_UNMAPPED) { 2405 if (allocbufkva(bp, maxsize, gbflags)) { 2406 defrag = 1; 2407 bp->b_flags |= B_INVAL; 2408 brelse(bp); 2409 goto restart; 2410 } 2411 atomic_add_int(&bufreusecnt, 1); 2412 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2413 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2414 /* 2415 * If the reused buffer has KVA allocated, 2416 * reassign b_kvaalloc to b_kvabase. 2417 */ 2418 bp->b_kvabase = bp->b_kvaalloc; 2419 bp->b_flags &= ~B_KVAALLOC; 2420 atomic_subtract_long(&unmapped_bufspace, 2421 bp->b_kvasize); 2422 atomic_add_int(&bufreusecnt, 1); 2423 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2424 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2425 GB_KVAALLOC)) { 2426 /* 2427 * The case of reused buffer already have KVA 2428 * mapped, but the request is for unmapped 2429 * buffer with KVA allocated. 2430 */ 2431 bp->b_kvaalloc = bp->b_kvabase; 2432 bp->b_data = bp->b_kvabase = unmapped_buf; 2433 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2434 atomic_add_long(&unmapped_bufspace, 2435 bp->b_kvasize); 2436 atomic_add_int(&bufreusecnt, 1); 2437 } 2438 if ((gbflags & GB_UNMAPPED) == 0) { 2439 bp->b_saveaddr = bp->b_kvabase; 2440 bp->b_data = bp->b_saveaddr; 2441 bp->b_flags &= ~B_UNMAPPED; 2442 BUF_CHECK_MAPPED(bp); 2443 } 2444 } 2445 return (bp); 2446} 2447 2448/* 2449 * buf_daemon: 2450 * 2451 * buffer flushing daemon. Buffers are normally flushed by the 2452 * update daemon but if it cannot keep up this process starts to 2453 * take the load in an attempt to prevent getnewbuf() from blocking. 2454 */ 2455 2456static struct kproc_desc buf_kp = { 2457 "bufdaemon", 2458 buf_daemon, 2459 &bufdaemonproc 2460}; 2461SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2462 2463static int 2464buf_do_flush(struct vnode *vp) 2465{ 2466 int flushed; 2467 2468 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2469 if (flushed == 0) { 2470 /* 2471 * Could not find any buffers without rollback 2472 * dependencies, so just write the first one 2473 * in the hopes of eventually making progress. 2474 */ 2475 flushbufqueues(vp, QUEUE_DIRTY, 1); 2476 } 2477 return (flushed); 2478} 2479 2480static void 2481buf_daemon() 2482{ 2483 int lodirtysave; 2484 2485 /* 2486 * This process needs to be suspended prior to shutdown sync. 2487 */ 2488 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2489 SHUTDOWN_PRI_LAST); 2490 2491 /* 2492 * This process is allowed to take the buffer cache to the limit 2493 */ 2494 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2495 mtx_lock(&bdlock); 2496 for (;;) { 2497 bd_request = 0; 2498 mtx_unlock(&bdlock); 2499 2500 kproc_suspend_check(bufdaemonproc); 2501 lodirtysave = lodirtybuffers; 2502 if (bd_speedupreq) { 2503 lodirtybuffers = numdirtybuffers / 2; 2504 bd_speedupreq = 0; 2505 } 2506 /* 2507 * Do the flush. Limit the amount of in-transit I/O we 2508 * allow to build up, otherwise we would completely saturate 2509 * the I/O system. Wakeup any waiting processes before we 2510 * normally would so they can run in parallel with our drain. 2511 */ 2512 while (numdirtybuffers > lodirtybuffers) { 2513 if (buf_do_flush(NULL) == 0) 2514 break; 2515 kern_yield(PRI_USER); 2516 } 2517 lodirtybuffers = lodirtysave; 2518 2519 /* 2520 * Only clear bd_request if we have reached our low water 2521 * mark. The buf_daemon normally waits 1 second and 2522 * then incrementally flushes any dirty buffers that have 2523 * built up, within reason. 2524 * 2525 * If we were unable to hit our low water mark and couldn't 2526 * find any flushable buffers, we sleep half a second. 2527 * Otherwise we loop immediately. 2528 */ 2529 mtx_lock(&bdlock); 2530 if (numdirtybuffers <= lodirtybuffers) { 2531 /* 2532 * We reached our low water mark, reset the 2533 * request and sleep until we are needed again. 2534 * The sleep is just so the suspend code works. 2535 */ 2536 bd_request = 0; 2537 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2538 } else { 2539 /* 2540 * We couldn't find any flushable dirty buffers but 2541 * still have too many dirty buffers, we 2542 * have to sleep and try again. (rare) 2543 */ 2544 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2545 } 2546 } 2547} 2548 2549/* 2550 * flushbufqueues: 2551 * 2552 * Try to flush a buffer in the dirty queue. We must be careful to 2553 * free up B_INVAL buffers instead of write them, which NFS is 2554 * particularly sensitive to. 2555 */ 2556static int flushwithdeps = 0; 2557SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2558 0, "Number of buffers flushed with dependecies that require rollbacks"); 2559 2560static int 2561flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2562{ 2563 struct buf *sentinel; 2564 struct vnode *vp; 2565 struct mount *mp; 2566 struct buf *bp; 2567 int hasdeps; 2568 int flushed; 2569 int target; 2570 2571 if (lvp == NULL) { 2572 target = numdirtybuffers - lodirtybuffers; 2573 if (flushdeps && target > 2) 2574 target /= 2; 2575 } else 2576 target = flushbufqtarget; 2577 flushed = 0; 2578 bp = NULL; 2579 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2580 sentinel->b_qindex = QUEUE_SENTINEL; 2581 mtx_lock(&bqlock); 2582 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2583 while (flushed != target) { 2584 bp = TAILQ_NEXT(sentinel, b_freelist); 2585 if (bp != NULL) { 2586 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2587 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2588 b_freelist); 2589 } else 2590 break; 2591 /* 2592 * Skip sentinels inserted by other invocations of the 2593 * flushbufqueues(), taking care to not reorder them. 2594 */ 2595 if (bp->b_qindex == QUEUE_SENTINEL) 2596 continue; 2597 /* 2598 * Only flush the buffers that belong to the 2599 * vnode locked by the curthread. 2600 */ 2601 if (lvp != NULL && bp->b_vp != lvp) 2602 continue; 2603 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2604 continue; 2605 if (bp->b_pin_count > 0) { 2606 BUF_UNLOCK(bp); 2607 continue; 2608 } 2609 /* 2610 * BKGRDINPROG can only be set with the buf and bufobj 2611 * locks both held. We tolerate a race to clear it here. 2612 */ 2613 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2614 (bp->b_flags & B_DELWRI) == 0) { 2615 BUF_UNLOCK(bp); 2616 continue; 2617 } 2618 if (bp->b_flags & B_INVAL) { 2619 bremfreel(bp); 2620 mtx_unlock(&bqlock); 2621 brelse(bp); 2622 flushed++; 2623 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2624 mtx_lock(&bqlock); 2625 continue; 2626 } 2627 2628 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2629 if (flushdeps == 0) { 2630 BUF_UNLOCK(bp); 2631 continue; 2632 } 2633 hasdeps = 1; 2634 } else 2635 hasdeps = 0; 2636 /* 2637 * We must hold the lock on a vnode before writing 2638 * one of its buffers. Otherwise we may confuse, or 2639 * in the case of a snapshot vnode, deadlock the 2640 * system. 2641 * 2642 * The lock order here is the reverse of the normal 2643 * of vnode followed by buf lock. This is ok because 2644 * the NOWAIT will prevent deadlock. 2645 */ 2646 vp = bp->b_vp; 2647 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2648 BUF_UNLOCK(bp); 2649 continue; 2650 } 2651 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2652 mtx_unlock(&bqlock); 2653 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2654 bp, bp->b_vp, bp->b_flags); 2655 if (curproc == bufdaemonproc) 2656 vfs_bio_awrite(bp); 2657 else { 2658 bremfree(bp); 2659 bwrite(bp); 2660 notbufdflashes++; 2661 } 2662 vn_finished_write(mp); 2663 VOP_UNLOCK(vp, 0); 2664 flushwithdeps += hasdeps; 2665 flushed++; 2666 2667 /* 2668 * Sleeping on runningbufspace while holding 2669 * vnode lock leads to deadlock. 