15 */
16
17/*
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
22 *
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
26 *
27 * see man buf(9) for more info.
28 */
29
30#include <sys/param.h>
31#include <sys/systm.h>
32#include <sys/stdint.h>
33#include <sys/bio.h>
34#include <sys/buf.h>
35#include <sys/eventhandler.h>
36#include <sys/lock.h>
37#include <sys/malloc.h>
38#include <sys/mount.h>
39#include <sys/mutex.h>
40#include <sys/kernel.h>
41#include <sys/kthread.h>
42#include <sys/ktr.h>
43#include <sys/proc.h>
44#include <sys/reboot.h>
45#include <sys/resourcevar.h>
46#include <sys/sysctl.h>
47#include <sys/vmmeter.h>
48#include <sys/vnode.h>
49#include <vm/vm.h>
50#include <vm/vm_param.h>
51#include <vm/vm_kern.h>
52#include <vm/vm_pageout.h>
53#include <vm/vm_page.h>
54#include <vm/vm_object.h>
55#include <vm/vm_extern.h>
56#include <vm/vm_map.h>
57
58static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
59
60struct bio_ops bioops; /* I/O operation notification */
61
62struct buf_ops buf_ops_bio = {
63 "buf_ops_bio",
64 bwrite
65};
66
67/*
68 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
69 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
70 */
71struct buf *buf; /* buffer header pool */
72struct mtx buftimelock; /* Interlock on setting prio and timo */
73
74static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
75 vm_offset_t to);
76static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
77 vm_offset_t to);
78static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
79 int pageno, vm_page_t m);
80static void vfs_clean_pages(struct buf * bp);
81static void vfs_setdirty(struct buf *bp);
82static void vfs_vmio_release(struct buf *bp);
83static void vfs_backgroundwritedone(struct buf *bp);
84static int flushbufqueues(void);
85static void buf_daemon(void);
86
87int vmiodirenable = TRUE;
88SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
89 "Use the VM system for directory writes");
90int runningbufspace;
91SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
92 "Amount of presently outstanding async buffer io");
93static int bufspace;
94SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
95 "KVA memory used for bufs");
96static int maxbufspace;
97SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
98 "Maximum allowed value of bufspace (including buf_daemon)");
99static int bufmallocspace;
100SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
101 "Amount of malloced memory for buffers");
102static int maxbufmallocspace;
103SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
104 "Maximum amount of malloced memory for buffers");
105static int lobufspace;
106SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
107 "Minimum amount of buffers we want to have");
108static int hibufspace;
109SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
110 "Maximum allowed value of bufspace (excluding buf_daemon)");
111static int bufreusecnt;
112SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
113 "Number of times we have reused a buffer");
114static int buffreekvacnt;
115SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
116 "Number of times we have freed the KVA space from some buffer");
117static int bufdefragcnt;
118SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
119 "Number of times we have had to repeat buffer allocation to defragment");
120static int lorunningspace;
121SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
122 "Minimum preferred space used for in-progress I/O");
123static int hirunningspace;
124SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
125 "Maximum amount of space to use for in-progress I/O");
126static int numdirtybuffers;
127SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
128 "Number of buffers that are dirty (has unwritten changes) at the moment");
129static int lodirtybuffers;
130SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
131 "How many buffers we want to have free before bufdaemon can sleep");
132static int hidirtybuffers;
133SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
134 "When the number of dirty buffers is considered severe");
135static int numfreebuffers;
136SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
137 "Number of free buffers");
138static int lofreebuffers;
139SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
140 "XXX Unused");
141static int hifreebuffers;
142SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
143 "XXX Complicatedly unused");
144static int getnewbufcalls;
145SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
146 "Number of calls to getnewbuf");
147static int getnewbufrestarts;
148SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
149 "Number of times getnewbuf has had to restart a buffer aquisition");
150static int dobkgrdwrite = 1;
151SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
152 "Do background writes (honoring the BX_BKGRDWRITE flag)?");
153
154/*
155 * Wakeup point for bufdaemon, as well as indicator of whether it is already
156 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
157 * is idling.
158 */
159static int bd_request;
160
161/*
162 * bogus page -- for I/O to/from partially complete buffers
163 * this is a temporary solution to the problem, but it is not
164 * really that bad. it would be better to split the buffer
165 * for input in the case of buffers partially already in memory,
166 * but the code is intricate enough already.
167 */
168vm_page_t bogus_page;
169
170/*
171 * Offset for bogus_page.
172 * XXX bogus_offset should be local to bufinit
173 */
174static vm_offset_t bogus_offset;
175
176/*
177 * Synchronization (sleep/wakeup) variable for active buffer space requests.
178 * Set when wait starts, cleared prior to wakeup().
179 * Used in runningbufwakeup() and waitrunningbufspace().
180 */
181static int runningbufreq;
182
183/*
184 * Synchronization (sleep/wakeup) variable for buffer requests.
185 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
186 * by and/or.
187 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
188 * getnewbuf(), and getblk().
189 */
190static int needsbuffer;
191
192#ifdef USE_BUFHASH
193/*
194 * Mask for index into the buffer hash table, which needs to be power of 2 in
195 * size. Set in kern_vfs_bio_buffer_alloc.
196 */
197static int bufhashmask;
198
199/*
200 * Hash table for all buffers, with a linked list hanging from each table
201 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init.
202 */
203static LIST_HEAD(bufhashhdr, buf) *bufhashtbl;
204
205/*
206 * Somewhere to store buffers when they are not in another list, to always
207 * have them in a list (and thus being able to use the same set of operations
208 * on them.)
209 */
210static struct bufhashhdr invalhash;
211
212#endif
213
214/*
215 * Definitions for the buffer free lists.
216 */
217#define BUFFER_QUEUES 6 /* number of free buffer queues */
218
219#define QUEUE_NONE 0 /* on no queue */
220#define QUEUE_LOCKED 1 /* locked buffers */
221#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */
222#define QUEUE_DIRTY 3 /* B_DELWRI buffers */
223#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */
224#define QUEUE_EMPTY 5 /* empty buffer headers */
225
226/* Queues for free buffers with various properties */
227static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
228/*
229 * Single global constant for BUF_WMESG, to avoid getting multiple references.
230 * buf_wmesg is referred from macros.
231 */
232const char *buf_wmesg = BUF_WMESG;
233
234#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
235#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
236#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
237#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
238
239#ifdef USE_BUFHASH
240/*
241 * Buffer hash table code. Note that the logical block scans linearly, which
242 * gives us some L1 cache locality.
243 */
244
245static __inline
246struct bufhashhdr *
247bufhash(struct vnode *vnp, daddr_t bn)
248{
249 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]);
250}
251
252#endif
253
254/*
255 * numdirtywakeup:
256 *
257 * If someone is blocked due to there being too many dirty buffers,
258 * and numdirtybuffers is now reasonable, wake them up.
259 */
260
261static __inline void
262numdirtywakeup(int level)
263{
264 if (numdirtybuffers <= level) {
265 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
266 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
267 wakeup(&needsbuffer);
268 }
269 }
270}
271
272/*
273 * bufspacewakeup:
274 *
275 * Called when buffer space is potentially available for recovery.
276 * getnewbuf() will block on this flag when it is unable to free
277 * sufficient buffer space. Buffer space becomes recoverable when
278 * bp's get placed back in the queues.
279 */
280
281static __inline void
282bufspacewakeup(void)
283{
284 /*
285 * If someone is waiting for BUF space, wake them up. Even
286 * though we haven't freed the kva space yet, the waiting
287 * process will be able to now.
288 */
289 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
290 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
291 wakeup(&needsbuffer);
292 }
293}
294
295/*
296 * runningbufwakeup() - in-progress I/O accounting.
297 *
298 */
299static __inline void
300runningbufwakeup(struct buf *bp)
301{
302 if (bp->b_runningbufspace) {
303 runningbufspace -= bp->b_runningbufspace;
304 bp->b_runningbufspace = 0;
305 if (runningbufreq && runningbufspace <= lorunningspace) {
306 runningbufreq = 0;
307 wakeup(&runningbufreq);
308 }
309 }
310}
311
312/*
313 * bufcountwakeup:
314 *
315 * Called when a buffer has been added to one of the free queues to
316 * account for the buffer and to wakeup anyone waiting for free buffers.
317 * This typically occurs when large amounts of metadata are being handled
318 * by the buffer cache ( else buffer space runs out first, usually ).
319 */
320
321static __inline void
322bufcountwakeup(void)
323{
324 ++numfreebuffers;
325 if (needsbuffer) {
326 needsbuffer &= ~VFS_BIO_NEED_ANY;
327 if (numfreebuffers >= hifreebuffers)
328 needsbuffer &= ~VFS_BIO_NEED_FREE;
329 wakeup(&needsbuffer);
330 }
331}
332
333/*
334 * waitrunningbufspace()
335 *
336 * runningbufspace is a measure of the amount of I/O currently
337 * running. This routine is used in async-write situations to
338 * prevent creating huge backups of pending writes to a device.
339 * Only asynchronous writes are governed by this function.
340 *
341 * Reads will adjust runningbufspace, but will not block based on it.
342 * The read load has a side effect of reducing the allowed write load.
343 *
344 * This does NOT turn an async write into a sync write. It waits
345 * for earlier writes to complete and generally returns before the
346 * caller's write has reached the device.
347 */
348static __inline void
349waitrunningbufspace(void)
350{
351 /*
352 * XXX race against wakeup interrupt, currently
353 * protected by Giant. FIXME!
354 */
355 while (runningbufspace > hirunningspace) {
356 ++runningbufreq;
357 tsleep(&runningbufreq, PVM, "wdrain", 0);
358 }
359}
360
361
362/*
363 * vfs_buf_test_cache:
364 *
365 * Called when a buffer is extended. This function clears the B_CACHE
366 * bit if the newly extended portion of the buffer does not contain
367 * valid data.
368 */
369static __inline__
370void
371vfs_buf_test_cache(struct buf *bp,
372 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
373 vm_page_t m)
374{
375 GIANT_REQUIRED;
376
377 if (bp->b_flags & B_CACHE) {
378 int base = (foff + off) & PAGE_MASK;
379 if (vm_page_is_valid(m, base, size) == 0)
380 bp->b_flags &= ~B_CACHE;
381 }
382}
383
384/* Wake up the buffer deamon if necessary */
385static __inline__
386void
387bd_wakeup(int dirtybuflevel)
388{
389 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
390 bd_request = 1;
391 wakeup(&bd_request);
392 }
393}
394
395/*
396 * bd_speedup - speedup the buffer cache flushing code
397 */
398
399static __inline__
400void
401bd_speedup(void)
402{
403 bd_wakeup(1);
404}
405
406/*
407 * Calculating buffer cache scaling values and reserve space for buffer
408 * headers. This is called during low level kernel initialization and
409 * may be called more then once. We CANNOT write to the memory area
410 * being reserved at this time.
411 */
412caddr_t
413kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
414{
415 /*
416 * physmem_est is in pages. Convert it to kilobytes (assumes
417 * PAGE_SIZE is >= 1K)
418 */
419 physmem_est = physmem_est * (PAGE_SIZE / 1024);
420
421 /*
422 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
423 * For the first 64MB of ram nominally allocate sufficient buffers to
424 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
425 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing
426 * the buffer cache we limit the eventual kva reservation to
427 * maxbcache bytes.
428 *
429 * factor represents the 1/4 x ram conversion.
430 */
431 if (nbuf == 0) {
432 int factor = 4 * BKVASIZE / 1024;
433
434 nbuf = 50;
435 if (physmem_est > 4096)
436 nbuf += min((physmem_est - 4096) / factor,
437 65536 / factor);
438 if (physmem_est > 65536)
439 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
440
441 if (maxbcache && nbuf > maxbcache / BKVASIZE)
442 nbuf = maxbcache / BKVASIZE;
443 }
444
445#if 0
446 /*
447 * Do not allow the buffer_map to be more then 1/2 the size of the
448 * kernel_map.
449 */
450 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
451 (BKVASIZE * 2)) {
452 nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
453 (BKVASIZE * 2);
454 printf("Warning: nbufs capped at %d\n", nbuf);
455 }
456#endif
457
458 /*
459 * swbufs are used as temporary holders for I/O, such as paging I/O.
460 * We have no less then 16 and no more then 256.
461 */
462 nswbuf = max(min(nbuf/4, 256), 16);
463
464 /*
465 * Reserve space for the buffer cache buffers
466 */
467 swbuf = (void *)v;
468 v = (caddr_t)(swbuf + nswbuf);
469 buf = (void *)v;
470 v = (caddr_t)(buf + nbuf);
471
472#ifdef USE_BUFHASH
473 /*
474 * Calculate the hash table size and reserve space
475 */
476 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
477 ;
478 bufhashtbl = (void *)v;
479 v = (caddr_t)(bufhashtbl + bufhashmask);
480 --bufhashmask;
481#endif
482 return(v);
483}
484
485/* Initialize the buffer subsystem. Called before use of any buffers. */
486void
487bufinit(void)
488{
489 struct buf *bp;
490 int i;
491
492 GIANT_REQUIRED;
493
494#ifdef USE_BUFHASH
495 LIST_INIT(&invalhash);
496#endif
497 mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF);
498
499#ifdef USE_BUFHASH
500 for (i = 0; i <= bufhashmask; i++)
501 LIST_INIT(&bufhashtbl[i]);
502#endif
503
504 /* next, make a null set of free lists */
505 for (i = 0; i < BUFFER_QUEUES; i++)
506 TAILQ_INIT(&bufqueues[i]);
507
508 /* finally, initialize each buffer header and stick on empty q */
509 for (i = 0; i < nbuf; i++) {
510 bp = &buf[i];
511 bzero(bp, sizeof *bp);
512 bp->b_flags = B_INVAL; /* we're just an empty header */
513 bp->b_dev = NODEV;
514 bp->b_rcred = NOCRED;
515 bp->b_wcred = NOCRED;
516 bp->b_qindex = QUEUE_EMPTY;
517 bp->b_xflags = 0;
518 LIST_INIT(&bp->b_dep);
519 BUF_LOCKINIT(bp);
520 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
521#ifdef USE_BUFHASH
522 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
523#endif
524 }
525
526 /*
527 * maxbufspace is the absolute maximum amount of buffer space we are
528 * allowed to reserve in KVM and in real terms. The absolute maximum
529 * is nominally used by buf_daemon. hibufspace is the nominal maximum
530 * used by most other processes. The differential is required to
531 * ensure that buf_daemon is able to run when other processes might
532 * be blocked waiting for buffer space.
533 *
534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
535 * this may result in KVM fragmentation which is not handled optimally
536 * by the system.
537 */
538 maxbufspace = nbuf * BKVASIZE;
539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
540 lobufspace = hibufspace - MAXBSIZE;
541
542 lorunningspace = 512 * 1024;
543 hirunningspace = 1024 * 1024;
544
545/*
546 * Limit the amount of malloc memory since it is wired permanently into
547 * the kernel space. Even though this is accounted for in the buffer
548 * allocation, we don't want the malloced region to grow uncontrolled.
549 * The malloc scheme improves memory utilization significantly on average
550 * (small) directories.
551 */
552 maxbufmallocspace = hibufspace / 20;
553
554/*
555 * Reduce the chance of a deadlock occuring by limiting the number
556 * of delayed-write dirty buffers we allow to stack up.
557 */
558 hidirtybuffers = nbuf / 4 + 20;
559 numdirtybuffers = 0;
560/*
561 * To support extreme low-memory systems, make sure hidirtybuffers cannot
562 * eat up all available buffer space. This occurs when our minimum cannot
563 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
564 * BKVASIZE'd (8K) buffers.
565 */
566 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
567 hidirtybuffers >>= 1;
568 }
569 lodirtybuffers = hidirtybuffers / 2;
570
571/*
572 * Try to keep the number of free buffers in the specified range,
573 * and give special processes (e.g. like buf_daemon) access to an
574 * emergency reserve.
575 */
576 lofreebuffers = nbuf / 18 + 5;
577 hifreebuffers = 2 * lofreebuffers;
578 numfreebuffers = nbuf;
579
580/*
581 * Maximum number of async ops initiated per buf_daemon loop. This is
582 * somewhat of a hack at the moment, we really need to limit ourselves
583 * based on the number of bytes of I/O in-transit that were initiated
584 * from buf_daemon.
585 */
586
587 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
588 bogus_page = vm_page_alloc(kernel_object,
589 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
590 VM_ALLOC_NORMAL);
591 cnt.v_wire_count++;
592}
593
594/*
595 * bfreekva() - free the kva allocation for a buffer.
596 *
597 * Must be called at splbio() or higher as this is the only locking for
598 * buffer_map.
599 *
600 * Since this call frees up buffer space, we call bufspacewakeup().
601 */
602static void
603bfreekva(struct buf * bp)
604{
605 GIANT_REQUIRED;
606
607 if (bp->b_kvasize) {
608 ++buffreekvacnt;
609 bufspace -= bp->b_kvasize;
610 vm_map_delete(buffer_map,
611 (vm_offset_t) bp->b_kvabase,
612 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
613 );
614 bp->b_kvasize = 0;
615 bufspacewakeup();
616 }
617}
618
619/*
620 * bremfree:
621 *
622 * Remove the buffer from the appropriate free list.
623 */
624void
625bremfree(struct buf * bp)
626{
627 int s = splbio();
628 int old_qindex = bp->b_qindex;
629
630 GIANT_REQUIRED;
631
632 if (bp->b_qindex != QUEUE_NONE) {
633 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
634 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
635 bp->b_qindex = QUEUE_NONE;
636 } else {
637 if (BUF_REFCNT(bp) <= 1)
638 panic("bremfree: removing a buffer not on a queue");
639 }
640
641 /*
642 * Fixup numfreebuffers count. If the buffer is invalid or not
643 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
644 * the buffer was free and we must decrement numfreebuffers.
645 */
646 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
647 switch(old_qindex) {
648 case QUEUE_DIRTY:
649 case QUEUE_CLEAN:
650 case QUEUE_EMPTY:
651 case QUEUE_EMPTYKVA:
652 --numfreebuffers;
653 break;
654 default:
655 break;
656 }
657 }
658 splx(s);
659}
660
661
662/*
663 * Get a buffer with the specified data. Look in the cache first. We
664 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
665 * is set, the buffer is valid and we do not have to do anything ( see
666 * getblk() ). This is really just a special case of breadn().
667 */
668int
669bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
670 struct buf ** bpp)
671{
672
673 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
674}
675
676/*
677 * Operates like bread, but also starts asynchronous I/O on
678 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior
679 * to initiating I/O . If B_CACHE is set, the buffer is valid
680 * and we do not have to do anything.
681 */
682int
683breadn(struct vnode * vp, daddr_t blkno, int size,
684 daddr_t * rablkno, int *rabsize,
685 int cnt, struct ucred * cred, struct buf ** bpp)
686{
687 struct buf *bp, *rabp;
688 int i;
689 int rv = 0, readwait = 0;
690
691 *bpp = bp = getblk(vp, blkno, size, 0, 0);
692
693 /* if not found in cache, do some I/O */
694 if ((bp->b_flags & B_CACHE) == 0) {
695 if (curthread != PCPU_GET(idlethread))
696 curthread->td_proc->p_stats->p_ru.ru_inblock++;
697 bp->b_iocmd = BIO_READ;
698 bp->b_flags &= ~B_INVAL;
699 bp->b_ioflags &= ~BIO_ERROR;
700 if (bp->b_rcred == NOCRED && cred != NOCRED)
701 bp->b_rcred = crhold(cred);
702 vfs_busy_pages(bp, 0);
703 VOP_STRATEGY(vp, bp);
704 ++readwait;
705 }
706
707 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
708 if (inmem(vp, *rablkno))
709 continue;
710 rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
711
712 if ((rabp->b_flags & B_CACHE) == 0) {
713 if (curthread != PCPU_GET(idlethread))
714 curthread->td_proc->p_stats->p_ru.ru_inblock++;
715 rabp->b_flags |= B_ASYNC;
716 rabp->b_flags &= ~B_INVAL;
717 rabp->b_ioflags &= ~BIO_ERROR;
718 rabp->b_iocmd = BIO_READ;
719 if (rabp->b_rcred == NOCRED && cred != NOCRED)
720 rabp->b_rcred = crhold(cred);
721 vfs_busy_pages(rabp, 0);
722 BUF_KERNPROC(rabp);
723 VOP_STRATEGY(vp, rabp);
724 } else {
725 brelse(rabp);
726 }
727 }
728
729 if (readwait) {
730 rv = bufwait(bp);
731 }
732 return (rv);
733}
734
735/*
736 * Write, release buffer on completion. (Done by iodone
737 * if async). Do not bother writing anything if the buffer
738 * is invalid.
739 *
740 * Note that we set B_CACHE here, indicating that buffer is
741 * fully valid and thus cacheable. This is true even of NFS
742 * now so we set it generally. This could be set either here
743 * or in biodone() since the I/O is synchronous. We put it
744 * here.
