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