2670 */ 2671 if (curproc == bufdaemonproc) 2672 waitrunningbufspace(); 2673 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2674 mtx_lock(&bqlock); 2675 continue; 2676 } 2677 vn_finished_write(mp); 2678 BUF_UNLOCK(bp); 2679 } 2680 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2681 mtx_unlock(&bqlock); 2682 free(sentinel, M_TEMP); 2683 return (flushed); 2684} 2685 2686/* 2687 * Check to see if a block is currently memory resident. 2688 */ 2689struct buf * 2690incore(struct bufobj *bo, daddr_t blkno) 2691{ 2692 struct buf *bp; 2693 2694 BO_RLOCK(bo); 2695 bp = gbincore(bo, blkno); 2696 BO_RUNLOCK(bo); 2697 return (bp); 2698} 2699 2700/* 2701 * Returns true if no I/O is needed to access the 2702 * associated VM object. This is like incore except 2703 * it also hunts around in the VM system for the data. 2704 */ 2705 2706static int 2707inmem(struct vnode * vp, daddr_t blkno) 2708{ 2709 vm_object_t obj; 2710 vm_offset_t toff, tinc, size; 2711 vm_page_t m; 2712 vm_ooffset_t off; 2713 2714 ASSERT_VOP_LOCKED(vp, "inmem"); 2715 2716 if (incore(&vp->v_bufobj, blkno)) 2717 return 1; 2718 if (vp->v_mount == NULL) 2719 return 0; 2720 obj = vp->v_object; 2721 if (obj == NULL) 2722 return (0); 2723 2724 size = PAGE_SIZE; 2725 if (size > vp->v_mount->mnt_stat.f_iosize) 2726 size = vp->v_mount->mnt_stat.f_iosize; 2727 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2728 2729 VM_OBJECT_RLOCK(obj); 2730 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2731 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2732 if (!m) 2733 goto notinmem; 2734 tinc = size; 2735 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2736 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2737 if (vm_page_is_valid(m, 2738 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2739 goto notinmem; 2740 } 2741 VM_OBJECT_RUNLOCK(obj); 2742 return 1; 2743 2744notinmem: 2745 VM_OBJECT_RUNLOCK(obj); 2746 return (0); 2747} 2748 2749/* 2750 * Set the dirty range for a buffer based on the status of the dirty 2751 * bits in the pages comprising the buffer. The range is limited 2752 * to the size of the buffer. 2753 * 2754 * Tell the VM system that the pages associated with this buffer 2755 * are clean. This is used for delayed writes where the data is 2756 * going to go to disk eventually without additional VM intevention. 2757 * 2758 * Note that while we only really need to clean through to b_bcount, we 2759 * just go ahead and clean through to b_bufsize. 2760 */ 2761static void 2762vfs_clean_pages_dirty_buf(struct buf *bp) 2763{ 2764 vm_ooffset_t foff, noff, eoff; 2765 vm_page_t m; 2766 int i; 2767 2768 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2769 return; 2770 2771 foff = bp->b_offset; 2772 KASSERT(bp->b_offset != NOOFFSET, 2773 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2774 2775 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2776 vfs_drain_busy_pages(bp); 2777 vfs_setdirty_locked_object(bp); 2778 for (i = 0; i < bp->b_npages; i++) { 2779 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2780 eoff = noff; 2781 if (eoff > bp->b_offset + bp->b_bufsize) 2782 eoff = bp->b_offset + bp->b_bufsize; 2783 m = bp->b_pages[i]; 2784 vfs_page_set_validclean(bp, foff, m); 2785 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2786 foff = noff; 2787 } 2788 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2789} 2790 2791static void 2792vfs_setdirty_locked_object(struct buf *bp) 2793{ 2794 vm_object_t object; 2795 int i; 2796 2797 object = bp->b_bufobj->bo_object; 2798 VM_OBJECT_ASSERT_WLOCKED(object); 2799 2800 /* 2801 * We qualify the scan for modified pages on whether the 2802 * object has been flushed yet. 2803 */ 2804 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2805 vm_offset_t boffset; 2806 vm_offset_t eoffset; 2807 2808 /* 2809 * test the pages to see if they have been modified directly 2810 * by users through the VM system. 2811 */ 2812 for (i = 0; i < bp->b_npages; i++) 2813 vm_page_test_dirty(bp->b_pages[i]); 2814 2815 /* 2816 * Calculate the encompassing dirty range, boffset and eoffset, 2817 * (eoffset - boffset) bytes. 2818 */ 2819 2820 for (i = 0; i < bp->b_npages; i++) { 2821 if (bp->b_pages[i]->dirty) 2822 break; 2823 } 2824 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2825 2826 for (i = bp->b_npages - 1; i >= 0; --i) { 2827 if (bp->b_pages[i]->dirty) { 2828 break; 2829 } 2830 } 2831 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2832 2833 /* 2834 * Fit it to the buffer. 2835 */ 2836 2837 if (eoffset > bp->b_bcount) 2838 eoffset = bp->b_bcount; 2839 2840 /* 2841 * If we have a good dirty range, merge with the existing 2842 * dirty range. 2843 */ 2844 2845 if (boffset < eoffset) { 2846 if (bp->b_dirtyoff > boffset) 2847 bp->b_dirtyoff = boffset; 2848 if (bp->b_dirtyend < eoffset) 2849 bp->b_dirtyend = eoffset; 2850 } 2851 } 2852} 2853 2854/* 2855 * Allocate the KVA mapping for an existing buffer. It handles the 2856 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2857 * KVA which is not mapped (B_KVAALLOC). 2858 */ 2859static void 2860bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2861{ 2862 struct buf *scratch_bp; 2863 int bsize, maxsize, need_mapping, need_kva; 2864 off_t offset; 2865 2866 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2867 (gbflags & GB_UNMAPPED) == 0; 2868 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2869 (gbflags & GB_KVAALLOC) != 0; 2870 if (!need_mapping && !need_kva) 2871 return; 2872 2873 BUF_CHECK_UNMAPPED(bp); 2874 2875 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2876 /* 2877 * Buffer is not mapped, but the KVA was already 2878 * reserved at the time of the instantiation. Use the 2879 * allocated space. 2880 */ 2881 bp->b_flags &= ~B_KVAALLOC; 2882 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2883 bp->b_kvabase = bp->b_kvaalloc; 2884 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2885 goto has_addr; 2886 } 2887 2888 /* 2889 * Calculate the amount of the address space we would reserve 2890 * if the buffer was mapped. 2891 */ 2892 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2893 offset = blkno * bsize; 2894 maxsize = size + (offset & PAGE_MASK); 2895 maxsize = imax(maxsize, bsize); 2896 2897mapping_loop: 2898 if (allocbufkva(bp, maxsize, gbflags)) { 2899 /* 2900 * Request defragmentation. getnewbuf() returns us the 2901 * allocated space by the scratch buffer KVA. 2902 */ 2903 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2904 (GB_UNMAPPED | GB_KVAALLOC)); 2905 if (scratch_bp == NULL) { 2906 if ((gbflags & GB_NOWAIT_BD) != 0) { 2907 /* 2908 * XXXKIB: defragmentation cannot 2909 * succeed, not sure what else to do. 2910 */ 2911 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 2912 } 2913 atomic_add_int(&mappingrestarts, 1); 2914 goto mapping_loop; 2915 } 2916 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 2917 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 2918 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 2919 scratch_bp->b_kvasize, gbflags); 2920 2921 /* Get rid of the scratch buffer. */ 2922 scratch_bp->b_kvasize = 0; 2923 scratch_bp->b_flags |= B_INVAL; 2924 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 2925 brelse(scratch_bp); 2926 } 2927 if (!need_mapping) 2928 return; 2929 2930has_addr: 2931 bp->b_saveaddr = bp->b_kvabase; 2932 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 2933 bp->b_flags &= ~B_UNMAPPED; 2934 BUF_CHECK_MAPPED(bp); 2935 bpmap_qenter(bp); 2936} 2937 2938/* 2939 * getblk: 2940 * 2941 * Get a block given a specified block and offset into a file/device. 2942 * The buffers B_DONE bit will be cleared on return, making it almost 2943 * ready for an I/O initiation. B_INVAL may or may not be set on 2944 * return. The caller should clear B_INVAL prior to initiating a 2945 * READ. 2946 * 2947 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2948 * an existing buffer. 