745 */
746
747int
748bwrite(struct buf * bp)
749{
750 int oldflags, s;
751 struct buf *newbp;
752
753 if (bp->b_flags & B_INVAL) {
754 brelse(bp);
755 return (0);
756 }
757
758 oldflags = bp->b_flags;
759
760 if (BUF_REFCNT(bp) == 0)
761 panic("bwrite: buffer is not busy???");
762 s = splbio();
763 /*
764 * If a background write is already in progress, delay
765 * writing this block if it is asynchronous. Otherwise
766 * wait for the background write to complete.
767 */
768 if (bp->b_xflags & BX_BKGRDINPROG) {
769 if (bp->b_flags & B_ASYNC) {
770 splx(s);
771 bdwrite(bp);
772 return (0);
773 }
774 bp->b_xflags |= BX_BKGRDWAIT;
775 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0);
776 if (bp->b_xflags & BX_BKGRDINPROG)
777 panic("bwrite: still writing");
778 }
779
780 /* Mark the buffer clean */
781 bundirty(bp);
782
783 /*
784 * If this buffer is marked for background writing and we
785 * do not have to wait for it, make a copy and write the
786 * copy so as to leave this buffer ready for further use.
787 *
788 * This optimization eats a lot of memory. If we have a page
789 * or buffer shortfall we can't do it.
790 */
791 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) &&
792 (bp->b_flags & B_ASYNC) &&
793 !vm_page_count_severe() &&
794 !buf_dirty_count_severe()) {
795 if (bp->b_iodone != NULL) {
796 printf("bp->b_iodone = %p\n", bp->b_iodone);
797 panic("bwrite: need chained iodone");
798 }
799
800 /* get a new block */
801 newbp = geteblk(bp->b_bufsize);
802
803 /*
804 * set it to be identical to the old block. We have to
805 * set b_lblkno and BKGRDMARKER before calling bgetvp()
806 * to avoid confusing the splay tree and gbincore().
807 */
808 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
809 newbp->b_lblkno = bp->b_lblkno;
810 newbp->b_xflags |= BX_BKGRDMARKER;
811 bgetvp(bp->b_vp, newbp);
812 newbp->b_blkno = bp->b_blkno;
813 newbp->b_offset = bp->b_offset;
814 newbp->b_iodone = vfs_backgroundwritedone;
815 newbp->b_flags |= B_ASYNC;
816 newbp->b_flags &= ~B_INVAL;
817
818 /* move over the dependencies */
819 if (LIST_FIRST(&bp->b_dep) != NULL)
820 buf_movedeps(bp, newbp);
821
822 /*
823 * Initiate write on the copy, release the original to
824 * the B_LOCKED queue so that it cannot go away until
825 * the background write completes. If not locked it could go
826 * away and then be reconstituted while it was being written.
827 * If the reconstituted buffer were written, we could end up
828 * with two background copies being written at the same time.
829 */
830 bp->b_xflags |= BX_BKGRDINPROG;
831 bp->b_flags |= B_LOCKED;
832 bqrelse(bp);
833 bp = newbp;
834 }
835
836 bp->b_flags &= ~B_DONE;
837 bp->b_ioflags &= ~BIO_ERROR;
838 bp->b_flags |= B_WRITEINPROG | B_CACHE;
839 bp->b_iocmd = BIO_WRITE;
840
841 bp->b_vp->v_numoutput++;
842 vfs_busy_pages(bp, 1);
843
844 /*
845 * Normal bwrites pipeline writes
846 */
847 bp->b_runningbufspace = bp->b_bufsize;
848 runningbufspace += bp->b_runningbufspace;
849
850 if (curthread != PCPU_GET(idlethread))
851 curthread->td_proc->p_stats->p_ru.ru_oublock++;
852 splx(s);
853 if (oldflags & B_ASYNC)
854 BUF_KERNPROC(bp);
855 BUF_STRATEGY(bp);
856
857 if ((oldflags & B_ASYNC) == 0) {
858 int rtval = bufwait(bp);
859 brelse(bp);
860 return (rtval);
861 } else if ((oldflags & B_NOWDRAIN) == 0) {
862 /*
863 * don't allow the async write to saturate the I/O
864 * system. Deadlocks can occur only if a device strategy
865 * routine (like in MD) turns around and issues another
866 * high-level write, in which case B_NOWDRAIN is expected
867 * to be set. Otherwise we will not deadlock here because
868 * we are blocking waiting for I/O that is already in-progress
869 * to complete.
870 */
871 waitrunningbufspace();
872 }
873
874 return (0);
875}
876
877/*
878 * Complete a background write started from bwrite.
879 */
880static void
881vfs_backgroundwritedone(bp)
882 struct buf *bp;
883{
884 struct buf *origbp;
885
886 /*
887 * Find the original buffer that we are writing.
888 */
889 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
890 panic("backgroundwritedone: lost buffer");
891 /*
892 * Process dependencies then return any unfinished ones.
893 */
894 if (LIST_FIRST(&bp->b_dep) != NULL)
895 buf_complete(bp);
896 if (LIST_FIRST(&bp->b_dep) != NULL)
897 buf_movedeps(bp, origbp);
898 /*
899 * Clear the BX_BKGRDINPROG flag in the original buffer
900 * and awaken it if it is waiting for the write to complete.
901 * If BX_BKGRDINPROG is not set in the original buffer it must
902 * have been released and re-instantiated - which is not legal.
903 */
904 KASSERT((origbp->b_xflags & BX_BKGRDINPROG),
905 ("backgroundwritedone: lost buffer2"));
906 origbp->b_xflags &= ~BX_BKGRDINPROG;
907 if (origbp->b_xflags & BX_BKGRDWAIT) {
908 origbp->b_xflags &= ~BX_BKGRDWAIT;
909 wakeup(&origbp->b_xflags);
910 }
911 /*
912 * Clear the B_LOCKED flag and remove it from the locked
913 * queue if it currently resides there.
914 */
915 origbp->b_flags &= ~B_LOCKED;
916 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
917 bremfree(origbp);
918 bqrelse(origbp);
919 }
920 /*
921 * This buffer is marked B_NOCACHE, so when it is released
922 * by biodone, it will be tossed. We mark it with BIO_READ
923 * to avoid biodone doing a second vwakeup.
924 */
925 bp->b_flags |= B_NOCACHE;
926 bp->b_iocmd = BIO_READ;
927 bp->b_flags &= ~(B_CACHE | B_DONE);
928 bp->b_iodone = 0;
929 bufdone(bp);
930}
931
932/*
933 * Delayed write. (Buffer is marked dirty). Do not bother writing
934 * anything if the buffer is marked invalid.
935 *
936 * Note that since the buffer must be completely valid, we can safely
937 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
938 * biodone() in order to prevent getblk from writing the buffer
939 * out synchronously.
940 */
941void
942bdwrite(struct buf * bp)
943{
944 GIANT_REQUIRED;
945
946 if (BUF_REFCNT(bp) == 0)
947 panic("bdwrite: buffer is not busy");
948
949 if (bp->b_flags & B_INVAL) {
950 brelse(bp);
951 return;
952 }
953 bdirty(bp);
954
955 /*
956 * Set B_CACHE, indicating that the buffer is fully valid. This is
957 * true even of NFS now.
958 */
959 bp->b_flags |= B_CACHE;
960
961 /*
962 * This bmap keeps the system from needing to do the bmap later,
963 * perhaps when the system is attempting to do a sync. Since it
964 * is likely that the indirect block -- or whatever other datastructure
965 * that the filesystem needs is still in memory now, it is a good
966 * thing to do this. Note also, that if the pageout daemon is
967 * requesting a sync -- there might not be enough memory to do
968 * the bmap then... So, this is important to do.
969 */
970 if (bp->b_lblkno == bp->b_blkno) {
971 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
972 }
973
974 /*
975 * Set the *dirty* buffer range based upon the VM system dirty pages.
976 */
977 vfs_setdirty(bp);
978
979 /*
980 * We need to do this here to satisfy the vnode_pager and the
981 * pageout daemon, so that it thinks that the pages have been
982 * "cleaned". Note that since the pages are in a delayed write
983 * buffer -- the VFS layer "will" see that the pages get written
984 * out on the next sync, or perhaps the cluster will be completed.
985 */
986 vfs_clean_pages(bp);
987 bqrelse(bp);
988
989 /*
990 * Wakeup the buffer flushing daemon if we have a lot of dirty
991 * buffers (midpoint between our recovery point and our stall
992 * point).
993 */
994 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
995
996 /*
997 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
998 * due to the softdep code.
999 */
1000}
1001
1002/*
1003 * bdirty:
1004 *
1005 * Turn buffer into delayed write request. We must clear BIO_READ and
1006 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1007 * itself to properly update it in the dirty/clean lists. We mark it
1008 * B_DONE to ensure that any asynchronization of the buffer properly
1009 * clears B_DONE ( else a panic will occur later ).
1010 *
1011 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1012 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1013 * should only be called if the buffer is known-good.
1014 *
1015 * Since the buffer is not on a queue, we do not update the numfreebuffers
1016 * count.
1017 *
1018 * Must be called at splbio().
1019 * The buffer must be on QUEUE_NONE.
1020 */
1021void
1022bdirty(bp)
1023 struct buf *bp;
1024{
1025 KASSERT(bp->b_qindex == QUEUE_NONE,
1026 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1027 bp->b_flags &= ~(B_RELBUF);
1028 bp->b_iocmd = BIO_WRITE;
1029
1030 if ((bp->b_flags & B_DELWRI) == 0) {
1031 bp->b_flags |= B_DONE | B_DELWRI;
1032 reassignbuf(bp, bp->b_vp);
1033 ++numdirtybuffers;
1034 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1035 }
1036}
1037
1038/*
1039 * bundirty:
1040 *
1041 * Clear B_DELWRI for buffer.
1042 *
1043 * Since the buffer is not on a queue, we do not update the numfreebuffers
1044 * count.
1045 *
1046 * Must be called at splbio().
1047 * The buffer must be on QUEUE_NONE.
1048 */
1049
1050void
1051bundirty(bp)
1052 struct buf *bp;
1053{
1054 KASSERT(bp->b_qindex == QUEUE_NONE,
1055 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1056
1057 if (bp->b_flags & B_DELWRI) {
1058 bp->b_flags &= ~B_DELWRI;
1059 reassignbuf(bp, bp->b_vp);
1060 --numdirtybuffers;
1061 numdirtywakeup(lodirtybuffers);
1062 }
1063 /*
1064 * Since it is now being written, we can clear its deferred write flag.
1065 */
1066 bp->b_flags &= ~B_DEFERRED;
1067}
1068
1069/*
1070 * bawrite:
1071 *
1072 * Asynchronous write. Start output on a buffer, but do not wait for
1073 * it to complete. The buffer is released when the output completes.
1074 *
1075 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1076 * B_INVAL buffers. Not us.
1077 */
1078void
1079bawrite(struct buf * bp)
1080{
1081 bp->b_flags |= B_ASYNC;
1082 (void) BUF_WRITE(bp);
1083}
1084
1085/*
1086 * bwillwrite:
1087 *
1088 * Called prior to the locking of any vnodes when we are expecting to
1089 * write. We do not want to starve the buffer cache with too many
1090 * dirty buffers so we block here. By blocking prior to the locking
1091 * of any vnodes we attempt to avoid the situation where a locked vnode
1092 * prevents the various system daemons from flushing related buffers.
1093 */
1094
1095void
1096bwillwrite(void)
1097{
1098 if (numdirtybuffers >= hidirtybuffers) {
1099 int s;
1100
1101 mtx_lock(&Giant);
1102 s = splbio();
1103 while (numdirtybuffers >= hidirtybuffers) {
1104 bd_wakeup(1);
1105 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1106 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0);
1107 }
1108 splx(s);
1109 mtx_unlock(&Giant);
1110 }
1111}
1112
1113/*
1114 * Return true if we have too many dirty buffers.
1115 */
1116int
1117buf_dirty_count_severe(void)
1118{
1119 return(numdirtybuffers >= hidirtybuffers);
1120}
1121
1122/*
1123 * brelse:
1124 *
1125 * Release a busy buffer and, if requested, free its resources. The
1126 * buffer will be stashed in the appropriate bufqueue[] allowing it
1127 * to be accessed later as a cache entity or reused for other purposes.
1128 */
1129void
1130brelse(struct buf * bp)
1131{
1132 int s;
1133
1134 GIANT_REQUIRED;
1135
1136 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1137 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1138
1139 s = splbio();
1140
1141 if (bp->b_flags & B_LOCKED)
1142 bp->b_ioflags &= ~BIO_ERROR;
1143
1144 if (bp->b_iocmd == BIO_WRITE &&
1145 (bp->b_ioflags & BIO_ERROR) &&
1146 !(bp->b_flags & B_INVAL)) {
1147 /*
1148 * Failed write, redirty. Must clear BIO_ERROR to prevent
1149 * pages from being scrapped. If B_INVAL is set then
1150 * this case is not run and the next case is run to
1151 * destroy the buffer. B_INVAL can occur if the buffer
1152 * is outside the range supported by the underlying device.
1153 */
1154 bp->b_ioflags &= ~BIO_ERROR;
1155 bdirty(bp);
1156 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1157 (bp->b_ioflags & BIO_ERROR) ||
1158 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) {
1159 /*
1160 * Either a failed I/O or we were asked to free or not
1161 * cache the buffer.
1162 */
1163 bp->b_flags |= B_INVAL;
1164 if (LIST_FIRST(&bp->b_dep) != NULL)
1165 buf_deallocate(bp);
1166 if (bp->b_flags & B_DELWRI) {
1167 --numdirtybuffers;
1168 numdirtywakeup(lodirtybuffers);
1169 }
1170 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1171 if ((bp->b_flags & B_VMIO) == 0) {
1172 if (bp->b_bufsize)
1173 allocbuf(bp, 0);
1174 if (bp->b_vp)
1175 brelvp(bp);
1176 }
1177 }
1178
1179 /*
1180 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1181 * is called with B_DELWRI set, the underlying pages may wind up
1182 * getting freed causing a previous write (bdwrite()) to get 'lost'
1183 * because pages associated with a B_DELWRI bp are marked clean.
1184 *
1185 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1186 * if B_DELWRI is set.
1187 *
1188 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1189 * on pages to return pages to the VM page queues.
1190 */
1191 if (bp->b_flags & B_DELWRI)
1192 bp->b_flags &= ~B_RELBUF;
1193 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1194 bp->b_flags |= B_RELBUF;
1195
1196 /*
1197 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1198 * constituted, not even NFS buffers now. Two flags effect this. If
1199 * B_INVAL, the struct buf is invalidated but the VM object is kept
1200 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1201 *
1202 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1203 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1204 * buffer is also B_INVAL because it hits the re-dirtying code above.
1205 *
1206 * Normally we can do this whether a buffer is B_DELWRI or not. If
1207 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1208 * the commit state and we cannot afford to lose the buffer. If the
1209 * buffer has a background write in progress, we need to keep it
1210 * around to prevent it from being reconstituted and starting a second
1211 * background write.
1212 */
1213 if ((bp->b_flags & B_VMIO)
| 15 */
16
17/*
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
22 *
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
26 *
27 * see man buf(9) for more info.
28 */
29
30#include <sys/param.h>
31#include <sys/systm.h>
32#include <sys/stdint.h>
33#include <sys/bio.h>
34#include <sys/buf.h>
35#include <sys/eventhandler.h>
36#include <sys/lock.h>
37#include <sys/malloc.h>
38#include <sys/mount.h>
39#include <sys/mutex.h>
40#include <sys/kernel.h>
41#include <sys/kthread.h>
42#include <sys/ktr.h>
43#include <sys/proc.h>
44#include <sys/reboot.h>
45#include <sys/resourcevar.h>
46#include <sys/sysctl.h>
47#include <sys/vmmeter.h>
48#include <sys/vnode.h>
49#include <vm/vm.h>
50#include <vm/vm_param.h>
51#include <vm/vm_kern.h>
52#include <vm/vm_pageout.h>
53#include <vm/vm_page.h>
54#include <vm/vm_object.h>
55#include <vm/vm_extern.h>
56#include <vm/vm_map.h>
57
58static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
59
60struct bio_ops bioops; /* I/O operation notification */
61
62struct buf_ops buf_ops_bio = {
63 "buf_ops_bio",
64 bwrite
65};
66
67/*
68 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
69 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
70 */
71struct buf *buf; /* buffer header pool */
72struct mtx buftimelock; /* Interlock on setting prio and timo */
73
74static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
75 vm_offset_t to);
76static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
77 vm_offset_t to);
78static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
79 int pageno, vm_page_t m);
80static void vfs_clean_pages(struct buf * bp);
81static void vfs_setdirty(struct buf *bp);
82static void vfs_vmio_release(struct buf *bp);
83static void vfs_backgroundwritedone(struct buf *bp);
84static int flushbufqueues(void);
85static void buf_daemon(void);
86
87int vmiodirenable = TRUE;
88SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
89 "Use the VM system for directory writes");
90int runningbufspace;
91SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
92 "Amount of presently outstanding async buffer io");
93static int bufspace;
94SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
95 "KVA memory used for bufs");
96static int maxbufspace;
97SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
98 "Maximum allowed value of bufspace (including buf_daemon)");
99static int bufmallocspace;
100SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
101 "Amount of malloced memory for buffers");
102static int maxbufmallocspace;
103SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
104 "Maximum amount of malloced memory for buffers");
105static int lobufspace;
106SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
107 "Minimum amount of buffers we want to have");
108static int hibufspace;
109SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
110 "Maximum allowed value of bufspace (excluding buf_daemon)");
111static int bufreusecnt;
112SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
113 "Number of times we have reused a buffer");
114static int buffreekvacnt;
115SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
116 "Number of times we have freed the KVA space from some buffer");
117static int bufdefragcnt;
118SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
119 "Number of times we have had to repeat buffer allocation to defragment");
120static int lorunningspace;
121SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
122 "Minimum preferred space used for in-progress I/O");
123static int hirunningspace;
124SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
125 "Maximum amount of space to use for in-progress I/O");
126static int numdirtybuffers;
127SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
128 "Number of buffers that are dirty (has unwritten changes) at the moment");
129static int lodirtybuffers;
130SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
131 "How many buffers we want to have free before bufdaemon can sleep");
132static int hidirtybuffers;
133SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
134 "When the number of dirty buffers is considered severe");
135static int numfreebuffers;
136SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
137 "Number of free buffers");
138static int lofreebuffers;
139SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
140 "XXX Unused");
141static int hifreebuffers;
142SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
143 "XXX Complicatedly unused");
144static int getnewbufcalls;
145SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
146 "Number of calls to getnewbuf");
147static int getnewbufrestarts;
148SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
149 "Number of times getnewbuf has had to restart a buffer aquisition");
150static int dobkgrdwrite = 1;
151SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
152 "Do background writes (honoring the BX_BKGRDWRITE flag)?");
153
154/*
155 * Wakeup point for bufdaemon, as well as indicator of whether it is already
156 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
157 * is idling.
158 */
159static int bd_request;
160
161/*
162 * bogus page -- for I/O to/from partially complete buffers
163 * this is a temporary solution to the problem, but it is not
164 * really that bad. it would be better to split the buffer
165 * for input in the case of buffers partially already in memory,
166 * but the code is intricate enough already.
167 */
168vm_page_t bogus_page;
169
170/*
171 * Offset for bogus_page.
172 * XXX bogus_offset should be local to bufinit
173 */
174static vm_offset_t bogus_offset;
175
176/*
177 * Synchronization (sleep/wakeup) variable for active buffer space requests.
178 * Set when wait starts, cleared prior to wakeup().
179 * Used in runningbufwakeup() and waitrunningbufspace().
180 */
181static int runningbufreq;
182
183/*
184 * Synchronization (sleep/wakeup) variable for buffer requests.
185 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
186 * by and/or.
187 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
188 * getnewbuf(), and getblk().
189 */
190static int needsbuffer;
191
192#ifdef USE_BUFHASH
193/*
194 * Mask for index into the buffer hash table, which needs to be power of 2 in
195 * size. Set in kern_vfs_bio_buffer_alloc.
196 */
197static int bufhashmask;
198
199/*
200 * Hash table for all buffers, with a linked list hanging from each table
201 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init.
202 */
203static LIST_HEAD(bufhashhdr, buf) *bufhashtbl;
204
205/*
206 * Somewhere to store buffers when they are not in another list, to always
207 * have them in a list (and thus being able to use the same set of operations
208 * on them.)
209 */
210static struct bufhashhdr invalhash;
211
212#endif
213
214/*
215 * Definitions for the buffer free lists.
216 */
217#define BUFFER_QUEUES 6 /* number of free buffer queues */
218
219#define QUEUE_NONE 0 /* on no queue */
220#define QUEUE_LOCKED 1 /* locked buffers */
221#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */
222#define QUEUE_DIRTY 3 /* B_DELWRI buffers */
223#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */
224#define QUEUE_EMPTY 5 /* empty buffer headers */
225
226/* Queues for free buffers with various properties */
227static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
228/*
229 * Single global constant for BUF_WMESG, to avoid getting multiple references.
230 * buf_wmesg is referred from macros.
231 */
232const char *buf_wmesg = BUF_WMESG;
233
234#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
235#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
236#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
237#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
238
239#ifdef USE_BUFHASH
240/*
241 * Buffer hash table code. Note that the logical block scans linearly, which
242 * gives us some L1 cache locality.
243 */
244
245static __inline
246struct bufhashhdr *
247bufhash(struct vnode *vnp, daddr_t bn)
248{
249 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]);
250}
251
252#endif
253
254/*
255 * numdirtywakeup:
256 *
257 * If someone is blocked due to there being too many dirty buffers,
258 * and numdirtybuffers is now reasonable, wake them up.
259 */
260
261static __inline void
262numdirtywakeup(int level)
263{
264 if (numdirtybuffers <= level) {
265 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
266 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
267 wakeup(&needsbuffer);
268 }
269 }
270}
271
272/*
273 * bufspacewakeup:
274 *
275 * Called when buffer space is potentially available for recovery.
276 * getnewbuf() will block on this flag when it is unable to free
277 * sufficient buffer space. Buffer space becomes recoverable when
278 * bp's get placed back in the queues.
279 */
280
281static __inline void
282bufspacewakeup(void)
283{
284 /*
285 * If someone is waiting for BUF space, wake them up. Even
286 * though we haven't freed the kva space yet, the waiting
287 * process will be able to now.
288 */
289 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
290 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
291 wakeup(&needsbuffer);
292 }
293}
294
295/*
296 * runningbufwakeup() - in-progress I/O accounting.
297 *
298 */
299static __inline void
300runningbufwakeup(struct buf *bp)
301{
302 if (bp->b_runningbufspace) {
303 runningbufspace -= bp->b_runningbufspace;
304 bp->b_runningbufspace = 0;
305 if (runningbufreq && runningbufspace <= lorunningspace) {
306 runningbufreq = 0;
307 wakeup(&runningbufreq);
308 }
309 }
310}
311
312/*
313 * bufcountwakeup:
314 *
315 * Called when a buffer has been added to one of the free queues to
316 * account for the buffer and to wakeup anyone waiting for free buffers.
317 * This typically occurs when large amounts of metadata are being handled
318 * by the buffer cache ( else buffer space runs out first, usually ).
319 */
320
321static __inline void
322bufcountwakeup(void)
323{
324 ++numfreebuffers;
325 if (needsbuffer) {
326 needsbuffer &= ~VFS_BIO_NEED_ANY;
327 if (numfreebuffers >= hifreebuffers)
328 needsbuffer &= ~VFS_BIO_NEED_FREE;
329 wakeup(&needsbuffer);
330 }
331}
332
333/*
334 * waitrunningbufspace()
335 *
336 * runningbufspace is a measure of the amount of I/O currently
337 * running. This routine is used in async-write situations to
338 * prevent creating huge backups of pending writes to a device.
339 * Only asynchronous writes are governed by this function.
340 *
341 * Reads will adjust runningbufspace, but will not block based on it.
342 * The read load has a side effect of reducing the allowed write load.
343 *
344 * This does NOT turn an async write into a sync write. It waits
345 * for earlier writes to complete and generally returns before the
346 * caller's write has reached the device.
347 */
348static __inline void
349waitrunningbufspace(void)
350{
351 /*
352 * XXX race against wakeup interrupt, currently
353 * protected by Giant. FIXME!
354 */
355 while (runningbufspace > hirunningspace) {
356 ++runningbufreq;
357 tsleep(&runningbufreq, PVM, "wdrain", 0);
358 }
359}
360
361
362/*
363 * vfs_buf_test_cache:
364 *
365 * Called when a buffer is extended. This function clears the B_CACHE
366 * bit if the newly extended portion of the buffer does not contain
367 * valid data.
368 */
369static __inline__
370void
371vfs_buf_test_cache(struct buf *bp,
372 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
373 vm_page_t m)
374{
375 GIANT_REQUIRED;
376
377 if (bp->b_flags & B_CACHE) {
378 int base = (foff + off) & PAGE_MASK;
379 if (vm_page_is_valid(m, base, size) == 0)
380 bp->b_flags &= ~B_CACHE;
381 }
382}
383
384/* Wake up the buffer deamon if necessary */
385static __inline__
386void
387bd_wakeup(int dirtybuflevel)
388{
389 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
390 bd_request = 1;
391 wakeup(&bd_request);
392 }
393}
394
395/*
396 * bd_speedup - speedup the buffer cache flushing code
397 */
398
399static __inline__
400void
401bd_speedup(void)
402{
403 bd_wakeup(1);
404}
405
406/*
407 * Calculating buffer cache scaling values and reserve space for buffer
408 * headers. This is called during low level kernel initialization and
409 * may be called more then once. We CANNOT write to the memory area
410 * being reserved at this time.
411 */
412caddr_t
413kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
414{
415 /*
416 * physmem_est is in pages. Convert it to kilobytes (assumes
417 * PAGE_SIZE is >= 1K)
418 */
419 physmem_est = physmem_est * (PAGE_SIZE / 1024);
420
421 /*
422 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
423 * For the first 64MB of ram nominally allocate sufficient buffers to
424 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
425 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing
426 * the buffer cache we limit the eventual kva reservation to
427 * maxbcache bytes.
428 *
429 * factor represents the 1/4 x ram conversion.
430 */
431 if (nbuf == 0) {
432 int factor = 4 * BKVASIZE / 1024;
433
434 nbuf = 50;
435 if (physmem_est > 4096)
436 nbuf += min((physmem_est - 4096) / factor,
437 65536 / factor);
438 if (physmem_est > 65536)
439 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
440
441 if (maxbcache && nbuf > maxbcache / BKVASIZE)
442 nbuf = maxbcache / BKVASIZE;
443 }
444
445#if 0
446 /*
447 * Do not allow the buffer_map to be more then 1/2 the size of the
448 * kernel_map.
449 */
450 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
451 (BKVASIZE * 2)) {
452 nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
453 (BKVASIZE * 2);
454 printf("Warning: nbufs capped at %d\n", nbuf);
455 }
456#endif
457
458 /*
459 * swbufs are used as temporary holders for I/O, such as paging I/O.
460 * We have no less then 16 and no more then 256.
461 */
462 nswbuf = max(min(nbuf/4, 256), 16);
463
464 /*
465 * Reserve space for the buffer cache buffers
466 */
467 swbuf = (void *)v;
468 v = (caddr_t)(swbuf + nswbuf);
469 buf = (void *)v;
470 v = (caddr_t)(buf + nbuf);
471
472#ifdef USE_BUFHASH
473 /*
474 * Calculate the hash table size and reserve space
475 */
476 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
477 ;
478 bufhashtbl = (void *)v;
479 v = (caddr_t)(bufhashtbl + bufhashmask);
480 --bufhashmask;
481#endif
482 return(v);
483}
484
485/* Initialize the buffer subsystem. Called before use of any buffers. */
486void
487bufinit(void)
488{
489 struct buf *bp;
490 int i;
491
492 GIANT_REQUIRED;
493
494#ifdef USE_BUFHASH
495 LIST_INIT(&invalhash);
496#endif
497 mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF);
498
499#ifdef USE_BUFHASH
500 for (i = 0; i <= bufhashmask; i++)
501 LIST_INIT(&bufhashtbl[i]);
502#endif
503
504 /* next, make a null set of free lists */
505 for (i = 0; i < BUFFER_QUEUES; i++)
506 TAILQ_INIT(&bufqueues[i]);
507
508 /* finally, initialize each buffer header and stick on empty q */
509 for (i = 0; i < nbuf; i++) {
510 bp = &buf[i];
511 bzero(bp, sizeof *bp);
512 bp->b_flags = B_INVAL; /* we're just an empty header */
513 bp->b_dev = NODEV;
514 bp->b_rcred = NOCRED;
515 bp->b_wcred = NOCRED;
516 bp->b_qindex = QUEUE_EMPTY;
517 bp->b_xflags = 0;
518 LIST_INIT(&bp->b_dep);
519 BUF_LOCKINIT(bp);
520 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
521#ifdef USE_BUFHASH
522 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
523#endif
524 }
525
526 /*
527 * maxbufspace is the absolute maximum amount of buffer space we are
528 * allowed to reserve in KVM and in real terms. The absolute maximum
529 * is nominally used by buf_daemon. hibufspace is the nominal maximum
530 * used by most other processes. The differential is required to
531 * ensure that buf_daemon is able to run when other processes might
532 * be blocked waiting for buffer space.
533 *
534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
535 * this may result in KVM fragmentation which is not handled optimally
536 * by the system.
537 */
538 maxbufspace = nbuf * BKVASIZE;
539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
540 lobufspace = hibufspace - MAXBSIZE;
541
542 lorunningspace = 512 * 1024;
543 hirunningspace = 1024 * 1024;
544
545/*
546 * Limit the amount of malloc memory since it is wired permanently into
547 * the kernel space. Even though this is accounted for in the buffer
548 * allocation, we don't want the malloced region to grow uncontrolled.
549 * The malloc scheme improves memory utilization significantly on average
550 * (small) directories.
551 */
552 maxbufmallocspace = hibufspace / 20;
553
554/*
555 * Reduce the chance of a deadlock occuring by limiting the number
556 * of delayed-write dirty buffers we allow to stack up.
557 */
558 hidirtybuffers = nbuf / 4 + 20;
559 numdirtybuffers = 0;
560/*
561 * To support extreme low-memory systems, make sure hidirtybuffers cannot
562 * eat up all available buffer space. This occurs when our minimum cannot
563 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
564 * BKVASIZE'd (8K) buffers.
565 */
566 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
567 hidirtybuffers >>= 1;
568 }
569 lodirtybuffers = hidirtybuffers / 2;
570
571/*
572 * Try to keep the number of free buffers in the specified range,
573 * and give special processes (e.g. like buf_daemon) access to an
574 * emergency reserve.
575 */
576 lofreebuffers = nbuf / 18 + 5;
577 hifreebuffers = 2 * lofreebuffers;
578 numfreebuffers = nbuf;
579
580/*
581 * Maximum number of async ops initiated per buf_daemon loop. This is
582 * somewhat of a hack at the moment, we really need to limit ourselves
583 * based on the number of bytes of I/O in-transit that were initiated
584 * from buf_daemon.
585 */
586
587 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
588 bogus_page = vm_page_alloc(kernel_object,
589 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
590 VM_ALLOC_NORMAL);
591 cnt.v_wire_count++;
592}
593
594/*
595 * bfreekva() - free the kva allocation for a buffer.
596 *
597 * Must be called at splbio() or higher as this is the only locking for
598 * buffer_map.
599 *
600 * Since this call frees up buffer space, we call bufspacewakeup().
601 */
602static void
603bfreekva(struct buf * bp)
604{
605 GIANT_REQUIRED;
606
607 if (bp->b_kvasize) {
608 ++buffreekvacnt;
609 bufspace -= bp->b_kvasize;
610 vm_map_delete(buffer_map,
611 (vm_offset_t) bp->b_kvabase,
612 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
613 );
614 bp->b_kvasize = 0;
615 bufspacewakeup();
616 }
617}
618
619/*
620 * bremfree:
621 *
622 * Remove the buffer from the appropriate free list.
623 */
624void
625bremfree(struct buf * bp)
626{
627 int s = splbio();
628 int old_qindex = bp->b_qindex;
629
630 GIANT_REQUIRED;
631
632 if (bp->b_qindex != QUEUE_NONE) {
633 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
634 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
635 bp->b_qindex = QUEUE_NONE;
636 } else {
637 if (BUF_REFCNT(bp) <= 1)
638 panic("bremfree: removing a buffer not on a queue");
639 }
640
641 /*
642 * Fixup numfreebuffers count. If the buffer is invalid or not
643 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
644 * the buffer was free and we must decrement numfreebuffers.
645 */
646 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
647 switch(old_qindex) {
648 case QUEUE_DIRTY:
649 case QUEUE_CLEAN:
650 case QUEUE_EMPTY:
651 case QUEUE_EMPTYKVA:
652 --numfreebuffers;
653 break;
654 default:
655 break;
656 }
657 }
658 splx(s);
659}
660
661
662/*
663 * Get a buffer with the specified data. Look in the cache first. We
664 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
665 * is set, the buffer is valid and we do not have to do anything ( see
666 * getblk() ). This is really just a special case of breadn().
667 */
668int
669bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
670 struct buf ** bpp)
671{
672
673 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
674}
675
676/*
677 * Operates like bread, but also starts asynchronous I/O on
678 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior
679 * to initiating I/O . If B_CACHE is set, the buffer is valid
680 * and we do not have to do anything.
681 */
682int
683breadn(struct vnode * vp, daddr_t blkno, int size,
684 daddr_t * rablkno, int *rabsize,
685 int cnt, struct ucred * cred, struct buf ** bpp)
686{
687 struct buf *bp, *rabp;
688 int i;
689 int rv = 0, readwait = 0;
690
691 *bpp = bp = getblk(vp, blkno, size, 0, 0);
692
693 /* if not found in cache, do some I/O */
694 if ((bp->b_flags & B_CACHE) == 0) {
695 if (curthread != PCPU_GET(idlethread))
696 curthread->td_proc->p_stats->p_ru.ru_inblock++;
697 bp->b_iocmd = BIO_READ;
698 bp->b_flags &= ~B_INVAL;
699 bp->b_ioflags &= ~BIO_ERROR;
700 if (bp->b_rcred == NOCRED && cred != NOCRED)
701 bp->b_rcred = crhold(cred);
702 vfs_busy_pages(bp, 0);
703 VOP_STRATEGY(vp, bp);
704 ++readwait;
705 }
706
707 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
708 if (inmem(vp, *rablkno))
709 continue;
710 rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
711
712 if ((rabp->b_flags & B_CACHE) == 0) {
713 if (curthread != PCPU_GET(idlethread))
714 curthread->td_proc->p_stats->p_ru.ru_inblock++;
715 rabp->b_flags |= B_ASYNC;
716 rabp->b_flags &= ~B_INVAL;
717 rabp->b_ioflags &= ~BIO_ERROR;
718 rabp->b_iocmd = BIO_READ;
719 if (rabp->b_rcred == NOCRED && cred != NOCRED)
720 rabp->b_rcred = crhold(cred);
721 vfs_busy_pages(rabp, 0);
722 BUF_KERNPROC(rabp);
723 VOP_STRATEGY(vp, rabp);
724 } else {
725 brelse(rabp);
726 }
727 }
728
729 if (readwait) {
730 rv = bufwait(bp);
731 }
732 return (rv);
733}
734
735/*
736 * Write, release buffer on completion. (Done by iodone
737 * if async). Do not bother writing anything if the buffer
738 * is invalid.
739 *
740 * Note that we set B_CACHE here, indicating that buffer is
741 * fully valid and thus cacheable. This is true even of NFS
742 * now so we set it generally. This could be set either here
743 * or in biodone() since the I/O is synchronous. We put it
744 * here.
745 */
746
747int
748bwrite(struct buf * bp)
749{
750 int oldflags, s;
751 struct buf *newbp;
752
753 if (bp->b_flags & B_INVAL) {
754 brelse(bp);
755 return (0);
756 }
757
758 oldflags = bp->b_flags;
759
760 if (BUF_REFCNT(bp) == 0)
761 panic("bwrite: buffer is not busy???");
762 s = splbio();
763 /*
764 * If a background write is already in progress, delay
765 * writing this block if it is asynchronous. Otherwise
766 * wait for the background write to complete.
767 */
768 if (bp->b_xflags & BX_BKGRDINPROG) {
769 if (bp->b_flags & B_ASYNC) {
770 splx(s);
771 bdwrite(bp);
772 return (0);
773 }
774 bp->b_xflags |= BX_BKGRDWAIT;
775 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0);
776 if (bp->b_xflags & BX_BKGRDINPROG)
777 panic("bwrite: still writing");
778 }
779
780 /* Mark the buffer clean */
781 bundirty(bp);
782
783 /*
784 * If this buffer is marked for background writing and we
785 * do not have to wait for it, make a copy and write the
786 * copy so as to leave this buffer ready for further use.
787 *
788 * This optimization eats a lot of memory. If we have a page
789 * or buffer shortfall we can't do it.
790 */
791 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) &&
792 (bp->b_flags & B_ASYNC) &&
793 !vm_page_count_severe() &&
794 !buf_dirty_count_severe()) {
795 if (bp->b_iodone != NULL) {
796 printf("bp->b_iodone = %p\n", bp->b_iodone);
797 panic("bwrite: need chained iodone");
798 }
799
800 /* get a new block */
801 newbp = geteblk(bp->b_bufsize);
802
803 /*
804 * set it to be identical to the old block. We have to
805 * set b_lblkno and BKGRDMARKER before calling bgetvp()
806 * to avoid confusing the splay tree and gbincore().
807 */
808 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
809 newbp->b_lblkno = bp->b_lblkno;
810 newbp->b_xflags |= BX_BKGRDMARKER;
811 bgetvp(bp->b_vp, newbp);
812 newbp->b_blkno = bp->b_blkno;
813 newbp->b_offset = bp->b_offset;
814 newbp->b_iodone = vfs_backgroundwritedone;
815 newbp->b_flags |= B_ASYNC;
816 newbp->b_flags &= ~B_INVAL;
817
818 /* move over the dependencies */
819 if (LIST_FIRST(&bp->b_dep) != NULL)
820 buf_movedeps(bp, newbp);
821
822 /*
823 * Initiate write on the copy, release the original to
824 * the B_LOCKED queue so that it cannot go away until
825 * the background write completes. If not locked it could go
826 * away and then be reconstituted while it was being written.
827 * If the reconstituted buffer were written, we could end up
828 * with two background copies being written at the same time.
829 */
830 bp->b_xflags |= BX_BKGRDINPROG;
831 bp->b_flags |= B_LOCKED;
832 bqrelse(bp);
833 bp = newbp;
834 }
835
836 bp->b_flags &= ~B_DONE;
837 bp->b_ioflags &= ~BIO_ERROR;
838 bp->b_flags |= B_WRITEINPROG | B_CACHE;
839 bp->b_iocmd = BIO_WRITE;
840
841 bp->b_vp->v_numoutput++;
842 vfs_busy_pages(bp, 1);
843
844 /*
845 * Normal bwrites pipeline writes
846 */
847 bp->b_runningbufspace = bp->b_bufsize;
848 runningbufspace += bp->b_runningbufspace;
849
850 if (curthread != PCPU_GET(idlethread))
851 curthread->td_proc->p_stats->p_ru.ru_oublock++;
852 splx(s);
853 if (oldflags & B_ASYNC)
854 BUF_KERNPROC(bp);
855 BUF_STRATEGY(bp);
856
857 if ((oldflags & B_ASYNC) == 0) {
858 int rtval = bufwait(bp);
859 brelse(bp);
860 return (rtval);
861 } else if ((oldflags & B_NOWDRAIN) == 0) {
862 /*
863 * don't allow the async write to saturate the I/O
864 * system. Deadlocks can occur only if a device strategy
865 * routine (like in MD) turns around and issues another
866 * high-level write, in which case B_NOWDRAIN is expected
867 * to be set. Otherwise we will not deadlock here because
868 * we are blocking waiting for I/O that is already in-progress
869 * to complete.
870 */
871 waitrunningbufspace();
872 }
873
874 return (0);
875}
876
877/*
878 * Complete a background write started from bwrite.
879 */
880static void
881vfs_backgroundwritedone(bp)
882 struct buf *bp;
883{
884 struct buf *origbp;
885
886 /*
887 * Find the original buffer that we are writing.
888 */
889 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
890 panic("backgroundwritedone: lost buffer");
891 /*
892 * Process dependencies then return any unfinished ones.
893 */
894 if (LIST_FIRST(&bp->b_dep) != NULL)
895 buf_complete(bp);
896 if (LIST_FIRST(&bp->b_dep) != NULL)
897 buf_movedeps(bp, origbp);
898 /*
899 * Clear the BX_BKGRDINPROG flag in the original buffer
900 * and awaken it if it is waiting for the write to complete.
901 * If BX_BKGRDINPROG is not set in the original buffer it must
902 * have been released and re-instantiated - which is not legal.
903 */
904 KASSERT((origbp->b_xflags & BX_BKGRDINPROG),
905 ("backgroundwritedone: lost buffer2"));
906 origbp->b_xflags &= ~BX_BKGRDINPROG;
907 if (origbp->b_xflags & BX_BKGRDWAIT) {
908 origbp->b_xflags &= ~BX_BKGRDWAIT;
909 wakeup(&origbp->b_xflags);
910 }
911 /*
912 * Clear the B_LOCKED flag and remove it from the locked
913 * queue if it currently resides there.
914 */
915 origbp->b_flags &= ~B_LOCKED;
916 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
917 bremfree(origbp);
918 bqrelse(origbp);
919 }
920 /*
921 * This buffer is marked B_NOCACHE, so when it is released
922 * by biodone, it will be tossed. We mark it with BIO_READ
923 * to avoid biodone doing a second vwakeup.
924 */
925 bp->b_flags |= B_NOCACHE;
926 bp->b_iocmd = BIO_READ;
927 bp->b_flags &= ~(B_CACHE | B_DONE);
928 bp->b_iodone = 0;
929 bufdone(bp);
930}
931
932/*
933 * Delayed write. (Buffer is marked dirty). Do not bother writing
934 * anything if the buffer is marked invalid.