2949 * 2950 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2951 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2952 * and then cleared based on the backing VM. If the previous buffer is 2953 * non-0-sized but invalid, B_CACHE will be cleared. 2954 * 2955 * If getblk() must create a new buffer, the new buffer is returned with 2956 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2957 * case it is returned with B_INVAL clear and B_CACHE set based on the 2958 * backing VM. 2959 * 2960 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2961 * B_CACHE bit is clear. 2962 * 2963 * What this means, basically, is that the caller should use B_CACHE to 2964 * determine whether the buffer is fully valid or not and should clear 2965 * B_INVAL prior to issuing a read. If the caller intends to validate 2966 * the buffer by loading its data area with something, the caller needs 2967 * to clear B_INVAL. If the caller does this without issuing an I/O, 2968 * the caller should set B_CACHE ( as an optimization ), else the caller 2969 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2970 * a write attempt or if it was a successfull read. If the caller 2971 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2972 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2973 */ 2974struct buf * 2975getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2976 int flags) 2977{ 2978 struct buf *bp; 2979 struct bufobj *bo; 2980 int bsize, error, maxsize, vmio; 2981 off_t offset; 2982 2983 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2984 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2985 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2986 ASSERT_VOP_LOCKED(vp, "getblk"); 2987 if (size > MAXBSIZE) 2988 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2989 if (!unmapped_buf_allowed) 2990 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2991 2992 bo = &vp->v_bufobj; 2993loop: 2994 /* 2995 * Block if we are low on buffers. Certain processes are allowed 2996 * to completely exhaust the buffer cache. 2997 * 2998 * If this check ever becomes a bottleneck it may be better to 2999 * move it into the else, when gbincore() fails. At the moment 3000 * it isn't a problem. 3001 */ 3002 if (numfreebuffers == 0) { 3003 if (TD_IS_IDLETHREAD(curthread)) 3004 return NULL; 3005 mtx_lock(&nblock); 3006 needsbuffer |= VFS_BIO_NEED_ANY; 3007 mtx_unlock(&nblock); 3008 } 3009 3010 BO_RLOCK(bo); 3011 bp = gbincore(bo, blkno); 3012 if (bp != NULL) { 3013 int lockflags; 3014 /* 3015 * Buffer is in-core. If the buffer is not busy nor managed, 3016 * it must be on a queue. 3017 */ 3018 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3019 3020 if (flags & GB_LOCK_NOWAIT) 3021 lockflags |= LK_NOWAIT; 3022 3023 error = BUF_TIMELOCK(bp, lockflags, 3024 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3025 3026 /* 3027 * If we slept and got the lock we have to restart in case 3028 * the buffer changed identities. 3029 */ 3030 if (error == ENOLCK) 3031 goto loop; 3032 /* We timed out or were interrupted. */ 3033 else if (error) 3034 return (NULL); 3035 /* If recursed, assume caller knows the rules. */ 3036 else if (BUF_LOCKRECURSED(bp)) 3037 goto end; 3038 3039 /* 3040 * The buffer is locked. B_CACHE is cleared if the buffer is 3041 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3042 * and for a VMIO buffer B_CACHE is adjusted according to the 3043 * backing VM cache. 3044 */ 3045 if (bp->b_flags & B_INVAL) 3046 bp->b_flags &= ~B_CACHE; 3047 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3048 bp->b_flags |= B_CACHE; 3049 if (bp->b_flags & B_MANAGED) 3050 MPASS(bp->b_qindex == QUEUE_NONE); 3051 else 3052 bremfree(bp); 3053 3054 /* 3055 * check for size inconsistencies for non-VMIO case. 3056 */ 3057 if (bp->b_bcount != size) { 3058 if ((bp->b_flags & B_VMIO) == 0 || 3059 (size > bp->b_kvasize)) { 3060 if (bp->b_flags & B_DELWRI) { 3061 /* 3062 * If buffer is pinned and caller does 3063 * not want sleep waiting for it to be 3064 * unpinned, bail out 3065 * */ 3066 if (bp->b_pin_count > 0) { 3067 if (flags & GB_LOCK_NOWAIT) { 3068 bqrelse(bp); 3069 return (NULL); 3070 } else { 3071 bunpin_wait(bp); 3072 } 3073 } 3074 bp->b_flags |= B_NOCACHE; 3075 bwrite(bp); 3076 } else { 3077 if (LIST_EMPTY(&bp->b_dep)) { 3078 bp->b_flags |= B_RELBUF; 3079 brelse(bp); 3080 } else { 3081 bp->b_flags |= B_NOCACHE; 3082 bwrite(bp); 3083 } 3084 } 3085 goto loop; 3086 } 3087 } 3088 3089 /* 3090 * Handle the case of unmapped buffer which should 3091 * become mapped, or the buffer for which KVA 3092 * reservation is requested. 3093 */ 3094 bp_unmapped_get_kva(bp, blkno, size, flags); 3095 3096 /* 3097 * If the size is inconsistant in the VMIO case, we can resize 3098 * the buffer. This might lead to B_CACHE getting set or 3099 * cleared. If the size has not changed, B_CACHE remains 3100 * unchanged from its previous state. 3101 */ 3102 if (bp->b_bcount != size) 3103 allocbuf(bp, size); 3104 3105 KASSERT(bp->b_offset != NOOFFSET, 3106 ("getblk: no buffer offset")); 3107 3108 /* 3109 * A buffer with B_DELWRI set and B_CACHE clear must 3110 * be committed before we can return the buffer in 3111 * order to prevent the caller from issuing a read 3112 * ( due to B_CACHE not being set ) and overwriting 3113 * it. 3114 * 3115 * Most callers, including NFS and FFS, need this to 3116 * operate properly either because they assume they 3117 * can issue a read if B_CACHE is not set, or because 3118 * ( for example ) an uncached B_DELWRI might loop due 3119 * to softupdates re-dirtying the buffer. In the latter 3120 * case, B_CACHE is set after the first write completes, 3121 * preventing further loops. 3122 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3123 * above while extending the buffer, we cannot allow the 3124 * buffer to remain with B_CACHE set after the write 3125 * completes or it will represent a corrupt state. To 3126 * deal with this we set B_NOCACHE to scrap the buffer 3127 * after the write. 3128 * 3129 * We might be able to do something fancy, like setting 3130 * B_CACHE in bwrite() except if B_DELWRI is already set, 3131 * so the below call doesn't set B_CACHE, but that gets real 3132 * confusing. This is much easier. 3133 */ 3134 3135 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3136 bp->b_flags |= B_NOCACHE; 3137 bwrite(bp); 3138 goto loop; 3139 } 3140 bp->b_flags &= ~B_DONE; 3141 } else { 3142 /* 3143 * Buffer is not in-core, create new buffer. The buffer 3144 * returned by getnewbuf() is locked. Note that the returned 3145 * buffer is also considered valid (not marked B_INVAL). 3146 */ 3147 BO_RUNLOCK(bo); 3148 /* 3149 * If the user does not want us to create the buffer, bail out 3150 * here. 3151 */ 3152 if (flags & GB_NOCREAT) 3153 return NULL; 3154 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3155 offset = blkno * bsize; 3156 vmio = vp->v_object != NULL; 3157 if (vmio) { 3158 maxsize = size + (offset & PAGE_MASK); 3159 } else { 3160 maxsize = size; 3161 /* Do not allow non-VMIO notmapped buffers. */ 3162 flags &= ~GB_UNMAPPED; 3163 } 3164 maxsize = imax(maxsize, bsize); 3165 3166 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3167 if (bp == NULL) { 3168 if (slpflag || slptimeo) 3169 return NULL; 3170 goto loop; 3171 } 3172 3173 /* 3174 * This code is used to make sure that a buffer is not 3175 * created while the getnewbuf routine is blocked. 3176 * This can be a problem whether the vnode is locked or not. 3177 * If the buffer is created out from under us, we have to 3178 * throw away the one we just created. 