935 *
936 * Note that since the buffer must be completely valid, we can safely
937 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
938 * biodone() in order to prevent getblk from writing the buffer
939 * out synchronously.
940 */
941void
942bdwrite(struct buf * bp)
943{
944 GIANT_REQUIRED;
945
946 if (BUF_REFCNT(bp) == 0)
947 panic("bdwrite: buffer is not busy");
948
949 if (bp->b_flags & B_INVAL) {
950 brelse(bp);
951 return;
952 }
953 bdirty(bp);
954
955 /*
956 * Set B_CACHE, indicating that the buffer is fully valid. This is
957 * true even of NFS now.
958 */
959 bp->b_flags |= B_CACHE;
960
961 /*
962 * This bmap keeps the system from needing to do the bmap later,
963 * perhaps when the system is attempting to do a sync. Since it
964 * is likely that the indirect block -- or whatever other datastructure
965 * that the filesystem needs is still in memory now, it is a good
966 * thing to do this. Note also, that if the pageout daemon is
967 * requesting a sync -- there might not be enough memory to do
968 * the bmap then... So, this is important to do.
969 */
970 if (bp->b_lblkno == bp->b_blkno) {
971 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
972 }
973
974 /*
975 * Set the *dirty* buffer range based upon the VM system dirty pages.
976 */
977 vfs_setdirty(bp);
978
979 /*
980 * We need to do this here to satisfy the vnode_pager and the
981 * pageout daemon, so that it thinks that the pages have been
982 * "cleaned". Note that since the pages are in a delayed write
983 * buffer -- the VFS layer "will" see that the pages get written
984 * out on the next sync, or perhaps the cluster will be completed.
985 */
986 vfs_clean_pages(bp);
987 bqrelse(bp);
988
989 /*
990 * Wakeup the buffer flushing daemon if we have a lot of dirty
991 * buffers (midpoint between our recovery point and our stall
992 * point).
993 */
994 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
995
996 /*
997 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
998 * due to the softdep code.
999 */
1000}
1001
1002/*
1003 * bdirty:
1004 *
1005 * Turn buffer into delayed write request. We must clear BIO_READ and
1006 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1007 * itself to properly update it in the dirty/clean lists. We mark it
1008 * B_DONE to ensure that any asynchronization of the buffer properly
1009 * clears B_DONE ( else a panic will occur later ).
1010 *
1011 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1012 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1013 * should only be called if the buffer is known-good.
1014 *
1015 * Since the buffer is not on a queue, we do not update the numfreebuffers
1016 * count.
1017 *
1018 * Must be called at splbio().
1019 * The buffer must be on QUEUE_NONE.
1020 */
1021void
1022bdirty(bp)
1023 struct buf *bp;
1024{
1025 KASSERT(bp->b_qindex == QUEUE_NONE,
1026 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1027 bp->b_flags &= ~(B_RELBUF);
1028 bp->b_iocmd = BIO_WRITE;
1029
1030 if ((bp->b_flags & B_DELWRI) == 0) {
1031 bp->b_flags |= B_DONE | B_DELWRI;
1032 reassignbuf(bp, bp->b_vp);
1033 ++numdirtybuffers;
1034 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1035 }
1036}
1037
1038/*
1039 * bundirty:
1040 *
1041 * Clear B_DELWRI for buffer.
1042 *
1043 * Since the buffer is not on a queue, we do not update the numfreebuffers
1044 * count.
1045 *
1046 * Must be called at splbio().
1047 * The buffer must be on QUEUE_NONE.
1048 */
1049
1050void
1051bundirty(bp)
1052 struct buf *bp;
1053{
1054 KASSERT(bp->b_qindex == QUEUE_NONE,
1055 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1056
1057 if (bp->b_flags & B_DELWRI) {
1058 bp->b_flags &= ~B_DELWRI;
1059 reassignbuf(bp, bp->b_vp);
1060 --numdirtybuffers;
1061 numdirtywakeup(lodirtybuffers);
1062 }
1063 /*
1064 * Since it is now being written, we can clear its deferred write flag.
1065 */
1066 bp->b_flags &= ~B_DEFERRED;
1067}
1068
1069/*
1070 * bawrite:
1071 *
1072 * Asynchronous write. Start output on a buffer, but do not wait for
1073 * it to complete. The buffer is released when the output completes.
1074 *
1075 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1076 * B_INVAL buffers. Not us.
1077 */
1078void
1079bawrite(struct buf * bp)
1080{
1081 bp->b_flags |= B_ASYNC;
1082 (void) BUF_WRITE(bp);
1083}
1084
1085/*
1086 * bwillwrite:
1087 *
1088 * Called prior to the locking of any vnodes when we are expecting to
1089 * write. We do not want to starve the buffer cache with too many
1090 * dirty buffers so we block here. By blocking prior to the locking
1091 * of any vnodes we attempt to avoid the situation where a locked vnode
1092 * prevents the various system daemons from flushing related buffers.
1093 */
1094
1095void
1096bwillwrite(void)
1097{
1098 if (numdirtybuffers >= hidirtybuffers) {
1099 int s;
1100
1101 mtx_lock(&Giant);
1102 s = splbio();
1103 while (numdirtybuffers >= hidirtybuffers) {
1104 bd_wakeup(1);
1105 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1106 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0);
1107 }
1108 splx(s);
1109 mtx_unlock(&Giant);
1110 }
1111}
1112
1113/*
1114 * Return true if we have too many dirty buffers.
1115 */
1116int
1117buf_dirty_count_severe(void)
1118{
1119 return(numdirtybuffers >= hidirtybuffers);
1120}
1121
1122/*
1123 * brelse:
1124 *
1125 * Release a busy buffer and, if requested, free its resources. The
1126 * buffer will be stashed in the appropriate bufqueue[] allowing it
1127 * to be accessed later as a cache entity or reused for other purposes.
1128 */
1129void
1130brelse(struct buf * bp)
1131{
1132 int s;
1133
1134 GIANT_REQUIRED;
1135
1136 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1137 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1138
1139 s = splbio();
1140
1141 if (bp->b_flags & B_LOCKED)
1142 bp->b_ioflags &= ~BIO_ERROR;
1143
1144 if (bp->b_iocmd == BIO_WRITE &&
1145 (bp->b_ioflags & BIO_ERROR) &&
1146 !(bp->b_flags & B_INVAL)) {
1147 /*
1148 * Failed write, redirty. Must clear BIO_ERROR to prevent
1149 * pages from being scrapped. If B_INVAL is set then
1150 * this case is not run and the next case is run to
1151 * destroy the buffer. B_INVAL can occur if the buffer
1152 * is outside the range supported by the underlying device.
1153 */
1154 bp->b_ioflags &= ~BIO_ERROR;
1155 bdirty(bp);
1156 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1157 (bp->b_ioflags & BIO_ERROR) ||
1158 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) {
1159 /*
1160 * Either a failed I/O or we were asked to free or not
1161 * cache the buffer.
1162 */
1163 bp->b_flags |= B_INVAL;
1164 if (LIST_FIRST(&bp->b_dep) != NULL)
1165 buf_deallocate(bp);
1166 if (bp->b_flags & B_DELWRI) {
1167 --numdirtybuffers;
1168 numdirtywakeup(lodirtybuffers);
1169 }
1170 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1171 if ((bp->b_flags & B_VMIO) == 0) {
1172 if (bp->b_bufsize)
1173 allocbuf(bp, 0);
1174 if (bp->b_vp)
1175 brelvp(bp);
1176 }
1177 }
1178
1179 /*
1180 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1181 * is called with B_DELWRI set, the underlying pages may wind up
1182 * getting freed causing a previous write (bdwrite()) to get 'lost'
1183 * because pages associated with a B_DELWRI bp are marked clean.
1184 *
1185 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1186 * if B_DELWRI is set.
1187 *
1188 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1189 * on pages to return pages to the VM page queues.
1190 */
1191 if (bp->b_flags & B_DELWRI)
1192 bp->b_flags &= ~B_RELBUF;
1193 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1194 bp->b_flags |= B_RELBUF;
1195
1196 /*
1197 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1198 * constituted, not even NFS buffers now. Two flags effect this. If
1199 * B_INVAL, the struct buf is invalidated but the VM object is kept
1200 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1201 *
1202 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1203 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1204 * buffer is also B_INVAL because it hits the re-dirtying code above.
1205 *
1206 * Normally we can do this whether a buffer is B_DELWRI or not. If
1207 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1208 * the commit state and we cannot afford to lose the buffer. If the
1209 * buffer has a background write in progress, we need to keep it
1210 * around to prevent it from being reconstituted and starting a second
1211 * background write.
1212 */
1213 if ((bp->b_flags & B_VMIO)
|
1215 !vn_isdisk(bp->b_vp, NULL) &&
1216 (bp->b_flags & B_DELWRI))
1217 ) {
1218
1219 int i, j, resid;
1220 vm_page_t m;
1221 off_t foff;
1222 vm_pindex_t poff;
1223 vm_object_t obj;
1224 struct vnode *vp;
1225
1226 vp = bp->b_vp;
1227 obj = bp->b_object;
1228
1229 /*
1230 * Get the base offset and length of the buffer. Note that
1231 * in the VMIO case if the buffer block size is not
1232 * page-aligned then b_data pointer may not be page-aligned.
1233 * But our b_pages[] array *IS* page aligned.
1234 *
1235 * block sizes less then DEV_BSIZE (usually 512) are not
1236 * supported due to the page granularity bits (m->valid,
1237 * m->dirty, etc...).
1238 *
1239 * See man buf(9) for more information
1240 */
1241 resid = bp->b_bufsize;
1242 foff = bp->b_offset;
1243
1244 for (i = 0; i < bp->b_npages; i++) {
1245 int had_bogus = 0;
1246
1247 m = bp->b_pages[i];
1248 vm_page_flag_clear(m, PG_ZERO);
1249
1250 /*
1251 * If we hit a bogus page, fixup *all* the bogus pages
1252 * now.
1253 */
1254 if (m == bogus_page) {
1255 poff = OFF_TO_IDX(bp->b_offset);
1256 had_bogus = 1;
1257
1258 for (j = i; j < bp->b_npages; j++) {
1259 vm_page_t mtmp;
1260 mtmp = bp->b_pages[j];
1261 if (mtmp == bogus_page) {
1262 mtmp = vm_page_lookup(obj, poff + j);
1263 if (!mtmp) {
1264 panic("brelse: page missing\n");
1265 }
1266 bp->b_pages[j] = mtmp;
1267 }
1268 }
1269
1270 if ((bp->b_flags & B_INVAL) == 0) {
1271 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1272 }
1273 m = bp->b_pages[i];
1274 }
1275 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) {
1276 int poffset = foff & PAGE_MASK;
1277 int presid = resid > (PAGE_SIZE - poffset) ?
1278 (PAGE_SIZE - poffset) : resid;
1279
1280 KASSERT(presid >= 0, ("brelse: extra page"));
1281 vm_page_set_invalid(m, poffset, presid);
1282 if (had_bogus)
1283 printf("avoided corruption bug in bogus_page/brelse code\n");
1284 }
1285 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1286 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1287 }
1288
1289 if (bp->b_flags & (B_INVAL | B_RELBUF))
1290 vfs_vmio_release(bp);
1291
1292 } else if (bp->b_flags & B_VMIO) {
1293
1294 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1295 vfs_vmio_release(bp);
1296 }
1297
1298 }
1299
1300 if (bp->b_qindex != QUEUE_NONE)
1301 panic("brelse: free buffer onto another queue???");
1302 if (BUF_REFCNT(bp) > 1) {
1303 /* do not release to free list */
1304 BUF_UNLOCK(bp);
1305 splx(s);
1306 return;
1307 }
1308
1309 /* enqueue */
1310
1311 /* buffers with no memory */
1312 if (bp->b_bufsize == 0) {
1313 bp->b_flags |= B_INVAL;
1314 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1315 if (bp->b_xflags & BX_BKGRDINPROG)
1316 panic("losing buffer 1");
1317 if (bp->b_kvasize) {
1318 bp->b_qindex = QUEUE_EMPTYKVA;
1319 } else {
1320 bp->b_qindex = QUEUE_EMPTY;
1321 }
1322 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1323#ifdef USE_BUFHASH
1324 LIST_REMOVE(bp, b_hash);
1325 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1326#endif
1327 bp->b_dev = NODEV;
1328 /* buffers with junk contents */
1329 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1330 (bp->b_ioflags & BIO_ERROR)) {
1331 bp->b_flags |= B_INVAL;
1332 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1333 if (bp->b_xflags & BX_BKGRDINPROG)
1334 panic("losing buffer 2");
1335 bp->b_qindex = QUEUE_CLEAN;
1336 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1337#ifdef USE_BUFHASH
1338 LIST_REMOVE(bp, b_hash);
1339 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1340#endif
1341 bp->b_dev = NODEV;
1342
1343 /* buffers that are locked */
1344 } else if (bp->b_flags & B_LOCKED) {
1345 bp->b_qindex = QUEUE_LOCKED;
1346 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1347
1348 /* remaining buffers */
1349 } else {
1350 if (bp->b_flags & B_DELWRI)
1351 bp->b_qindex = QUEUE_DIRTY;
1352 else
1353 bp->b_qindex = QUEUE_CLEAN;
1354 if (bp->b_flags & B_AGE)
1355 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1356 else
1357 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1358 }
1359
1360 /*
1361 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already
1362 * placed the buffer on the correct queue. We must also disassociate
1363 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1364 * find it.
1365 */
1366 if (bp->b_flags & B_INVAL) {
1367 if (bp->b_flags & B_DELWRI)
1368 bundirty(bp);
1369 if (bp->b_vp)
1370 brelvp(bp);
1371 }
1372
1373 /*
1374 * Fixup numfreebuffers count. The bp is on an appropriate queue
1375 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1376 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1377 * if B_INVAL is set ).
1378 */
1379
1380 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1381 bufcountwakeup();
1382
1383 /*
1384 * Something we can maybe free or reuse
1385 */
1386 if (bp->b_bufsize || bp->b_kvasize)
1387 bufspacewakeup();
1388
1389 /* unlock */
1390 BUF_UNLOCK(bp);
1391 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1392 B_DIRECT | B_NOWDRAIN);
1393 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1394 panic("brelse: not dirty");
1395 splx(s);
1396}
1397
1398/*
1399 * Release a buffer back to the appropriate queue but do not try to free
1400 * it. The buffer is expected to be used again soon.
1401 *
1402 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1403 * biodone() to requeue an async I/O on completion. It is also used when
1404 * known good buffers need to be requeued but we think we may need the data
1405 * again soon.
1406 *
1407 * XXX we should be able to leave the B_RELBUF hint set on completion.
1408 */
1409void
1410bqrelse(struct buf * bp)
1411{
1412 int s;
1413
1414 s = splbio();
1415
1416 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1417
1418 if (bp->b_qindex != QUEUE_NONE)
1419 panic("bqrelse: free buffer onto another queue???");
1420 if (BUF_REFCNT(bp) > 1) {
1421 /* do not release to free list */
1422 BUF_UNLOCK(bp);
1423 splx(s);
1424 return;
1425 }
1426 if (bp->b_flags & B_LOCKED) {
1427 bp->b_ioflags &= ~BIO_ERROR;
1428 bp->b_qindex = QUEUE_LOCKED;
1429 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1430 /* buffers with stale but valid contents */
1431 } else if (bp->b_flags & B_DELWRI) {
1432 bp->b_qindex = QUEUE_DIRTY;
1433 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1434 } else if (vm_page_count_severe()) {
1435 /*
1436 * We are too low on memory, we have to try to free the
1437 * buffer (most importantly: the wired pages making up its
1438 * backing store) *now*.
1439 */
1440 splx(s);
1441 brelse(bp);
1442 return;
1443 } else {
1444 bp->b_qindex = QUEUE_CLEAN;
1445 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1446 }
1447
1448 if ((bp->b_flags & B_LOCKED) == 0 &&
1449 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1450 bufcountwakeup();
1451 }
1452
1453 /*
1454 * Something we can maybe free or reuse.
1455 */
1456 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1457 bufspacewakeup();
1458
1459 /* unlock */
1460 BUF_UNLOCK(bp);
1461 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1462 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1463 panic("bqrelse: not dirty");
1464 splx(s);
1465}
1466
1467/* Give pages used by the bp back to the VM system (where possible) */
1468static void
1469vfs_vmio_release(bp)
1470 struct buf *bp;
1471{
1472 int i;
1473 vm_page_t m;
1474
1475 GIANT_REQUIRED;
1476 vm_page_lock_queues();
1477 for (i = 0; i < bp->b_npages; i++) {
1478 m = bp->b_pages[i];
1479 bp->b_pages[i] = NULL;
1480 /*
1481 * In order to keep page LRU ordering consistent, put
1482 * everything on the inactive queue.
1483 */
1484 vm_page_unwire(m, 0);
1485 /*
1486 * We don't mess with busy pages, it is
1487 * the responsibility of the process that
1488 * busied the pages to deal with them.
1489 */
1490 if ((m->flags & PG_BUSY) || (m->busy != 0))
1491 continue;
1492
1493 if (m->wire_count == 0) {
1494 vm_page_flag_clear(m, PG_ZERO);
1495 /*
1496 * Might as well free the page if we can and it has
1497 * no valid data. We also free the page if the
1498 * buffer was used for direct I/O
1499 */
1500 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1501 m->hold_count == 0) {
1502 vm_page_busy(m);
1503 vm_page_protect(m, VM_PROT_NONE);
1504 vm_page_free(m);
1505 } else if (bp->b_flags & B_DIRECT) {
1506 vm_page_try_to_free(m);
1507 } else if (vm_page_count_severe()) {
1508 vm_page_try_to_cache(m);
1509 }
1510 }
1511 }
1512 vm_page_unlock_queues();
1513 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1514
1515 if (bp->b_bufsize) {
1516 bufspacewakeup();
1517 bp->b_bufsize = 0;
1518 }
1519 bp->b_npages = 0;
1520 bp->b_flags &= ~B_VMIO;
1521 if (bp->b_vp)
1522 brelvp(bp);
1523}
1524
1525#ifdef USE_BUFHASH
1526/*
1527 * XXX MOVED TO VFS_SUBR.C
1528 *
1529 * Check to see if a block is currently memory resident.
1530 */
1531struct buf *
1532gbincore(struct vnode * vp, daddr_t blkno)
1533{
1534 struct buf *bp;
1535 struct bufhashhdr *bh;
1536
1537 bh = bufhash(vp, blkno);
1538
1539 /* Search hash chain */
1540 LIST_FOREACH(bp, bh, b_hash) {
1541 /* hit */
1542 if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1543 (bp->b_flags & B_INVAL) == 0) {
1544 break;
1545 }
1546 }
1547 return (bp);
1548}
1549#endif
1550
1551/*
1552 * vfs_bio_awrite:
1553 *
1554 * Implement clustered async writes for clearing out B_DELWRI buffers.
1555 * This is much better then the old way of writing only one buffer at
1556 * a time. Note that we may not be presented with the buffers in the
1557 * correct order, so we search for the cluster in both directions.
1558 */
1559int
1560vfs_bio_awrite(struct buf * bp)
1561{
1562 int i;
1563 int j;
1564 daddr_t lblkno = bp->b_lblkno;
1565 struct vnode *vp = bp->b_vp;
1566 int s;
1567 int ncl;
1568 struct buf *bpa;
1569 int nwritten;
1570 int size;
1571 int maxcl;
1572
1573 s = splbio();
1574 /*
1575 * right now we support clustered writing only to regular files. If
1576 * we find a clusterable block we could be in the middle of a cluster
1577 * rather then at the beginning.
1578 */
1579 if ((vp->v_type == VREG) &&
1580 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1581 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1582
1583 size = vp->v_mount->mnt_stat.f_iosize;
1584 maxcl = MAXPHYS / size;
1585
1586 for (i = 1; i < maxcl; i++) {
1587 if ((bpa = gbincore(vp, lblkno + i)) &&
1588 BUF_REFCNT(bpa) == 0 &&
1589 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1590 (B_DELWRI | B_CLUSTEROK)) &&
1591 (bpa->b_bufsize == size)) {
1592 if ((bpa->b_blkno == bpa->b_lblkno) ||
1593 (bpa->b_blkno !=
1594 bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1595 break;
1596 } else {
1597 break;
1598 }
1599 }
1600 for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1601 if ((bpa = gbincore(vp, lblkno - j)) &&
1602 BUF_REFCNT(bpa) == 0 &&
1603 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1604 (B_DELWRI | B_CLUSTEROK)) &&
1605 (bpa->b_bufsize == size)) {
1606 if ((bpa->b_blkno == bpa->b_lblkno) ||
1607 (bpa->b_blkno !=
1608 bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1609 break;
1610 } else {
1611 break;
1612 }
1613 }
1614 --j;
1615 ncl = i + j;
1616 /*
1617 * this is a possible cluster write
1618 */
1619 if (ncl != 1) {
1620 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1621 splx(s);
1622 return nwritten;
1623 }
1624 }
1625
1626 BUF_LOCK(bp, LK_EXCLUSIVE);
1627 bremfree(bp);
1628 bp->b_flags |= B_ASYNC;
1629
1630 splx(s);
1631 /*
1632 * default (old) behavior, writing out only one block
1633 *
1634 * XXX returns b_bufsize instead of b_bcount for nwritten?