3179 * 3180 * Note: this must occur before we associate the buffer 3181 * with the vp especially considering limitations in 3182 * the splay tree implementation when dealing with duplicate 3183 * lblkno's. 3184 */ 3185 BO_LOCK(bo); 3186 if (gbincore(bo, blkno)) { 3187 BO_UNLOCK(bo); 3188 bp->b_flags |= B_INVAL; 3189 brelse(bp); 3190 goto loop; 3191 } 3192 3193 /* 3194 * Insert the buffer into the hash, so that it can 3195 * be found by incore. 3196 */ 3197 bp->b_blkno = bp->b_lblkno = blkno; 3198 bp->b_offset = offset; 3199 bgetvp(vp, bp); 3200 BO_UNLOCK(bo); 3201 3202 /* 3203 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3204 * buffer size starts out as 0, B_CACHE will be set by 3205 * allocbuf() for the VMIO case prior to it testing the 3206 * backing store for validity. 3207 */ 3208 3209 if (vmio) { 3210 bp->b_flags |= B_VMIO; 3211 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3212 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3213 bp, vp->v_object, bp->b_bufobj->bo_object)); 3214 } else { 3215 bp->b_flags &= ~B_VMIO; 3216 KASSERT(bp->b_bufobj->bo_object == NULL, 3217 ("ARGH! has b_bufobj->bo_object %p %p\n", 3218 bp, bp->b_bufobj->bo_object)); 3219 BUF_CHECK_MAPPED(bp); 3220 } 3221 3222 allocbuf(bp, size); 3223 bp->b_flags &= ~B_DONE; 3224 } 3225 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3226 BUF_ASSERT_HELD(bp); 3227end: 3228 KASSERT(bp->b_bufobj == bo, 3229 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3230 return (bp); 3231} 3232 3233/* 3234 * Get an empty, disassociated buffer of given size. The buffer is initially 3235 * set to B_INVAL. 3236 */ 3237struct buf * 3238geteblk(int size, int flags) 3239{ 3240 struct buf *bp; 3241 int maxsize; 3242 3243 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3244 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3245 if ((flags & GB_NOWAIT_BD) && 3246 (curthread->td_pflags & TDP_BUFNEED) != 0) 3247 return (NULL); 3248 } 3249 allocbuf(bp, size); 3250 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3251 BUF_ASSERT_HELD(bp); 3252 return (bp); 3253} 3254 3255 3256/* 3257 * This code constitutes the buffer memory from either anonymous system 3258 * memory (in the case of non-VMIO operations) or from an associated 3259 * VM object (in the case of VMIO operations). This code is able to 3260 * resize a buffer up or down. 3261 * 3262 * Note that this code is tricky, and has many complications to resolve 3263 * deadlock or inconsistant data situations. Tread lightly!!! 3264 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3265 * the caller. Calling this code willy nilly can result in the loss of data. 3266 * 3267 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3268 * B_CACHE for the non-VMIO case. 3269 */ 3270 3271int 3272allocbuf(struct buf *bp, int size) 3273{ 3274 int newbsize, mbsize; 3275 int i; 3276 3277 BUF_ASSERT_HELD(bp); 3278 3279 if (bp->b_kvasize < size) 3280 panic("allocbuf: buffer too small"); 3281 3282 if ((bp->b_flags & B_VMIO) == 0) { 3283 caddr_t origbuf; 3284 int origbufsize; 3285 /* 3286 * Just get anonymous memory from the kernel. Don't 3287 * mess with B_CACHE. 3288 */ 3289 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3290 if (bp->b_flags & B_MALLOC) 3291 newbsize = mbsize; 3292 else 3293 newbsize = round_page(size); 3294 3295 if (newbsize < bp->b_bufsize) { 3296 /* 3297 * malloced buffers are not shrunk 3298 */ 3299 if (bp->b_flags & B_MALLOC) { 3300 if (newbsize) { 3301 bp->b_bcount = size; 3302 } else { 3303 free(bp->b_data, M_BIOBUF); 3304 if (bp->b_bufsize) { 3305 atomic_subtract_long( 3306 &bufmallocspace, 3307 bp->b_bufsize); 3308 bufspacewakeup(); 3309 bp->b_bufsize = 0; 3310 } 3311 bp->b_saveaddr = bp->b_kvabase; 3312 bp->b_data = bp->b_saveaddr; 3313 bp->b_bcount = 0; 3314 bp->b_flags &= ~B_MALLOC; 3315 } 3316 return 1; 3317 } 3318 vm_hold_free_pages(bp, newbsize); 3319 } else if (newbsize > bp->b_bufsize) { 3320 /* 3321 * We only use malloced memory on the first allocation. 3322 * and revert to page-allocated memory when the buffer 3323 * grows. 3324 */ 3325 /* 3326 * There is a potential smp race here that could lead 3327 * to bufmallocspace slightly passing the max. It 3328 * is probably extremely rare and not worth worrying 3329 * over. 3330 */ 3331 if ( (bufmallocspace < maxbufmallocspace) && 3332 (bp->b_bufsize == 0) && 3333 (mbsize <= PAGE_SIZE/2)) { 3334 3335 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3336 bp->b_bufsize = mbsize; 3337 bp->b_bcount = size; 3338 bp->b_flags |= B_MALLOC; 3339 atomic_add_long(&bufmallocspace, mbsize); 3340 return 1; 3341 } 3342 origbuf = NULL; 3343 origbufsize = 0; 3344 /* 3345 * If the buffer is growing on its other-than-first allocation, 3346 * then we revert to the page-allocation scheme. 3347 */ 3348 if (bp->b_flags & B_MALLOC) { 3349 origbuf = bp->b_data; 3350 origbufsize = bp->b_bufsize; 3351 bp->b_data = bp->b_kvabase; 3352 if (bp->b_bufsize) { 3353 atomic_subtract_long(&bufmallocspace, 3354 bp->b_bufsize); 3355 bufspacewakeup(); 3356 bp->b_bufsize = 0; 3357 } 3358 bp->b_flags &= ~B_MALLOC; 3359 newbsize = round_page(newbsize); 3360 } 3361 vm_hold_load_pages( 3362 bp, 3363 (vm_offset_t) bp->b_data + bp->b_bufsize, 3364 (vm_offset_t) bp->b_data + newbsize); 3365 if (origbuf) { 3366 bcopy(origbuf, bp->b_data, origbufsize); 3367 free(origbuf, M_BIOBUF); 3368 } 3369 } 3370 } else { 3371 int desiredpages; 3372 3373 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3374 desiredpages = (size == 0) ? 0 : 3375 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3376 3377 if (bp->b_flags & B_MALLOC) 3378 panic("allocbuf: VMIO buffer can't be malloced"); 3379 /* 3380 * Set B_CACHE initially if buffer is 0 length or will become 3381 * 0-length. 3382 */ 3383 if (size == 0 || bp->b_bufsize == 0) 3384 bp->b_flags |= B_CACHE; 3385 3386 if (newbsize < bp->b_bufsize) { 3387 /* 3388 * DEV_BSIZE aligned new buffer size is less then the 3389 * DEV_BSIZE aligned existing buffer size. Figure out 3390 * if we have to remove any pages. 3391 */ 3392 if (desiredpages < bp->b_npages) { 3393 vm_page_t m; 3394 3395 if ((bp->b_flags & B_UNMAPPED) == 0) { 3396 BUF_CHECK_MAPPED(bp); 3397 pmap_qremove((vm_offset_t)trunc_page( 3398 (vm_offset_t)bp->b_data) + 3399 (desiredpages << PAGE_SHIFT), 3400 (bp->b_npages - desiredpages)); 3401 } else 3402 BUF_CHECK_UNMAPPED(bp); 3403 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3404 for (i = desiredpages; i < bp->b_npages; i++) { 3405 /* 3406 * the page is not freed here -- it 3407 * is the responsibility of 3408 * vnode_pager_setsize 3409 */ 3410 m = bp->b_pages[i]; 3411 KASSERT(m != bogus_page, 3412 ("allocbuf: bogus page found")); 3413 while (vm_page_sleep_if_busy(m, TRUE, 3414 "biodep")) 3415 continue; 3416 3417 bp->b_pages[i] = NULL; 3418 vm_page_lock(m); 3419 vm_page_unwire(m, 0); 3420 vm_page_unlock(m); 3421 } 3422 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3423 bp->b_npages = desiredpages; 3424 } 3425 } else if (size > bp->b_bcount) { 3426 /* 3427 * We are growing the buffer, possibly in a 3428 * byte-granular fashion. 3429 */ 3430 vm_object_t obj; 3431 vm_offset_t toff; 3432 vm_offset_t tinc; 3433 3434 /* 3435 * Step 1, bring in the VM pages from the object, 3436 * allocating them if necessary. We must clear 3437 * B_CACHE if these pages are not valid for the 3438 * range covered by the buffer. 3439 */ 3440 3441 obj = bp->b_bufobj->bo_object; 3442 3443 VM_OBJECT_WLOCK(obj); 3444 while (bp->b_npages < desiredpages) { 3445 vm_page_t m; 3446 3447 /* 3448 * We must allocate system pages since blocking 3449 * here could interfere with paging I/O, no 3450 * matter which process we are. 3451 * 3452 * We can only test VPO_BUSY here. Blocking on 3453 * m->busy might lead to a deadlock: 3454 * vm_fault->getpages->cluster_read->allocbuf 3455 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3456 */ 3457 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3458 bp->b_npages, VM_ALLOC_NOBUSY | 3459 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3460 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3461 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3462 if (m->valid == 0) 3463 bp->b_flags &= ~B_CACHE; 3464 bp->b_pages[bp->b_npages] = m; 3465 ++bp->b_npages; 3466 } 3467 3468 /* 3469 * Step 2. We've loaded the pages into the buffer, 3470 * we have to figure out if we can still have B_CACHE 3471 * set. Note that B_CACHE is set according to the 3472 * byte-granular range ( bcount and size ), new the 3473 * aligned range ( newbsize ). 3474 * 3475 * The VM test is against m->valid, which is DEV_BSIZE 3476 * aligned. Needless to say, the validity of the data 3477 * needs to also be DEV_BSIZE aligned. Note that this 3478 * fails with NFS if the server or some other client 3479 * extends the file's EOF. If our buffer is resized, 3480 * B_CACHE may remain set! XXX 3481 */ 3482 3483 toff = bp->b_bcount; 3484 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3485 3486 while ((bp->b_flags & B_CACHE) && toff < size) { 3487 vm_pindex_t pi; 3488 3489 if (tinc > (size - toff)) 3490 tinc = size - toff; 3491 3492 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3493 PAGE_SHIFT; 3494 3495 vfs_buf_test_cache( 3496 bp, 3497 bp->b_offset, 3498 toff, 3499 tinc, 3500 bp->b_pages[pi] 3501 ); 3502 toff += tinc; 3503 tinc = PAGE_SIZE; 3504 } 3505 VM_OBJECT_WUNLOCK(obj); 3506 3507 /* 3508 * Step 3, fixup the KVM pmap. 3509 */ 3510 if ((bp->b_flags & B_UNMAPPED) == 0) 3511 bpmap_qenter(bp); 3512 else 3513 BUF_CHECK_UNMAPPED(bp); 3514 } 3515 } 3516 if (newbsize < bp->b_bufsize) 3517 bufspacewakeup(); 3518 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3519 bp->b_bcount = size; /* requested buffer size */ 3520 return 1; 3521} 3522 3523extern int inflight_transient_maps; 3524 3525void 3526biodone(struct bio *bp) 3527{ 3528 struct mtx *mtxp; 3529 void (*done)(struct bio *); 3530 vm_offset_t start, end; 3531 int transient; 3532 3533 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3534 mtx_lock(mtxp); 3535 bp->bio_flags |= BIO_DONE; 3536 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3537 start = trunc_page((vm_offset_t)bp->bio_data); 3538 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3539 transient = 1; 3540 } else { 3541 transient = 0; 3542 start = end = 0; 3543 } 3544 done = bp->bio_done; 3545 if (done == NULL) 3546 wakeup(bp); 3547 mtx_unlock(mtxp); 3548 if (done != NULL) 3549 done(bp); 3550 if (transient) { 3551 pmap_qremove(start, OFF_TO_IDX(end - start)); 3552 vm_map_remove(bio_transient_map, start, end); 3553 atomic_add_int(&inflight_transient_maps, -1); 3554 } 3555} 3556 3557/* 3558 * Wait for a BIO to finish. 3559 * 3560 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3561 * case is not yet clear. 3562 */ 3563int 3564biowait(struct bio *bp, const char *wchan) 3565{ 3566 struct mtx *mtxp; 3567 3568 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3569 mtx_lock(mtxp); 3570 while ((bp->bio_flags & BIO_DONE) == 0) 3571 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3572 mtx_unlock(mtxp); 3573 if (bp->bio_error != 0) 3574 return (bp->bio_error); 3575 if (!(bp->bio_flags & BIO_ERROR)) 3576 return (0); 3577 return (EIO); 3578} 3579 3580void 3581biofinish(struct bio *bp, struct devstat *stat, int error) 3582{ 3583 3584 if (error) { 3585 bp->bio_error = error; 3586 bp->bio_flags |= BIO_ERROR; 3587 } 3588 if (stat != NULL) 3589 devstat_end_transaction_bio(stat, bp); 3590 biodone(bp); 3591} 3592 3593/* 3594 * bufwait: 3595 * 3596 * Wait for buffer I/O completion, returning error status. The buffer 3597 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3598 * error and cleared. 3599 */ 3600int 3601bufwait(struct buf *bp) 3602{ 3603 if (bp->b_iocmd == BIO_READ) 3604 bwait(bp, PRIBIO, "biord"); 3605 else 3606 bwait(bp, PRIBIO, "biowr"); 3607 if (bp->b_flags & B_EINTR) { 3608 bp->b_flags &= ~B_EINTR; 3609 return (EINTR); 3610 } 3611 if (bp->b_ioflags & BIO_ERROR) { 3612 return (bp->b_error ? bp->b_error : EIO); 3613 } else { 3614 return (0); 3615 } 3616} 3617 3618 /* 3619 * Call back function from struct bio back up to struct buf. 3620 */ 3621static void 3622bufdonebio(struct bio *bip) 3623{ 3624 struct buf *bp; 3625 3626 bp = bip->bio_caller2; 3627 bp->b_resid = bp->b_bcount - bip->bio_completed; 3628 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3629 bp->b_ioflags = bip->bio_flags; 3630 bp->b_error = bip->bio_error; 3631 if (bp->b_error) 3632 bp->b_ioflags |= BIO_ERROR; 3633 bufdone(bp); 3634 g_destroy_bio(bip); 3635} 3636 3637void 3638dev_strategy(struct cdev *dev, struct buf *bp) 3639{ 3640 struct cdevsw *csw; 3641 int ref; 3642 3643 KASSERT(dev->si_refcount > 0, 3644 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3645 devtoname(dev), dev)); 3646 3647 csw = dev_refthread(dev, &ref); 3648 dev_strategy_csw(dev, csw, bp); 3649 dev_relthread(dev, ref); 3650} 3651 3652void 3653dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3654{ 3655 struct bio *bip; 3656 3657 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3658 ("b_iocmd botch")); 3659 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3660 dev->si_threadcount > 0, 3661 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3662 dev)); 3663 if (csw == NULL) { 3664 bp->b_error = ENXIO; 3665 bp->b_ioflags = BIO_ERROR; 3666 bufdone(bp); 3667 return; 3668 } 3669 for (;;) { 3670 bip = g_new_bio(); 3671 if (bip != NULL) 3672 break; 3673 /* Try again later */ 3674 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3675 } 3676 bip->bio_cmd = bp->b_iocmd; 3677 bip->bio_offset = bp->b_iooffset; 3678 bip->bio_length = bp->b_bcount; 3679 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3680 bdata2bio(bp, bip); 3681 bip->bio_done = bufdonebio; 3682 bip->bio_caller2 = bp; 3683 bip->bio_dev = dev; 3684 (*csw->d_strategy)(bip); 3685} 3686 3687/* 3688 * bufdone: 3689 * 3690 * Finish I/O on a buffer, optionally calling a completion function. 3691 * This is usually called from an interrupt so process blocking is 3692 * not allowed. 3693 * 3694 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3695 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3696 * assuming B_INVAL is clear. 3697 * 3698 * For the VMIO case, we set B_CACHE if the op was a read and no 3699 * read error occured, or if the op was a write. B_CACHE is never 3700 * set if the buffer is invalid or otherwise uncacheable. 3701 * 3702 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3703 * initiator to leave B_INVAL set to brelse the buffer out of existance 3704 * in the biodone routine. 