1635 */
1636 nwritten = bp->b_bufsize;
1637 (void) BUF_WRITE(bp);
1638
1639 return nwritten;
1640}
1641
1642/*
1643 * getnewbuf:
1644 *
1645 * Find and initialize a new buffer header, freeing up existing buffers
1646 * in the bufqueues as necessary. The new buffer is returned locked.
1647 *
1648 * Important: B_INVAL is not set. If the caller wishes to throw the
1649 * buffer away, the caller must set B_INVAL prior to calling brelse().
1650 *
1651 * We block if:
1652 * We have insufficient buffer headers
1653 * We have insufficient buffer space
1654 * buffer_map is too fragmented ( space reservation fails )
1655 * If we have to flush dirty buffers ( but we try to avoid this )
1656 *
1657 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1658 * Instead we ask the buf daemon to do it for us. We attempt to
1659 * avoid piecemeal wakeups of the pageout daemon.
1660 */
1661
1662static struct buf *
1663getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1664{
1665 struct buf *bp;
1666 struct buf *nbp;
1667 int defrag = 0;
1668 int nqindex;
1669 static int flushingbufs;
1670
1671 GIANT_REQUIRED;
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 ++getnewbufcalls;
1681 --getnewbufrestarts;
1682restart:
1683 ++getnewbufrestarts;
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 nqindex = QUEUE_EMPTYKVA;
1698 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1699
1700 if (nbp == NULL) {
1701 /*
1702 * If no EMPTYKVA buffers and we are either
1703 * defragging or reusing, locate a CLEAN buffer
1704 * to free or reuse. If bufspace useage is low
1705 * skip this step so we can allocate a new buffer.
1706 */
1707 if (defrag || bufspace >= lobufspace) {
1708 nqindex = QUEUE_CLEAN;
1709 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1710 }
1711
1712 /*
1713 * If we could not find or were not allowed to reuse a
1714 * CLEAN buffer, check to see if it is ok to use an EMPTY
1715 * buffer. We can only use an EMPTY buffer if allocating
1716 * its KVA would not otherwise run us out of buffer space.
1717 */
1718 if (nbp == NULL && defrag == 0 &&
1719 bufspace + maxsize < hibufspace) {
1720 nqindex = QUEUE_EMPTY;
1721 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1722 }
1723 }
1724
1725 /*
1726 * Run scan, possibly freeing data and/or kva mappings on the fly
1727 * depending.
1728 */
1729
1730 while ((bp = nbp) != NULL) {
1731 int qindex = nqindex;
1732
1733 /*
1734 * Calculate next bp ( we can only use it if we do not block
1735 * or do other fancy things ).
1736 */
1737 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1738 switch(qindex) {
1739 case QUEUE_EMPTY:
1740 nqindex = QUEUE_EMPTYKVA;
1741 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1742 break;
1743 /* FALLTHROUGH */
1744 case QUEUE_EMPTYKVA:
1745 nqindex = QUEUE_CLEAN;
1746 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1747 break;
1748 /* FALLTHROUGH */
1749 case QUEUE_CLEAN:
1750 /*
1751 * nbp is NULL.
1752 */
1753 break;
1754 }
1755 }
1756
1757 /*
1758 * Sanity Checks
1759 */
1760 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1761
1762 /*
1763 * Note: we no longer distinguish between VMIO and non-VMIO
1764 * buffers.
1765 */
1766
1767 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1768
1769 /*
1770 * If we are defragging then we need a buffer with
1771 * b_kvasize != 0. XXX this situation should no longer
1772 * occur, if defrag is non-zero the buffer's b_kvasize
1773 * should also be non-zero at this point. XXX
1774 */
1775 if (defrag && bp->b_kvasize == 0) {
1776 printf("Warning: defrag empty buffer %p\n", bp);
1777 continue;
1778 }
1779
1780 /*
1781 * Start freeing the bp. This is somewhat involved. nbp
1782 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1783 */
1784
1785 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1786 panic("getnewbuf: locked buf");
1787 bremfree(bp);
1788
1789 if (qindex == QUEUE_CLEAN) {
1790 if (bp->b_flags & B_VMIO) {
1791 bp->b_flags &= ~B_ASYNC;
1792 vfs_vmio_release(bp);
1793 }
1794 if (bp->b_vp)
1795 brelvp(bp);
1796 }
1797
1798 /*
1799 * NOTE: nbp is now entirely invalid. We can only restart
1800 * the scan from this point on.
1801 *
1802 * Get the rest of the buffer freed up. b_kva* is still
1803 * valid after this operation.
1804 */
1805
1806 if (bp->b_rcred != NOCRED) {
1807 crfree(bp->b_rcred);
1808 bp->b_rcred = NOCRED;
1809 }
1810 if (bp->b_wcred != NOCRED) {
1811 crfree(bp->b_wcred);
1812 bp->b_wcred = NOCRED;
1813 }
1814 if (LIST_FIRST(&bp->b_dep) != NULL)
1815 buf_deallocate(bp);
1816 if (bp->b_xflags & BX_BKGRDINPROG)
1817 panic("losing buffer 3");
1818#ifdef USE_BUFHASH
1819 LIST_REMOVE(bp, b_hash);
1820 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1821#endif
1822
1823 if (bp->b_bufsize)
1824 allocbuf(bp, 0);
1825
1826 bp->b_flags = 0;
1827 bp->b_ioflags = 0;
1828 bp->b_xflags = 0;
1829 bp->b_dev = NODEV;
1830 bp->b_vp = NULL;
1831 bp->b_blkno = bp->b_lblkno = 0;
1832 bp->b_offset = NOOFFSET;
1833 bp->b_iodone = 0;
1834 bp->b_error = 0;
1835 bp->b_resid = 0;
1836 bp->b_bcount = 0;
1837 bp->b_npages = 0;
1838 bp->b_dirtyoff = bp->b_dirtyend = 0;
1839 bp->b_magic = B_MAGIC_BIO;
1840 bp->b_op = &buf_ops_bio;
1841 bp->b_object = NULL;
1842
1843 LIST_INIT(&bp->b_dep);
1844
1845 /*
1846 * If we are defragging then free the buffer.
1847 */
1848 if (defrag) {
1849 bp->b_flags |= B_INVAL;
1850 bfreekva(bp);
1851 brelse(bp);
1852 defrag = 0;
1853 goto restart;
1854 }
1855
1856 /*
1857 * If we are overcomitted then recover the buffer and its
1858 * KVM space. This occurs in rare situations when multiple
1859 * processes are blocked in getnewbuf() or allocbuf().
1860 */
1861 if (bufspace >= hibufspace)
1862 flushingbufs = 1;
1863 if (flushingbufs && bp->b_kvasize != 0) {
1864 bp->b_flags |= B_INVAL;
1865 bfreekva(bp);
1866 brelse(bp);
1867 goto restart;
1868 }
1869 if (bufspace < lobufspace)
1870 flushingbufs = 0;
1871 break;
1872 }
1873
1874 /*
1875 * If we exhausted our list, sleep as appropriate. We may have to
1876 * wakeup various daemons and write out some dirty buffers.
1877 *
1878 * Generally we are sleeping due to insufficient buffer space.
1879 */
1880
1881 if (bp == NULL) {
1882 int flags;
1883 char *waitmsg;
1884
1885 if (defrag) {
1886 flags = VFS_BIO_NEED_BUFSPACE;
1887 waitmsg = "nbufkv";
1888 } else if (bufspace >= hibufspace) {
1889 waitmsg = "nbufbs";
1890 flags = VFS_BIO_NEED_BUFSPACE;
1891 } else {
1892 waitmsg = "newbuf";
1893 flags = VFS_BIO_NEED_ANY;
1894 }
1895
1896 bd_speedup(); /* heeeelp */
1897
1898 needsbuffer |= flags;
1899 while (needsbuffer & flags) {
1900 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag,
1901 waitmsg, slptimeo))
1902 return (NULL);
1903 }
1904 } else {
1905 /*
1906 * We finally have a valid bp. We aren't quite out of the
1907 * woods, we still have to reserve kva space. In order
1908 * to keep fragmentation sane we only allocate kva in
1909 * BKVASIZE chunks.
1910 */
1911 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1912
1913 if (maxsize != bp->b_kvasize) {
1914 vm_offset_t addr = 0;
1915
1916 bfreekva(bp);
1917
1918 if (vm_map_findspace(buffer_map,
1919 vm_map_min(buffer_map), maxsize, &addr)) {
1920 /*
1921 * Uh oh. Buffer map is to fragmented. We
1922 * must defragment the map.
1923 */
1924 ++bufdefragcnt;
1925 defrag = 1;
1926 bp->b_flags |= B_INVAL;
1927 brelse(bp);
1928 goto restart;
1929 }
1930 if (addr) {
1931 vm_map_insert(buffer_map, NULL, 0,
1932 addr, addr + maxsize,
1933 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1934
1935 bp->b_kvabase = (caddr_t) addr;
1936 bp->b_kvasize = maxsize;
1937 bufspace += bp->b_kvasize;
1938 ++bufreusecnt;
1939 }
1940 }
1941 bp->b_data = bp->b_kvabase;
1942 }
1943 return(bp);
1944}
1945
1946/*
1947 * buf_daemon:
1948 *
1949 * buffer flushing daemon. Buffers are normally flushed by the
1950 * update daemon but if it cannot keep up this process starts to
1951 * take the load in an attempt to prevent getnewbuf() from blocking.
1952 */
1953
1954static struct proc *bufdaemonproc;
1955
1956static struct kproc_desc buf_kp = {
1957 "bufdaemon",
1958 buf_daemon,
1959 &bufdaemonproc
1960};
1961SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1962
1963static void
1964buf_daemon()
1965{
1966 int s;
1967
1968 mtx_lock(&Giant);
1969
1970 /*
1971 * This process needs to be suspended prior to shutdown sync.
1972 */
1973 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
1974 SHUTDOWN_PRI_LAST);
1975
1976 /*
1977 * This process is allowed to take the buffer cache to the limit
1978 */
1979 s = splbio();
1980
1981 for (;;) {
1982 kthread_suspend_check(bufdaemonproc);
1983
1984 bd_request = 0;
1985
1986 /*
1987 * Do the flush. Limit the amount of in-transit I/O we
1988 * allow to build up, otherwise we would completely saturate
1989 * the I/O system. Wakeup any waiting processes before we
1990 * normally would so they can run in parallel with our drain.
1991 */
1992 while (numdirtybuffers > lodirtybuffers) {
1993 if (flushbufqueues() == 0)
1994 break;
1995 waitrunningbufspace();
1996 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1997 }
1998
1999 /*
2000 * Only clear bd_request if we have reached our low water
2001 * mark. The buf_daemon normally waits 1 second and
2002 * then incrementally flushes any dirty buffers that have
2003 * built up, within reason.
2004 *
2005 * If we were unable to hit our low water mark and couldn't
2006 * find any flushable buffers, we sleep half a second.
2007 * Otherwise we loop immediately.
2008 */
2009 if (numdirtybuffers <= lodirtybuffers) {
2010 /*
2011 * We reached our low water mark, reset the
2012 * request and sleep until we are needed again.
2013 * The sleep is just so the suspend code works.
2014 */
2015 bd_request = 0;
2016 tsleep(&bd_request, PVM, "psleep", hz);
2017 } else {
2018 /*
2019 * We couldn't find any flushable dirty buffers but
2020 * still have too many dirty buffers, we
2021 * have to sleep and try again. (rare)
2022 */
2023 tsleep(&bd_request, PVM, "qsleep", hz / 2);
2024 }
2025 }
2026}
2027
2028/*
2029 * flushbufqueues:
2030 *
2031 * Try to flush a buffer in the dirty queue. We must be careful to
2032 * free up B_INVAL buffers instead of write them, which NFS is
2033 * particularly sensitive to.
2034 */
2035
2036static int
2037flushbufqueues(void)
2038{
2039 struct buf *bp;
2040 int r = 0;
2041
2042 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
2043
2044 while (bp) {
2045 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
2046 if ((bp->b_flags & B_DELWRI) != 0 &&
2047 (bp->b_xflags & BX_BKGRDINPROG) == 0) {
2048 if (bp->b_flags & B_INVAL) {
2049 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2050 panic("flushbufqueues: locked buf");
2051 bremfree(bp);
2052 brelse(bp);
2053 ++r;
2054 break;
2055 }
2056 if (LIST_FIRST(&bp->b_dep) != NULL &&
2057 (bp->b_flags & B_DEFERRED) == 0 &&
2058 buf_countdeps(bp, 0)) {
2059 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
2060 bp, b_freelist);
2061 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
2062 bp, b_freelist);
2063 bp->b_flags |= B_DEFERRED;
2064 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
2065 continue;
2066 }
2067 vfs_bio_awrite(bp);
2068 ++r;
2069 break;
2070 }
2071 bp = TAILQ_NEXT(bp, b_freelist);
2072 }
2073 return (r);
2074}
2075
2076/*
2077 * Check to see if a block is currently memory resident.
2078 */
2079struct buf *
2080incore(struct vnode * vp, daddr_t blkno)
2081{
2082 struct buf *bp;
2083
2084 int s = splbio();
2085 bp = gbincore(vp, blkno);
2086 splx(s);
2087 return (bp);
2088}
2089
2090/*
2091 * Returns true if no I/O is needed to access the
2092 * associated VM object. This is like incore except
2093 * it also hunts around in the VM system for the data.
2094 */
2095
2096int
2097inmem(struct vnode * vp, daddr_t blkno)
2098{
2099 vm_object_t obj;
2100 vm_offset_t toff, tinc, size;
2101 vm_page_t m;
2102 vm_ooffset_t off;
2103
2104 GIANT_REQUIRED;
2105
2106 if (incore(vp, blkno))
2107 return 1;
2108 if (vp->v_mount == NULL)
2109 return 0;
2110 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0)
2111 return 0;
2112
2113 size = PAGE_SIZE;
2114 if (size > vp->v_mount->mnt_stat.f_iosize)
2115 size = vp->v_mount->mnt_stat.f_iosize;
2116 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2117
2118 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2119 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2120 if (!m)
2121 goto notinmem;
2122 tinc = size;
2123 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2124 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2125 if (vm_page_is_valid(m,
2126 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2127 goto notinmem;
2128 }
2129 return 1;
2130
2131notinmem:
2132 return (0);
2133}
2134
2135/*
2136 * vfs_setdirty:
2137 *
2138 * Sets the dirty range for a buffer based on the status of the dirty
2139 * bits in the pages comprising the buffer.
2140 *
2141 * The range is limited to the size of the buffer.
2142 *
2143 * This routine is primarily used by NFS, but is generalized for the
2144 * B_VMIO case.
2145 */
2146static void
2147vfs_setdirty(struct buf *bp)
2148{
2149 int i;
2150 vm_object_t object;
2151
2152 GIANT_REQUIRED;
2153 /*
2154 * Degenerate case - empty buffer
2155 */
2156
2157 if (bp->b_bufsize == 0)
2158 return;
2159
2160 /*
2161 * We qualify the scan for modified pages on whether the
2162 * object has been flushed yet. The OBJ_WRITEABLE flag
2163 * is not cleared simply by protecting pages off.
2164 */
2165
2166 if ((bp->b_flags & B_VMIO) == 0)
2167 return;
2168
2169 object = bp->b_pages[0]->object;
2170
2171 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2172 printf("Warning: object %p writeable but not mightbedirty\n", object);
2173 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2174 printf("Warning: object %p mightbedirty but not writeable\n", object);
2175
2176 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2177 vm_offset_t boffset;
2178 vm_offset_t eoffset;
2179
2180 /*
2181 * test the pages to see if they have been modified directly
2182 * by users through the VM system.
2183 */
2184 for (i = 0; i < bp->b_npages; i++) {
2185 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
2186 vm_page_test_dirty(bp->b_pages[i]);
2187 }
2188
2189 /*
2190 * Calculate the encompassing dirty range, boffset and eoffset,
2191 * (eoffset - boffset) bytes.
2192 */
2193
2194 for (i = 0; i < bp->b_npages; i++) {
2195 if (bp->b_pages[i]->dirty)
2196 break;
2197 }
2198 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2199
2200 for (i = bp->b_npages - 1; i >= 0; --i) {
2201 if (bp->b_pages[i]->dirty) {
2202 break;
2203 }
2204 }
2205 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2206
2207 /*
2208 * Fit it to the buffer.
2209 */
2210
2211 if (eoffset > bp->b_bcount)
2212 eoffset = bp->b_bcount;
2213
2214 /*
2215 * If we have a good dirty range, merge with the existing
2216 * dirty range.
2217 */
2218
2219 if (boffset < eoffset) {
2220 if (bp->b_dirtyoff > boffset)
2221 bp->b_dirtyoff = boffset;
2222 if (bp->b_dirtyend < eoffset)
2223 bp->b_dirtyend = eoffset;
2224 }
2225 }
2226}
2227
2228/*
2229 * getblk:
2230 *
2231 * Get a block given a specified block and offset into a file/device.
2232 * The buffers B_DONE bit will be cleared on return, making it almost
2233 * ready for an I/O initiation. B_INVAL may or may not be set on
2234 * return. The caller should clear B_INVAL prior to initiating a
2235 * READ.
2236 *
2237 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2238 * an existing buffer.
2239 *
2240 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2241 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2242 * and then cleared based on the backing VM. If the previous buffer is
2243 * non-0-sized but invalid, B_CACHE will be cleared.
2244 *
2245 * If getblk() must create a new buffer, the new buffer is returned with
2246 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2247 * case it is returned with B_INVAL clear and B_CACHE set based on the
2248 * backing VM.
2249 *
2250 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos
2251 * B_CACHE bit is clear.
2252 *
2253 * What this means, basically, is that the caller should use B_CACHE to
2254 * determine whether the buffer is fully valid or not and should clear
2255 * B_INVAL prior to issuing a read. If the caller intends to validate
2256 * the buffer by loading its data area with something, the caller needs
2257 * to clear B_INVAL. If the caller does this without issuing an I/O,
2258 * the caller should set B_CACHE ( as an optimization ), else the caller
2259 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2260 * a write attempt or if it was a successfull read. If the caller
2261 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2262 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2263 */
2264struct buf *
2265getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2266{
2267 struct buf *bp;
2268 int s;
2269#ifdef USE_BUFHASH
2270 struct bufhashhdr *bh;
2271#endif
2272
2273 if (size > MAXBSIZE)
2274 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2275
2276 s = splbio();
2277loop:
2278 /*
2279 * Block if we are low on buffers. Certain processes are allowed
2280 * to completely exhaust the buffer cache.
2281 *
2282 * If this check ever becomes a bottleneck it may be better to
2283 * move it into the else, when gbincore() fails. At the moment
2284 * it isn't a problem.
2285 *
2286 * XXX remove if 0 sections (clean this up after its proven)
2287 */
2288 if (numfreebuffers == 0) {
2289 if (curthread == PCPU_GET(idlethread))
2290 return NULL;
2291 needsbuffer |= VFS_BIO_NEED_ANY;
2292 }
2293
2294 if ((bp = gbincore(vp, blkno))) {
2295 /*
2296 * Buffer is in-core. If the buffer is not busy, it must
2297 * be on a queue.
2298 */
2299
2300 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2301 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2302 "getblk", slpflag, slptimeo) == ENOLCK)
2303 goto loop;
2304 splx(s);
2305 return (struct buf *) NULL;
2306 }
2307
2308 /*
2309 * The buffer is locked. B_CACHE is cleared if the buffer is
2310 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2311 * and for a VMIO buffer B_CACHE is adjusted according to the
2312 * backing VM cache.
2313 */
2314 if (bp->b_flags & B_INVAL)
2315 bp->b_flags &= ~B_CACHE;
2316 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2317 bp->b_flags |= B_CACHE;
2318 bremfree(bp);
2319
2320 /*
2321 * check for size inconsistancies for non-VMIO case.
2322 */
2323
2324 if (bp->b_bcount != size) {
2325 if ((bp->b_flags & B_VMIO) == 0 ||
2326 (size > bp->b_kvasize)) {
2327 if (bp->b_flags & B_DELWRI) {
2328 bp->b_flags |= B_NOCACHE;
2329 BUF_WRITE(bp);
2330 } else {
2331 if ((bp->b_flags & B_VMIO) &&
2332 (LIST_FIRST(&bp->b_dep) == NULL)) {
2333 bp->b_flags |= B_RELBUF;
2334 brelse(bp);
2335 } else {
2336 bp->b_flags |= B_NOCACHE;
2337 BUF_WRITE(bp);
2338 }
2339 }
2340 goto loop;
2341 }
2342 }
2343
2344 /*
2345 * If the size is inconsistant in the VMIO case, we can resize
2346 * the buffer. This might lead to B_CACHE getting set or
2347 * cleared. If the size has not changed, B_CACHE remains
2348 * unchanged from its previous state.