3705 */ 3706void 3707bufdone(struct buf *bp) 3708{ 3709 struct bufobj *dropobj; 3710 void (*biodone)(struct buf *); 3711 3712 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3713 dropobj = NULL; 3714 3715 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3716 BUF_ASSERT_HELD(bp); 3717 3718 runningbufwakeup(bp); 3719 if (bp->b_iocmd == BIO_WRITE) 3720 dropobj = bp->b_bufobj; 3721 /* call optional completion function if requested */ 3722 if (bp->b_iodone != NULL) { 3723 biodone = bp->b_iodone; 3724 bp->b_iodone = NULL; 3725 (*biodone) (bp); 3726 if (dropobj) 3727 bufobj_wdrop(dropobj); 3728 return; 3729 } 3730 3731 bufdone_finish(bp); 3732 3733 if (dropobj) 3734 bufobj_wdrop(dropobj); 3735} 3736 3737void 3738bufdone_finish(struct buf *bp) 3739{ 3740 BUF_ASSERT_HELD(bp); 3741 3742 if (!LIST_EMPTY(&bp->b_dep)) 3743 buf_complete(bp); 3744 3745 if (bp->b_flags & B_VMIO) { 3746 vm_ooffset_t foff; 3747 vm_page_t m; 3748 vm_object_t obj; 3749 struct vnode *vp; 3750 int bogus, i, iosize; 3751 3752 obj = bp->b_bufobj->bo_object; 3753 KASSERT(obj->paging_in_progress >= bp->b_npages, 3754 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3755 obj->paging_in_progress, bp->b_npages)); 3756 3757 vp = bp->b_vp; 3758 KASSERT(vp->v_holdcnt > 0, 3759 ("biodone_finish: vnode %p has zero hold count", vp)); 3760 KASSERT(vp->v_object != NULL, 3761 ("biodone_finish: vnode %p has no vm_object", vp)); 3762 3763 foff = bp->b_offset; 3764 KASSERT(bp->b_offset != NOOFFSET, 3765 ("biodone_finish: bp %p has no buffer offset", bp)); 3766 3767 /* 3768 * Set B_CACHE if the op was a normal read and no error 3769 * occured. B_CACHE is set for writes in the b*write() 3770 * routines. 3771 */ 3772 iosize = bp->b_bcount - bp->b_resid; 3773 if (bp->b_iocmd == BIO_READ && 3774 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3775 !(bp->b_ioflags & BIO_ERROR)) { 3776 bp->b_flags |= B_CACHE; 3777 } 3778 bogus = 0; 3779 VM_OBJECT_WLOCK(obj); 3780 for (i = 0; i < bp->b_npages; i++) { 3781 int bogusflag = 0; 3782 int resid; 3783 3784 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3785 if (resid > iosize) 3786 resid = iosize; 3787 3788 /* 3789 * cleanup bogus pages, restoring the originals 3790 */ 3791 m = bp->b_pages[i]; 3792 if (m == bogus_page) { 3793 bogus = bogusflag = 1; 3794 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3795 if (m == NULL) 3796 panic("biodone: page disappeared!"); 3797 bp->b_pages[i] = m; 3798 } 3799 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3800 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3801 (intmax_t)foff, (uintmax_t)m->pindex)); 3802 3803 /* 3804 * In the write case, the valid and clean bits are 3805 * already changed correctly ( see bdwrite() ), so we 3806 * only need to do this here in the read case. 3807 */ 3808 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3809 KASSERT((m->dirty & vm_page_bits(foff & 3810 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3811 " page %p has unexpected dirty bits", m)); 3812 vfs_page_set_valid(bp, foff, m); 3813 } 3814 3815 vm_page_io_finish(m); 3816 vm_object_pip_subtract(obj, 1); 3817 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3818 iosize -= resid; 3819 } 3820 vm_object_pip_wakeupn(obj, 0); 3821 VM_OBJECT_WUNLOCK(obj); 3822 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3823 BUF_CHECK_MAPPED(bp); 3824 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3825 bp->b_pages, bp->b_npages); 3826 } 3827 } 3828 3829 /* 3830 * For asynchronous completions, release the buffer now. The brelse 3831 * will do a wakeup there if necessary - so no need to do a wakeup 3832 * here in the async case. The sync case always needs to do a wakeup. 3833 */ 3834 3835 if (bp->b_flags & B_ASYNC) { 3836 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3837 brelse(bp); 3838 else 3839 bqrelse(bp); 3840 } else 3841 bdone(bp); 3842} 3843 3844/* 3845 * This routine is called in lieu of iodone in the case of 3846 * incomplete I/O. This keeps the busy status for pages 3847 * consistant. 3848 */ 3849void 3850vfs_unbusy_pages(struct buf *bp) 3851{ 3852 int i; 3853 vm_object_t obj; 3854 vm_page_t m; 3855 3856 runningbufwakeup(bp); 3857 if (!(bp->b_flags & B_VMIO)) 3858 return; 3859 3860 obj = bp->b_bufobj->bo_object; 3861 VM_OBJECT_WLOCK(obj); 3862 for (i = 0; i < bp->b_npages; i++) { 3863 m = bp->b_pages[i]; 3864 if (m == bogus_page) { 3865 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3866 if (!m) 3867 panic("vfs_unbusy_pages: page missing\n"); 3868 bp->b_pages[i] = m; 3869 if ((bp->b_flags & B_UNMAPPED) == 0) { 3870 BUF_CHECK_MAPPED(bp); 3871 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3872 bp->b_pages, bp->b_npages); 3873 } else 3874 BUF_CHECK_UNMAPPED(bp); 3875 } 3876 vm_object_pip_subtract(obj, 1); 3877 vm_page_io_finish(m); 3878 } 3879 vm_object_pip_wakeupn(obj, 0); 3880 VM_OBJECT_WUNLOCK(obj); 3881} 3882 3883/* 3884 * vfs_page_set_valid: 3885 * 3886 * Set the valid bits in a page based on the supplied offset. The 3887 * range is restricted to the buffer's size. 3888 * 3889 * This routine is typically called after a read completes. 3890 */ 3891static void 3892vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3893{ 3894 vm_ooffset_t eoff; 3895 3896 /* 3897 * Compute the end offset, eoff, such that [off, eoff) does not span a 3898 * page boundary and eoff is not greater than the end of the buffer. 3899 * The end of the buffer, in this case, is our file EOF, not the 3900 * allocation size of the buffer. 3901 */ 3902 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3903 if (eoff > bp->b_offset + bp->b_bcount) 3904 eoff = bp->b_offset + bp->b_bcount; 3905 3906 /* 3907 * Set valid range. This is typically the entire buffer and thus the 3908 * entire page. 3909 */ 3910 if (eoff > off) 3911 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3912} 3913 3914/* 3915 * vfs_page_set_validclean: 3916 * 3917 * Set the valid bits and clear the dirty bits in a page based on the 3918 * supplied offset. The range is restricted to the buffer's size. 3919 */ 3920static void 3921vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3922{ 3923 vm_ooffset_t soff, eoff; 3924 3925 /* 3926 * Start and end offsets in buffer. eoff - soff may not cross a 3927 * page boundry or cross the end of the buffer. The end of the 3928 * buffer, in this case, is our file EOF, not the allocation size 3929 * of the buffer. 3930 */ 3931 soff = off; 3932 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3933 if (eoff > bp->b_offset + bp->b_bcount) 3934 eoff = bp->b_offset + bp->b_bcount; 3935 3936 /* 3937 * Set valid range. This is typically the entire buffer and thus the 3938 * entire page. 3939 */ 3940 if (eoff > soff) { 3941 vm_page_set_validclean( 3942 m, 3943 (vm_offset_t) (soff & PAGE_MASK), 3944 (vm_offset_t) (eoff - soff) 3945 ); 3946 } 3947} 3948 3949/* 3950 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3951 * any page is busy, drain the flag. 3952 */ 3953static void 3954vfs_drain_busy_pages(struct buf *bp) 3955{ 3956 vm_page_t m; 3957 int i, last_busied; 3958 3959 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3960 last_busied = 0; 3961 for (i = 0; i < bp->b_npages; i++) { 3962 m = bp->b_pages[i]; 3963 if ((m->oflags & VPO_BUSY) != 0) { 3964 for (; last_busied < i; last_busied++) 3965 vm_page_busy(bp->b_pages[last_busied]); 3966 while ((m->oflags & VPO_BUSY) != 0) 3967 vm_page_sleep(m, "vbpage"); 3968 } 3969 } 3970 for (i = 0; i < last_busied; i++) 3971 vm_page_wakeup(bp->b_pages[i]); 3972} 3973 3974/* 3975 * This routine is called before a device strategy routine. 3976 * It is used to tell the VM system that paging I/O is in 3977 * progress, and treat the pages associated with the buffer 3978 * almost as being VPO_BUSY. Also the object paging_in_progress 3979 * flag is handled to make sure that the object doesn't become 3980 * inconsistant. 3981 * 3982 * Since I/O has not been initiated yet, certain buffer flags 3983 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3984 * and should be ignored. 3985 */ 3986void 3987vfs_busy_pages(struct buf *bp, int clear_modify) 3988{ 3989 int i, bogus; 3990 vm_object_t obj; 3991 vm_ooffset_t foff; 3992 vm_page_t m; 3993 3994 if (!