2349 */
2350
2351 if (bp->b_bcount != size)
2352 allocbuf(bp, size);
2353
2354 KASSERT(bp->b_offset != NOOFFSET,
2355 ("getblk: no buffer offset"));
2356
2357 /*
2358 * A buffer with B_DELWRI set and B_CACHE clear must
2359 * be committed before we can return the buffer in
2360 * order to prevent the caller from issuing a read
2361 * ( due to B_CACHE not being set ) and overwriting
2362 * it.
2363 *
2364 * Most callers, including NFS and FFS, need this to
2365 * operate properly either because they assume they
2366 * can issue a read if B_CACHE is not set, or because
2367 * ( for example ) an uncached B_DELWRI might loop due
2368 * to softupdates re-dirtying the buffer. In the latter
2369 * case, B_CACHE is set after the first write completes,
2370 * preventing further loops.
2371 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2372 * above while extending the buffer, we cannot allow the
2373 * buffer to remain with B_CACHE set after the write
2374 * completes or it will represent a corrupt state. To
2375 * deal with this we set B_NOCACHE to scrap the buffer
2376 * after the write.
2377 *
2378 * We might be able to do something fancy, like setting
2379 * B_CACHE in bwrite() except if B_DELWRI is already set,
2380 * so the below call doesn't set B_CACHE, but that gets real
2381 * confusing. This is much easier.
2382 */
2383
2384 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2385 bp->b_flags |= B_NOCACHE;
2386 BUF_WRITE(bp);
2387 goto loop;
2388 }
2389
2390 splx(s);
2391 bp->b_flags &= ~B_DONE;
2392 } else {
2393 /*
2394 * Buffer is not in-core, create new buffer. The buffer
2395 * returned by getnewbuf() is locked. Note that the returned
2396 * buffer is also considered valid (not marked B_INVAL).
2397 */
2398 int bsize, maxsize, vmio;
2399 off_t offset;
2400
2401 if (vn_isdisk(vp, NULL))
2402 bsize = DEV_BSIZE;
2403 else if (vp->v_mountedhere)
2404 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2405 else if (vp->v_mount)
2406 bsize = vp->v_mount->mnt_stat.f_iosize;
2407 else
2408 bsize = size;
2409
2410 offset = blkno * bsize;
2411 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) &&
2412 (vp->v_vflag & VV_OBJBUF);
2413 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2414 maxsize = imax(maxsize, bsize);
2415
2416 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2417 if (slpflag || slptimeo) {
2418 splx(s);
2419 return NULL;
2420 }
2421 goto loop;
2422 }
2423
2424 /*
2425 * This code is used to make sure that a buffer is not
2426 * created while the getnewbuf routine is blocked.
2427 * This can be a problem whether the vnode is locked or not.
2428 * If the buffer is created out from under us, we have to
2429 * throw away the one we just created. There is now window
2430 * race because we are safely running at splbio() from the
2431 * point of the duplicate buffer creation through to here,
2432 * and we've locked the buffer.
2433 *
2434 * Note: this must occur before we associate the buffer
2435 * with the vp especially considering limitations in
2436 * the splay tree implementation when dealing with duplicate
2437 * lblkno's.
2438 */
2439 if (gbincore(vp, blkno)) {
2440 bp->b_flags |= B_INVAL;
2441 brelse(bp);
2442 goto loop;
2443 }
2444
2445 /*
2446 * Insert the buffer into the hash, so that it can
2447 * be found by incore.
2448 */
2449 bp->b_blkno = bp->b_lblkno = blkno;
2450 bp->b_offset = offset;
2451
2452 bgetvp(vp, bp);
2453#ifdef USE_BUFHASH
2454 LIST_REMOVE(bp, b_hash);
2455 bh = bufhash(vp, blkno);
2456 LIST_INSERT_HEAD(bh, bp, b_hash);
2457#endif
2458
2459 /*
2460 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2461 * buffer size starts out as 0, B_CACHE will be set by
2462 * allocbuf() for the VMIO case prior to it testing the
2463 * backing store for validity.
2464 */
2465
2466 if (vmio) {
2467 bp->b_flags |= B_VMIO;
2468#if defined(VFS_BIO_DEBUG)
2469 if (vp->v_type != VREG)
2470 printf("getblk: vmioing file type %d???\n", vp->v_type);
2471#endif
2472 VOP_GETVOBJECT(vp, &bp->b_object);
2473 } else {
2474 bp->b_flags &= ~B_VMIO;
2475 bp->b_object = NULL;
2476 }
2477
2478 allocbuf(bp, size);
2479
2480 splx(s);
2481 bp->b_flags &= ~B_DONE;
2482 }
2483 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2484 return (bp);
2485}
2486
2487/*
2488 * Get an empty, disassociated buffer of given size. The buffer is initially
2489 * set to B_INVAL.
2490 */
2491struct buf *
2492geteblk(int size)
2493{
2494 struct buf *bp;
2495 int s;
2496 int maxsize;
2497
2498 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2499
2500 s = splbio();
2501 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2502 splx(s);
2503 allocbuf(bp, size);
2504 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2505 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2506 return (bp);
2507}
2508
2509
2510/*
2511 * This code constitutes the buffer memory from either anonymous system
2512 * memory (in the case of non-VMIO operations) or from an associated
2513 * VM object (in the case of VMIO operations). This code is able to
2514 * resize a buffer up or down.
2515 *
2516 * Note that this code is tricky, and has many complications to resolve
2517 * deadlock or inconsistant data situations. Tread lightly!!!
2518 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2519 * the caller. Calling this code willy nilly can result in the loss of data.
2520 *
2521 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2522 * B_CACHE for the non-VMIO case.
2523 */
2524
2525int
2526allocbuf(struct buf *bp, int size)
2527{
2528 int newbsize, mbsize;
2529 int i;
2530
2531 GIANT_REQUIRED;
2532
2533 if (BUF_REFCNT(bp) == 0)
2534 panic("allocbuf: buffer not busy");
2535
2536 if (bp->b_kvasize < size)
2537 panic("allocbuf: buffer too small");
2538
2539 if ((bp->b_flags & B_VMIO) == 0) {
2540 caddr_t origbuf;
2541 int origbufsize;
2542 /*
2543 * Just get anonymous memory from the kernel. Don't
2544 * mess with B_CACHE.
2545 */
2546 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2547 if (bp->b_flags & B_MALLOC)
2548 newbsize = mbsize;
2549 else
2550 newbsize = round_page(size);
2551
2552 if (newbsize < bp->b_bufsize) {
2553 /*
2554 * malloced buffers are not shrunk
2555 */
2556 if (bp->b_flags & B_MALLOC) {
2557 if (newbsize) {
2558 bp->b_bcount = size;
2559 } else {
2560 free(bp->b_data, M_BIOBUF);
2561 if (bp->b_bufsize) {
2562 bufmallocspace -= bp->b_bufsize;
2563 bufspacewakeup();
2564 bp->b_bufsize = 0;
2565 }
2566 bp->b_data = bp->b_kvabase;
2567 bp->b_bcount = 0;
2568 bp->b_flags &= ~B_MALLOC;
2569 }
2570 return 1;
2571 }
2572 vm_hold_free_pages(
2573 bp,
2574 (vm_offset_t) bp->b_data + newbsize,
2575 (vm_offset_t) bp->b_data + bp->b_bufsize);
2576 } else if (newbsize > bp->b_bufsize) {
2577 /*
2578 * We only use malloced memory on the first allocation.
2579 * and revert to page-allocated memory when the buffer
2580 * grows.
2581 */
2582 if ( (bufmallocspace < maxbufmallocspace) &&
2583 (bp->b_bufsize == 0) &&
2584 (mbsize <= PAGE_SIZE/2)) {
2585
2586 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2587 bp->b_bufsize = mbsize;
2588 bp->b_bcount = size;
2589 bp->b_flags |= B_MALLOC;
2590 bufmallocspace += mbsize;
2591 return 1;
2592 }
2593 origbuf = NULL;
2594 origbufsize = 0;
2595 /*
2596 * If the buffer is growing on its other-than-first allocation,
2597 * then we revert to the page-allocation scheme.
2598 */
2599 if (bp->b_flags & B_MALLOC) {
2600 origbuf = bp->b_data;
2601 origbufsize = bp->b_bufsize;
2602 bp->b_data = bp->b_kvabase;
2603 if (bp->b_bufsize) {
2604 bufmallocspace -= bp->b_bufsize;
2605 bufspacewakeup();
2606 bp->b_bufsize = 0;
2607 }
2608 bp->b_flags &= ~B_MALLOC;
2609 newbsize = round_page(newbsize);
2610 }
2611 vm_hold_load_pages(
2612 bp,
2613 (vm_offset_t) bp->b_data + bp->b_bufsize,
2614 (vm_offset_t) bp->b_data + newbsize);
2615 if (origbuf) {
2616 bcopy(origbuf, bp->b_data, origbufsize);
2617 free(origbuf, M_BIOBUF);
2618 }
2619 }
2620 } else {
2621 vm_page_t m;
2622 int desiredpages;
2623
2624 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2625 desiredpages = (size == 0) ? 0 :
2626 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2627
2628 if (bp->b_flags & B_MALLOC)
2629 panic("allocbuf: VMIO buffer can't be malloced");
2630 /*
2631 * Set B_CACHE initially if buffer is 0 length or will become
2632 * 0-length.
2633 */
2634 if (size == 0 || bp->b_bufsize == 0)
2635 bp->b_flags |= B_CACHE;
2636
2637 if (newbsize < bp->b_bufsize) {
2638 /*
2639 * DEV_BSIZE aligned new buffer size is less then the
2640 * DEV_BSIZE aligned existing buffer size. Figure out
2641 * if we have to remove any pages.
2642 */
2643 if (desiredpages < bp->b_npages) {
2644 vm_page_lock_queues();
2645 for (i = desiredpages; i < bp->b_npages; i++) {
2646 /*
2647 * the page is not freed here -- it
2648 * is the responsibility of
2649 * vnode_pager_setsize
2650 */
2651 m = bp->b_pages[i];
2652 KASSERT(m != bogus_page,
2653 ("allocbuf: bogus page found"));
2654 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2655 vm_page_lock_queues();
2656
2657 bp->b_pages[i] = NULL;
2658 vm_page_unwire(m, 0);
2659 }
2660 vm_page_unlock_queues();
2661 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2662 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2663 bp->b_npages = desiredpages;
2664 }
2665 } else if (size > bp->b_bcount) {
2666 /*
2667 * We are growing the buffer, possibly in a
2668 * byte-granular fashion.
2669 */
2670 struct vnode *vp;
2671 vm_object_t obj;
2672 vm_offset_t toff;
2673 vm_offset_t tinc;
2674
2675 /*
2676 * Step 1, bring in the VM pages from the object,
2677 * allocating them if necessary. We must clear
2678 * B_CACHE if these pages are not valid for the
2679 * range covered by the buffer.
2680 */
2681
2682 vp = bp->b_vp;
2683 obj = bp->b_object;
2684
2685 while (bp->b_npages < desiredpages) {
2686 vm_page_t m;
2687 vm_pindex_t pi;
2688
2689 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2690 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2691 /*
2692 * note: must allocate system pages
2693 * since blocking here could intefere
2694 * with paging I/O, no matter which
2695 * process we are.
2696 */
2697 m = vm_page_alloc(obj, pi,
2698 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
2699 if (m == NULL) {
2700 VM_WAIT;
2701 vm_pageout_deficit += desiredpages - bp->b_npages;
2702 } else {
2703 vm_page_lock_queues();
2704 vm_page_wakeup(m);
2705 vm_page_unlock_queues();
2706 bp->b_flags &= ~B_CACHE;
2707 bp->b_pages[bp->b_npages] = m;
2708 ++bp->b_npages;
2709 }
2710 continue;
2711 }
2712
2713 /*
2714 * We found a page. If we have to sleep on it,
2715 * retry because it might have gotten freed out
2716 * from under us.
2717 *
2718 * We can only test PG_BUSY here. Blocking on
2719 * m->busy might lead to a deadlock:
2720 *
2721 * vm_fault->getpages->cluster_read->allocbuf
2722 *
2723 */
2724 vm_page_lock_queues();
2725 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2726 continue;
2727
2728 /*
2729 * We have a good page. Should we wakeup the
2730 * page daemon?
2731 */
2732 if ((curproc != pageproc) &&
2733 ((m->queue - m->pc) == PQ_CACHE) &&
2734 ((cnt.v_free_count + cnt.v_cache_count) <
2735 (cnt.v_free_min + cnt.v_cache_min))) {
2736 pagedaemon_wakeup();
2737 }
2738 vm_page_flag_clear(m, PG_ZERO);
2739 vm_page_wire(m);
2740 vm_page_unlock_queues();
2741 bp->b_pages[bp->b_npages] = m;
2742 ++bp->b_npages;
2743 }
2744
2745 /*
2746 * Step 2. We've loaded the pages into the buffer,
2747 * we have to figure out if we can still have B_CACHE
2748 * set. Note that B_CACHE is set according to the
2749 * byte-granular range ( bcount and size ), new the
2750 * aligned range ( newbsize ).
2751 *
2752 * The VM test is against m->valid, which is DEV_BSIZE
2753 * aligned. Needless to say, the validity of the data
2754 * needs to also be DEV_BSIZE aligned. Note that this
2755 * fails with NFS if the server or some other client
2756 * extends the file's EOF. If our buffer is resized,
2757 * B_CACHE may remain set! XXX
2758 */
2759
2760 toff = bp->b_bcount;
2761 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2762
2763 while ((bp->b_flags & B_CACHE) && toff < size) {
2764 vm_pindex_t pi;
2765
2766 if (tinc > (size - toff))
2767 tinc = size - toff;
2768
2769 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2770 PAGE_SHIFT;
2771
2772 vfs_buf_test_cache(
2773 bp,
2774 bp->b_offset,
2775 toff,
2776 tinc,
2777 bp->b_pages[pi]
2778 );
2779 toff += tinc;
2780 tinc = PAGE_SIZE;
2781 }
2782
2783 /*
2784 * Step 3, fixup the KVM pmap. Remember that
2785 * bp->b_data is relative to bp->b_offset, but
2786 * bp->b_offset may be offset into the first page.
2787 */
2788
2789 bp->b_data = (caddr_t)
2790 trunc_page((vm_offset_t)bp->b_data);
2791 pmap_qenter(
2792 (vm_offset_t)bp->b_data,
2793 bp->b_pages,
2794 bp->b_npages
2795 );
2796
2797 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2798 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2799 }
2800 }
2801 if (newbsize < bp->b_bufsize)
2802 bufspacewakeup();
2803 bp->b_bufsize = newbsize; /* actual buffer allocation */
2804 bp->b_bcount = size; /* requested buffer size */
2805 return 1;
2806}
2807
2808void
2809biodone(struct bio *bp)
2810{
2811 bp->bio_flags |= BIO_DONE;
2812 if (bp->bio_done != NULL)
2813 bp->bio_done(bp);
2814 else
2815 wakeup(bp);
2816}
2817
2818/*
2819 * Wait for a BIO to finish.
2820 * XXX: For now resort to a timeout, the optimal locking (if any) for this
2821 * case is not at this point obvious.
2822 */
2823int
2824biowait(struct bio *bp, const char *wchan)
2825{
2826
2827 while ((bp->bio_flags & BIO_DONE) == 0)
2828 msleep(bp, NULL, 0, wchan, hz);
2829 if (!(bp->bio_flags & BIO_ERROR))
2830 return (0);
2831 if (bp->bio_error)
2832 return (bp->bio_error);
2833 return (EIO);
2834}
2835
2836/*
2837 * bufwait:
2838 *
2839 * Wait for buffer I/O completion, returning error status. The buffer
2840 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR
2841 * error and cleared.
2842 */
2843int
2844bufwait(register struct buf * bp)
2845{
2846 int s;
2847
2848 s = splbio();
2849 while ((bp->b_flags & B_DONE) == 0) {
2850 if (bp->b_iocmd == BIO_READ)
2851 tsleep(bp, PRIBIO, "biord", 0);
2852 else
2853 tsleep(bp, PRIBIO, "biowr", 0);
2854 }
2855 splx(s);
2856 if (bp->b_flags & B_EINTR) {
2857 bp->b_flags &= ~B_EINTR;
2858 return (EINTR);
2859 }
2860 if (bp->b_ioflags & BIO_ERROR) {
2861 return (bp->b_error ? bp->b_error : EIO);
2862 } else {
2863 return (0);
2864 }
2865}
2866
2867 /*
2868 * Call back function from struct bio back up to struct buf.
2869 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY().
2870 */
2871void
2872bufdonebio(struct bio *bp)
2873{
2874 bufdone(bp->bio_caller2);
2875}
2876
2877/*
2878 * bufdone:
2879 *
2880 * Finish I/O on a buffer, optionally calling a completion function.
2881 * This is usually called from an interrupt so process blocking is
2882 * not allowed.
2883 *
2884 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2885 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2886 * assuming B_INVAL is clear.
2887 *
2888 * For the VMIO case, we set B_CACHE if the op was a read and no
2889 * read error occured, or if the op was a write. B_CACHE is never
2890 * set if the buffer is invalid or otherwise uncacheable.
2891 *
2892 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2893 * initiator to leave B_INVAL set to brelse the buffer out of existance
2894 * in the biodone routine.
2895 */
2896void
2897bufdone(struct buf *bp)
2898{
2899 int s;
2900 void (*biodone)(struct buf *);
2901
2902 GIANT_REQUIRED;
2903
2904 s = splbio();
2905
2906 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
2907 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2908
2909 bp->b_flags |= B_DONE;
2910 runningbufwakeup(bp);
2911
2912 if (bp->b_iocmd == BIO_DELETE) {
2913 brelse(bp);
2914 splx(s);
2915 return;
2916 }
2917
2918 if (bp->b_iocmd == BIO_WRITE) {
2919 vwakeup(bp);
2920 }
2921
2922 /* call optional completion function if requested */
2923 if (bp->b_iodone != NULL) {
2924 biodone = bp->b_iodone;
2925 bp->b_iodone = NULL;
2926 (*biodone) (bp);
2927 splx(s);
2928 return;
2929 }
2930 if (LIST_FIRST(&bp->b_dep) != NULL)
2931 buf_complete(bp);
2932
2933 if (bp->b_flags & B_VMIO) {
2934 int i;
2935 vm_ooffset_t foff;
2936 vm_page_t m;
2937 vm_object_t obj;
2938 int iosize;
2939 struct vnode *vp = bp->b_vp;
2940
2941 obj = bp->b_object;
2942
2943#if defined(VFS_BIO_DEBUG)
2944 mp_fixme("usecount and vflag accessed without locks.");
2945 if (vp->v_usecount == 0) {
2946 panic("biodone: zero vnode ref count");
2947 }
2948
2949 if ((vp->v_vflag & VV_OBJBUF) == 0) {
2950 panic("biodone: vnode is not setup for merged cache");
2951 }
2952#endif
2953
2954 foff = bp->b_offset;
2955 KASSERT(bp->b_offset != NOOFFSET,
2956 ("biodone: no buffer offset"));
2957
2958#if defined(VFS_BIO_DEBUG)
2959 if (obj->paging_in_progress < bp->b_npages) {
2960 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
2961 obj->paging_in_progress, bp->b_npages);
2962 }
2963#endif
2964
2965 /*
2966 * Set B_CACHE if the op was a normal read and no error
2967 * occured. B_CACHE is set for writes in the b*write()
2968 * routines.
2969 */
2970 iosize = bp->b_bcount - bp->b_resid;
2971 if (bp->b_iocmd == BIO_READ &&
2972 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
2973 !(bp->b_ioflags & BIO_ERROR)) {
2974 bp->b_flags |= B_CACHE;
2975 }
2976 vm_page_lock_queues();
2977 for (i = 0; i < bp->b_npages; i++) {
2978 int bogusflag = 0;
2979 int resid;
2980
2981 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2982 if (resid > iosize)
2983 resid = iosize;
2984
2985 /*
2986 * cleanup bogus pages, restoring the originals
2987 */
2988 m = bp->b_pages[i];
2989 if (m == bogus_page) {
2990 bogusflag = 1;
2991 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2992 if (m == NULL)
2993 panic("biodone: page disappeared!");
2994 bp->b_pages[i] = m;
2995 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2996 }
2997#if defined(VFS_BIO_DEBUG)
2998 if (OFF_TO_IDX(foff) != m->pindex) {
2999 printf(
3000"biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3001 (intmax_t)foff, (uintmax_t)m->pindex);
3002 }
3003#endif
3004
3005 /*
3006 * In the write case, the valid and clean bits are
3007 * already changed correctly ( see bdwrite() ), so we
3008 * only need to do this here in the read case.
3009 */
3010 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3011 vfs_page_set_valid(bp, foff, i, m);
3012 }
3013 vm_page_flag_clear(m, PG_ZERO);
3014
3015 /*
3016 * when debugging new filesystems or buffer I/O methods, this
3017 * is the most common error that pops up. if you see this, you
3018 * have not set the page busy flag correctly!!!
3019 */
3020 if (m->busy == 0) {
3021 printf("biodone: page busy < 0, "
3022 "pindex: %d, foff: 0x(%x,%x), "
3023 "resid: %d, index: %d\n",
3024 (int) m->pindex, (int)(foff >> 32),
3025 (int) foff & 0xffffffff, resid, i);
3026 if (!vn_isdisk(vp, NULL))
3027 printf(" iosize: %ld, lblkno: %jd, flags: 0x%lx, npages: %d\n",
3028 bp->b_vp->v_mount->mnt_stat.f_iosize,
3029 (intmax_t) bp->b_lblkno,
3030 bp->b_flags, bp->b_npages);
3031 else
3032 printf(" VDEV, lblkno: %jd, flags: 0x%lx, npages: %d\n",
3033 (intmax_t) bp->b_lblkno,
3034 bp->b_flags, bp->b_npages);
3035 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3036 m->valid, m->dirty, m->wire_count);
3037 panic("biodone: page busy < 0\n");
3038 }
3039 vm_page_io_finish(m);
3040 vm_object_pip_subtract(obj, 1);
3041 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3042 iosize -= resid;
3043 }
3044 vm_page_unlock_queues();
3045 if (obj)
3046 vm_object_pip_wakeupn(obj, 0);
3047 }
3048
3049 /*
3050 * For asynchronous completions, release the buffer now. The brelse
3051 * will do a wakeup there if necessary - so no need to do a wakeup
3052 * here in the async case. The sync case always needs to do a wakeup.
3053 */
3054
3055 if (bp->b_flags & B_ASYNC) {
3056 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3057 brelse(bp);
3058 else
3059 bqrelse(bp);
3060 } else {
3061 wakeup(bp);
3062 }
3063 splx(s);
3064}
3065
3066/*
3067 * This routine is called in lieu of iodone in the case of
3068 * incomplete I/O. This keeps the busy status for pages
3069 * consistant.
3070 */
3071void
3072vfs_unbusy_pages(struct buf * bp)
3073{
3074 int i;
3075
3076 GIANT_REQUIRED;
3077
3078 runningbufwakeup(bp);
3079 if (bp->b_flags & B_VMIO) {
3080 vm_object_t obj;
3081
3082 obj = bp->b_object;
3083 vm_page_lock_queues();
3084 for (i = 0; i < bp->b_npages; i++) {
3085 vm_page_t m = bp->b_pages[i];
3086
3087 if (m == bogus_page) {
3088 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3089 if (!m) {
3090 panic("vfs_unbusy_pages: page missing\n");
3091 }
3092 bp->b_pages[i] = m;
3093 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3094 }
3095 vm_object_pip_subtract(obj, 1);
3096 vm_page_flag_clear(m, PG_ZERO);
3097 vm_page_io_finish(m);
3098 }
3099 vm_page_unlock_queues();
3100 vm_object_pip_wakeupn(obj, 0);
3101 }
3102}
3103
3104/*
3105 * vfs_page_set_valid:
3106 *
3107 * Set the valid bits in a page based on the supplied offset. The
3108 * range is restricted to the buffer's size.
3109 *
3110 * This routine is typically called after a read completes.
3111 */
3112static void
3113vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3114{
3115 vm_ooffset_t soff, eoff;
3116
3117 GIANT_REQUIRED;
3118 /*
3119 * Start and end offsets in buffer. eoff - soff may not cross a
3120 * page boundry or cross the end of the buffer. The end of the
3121 * buffer, in this case, is our file EOF, not the allocation size
3122 * of the buffer.
3123 */
3124 soff = off;
3125 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3126 if (eoff > bp->b_offset + bp->b_bcount)
3127 eoff = bp->b_offset + bp->b_bcount;
3128
3129 /*
3130 * Set valid range. This is typically the entire buffer and thus the
3131 * entire page.
3132 */
3133 if (eoff > soff) {
3134 vm_page_set_validclean(
3135 m,
3136 (vm_offset_t) (soff & PAGE_MASK),
3137 (vm_offset_t) (eoff - soff)
3138 );
3139 }
3140}
3141
3142/*
3143 * This routine is called before a device strategy routine.
3144 * It is used to tell the VM system that paging I/O is in
3145 * progress, and treat the pages associated with the buffer
3146 * almost as being PG_BUSY. Also the object paging_in_progress
3147 * flag is handled to make sure that the object doesn't become
3148 * inconsistant.
3149 *
3150 * Since I/O has not been initiated yet, certain buffer flags
3151 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3152 * and should be ignored.
3153 */
3154void
3155vfs_busy_pages(struct buf * bp, int clear_modify)
3156{
3157 int i, bogus;
3158
3159 if (bp->b_flags & B_VMIO) {
3160 vm_object_t obj;
3161 vm_ooffset_t foff;
3162
3163 obj = bp->b_object;
3164 foff = bp->b_offset;
3165 KASSERT(bp->b_offset != NOOFFSET,
3166 ("vfs_busy_pages: no buffer offset"));
3167 vfs_setdirty(bp);
3168retry:
3169 vm_page_lock_queues();
3170 for (i = 0; i < bp->b_npages; i++) {
3171 vm_page_t m = bp->b_pages[i];
3172
3173 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3174 goto retry;
3175 }
3176 bogus = 0;
3177 for (i = 0; i < bp->b_npages; i++) {
3178 vm_page_t m = bp->b_pages[i];
3179
3180 vm_page_flag_clear(m, PG_ZERO);
3181 if ((bp->b_flags & B_CLUSTER) == 0) {
3182 vm_object_pip_add(obj, 1);
3183 vm_page_io_start(m);
3184 }
3185 /*
3186 * When readying a buffer for a read ( i.e
3187 * clear_modify == 0 ), it is important to do
3188 * bogus_page replacement for valid pages in
3189 * partially instantiated buffers. Partially
3190 * instantiated buffers can, in turn, occur when
3191 * reconstituting a buffer from its VM backing store
3192 * base. We only have to do this if B_CACHE is
3193 * clear ( which causes the I/O to occur in the
3194 * first place ). The replacement prevents the read
3195 * I/O from overwriting potentially dirty VM-backed
3196 * pages. XXX bogus page replacement is, uh, bogus.
3197 * It may not work properly with small-block devices.
3198 * We need to find a better way.
3199 */
3200 vm_page_protect(m, VM_PROT_NONE);
3201 if (clear_modify)
3202 vfs_page_set_valid(bp, foff, i, m);
3203 else if (m->valid == VM_PAGE_BITS_ALL &&
3204 (bp->b_flags & B_CACHE) == 0) {
3205 bp->b_pages[i] = bogus_page;
3206 bogus++;
3207 }
3208 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3209 }
3210 vm_page_unlock_queues();
3211 if (bogus)
3212 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3213 }
3214}
3215
3216/*
3217 * Tell the VM system that the pages associated with this buffer
3218 * are clean. This is used for delayed writes where the data is
3219 * going to go to disk eventually without additional VM intevention.
3220 *
3221 * Note that while we only really need to clean through to b_bcount, we
3222 * just go ahead and clean through to b_bufsize.
3223 */
3224static void
3225vfs_clean_pages(struct buf * bp)
3226{
3227 int i;
3228
3229 GIANT_REQUIRED;
3230
3231 if (bp->b_flags & B_VMIO) {
3232 vm_ooffset_t foff;
3233
3234 foff = bp->b_offset;
3235 KASSERT(bp->b_offset != NOOFFSET,
3236 ("vfs_clean_pages: no buffer offset"));
3237 for (i = 0; i < bp->b_npages; i++) {
3238 vm_page_t m = bp->b_pages[i];
3239 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3240 vm_ooffset_t eoff = noff;
3241
3242 if (eoff > bp->b_offset + bp->b_bufsize)
3243 eoff = bp->b_offset + bp->b_bufsize;
3244 vfs_page_set_valid(bp, foff, i, m);
3245 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3246 foff = noff;
3247 }
3248 }
3249}
3250
3251/*
3252 * vfs_bio_set_validclean:
3253 *
3254 * Set the range within the buffer to valid and clean. The range is
3255 * relative to the beginning of the buffer, b_offset. Note that b_offset
3256 * itself may be offset from the beginning of the first page.
3257 *
3258 */
3259
3260void
3261vfs_bio_set_validclean(struct buf *bp, int base, int size)
3262{
3263 if (bp->b_flags & B_VMIO) {
3264 int i;
3265 int n;
3266
3267 /*
3268 * Fixup base to be relative to beginning of first page.
3269 * Set initial n to be the maximum number of bytes in the
3270 * first page that can be validated.
3271 */
3272
3273 base += (bp->b_offset & PAGE_MASK);
3274 n = PAGE_SIZE - (base & PAGE_MASK);
3275
3276 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3277 vm_page_t m = bp->b_pages[i];
3278
3279 if (n > size)
3280 n = size;
3281
3282 vm_page_set_validclean(m, base & PAGE_MASK, n);
3283 base += n;
3284 size -= n;
3285 n = PAGE_SIZE;
3286 }
3287 }
3288}
3289
3290/*
3291 * vfs_bio_clrbuf:
3292 *
3293 * clear a buffer. This routine essentially fakes an I/O, so we need
3294 * to clear BIO_ERROR and B_INVAL.
3295 *
3296 * Note that while we only theoretically need to clear through b_bcount,
3297 * we go ahead and clear through b_bufsize.
3298 */
3299
3300void
3301vfs_bio_clrbuf(struct buf *bp)
3302{
3303 int i, mask = 0;
3304 caddr_t sa, ea;
3305
3306 GIANT_REQUIRED;
3307
3308 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3309 bp->b_flags &= ~B_INVAL;
3310 bp->b_ioflags &= ~BIO_ERROR;
3311 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3312 (bp->b_offset & PAGE_MASK) == 0) {
3313 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3314 if ((bp->b_pages[0]->valid & mask) == mask) {
3315 bp->b_resid = 0;
3316 return;
3317 }
3318 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3319 ((bp->b_pages[0]->valid & mask) == 0)) {
3320 bzero(bp->b_data, bp->b_bufsize);
3321 bp->b_pages[0]->valid |= mask;
3322 bp->b_resid = 0;
3323 return;
3324 }
3325 }
3326 ea = sa = bp->b_data;
3327 for(i=0;i<bp->b_npages;i++,sa=ea) {
3328 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3329 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3330 ea = (caddr_t)(vm_offset_t)ulmin(
3331 (u_long)(vm_offset_t)ea,
3332 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3333 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3334 if ((bp->b_pages[i]->valid & mask) == mask)
3335 continue;
3336 if ((bp->b_pages[i]->valid & mask) == 0) {
3337 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3338 bzero(sa, ea - sa);
3339 }
3340 } else {
3341 for (; sa < ea; sa += DEV_BSIZE, j++) {
3342 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3343 (bp->b_pages[i]->valid & (1<<j)) == 0)
3344 bzero(sa, DEV_BSIZE);
3345 }
3346 }
3347 bp->b_pages[i]->valid |= mask;
3348 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3349 }
3350 bp->b_resid = 0;
3351 } else {
3352 clrbuf(bp);
3353 }
3354}
3355
3356/*
3357 * vm_hold_load_pages and vm_hold_free_pages get pages into
3358 * a buffers address space. The pages are anonymous and are
3359 * not associated with a file object.
3360 */
3361static void
3362vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3363{
3364 vm_offset_t pg;
3365 vm_page_t p;
3366 int index;
3367
3368 GIANT_REQUIRED;
3369
3370 to = round_page(to);
3371 from = round_page(from);
3372 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3373
3374 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3375tryagain:
3376 /*
3377 * note: must allocate system pages since blocking here
3378 * could intefere with paging I/O, no matter which
3379 * process we are.
3380 */
3381 p = vm_page_alloc(kernel_object,
3382 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3383 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3384 if (!p) {
3385 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3386 VM_WAIT;
3387 goto tryagain;
3388 }
3389 vm_page_lock_queues();
3390 p->valid = VM_PAGE_BITS_ALL;
3391 vm_page_flag_clear(p, PG_ZERO);
3392 vm_page_unlock_queues();
3393 pmap_qenter(pg, &p, 1);
3394 bp->b_pages[index] = p;
3395 vm_page_wakeup(p);
3396 }
3397 bp->b_npages = index;
3398}
3399
3400/* Return pages associated with this buf to the vm system */
3401void
3402vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3403{
3404 vm_offset_t pg;
3405 vm_page_t p;
3406 int index, newnpages;
3407
3408 GIANT_REQUIRED;
3409
3410 from = round_page(from);
3411 to = round_page(to);
3412 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3413
3414 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3415 p = bp->b_pages[index];
3416 if (p && (index < bp->b_npages)) {
3417 if (p->busy) {
3418 printf(
3419 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3420 (intmax_t)bp->b_blkno,
3421 (intmax_t)bp->b_lblkno);
3422 }
3423 bp->b_pages[index] = NULL;
3424 pmap_qremove(pg, 1);
3425 vm_page_lock_queues();
3426 vm_page_busy(p);
3427 vm_page_unwire(p, 0);
3428 vm_page_free(p);
3429 vm_page_unlock_queues();
3430 }
3431 }
3432 bp->b_npages = newnpages;
3433}
3434
3435
3436#include "opt_ddb.h"
3437#ifdef DDB
3438#include <ddb/ddb.h>
3439
3440/* DDB command to show buffer data */
3441DB_SHOW_COMMAND(buffer, db_show_buffer)
3442{
3443 /* get args */
3444 struct buf *bp = (struct buf *)addr;
3445
3446 if (!have_addr) {
3447 db_printf("usage: show buffer <addr>\n");
3448 return;
3449 }
3450
3451 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3452 db_printf(
3453 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3454 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n",
3455 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3456 major(bp->b_dev), minor(bp->b_dev), bp->b_data,
3457 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno);
3458 if (bp->b_npages) {
3459 int i;
3460 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3461 for (i = 0; i < bp->b_npages; i++) {
3462 vm_page_t m;
3463 m = bp->b_pages[i];
3464 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3465 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3466 if ((i + 1) < bp->b_npages)
3467 db_printf(",");
3468 }
3469 db_printf("\n");
3470 }
3471}
3472#endif /* DDB */
| 1216 !vn_isdisk(bp->b_vp, NULL) &&
1217 (bp->b_flags & B_DELWRI))
1218 ) {
1219
1220 int i, j, resid;
1221 vm_page_t m;
1222 off_t foff;
1223 vm_pindex_t poff;
1224 vm_object_t obj;
1225 struct vnode *vp;
1226
1227 vp = bp->b_vp;
1228 obj = bp->b_object;
1229
1230 /*
1231 * Get the base offset and length of the buffer. Note that
1232 * in the VMIO case if the buffer block size is not
1233 * page-aligned then b_data pointer may not be page-aligned.
1234 * But our b_pages[] array *IS* page aligned.
1235 *
1236 * block sizes less then DEV_BSIZE (usually 512) are not
1237 * supported due to the page granularity bits (m->valid,
1238 * m->dirty, etc...).
1239 *
1240 * See man buf(9) for more information
1241 */
1242 resid = bp->b_bufsize;
1243 foff = bp->b_offset;
1244
1245 for (i = 0; i < bp->b_npages; i++) {
1246 int had_bogus = 0;
1247
1248 m = bp->b_pages[i];
1249 vm_page_flag_clear(m, PG_ZERO);
1250
1251 /*
1252 * If we hit a bogus page, fixup *all* the bogus pages
1253 * now.
1254 */
1255 if (m == bogus_page) {
1256 poff = OFF_TO_IDX(bp->b_offset);
1257 had_bogus = 1;
1258
1259 for (j = i; j < bp->b_npages; j++) {
1260 vm_page_t mtmp;
1261 mtmp = bp->b_pages[j];
1262 if (mtmp == bogus_page) {
1263 mtmp = vm_page_lookup(obj, poff + j);
1264 if (!mtmp) {
1265 panic("brelse: page missing\n");
1266 }
1267 bp->b_pages[j] = mtmp;
1268 }
1269 }
1270
1271 if ((bp->b_flags & B_INVAL) == 0) {
1272 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1273 }
1274 m = bp->b_pages[i];
1275 }
1276 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) {
1277 int poffset = foff & PAGE_MASK;
1278 int presid = resid > (PAGE_SIZE - poffset) ?
1279 (PAGE_SIZE - poffset) : resid;
1280
1281 KASSERT(presid >= 0, ("brelse: extra page"));
1282 vm_page_set_invalid(m, poffset, presid);
1283 if (had_bogus)
1284 printf("avoided corruption bug in bogus_page/brelse code\n");
1285 }
1286 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1287 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1288 }
1289
1290 if (bp->b_flags & (B_INVAL | B_RELBUF))
1291 vfs_vmio_release(bp);
1292
1293 } else if (bp->b_flags & B_VMIO) {
1294
1295 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1296 vfs_vmio_release(bp);
1297 }
1298
1299 }
1300
1301 if (bp->b_qindex != QUEUE_NONE)
1302 panic("brelse: free buffer onto another queue???");
1303 if (BUF_REFCNT(bp) > 1) {
1304 /* do not release to free list */
1305 BUF_UNLOCK(bp);
1306 splx(s);
1307 return;
1308 }
1309
1310 /* enqueue */
1311
1312 /* buffers with no memory */
1313 if (bp->b_bufsize == 0) {
1314 bp->b_flags |= B_INVAL;
1315 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1316 if (bp->b_xflags & BX_BKGRDINPROG)
1317 panic("losing buffer 1");
1318 if (bp->b_kvasize) {
1319 bp->b_qindex = QUEUE_EMPTYKVA;
1320 } else {
1321 bp->b_qindex = QUEUE_EMPTY;
1322 }
1323 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1324#ifdef USE_BUFHASH
1325 LIST_REMOVE(bp, b_hash);
1326 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1327#endif
1328 bp->b_dev = NODEV;
1329 /* buffers with junk contents */
1330 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1331 (bp->b_ioflags & BIO_ERROR)) {
1332 bp->b_flags |= B_INVAL;
1333 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1334 if (bp->b_xflags & BX_BKGRDINPROG)
1335 panic("losing buffer 2");
1336 bp->b_qindex = QUEUE_CLEAN;
1337 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1338#ifdef USE_BUFHASH
1339 LIST_REMOVE(bp, b_hash);
1340 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1341#endif
1342 bp->b_dev = NODEV;
1343
1344 /* buffers that are locked */
1345 } else if (bp->b_flags & B_LOCKED) {
1346 bp->b_qindex = QUEUE_LOCKED;
1347 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1348
1349 /* remaining buffers */
1350 } else {
1351 if (bp->b_flags & B_DELWRI)
1352 bp->b_qindex = QUEUE_DIRTY;
1353 else
1354 bp->b_qindex = QUEUE_CLEAN;
1355 if (bp->b_flags & B_AGE)
1356 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1357 else
1358 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1359 }
1360
1361 /*
1362 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already
1363 * placed the buffer on the correct queue. We must also disassociate
1364 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1365 * find it.
1366 */
1367 if (bp->b_flags & B_INVAL) {
1368 if (bp->b_flags & B_DELWRI)
1369 bundirty(bp);
1370 if (bp->b_vp)
1371 brelvp(bp);
1372 }
1373
1374 /*
1375 * Fixup numfreebuffers count. The bp is on an appropriate queue
1376 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1377 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1378 * if B_INVAL is set ).
1379 */
1380
1381 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1382 bufcountwakeup();
1383
1384 /*
1385 * Something we can maybe free or reuse
1386 */
1387 if (bp->b_bufsize || bp->b_kvasize)
1388 bufspacewakeup();
1389
1390 /* unlock */
1391 BUF_UNLOCK(bp);
1392 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1393 B_DIRECT | B_NOWDRAIN);
1394 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1395 panic("brelse: not dirty");
1396 splx(s);
1397}
1398
1399/*
1400 * Release a buffer back to the appropriate queue but do not try to free
1401 * it. The buffer is expected to be used again soon.
1402 *
1403 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1404 * biodone() to requeue an async I/O on completion. It is also used when
1405 * known good buffers need to be requeued but we think we may need the data
1406 * again soon.
1407 *
1408 * XXX we should be able to leave the B_RELBUF hint set on completion.
1409 */
1410void
1411bqrelse(struct buf * bp)
1412{
1413 int s;
1414
1415 s = splbio();
1416
1417 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1418
1419 if (bp->b_qindex != QUEUE_NONE)
1420 panic("bqrelse: free buffer onto another queue???");
1421 if (BUF_REFCNT(bp) > 1) {
1422 /* do not release to free list */
1423 BUF_UNLOCK(bp);
1424 splx(s);
1425 return;
1426 }
1427 if (bp->b_flags & B_LOCKED) {
1428 bp->b_ioflags &= ~BIO_ERROR;
1429 bp->b_qindex = QUEUE_LOCKED;
1430 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1431 /* buffers with stale but valid contents */
1432 } else if (bp->b_flags & B_DELWRI) {
1433 bp->b_qindex = QUEUE_DIRTY;
1434 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1435 } else if (vm_page_count_severe()) {
1436 /*
1437 * We are too low on memory, we have to try to free the
1438 * buffer (most importantly: the wired pages making up its
1439 * backing store) *now*.