(bp->b_flags & B_VMIO)) 3995 return; 3996 3997 obj = bp->b_bufobj->bo_object; 3998 foff = bp->b_offset; 3999 KASSERT(bp->b_offset != NOOFFSET, 4000 ("vfs_busy_pages: no buffer offset")); 4001 VM_OBJECT_WLOCK(obj); 4002 vfs_drain_busy_pages(bp); 4003 if (bp->b_bufsize != 0) 4004 vfs_setdirty_locked_object(bp); 4005 bogus = 0; 4006 for (i = 0; i < bp->b_npages; i++) { 4007 m = bp->b_pages[i]; 4008 4009 if ((bp->b_flags & B_CLUSTER) == 0) { 4010 vm_object_pip_add(obj, 1); 4011 vm_page_io_start(m); 4012 } 4013 /* 4014 * When readying a buffer for a read ( i.e 4015 * clear_modify == 0 ), it is important to do 4016 * bogus_page replacement for valid pages in 4017 * partially instantiated buffers. Partially 4018 * instantiated buffers can, in turn, occur when 4019 * reconstituting a buffer from its VM backing store 4020 * base. We only have to do this if B_CACHE is 4021 * clear ( which causes the I/O to occur in the 4022 * first place ). The replacement prevents the read 4023 * I/O from overwriting potentially dirty VM-backed 4024 * pages. XXX bogus page replacement is, uh, bogus. 4025 * It may not work properly with small-block devices. 4026 * We need to find a better way. 4027 */ 4028 if (clear_modify) { 4029 pmap_remove_write(m); 4030 vfs_page_set_validclean(bp, foff, m); 4031 } else if (m->valid == VM_PAGE_BITS_ALL && 4032 (bp->b_flags & B_CACHE) == 0) { 4033 bp->b_pages[i] = bogus_page; 4034 bogus++; 4035 } 4036 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4037 } 4038 VM_OBJECT_WUNLOCK(obj); 4039 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4040 BUF_CHECK_MAPPED(bp); 4041 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4042 bp->b_pages, bp->b_npages); 4043 } 4044} 4045 4046/* 4047 * vfs_bio_set_valid: 4048 * 4049 * Set the range within the buffer to valid. The range is 4050 * relative to the beginning of the buffer, b_offset. Note that 4051 * b_offset itself may be offset from the beginning of the first 4052 * page. 4053 */ 4054void 4055vfs_bio_set_valid(struct buf *bp, int base, int size) 4056{ 4057 int i, n; 4058 vm_page_t m; 4059 4060 if (!(bp->b_flags & B_VMIO)) 4061 return; 4062 4063 /* 4064 * Fixup base to be relative to beginning of first page. 4065 * Set initial n to be the maximum number of bytes in the 4066 * first page that can be validated. 4067 */ 4068 base += (bp->b_offset & PAGE_MASK); 4069 n = PAGE_SIZE - (base & PAGE_MASK); 4070 4071 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4072 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4073 m = bp->b_pages[i]; 4074 if (n > size) 4075 n = size; 4076 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4077 base += n; 4078 size -= n; 4079 n = PAGE_SIZE; 4080 } 4081 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4082} 4083 4084/* 4085 * vfs_bio_clrbuf: 4086 * 4087 * If the specified buffer is a non-VMIO buffer, clear the entire 4088 * buffer. If the specified buffer is a VMIO buffer, clear and 4089 * validate only the previously invalid portions of the buffer. 4090 * This routine essentially fakes an I/O, so we need to clear 4091 * BIO_ERROR and B_INVAL. 4092 * 4093 * Note that while we only theoretically need to clear through b_bcount, 4094 * we go ahead and clear through b_bufsize. 4095 */ 4096void 4097vfs_bio_clrbuf(struct buf *bp) 4098{ 4099 int i, j, mask, sa, ea, slide; 4100 4101 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4102 clrbuf(bp); 4103 return; 4104 } 4105 bp->b_flags &= ~B_INVAL; 4106 bp->b_ioflags &= ~BIO_ERROR; 4107 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4108 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4109 (bp->b_offset & PAGE_MASK) == 0) { 4110 if (bp->b_pages[0] == bogus_page) 4111 goto unlock; 4112 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4113 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4114 if ((bp->b_pages[0]->valid & mask) == mask) 4115 goto unlock; 4116 if ((bp->b_pages[0]->valid & mask) == 0) { 4117 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4118 bp->b_pages[0]->valid |= mask; 4119 goto unlock; 4120 } 4121 } 4122 sa = bp->b_offset & PAGE_MASK; 4123 slide = 0; 4124 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4125 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4126 ea = slide & PAGE_MASK; 4127 if (ea == 0) 4128 ea = PAGE_SIZE; 4129 if (bp->b_pages[i] == bogus_page) 4130 continue; 4131 j = sa / DEV_BSIZE; 4132 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4133 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4134 if ((bp->b_pages[i]->valid & mask) == mask) 4135 continue; 4136 if ((bp->b_pages[i]->valid & mask) == 0) 4137 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4138 else { 4139 for (; sa < ea; sa += DEV_BSIZE, j++) { 4140 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4141 pmap_zero_page_area(bp->b_pages[i], 4142 sa, DEV_BSIZE); 4143 } 4144 } 4145 } 4146 bp->b_pages[i]->valid |= mask; 4147 } 4148unlock: 4149 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4150 bp->b_resid = 0; 4151} 4152 4153void 4154vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4155{ 4156 vm_page_t m; 4157 int i, n; 4158 4159 if ((bp->b_flags & B_UNMAPPED) == 0) { 4160 BUF_CHECK_MAPPED(bp); 4161 bzero(bp->b_data + base, size); 4162 } else { 4163 BUF_CHECK_UNMAPPED(bp); 4164 n = PAGE_SIZE - (base & PAGE_MASK); 4165 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4166 m = bp->b_pages[i]; 4167 if (n > size) 4168 n = size; 4169 pmap_zero_page_area(m, base & PAGE_MASK, n); 4170 base += n; 4171 size -= n; 4172 n = PAGE_SIZE; 4173 } 4174 } 4175} 4176 4177/* 4178 * vm_hold_load_pages and vm_hold_free_pages get pages into 4179 * a buffers address space. The pages are anonymous and are 4180 * not associated with a file object. 4181 */ 4182static void 4183vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4184{ 4185 vm_offset_t pg; 4186 vm_page_t p; 4187 int index; 4188 4189 BUF_CHECK_MAPPED(bp); 4190 4191 to = round_page(to); 4192 from = round_page(from); 4193 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4194 4195 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4196tryagain: 4197 /* 4198 * note: must allocate system pages since blocking here 4199 * could interfere with paging I/O, no matter which 4200 * process we are. 4201 */ 4202 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4203 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4204 if (p == NULL) { 4205 VM_WAIT; 4206 goto tryagain; 4207 } 4208 pmap_qenter(pg, &p, 1); 4209 bp->b_pages[index] = p; 4210 } 4211 bp->b_npages = index; 4212} 4213 4214/* Return pages associated with this buf to the vm system */ 4215static void 4216vm_hold_free_pages(struct buf *bp, int newbsize) 4217{ 4218 vm_offset_t from; 4219 vm_page_t p; 4220 int index, newnpages; 4221 4222 BUF_CHECK_MAPPED(bp); 4223 4224 from = round_page((vm_offset_t)bp->b_data + newbsize); 4225 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4226 if (bp->b_npages > newnpages) 4227 pmap_qremove(from, bp->b_npages - newnpages); 4228 for (index = newnpages; index < bp->b_npages; index++) { 4229 p = bp->b_pages[index]; 4230 bp->b_pages[index] = NULL; 4231 if (p->busy != 0) 4232 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4233 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4234 p->wire_count--; 4235 vm_page_free(p); 4236 atomic_subtract_int(&cnt.v_wire_count, 1); 4237 } 4238 bp->b_npages = newnpages; 4239} 4240 4241/* 4242 * Map an IO request into kernel virtual address space. 4243 * 4244 * All requests are (re)mapped into kernel VA space. 4245 * Notice that we use b_bufsize for the size of the buffer 4246 * to be mapped. b_bcount might be modified by the driver. 4247 * 4248 * Note that even if the caller determines that the address space should 4249 * be valid, a race or a smaller-file mapped into a larger space may 4250 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4251 * check the return value. 4252 */ 4253int 4254vmapbuf(struct buf *bp, int mapbuf) 4255{ 4256 caddr_t kva; 4257 vm_prot_t prot; 4258 int pidx; 4259 4260 if (bp->b_bufsize < 0) 4261 return (-1); 4262 prot = VM_PROT_READ; 4263 if (bp->b_iocmd == BIO_READ) 4264 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4265 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4266 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4267 btoc(MAXPHYS))) < 0) 4268 return (-1); 4269 bp->b_npages = pidx; 4270 if (mapbuf || !unmapped_buf_allowed) { 4271 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4272 kva = bp->b_saveaddr; 4273 bp->b_saveaddr = bp->b_data; 4274 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4275 bp->b_flags &= ~B_UNMAPPED; 4276 } else { 4277 bp->b_flags |= B_UNMAPPED; 4278 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4279 bp->b_saveaddr = bp->b_data; 4280 bp->b_data = unmapped_buf; 4281 } 4282 return(0); 4283} 4284 4285/* 4286 * Free the io map PTEs associated with this IO operation. 4287 * We also invalidate the TLB entries and restore the original b_addr. 4288 */ 4289void 4290vunmapbuf(struct buf *bp) 4291{ 4292 int npages; 4293 4294 npages = bp->b_npages; 4295 if (bp->b_flags & B_UNMAPPED) 4296 bp->b_flags &= ~B_UNMAPPED; 4297 else 4298 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4299 vm_page_unhold_pages(bp->b_pages, npages); 4300 4301 bp->b_data = bp->b_saveaddr; 4302} 4303 4304void 4305bdone(struct buf *bp) 4306{ 4307 struct mtx *mtxp; 4308 4309 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4310 mtx_lock(mtxp); 4311 bp->b_flags |= B_DONE; 4312 wakeup(bp); 4313 mtx_unlock(mtxp); 4314} 4315 4316void 4317bwait(struct buf *bp, u_char pri, const char *wchan) 4318{ 4319 struct mtx *mtxp; 4320 4321 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4322 mtx_lock(mtxp); 4323 while ((bp->b_flags & B_DONE) == 0) 4324 msleep(bp, mtxp, pri, wchan, 0); 4325 mtx_unlock(mtxp); 4326} 4327 4328int 4329bufsync(struct bufobj *bo, int waitfor) 4330{ 4331 4332 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4333} 4334 4335void 4336bufstrategy(struct bufobj *bo, struct buf *bp) 4337{ 4338 int i = 0; 4339 struct vnode *vp; 4340 4341 vp = bp->b_vp; 4342 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4343 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4344 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4345 i = VOP_STRATEGY(vp, bp); 4346 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4347} 4348 4349void 4350bufobj_wrefl(struct bufobj *bo) 4351{ 4352 4353 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4354 ASSERT_BO_WLOCKED(bo); 4355 bo->bo_numoutput++; 4356} 4357 4358void 4359bufobj_wref(struct bufobj *bo) 4360{ 4361 4362 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4363 BO_LOCK(bo); 4364 bo->bo_numoutput++; 4365 BO_UNLOCK(bo); 4366} 4367 4368void 4369bufobj_wdrop(struct bufobj *bo) 4370{ 4371 4372 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4373 BO_LOCK(bo); 4374 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4375 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4376 bo->bo_flag &= ~BO_WWAIT; 4377 wakeup(&bo->bo_numoutput); 4378 } 4379 BO_UNLOCK(bo); 4380} 4381 4382int 4383bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4384{ 4385 int error; 4386 4387 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4388 ASSERT_BO_WLOCKED(bo); 4389 error = 0; 4390 while (bo->bo_numoutput) { 4391 bo->bo_flag |= BO_WWAIT; 4392 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4393 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4394 if (error) 4395 break; 4396 } 4397 return (error); 4398} 4399 4400void 4401bpin(struct buf *bp) 4402{ 4403 struct mtx *mtxp; 4404 4405 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4406 mtx_lock(mtxp); 4407 bp->b_pin_count++; 4408 mtx_unlock(mtxp); 4409} 4410 4411void 4412bunpin(struct buf *bp) 4413{ 4414 struct mtx *mtxp; 4415 4416 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4417 mtx_lock(mtxp); 4418 if (--bp->b_pin_count == 0) 4419 wakeup(bp); 4420 mtx_unlock(mtxp); 4421} 4422 4423void 4424bunpin_wait(struct buf *bp) 4425{ 4426 struct mtx *mtxp; 4427 4428 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4429 mtx_lock(mtxp); 4430 while (bp->b_pin_count > 0) 4431 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4432 mtx_unlock(mtxp); 4433} 4434 4435/* 4436 * Set bio_data or bio_ma for struct bio from the struct buf. 4437 */ 4438void 4439bdata2bio(struct buf *bp, struct bio *bip) 4440{ 4441 4442 if ((bp->b_flags & B_UNMAPPED) != 0) { 4443 KASSERT(unmapped_buf_allowed, ("unmapped")); 4444 bip->bio_ma = bp->b_pages; 4445 bip->bio_ma_n = bp->b_npages; 4446 bip->bio_data = unmapped_buf; 4447 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4448 bip->bio_flags |= BIO_UNMAPPED; 4449 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4450 PAGE_SIZE == bp->b_npages, 4451 ("Buffer %p too short: %d %d %d", bp, bip->bio_ma_offset, 4452 bip->bio_length, bip->bio_ma_n)); 4453 } else { 4454 bip->bio_data = bp->b_data; 4455 bip->bio_ma = NULL; 4456 } 4457} 4458 4459#include "opt_ddb.h" 4460#ifdef DDB 4461#include <ddb/ddb.h> 4462 4463/* DDB command to show buffer data */ 4464DB_SHOW_COMMAND(buffer, db_show_buffer) 4465{ 4466 /* get args */ 4467 struct buf *bp = (struct buf *)addr; 4468 4469 if (!have_addr) { 4470 db_printf("usage: show buffer <addr>\n"); 4471 return; 4472 } 4473 4474 db_printf("buf at %p\n", bp); 4475 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4476 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4477 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4478 db_printf( 4479 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4480 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4481 "b_dep = %p\n", 4482 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4483 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4484 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4485 if (bp->b_npages) { 4486 int i; 4487 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4488 for (i = 0; i < bp->b_npages; i++) { 4489 vm_page_t m; 4490 m = bp->b_pages[i]; 4491 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4492 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4493 if ((i + 1) < bp->b_npages) 4494 db_printf(","); 4495 } 4496 db_printf("\n"); 4497 } 4498 db_printf(" "); 4499 BUF_LOCKPRINTINFO(bp); 4500} 4501 4502DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4503{ 4504 struct buf *bp; 4505 int i; 4506 4507 for (i = 0; i < nbuf; i++) { 4508 bp = &buf[i]; 4509 if (BUF_ISLOCKED(bp)) { 4510 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4511 db_printf("\n"); 4512 } 4513 } 4514} 4515 4516DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4517{ 4518 struct vnode *vp; 4519 struct buf *bp; 4520 4521 if (!have_addr) { 4522 db_printf("usage: show vnodebufs <addr>\n"); 4523 return; 4524 } 4525 vp = (struct vnode *)addr; 4526 db_printf("Clean buffers:\n"); 4527 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4528 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4529 db_printf("\n"); 4530 } 4531 db_printf("Dirty buffers:\n"); 4532 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4533 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4534 db_printf("\n"); 4535 } 4536} 4537 4538DB_COMMAND(countfreebufs, db_coundfreebufs) 4539{ 4540 struct buf *bp; 4541 int i, used = 0, nfree = 0; 4542 4543 if (have_addr) { 4544 db_printf("usage: countfreebufs\n"); 4545 return; 4546 } 4547 4548 for (i = 0; i < nbuf; i++) { 4549 bp = &buf[i]; 4550 if ((bp->b_flags & B_INFREECNT) != 0) 4551 nfree++; 4552 else 4553 used++; 4554 } 4555 4556 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4557 nfree + used); 4558 db_printf("numfreebuffers is %d\n", numfreebuffers); 4559} 4560#endif /* DDB */ 4561