1440 */
1441 splx(s);
1442 brelse(bp);
1443 return;
1444 } else {
1445 bp->b_qindex = QUEUE_CLEAN;
1446 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1447 }
1448
1449 if ((bp->b_flags & B_LOCKED) == 0 &&
1450 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1451 bufcountwakeup();
1452 }
1453
1454 /*
1455 * Something we can maybe free or reuse.
1456 */
1457 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1458 bufspacewakeup();
1459
1460 /* unlock */
1461 BUF_UNLOCK(bp);
1462 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1463 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1464 panic("bqrelse: not dirty");
1465 splx(s);
1466}
1467
1468/* Give pages used by the bp back to the VM system (where possible) */
1469static void
1470vfs_vmio_release(bp)
1471 struct buf *bp;
1472{
1473 int i;
1474 vm_page_t m;
1475
1476 GIANT_REQUIRED;
1477 vm_page_lock_queues();
1478 for (i = 0; i < bp->b_npages; i++) {
1479 m = bp->b_pages[i];
1480 bp->b_pages[i] = NULL;
1481 /*
1482 * In order to keep page LRU ordering consistent, put
1483 * everything on the inactive queue.
1484 */
1485 vm_page_unwire(m, 0);
1486 /*
1487 * We don't mess with busy pages, it is
1488 * the responsibility of the process that
1489 * busied the pages to deal with them.
1490 */
1491 if ((m->flags & PG_BUSY) || (m->busy != 0))
1492 continue;
1493
1494 if (m->wire_count == 0) {
1495 vm_page_flag_clear(m, PG_ZERO);
1496 /*
1497 * Might as well free the page if we can and it has
1498 * no valid data. We also free the page if the
1499 * buffer was used for direct I/O
1500 */
1501 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1502 m->hold_count == 0) {
1503 vm_page_busy(m);
1504 vm_page_protect(m, VM_PROT_NONE);
1505 vm_page_free(m);
1506 } else if (bp->b_flags & B_DIRECT) {
1507 vm_page_try_to_free(m);
1508 } else if (vm_page_count_severe()) {
1509 vm_page_try_to_cache(m);
1510 }
1511 }
1512 }
1513 vm_page_unlock_queues();
1514 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1515
1516 if (bp->b_bufsize) {
1517 bufspacewakeup();
1518 bp->b_bufsize = 0;
1519 }
1520 bp->b_npages = 0;
1521 bp->b_flags &= ~B_VMIO;
1522 if (bp->b_vp)
1523 brelvp(bp);
1524}
1525
1526#ifdef USE_BUFHASH
1527/*
1528 * XXX MOVED TO VFS_SUBR.C
1529 *
1530 * Check to see if a block is currently memory resident.
1531 */
1532struct buf *
1533gbincore(struct vnode * vp, daddr_t blkno)
1534{
1535 struct buf *bp;
1536 struct bufhashhdr *bh;
1537
1538 bh = bufhash(vp, blkno);
1539
1540 /* Search hash chain */
1541 LIST_FOREACH(bp, bh, b_hash) {
1542 /* hit */
1543 if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1544 (bp->b_flags & B_INVAL) == 0) {
1545 break;
1546 }
1547 }
1548 return (bp);
1549}
1550#endif
1551
1552/*
1553 * vfs_bio_awrite:
1554 *
1555 * Implement clustered async writes for clearing out B_DELWRI buffers.
1556 * This is much better then the old way of writing only one buffer at
1557 * a time. Note that we may not be presented with the buffers in the
1558 * correct order, so we search for the cluster in both directions.
1559 */
1560int
1561vfs_bio_awrite(struct buf * bp)
1562{
1563 int i;
1564 int j;
1565 daddr_t lblkno = bp->b_lblkno;
1566 struct vnode *vp = bp->b_vp;
1567 int s;
1568 int ncl;
1569 struct buf *bpa;
1570 int nwritten;
1571 int size;
1572 int maxcl;
1573
1574 s = splbio();
1575 /*
1576 * right now we support clustered writing only to regular files. If
1577 * we find a clusterable block we could be in the middle of a cluster
1578 * rather then at the beginning.
1579 */
1580 if ((vp->v_type == VREG) &&
1581 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1582 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1583
1584 size = vp->v_mount->mnt_stat.f_iosize;
1585 maxcl = MAXPHYS / size;
1586
1587 for (i = 1; i < maxcl; i++) {
1588 if ((bpa = gbincore(vp, lblkno + i)) &&
1589 BUF_REFCNT(bpa) == 0 &&
1590 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1591 (B_DELWRI | B_CLUSTEROK)) &&
1592 (bpa->b_bufsize == size)) {
1593 if ((bpa->b_blkno == bpa->b_lblkno) ||
1594 (bpa->b_blkno !=
1595 bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1596 break;
1597 } else {
1598 break;
1599 }
1600 }
1601 for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1602 if ((bpa = gbincore(vp, lblkno - j)) &&
1603 BUF_REFCNT(bpa) == 0 &&
1604 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1605 (B_DELWRI | B_CLUSTEROK)) &&
1606 (bpa->b_bufsize == size)) {
1607 if ((bpa->b_blkno == bpa->b_lblkno) ||
1608 (bpa->b_blkno !=
1609 bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1610 break;
1611 } else {
1612 break;
1613 }
1614 }
1615 --j;
1616 ncl = i + j;
1617 /*
1618 * this is a possible cluster write
1619 */
1620 if (ncl != 1) {
1621 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1622 splx(s);
1623 return nwritten;
1624 }
1625 }
1626
1627 BUF_LOCK(bp, LK_EXCLUSIVE);
1628 bremfree(bp);
1629 bp->b_flags |= B_ASYNC;
1630
1631 splx(s);
1632 /*
1633 * default (old) behavior, writing out only one block
1634 *
1635 * XXX returns b_bufsize instead of b_bcount for nwritten?
1636 */
1637 nwritten = bp->b_bufsize;
1638 (void) BUF_WRITE(bp);
1639
1640 return nwritten;
1641}
1642
1643/*
1644 * getnewbuf:
1645 *
1646 * Find and initialize a new buffer header, freeing up existing buffers
1647 * in the bufqueues as necessary. The new buffer is returned locked.
1648 *
1649 * Important: B_INVAL is not set. If the caller wishes to throw the
1650 * buffer away, the caller must set B_INVAL prior to calling brelse().
1651 *
1652 * We block if:
1653 * We have insufficient buffer headers
1654 * We have insufficient buffer space
1655 * buffer_map is too fragmented ( space reservation fails )
1656 * If we have to flush dirty buffers ( but we try to avoid this )
1657 *
1658 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1659 * Instead we ask the buf daemon to do it for us. We attempt to
1660 * avoid piecemeal wakeups of the pageout daemon.
1661 */
1662
1663static struct buf *
1664getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1665{
1666 struct buf *bp;
1667 struct buf *nbp;
1668 int defrag = 0;
1669 int nqindex;
1670 static int flushingbufs;
1671
1672 GIANT_REQUIRED;
1673
1674 /*
1675 * We can't afford to block since we might be holding a vnode lock,
1676 * which may prevent system daemons from running. We deal with
1677 * low-memory situations by proactively returning memory and running
1678 * async I/O rather then sync I/O.
1679 */
1680
1681 ++getnewbufcalls;
1682 --getnewbufrestarts;
1683restart:
1684 ++getnewbufrestarts;
1685
1686 /*
1687 * Setup for scan. If we do not have enough free buffers,
1688 * we setup a degenerate case that immediately fails. Note
1689 * that if we are specially marked process, we are allowed to
1690 * dip into our reserves.
1691 *
1692 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1693 *
1694 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1695 * However, there are a number of cases (defragging, reusing, ...)
1696 * where we cannot backup.
1697 */
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 /*
1759 * Sanity Checks
1760 */
1761 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1762
1763 /*
1764 * Note: we no longer distinguish between VMIO and non-VMIO
1765 * buffers.
1766 */
1767
1768 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1769
1770 /*
1771 * If we are defragging then we need a buffer with
1772 * b_kvasize != 0. XXX this situation should no longer
1773 * occur, if defrag is non-zero the buffer's b_kvasize
1774 * should also be non-zero at this point. XXX
1775 */
1776 if (defrag && bp->b_kvasize == 0) {
1777 printf("Warning: defrag empty buffer %p\n", bp);
1778 continue;
1779 }
1780
1781 /*
1782 * Start freeing the bp. This is somewhat involved. nbp
1783 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1784 */
1785
1786 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1787 panic("getnewbuf: locked buf");
1788 bremfree(bp);
1789
1790 if (qindex == QUEUE_CLEAN) {
1791 if (bp->b_flags & B_VMIO) {
1792 bp->b_flags &= ~B_ASYNC;
1793 vfs_vmio_release(bp);
1794 }
1795 if (bp->b_vp)
1796 brelvp(bp);
1797 }
1798
1799 /*
1800 * NOTE: nbp is now entirely invalid. We can only restart
1801 * the scan from this point on.
1802 *
1803 * Get the rest of the buffer freed up. b_kva* is still
1804 * valid after this operation.
1805 */
1806
1807 if (bp->b_rcred != NOCRED) {
1808 crfree(bp->b_rcred);
1809 bp->b_rcred = NOCRED;
1810 }
1811 if (bp->b_wcred != NOCRED) {
1812 crfree(bp->b_wcred);
1813 bp->b_wcred = NOCRED;
1814 }
1815 if (LIST_FIRST(&bp->b_dep) != NULL)
1816 buf_deallocate(bp);
1817 if (bp->b_xflags & BX_BKGRDINPROG)
1818 panic("losing buffer 3");
1819#ifdef USE_BUFHASH
1820 LIST_REMOVE(bp, b_hash);
1821 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1822#endif
1823
1824 if (bp->b_bufsize)
1825 allocbuf(bp, 0);
1826
1827 bp->b_flags = 0;
1828 bp->b_ioflags = 0;
1829 bp->b_xflags = 0;
1830 bp->b_dev = NODEV;
1831 bp->b_vp = NULL;
1832 bp->b_blkno = bp->b_lblkno = 0;
1833 bp->b_offset = NOOFFSET;
1834 bp->b_iodone = 0;
1835 bp->b_error = 0;
1836 bp->b_resid = 0;
1837 bp->b_bcount = 0;
1838 bp->b_npages = 0;
1839 bp->b_dirtyoff = bp->b_dirtyend = 0;
1840 bp->b_magic = B_MAGIC_BIO;
1841 bp->b_op = &buf_ops_bio;
1842 bp->b_object = NULL;
1843
1844 LIST_INIT(&bp->b_dep);
1845
1846 /*
1847 * If we are defragging then free the buffer.
1848 */
1849 if (defrag) {
1850 bp->b_flags |= B_INVAL;
1851 bfreekva(bp);
1852 brelse(bp);
1853 defrag = 0;
1854 goto restart;
1855 }
1856
1857 /*
1858 * If we are overcomitted then recover the buffer and its
1859 * KVM space. This occurs in rare situations when multiple
1860 * processes are blocked in getnewbuf() or allocbuf().
1861 */
1862 if (bufspace >= hibufspace)
1863 flushingbufs = 1;
1864 if (flushingbufs && bp->b_kvasize != 0) {
1865 bp->b_flags |= B_INVAL;
1866 bfreekva(bp);
1867 brelse(bp);
1868 goto restart;
1869 }
1870 if (bufspace < lobufspace)
1871 flushingbufs = 0;
1872 break;
1873 }
1874
1875 /*
1876 * If we exhausted our list, sleep as appropriate. We may have to
1877 * wakeup various daemons and write out some dirty buffers.
1878 *
1879 * Generally we are sleeping due to insufficient buffer space.
1880 */
1881
1882 if (bp == NULL) {
1883 int flags;
1884 char *waitmsg;
1885
1886 if (defrag) {
1887 flags = VFS_BIO_NEED_BUFSPACE;
1888 waitmsg = "nbufkv";
1889 } else if (bufspace >= hibufspace) {
1890 waitmsg = "nbufbs";
1891 flags = VFS_BIO_NEED_BUFSPACE;
1892 } else {
1893 waitmsg = "newbuf";
1894 flags = VFS_BIO_NEED_ANY;
1895 }
1896
1897 bd_speedup(); /* heeeelp */
1898
1899 needsbuffer |= flags;
1900 while (needsbuffer & flags) {
1901 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag,
1902 waitmsg, slptimeo))
1903 return (NULL);
1904 }
1905 } else {
1906 /*
1907 * We finally have a valid bp. We aren't quite out of the
1908 * woods, we still have to reserve kva space. In order
1909 * to keep fragmentation sane we only allocate kva in
1910 * BKVASIZE chunks.
1911 */
1912 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1913
1914 if (maxsize != bp->b_kvasize) {
1915 vm_offset_t addr = 0;
1916
1917 bfreekva(bp);
1918
1919 if (vm_map_findspace(buffer_map,
1920 vm_map_min(buffer_map), maxsize, &addr)) {
1921 /*
1922 * Uh oh. Buffer map is to fragmented. We
1923 * must defragment the map.
1924 */
1925 ++bufdefragcnt;
1926 defrag = 1;
1927 bp->b_flags |= B_INVAL;
1928 brelse(bp);
1929 goto restart;
1930 }
1931 if (addr) {
1932 vm_map_insert(buffer_map, NULL, 0,
1933 addr, addr + maxsize,
1934 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1935
1936 bp->b_kvabase = (caddr_t) addr;
1937 bp->b_kvasize = maxsize;
1938 bufspace += bp->b_kvasize;
1939 ++bufreusecnt;
1940 }
1941 }
1942 bp->b_data = bp->b_kvabase;
1943 }
1944 return(bp);
1945}
1946
1947/*
1948 * buf_daemon:
1949 *
1950 * buffer flushing daemon. Buffers are normally flushed by the
1951 * update daemon but if it cannot keep up this process starts to
1952 * take the load in an attempt to prevent getnewbuf() from blocking.
1953 */
1954
1955static struct proc *bufdaemonproc;
1956
1957static struct kproc_desc buf_kp = {
1958 "bufdaemon",
1959 buf_daemon,
1960 &bufdaemonproc
1961};
1962SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1963
1964static void
1965buf_daemon()
1966{
1967 int s;
1968
1969 mtx_lock(&Giant);
1970
1971 /*
1972 * This process needs to be suspended prior to shutdown sync.
1973 */
1974 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
1975 SHUTDOWN_PRI_LAST);
1976
1977 /*
1978 * This process is allowed to take the buffer cache to the limit
1979 */
1980 s = splbio();
1981
1982 for (;;) {
1983 kthread_suspend_check(bufdaemonproc);
1984
1985 bd_request = 0;
1986
1987 /*
1988 * Do the flush. Limit the amount of in-transit I/O we
1989 * allow to build up, otherwise we would completely saturate
1990 * the I/O system. Wakeup any waiting processes before we
1991 * normally would so they can run in parallel with our drain.
1992 */
1993 while (numdirtybuffers > lodirtybuffers) {
1994 if (flushbufqueues() == 0)
1995 break;
1996 waitrunningbufspace();
1997 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1998 }
1999
2000 /*
2001 * Only clear bd_request if we have reached our low water
2002 * mark. The buf_daemon normally waits 1 second and
2003 * then incrementally flushes any dirty buffers that have
2004 * built up, within reason.
2005 *
2006 * If we were unable to hit our low water mark and couldn't
2007 * find any flushable buffers, we sleep half a second.
2008 * Otherwise we loop immediately.
2009 */
2010 if (numdirtybuffers <= lodirtybuffers) {
2011 /*
2012 * We reached our low water mark, reset the
2013 * request and sleep until we are needed again.
2014 * The sleep is just so the suspend code works.
2015 */
2016 bd_request = 0;
2017 tsleep(&bd_request, PVM, "psleep", hz);
2018 } else {
2019 /*
2020 * We couldn't find any flushable dirty buffers but
2021 * still have too many dirty buffers, we
2022 * have to sleep and try again. (rare)
2023 */
2024 tsleep(&bd_request, PVM, "qsleep", hz / 2);
2025 }
2026 }
2027}
2028
2029/*
2030 * flushbufqueues:
2031 *
2032 * Try to flush a buffer in the dirty queue. We must be careful to
2033 * free up B_INVAL buffers instead of write them, which NFS is
2034 * particularly sensitive to.
2035 */
2036
2037static int
2038flushbufqueues(void)
2039{
2040 struct buf *bp;
2041 int r = 0;
2042
2043 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
2044
2045 while (bp) {
2046 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
2047 if ((bp->b_flags & B_DELWRI) != 0 &&
2048 (bp->b_xflags & BX_BKGRDINPROG) == 0) {
2049 if (bp->b_flags & B_INVAL) {
2050 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2051 panic("flushbufqueues: locked buf");
2052 bremfree(bp);
2053 brelse(bp);
2054 ++r;
2055 break;
2056 }
2057 if (LIST_FIRST(&bp->b_dep) != NULL &&
2058 (bp->b_flags & B_DEFERRED) == 0 &&
2059 buf_countdeps(bp, 0)) {
2060 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
2061 bp, b_freelist);
2062 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
2063 bp, b_freelist);
2064 bp->b_flags |= B_DEFERRED;
2065 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
2066 continue;
2067 }
2068 vfs_bio_awrite(bp);
2069 ++r;
2070 break;
2071 }
2072 bp = TAILQ_NEXT(bp, b_freelist);
2073 }
2074 return (r);
2075}
2076
2077/*
2078 * Check to see if a block is currently memory resident.
2079 */
2080struct buf *
2081incore(struct vnode * vp, daddr_t blkno)
2082{
2083 struct buf *bp;
2084
2085 int s = splbio();
2086 bp = gbincore(vp, blkno);
2087 splx(s);
2088 return (bp);
2089}
2090
2091/*
2092 * Returns true if no I/O is needed to access the
2093 * associated VM object. This is like incore except
2094 * it also hunts around in the VM system for the data.
2095 */
2096
2097int
2098inmem(struct vnode * vp, daddr_t blkno)
2099{
2100 vm_object_t obj;
2101 vm_offset_t toff, tinc, size;
2102 vm_page_t m;
2103 vm_ooffset_t off;
2104
2105 GIANT_REQUIRED;
2106
2107 if (incore(vp, blkno))
2108 return 1;
2109 if (vp->v_mount == NULL)
2110 return 0;
2111 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0)
2112 return 0;
2113
2114 size = PAGE_SIZE;
2115 if (size > vp->v_mount->mnt_stat.f_iosize)
2116 size = vp->v_mount->mnt_stat.f_iosize;
2117 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2118
2119 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2120 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2121 if (!m)
2122 goto notinmem;
2123 tinc = size;
2124 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2125 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2126 if (vm_page_is_valid(m,
2127 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2128 goto notinmem;
2129 }
2130 return 1;
2131
2132notinmem:
2133 return (0);
2134}
2135
2136/*
2137 * vfs_setdirty:
2138 *
2139 * Sets the dirty range for a buffer based on the status of the dirty
2140 * bits in the pages comprising the buffer.
2141 *
2142 * The range is limited to the size of the buffer.
2143 *
2144 * This routine is primarily used by NFS, but is generalized for the
2145 * B_VMIO case.
2146 */
2147static void
2148vfs_setdirty(struct buf *bp)
2149{
2150 int i;
2151 vm_object_t object;
2152
2153 GIANT_REQUIRED;
2154 /*
2155 * Degenerate case - empty buffer
2156 */
2157
2158 if (bp->b_bufsize == 0)
2159 return;
2160
2161 /*
2162 * We qualify the scan for modified pages on whether the
2163 * object has been flushed yet. The OBJ_WRITEABLE flag
2164 * is not cleared simply by protecting pages off.
2165 */
2166
2167 if ((bp->b_flags & B_VMIO) == 0)
2168 return;
2169
2170 object = bp->b_pages[0]->object;
2171
2172 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2173 printf("Warning: object %p writeable but not mightbedirty\n", object);
2174 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2175 printf("Warning: object %p mightbedirty but not writeable\n", object);
2176
2177 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2178 vm_offset_t boffset;
2179 vm_offset_t eoffset;
2180
2181 /*
2182 * test the pages to see if they have been modified directly
2183 * by users through the VM system.
2184 */
2185 for (i = 0; i < bp->b_npages; i++) {
2186 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
2187 vm_page_test_dirty(bp->b_pages[i]);
2188 }
2189
2190 /*
2191 * Calculate the encompassing dirty range, boffset and eoffset,
2192 * (eoffset - boffset) bytes.
2193 */
2194
2195 for (i = 0; i < bp->b_npages; i++) {
2196 if (bp->b_pages[i]->dirty)
2197 break;
2198 }
2199 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2200
2201 for (i = bp->b_npages - 1; i >= 0; --i) {
2202 if (bp->b_pages[i]->dirty) {
2203 break;
2204 }
2205 }
2206 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2207
2208 /*
2209 * Fit it to the buffer.
2210 */
2211
2212 if (eoffset > bp->b_bcount)
2213 eoffset = bp->b_bcount;
2214
2215 /*
2216 * If we have a good dirty range, merge with the existing
2217 * dirty range.
2218 */
2219
2220 if (boffset < eoffset) {
2221 if (bp->b_dirtyoff > boffset)
2222 bp->b_dirtyoff = boffset;
2223 if (bp->b_dirtyend < eoffset)
2224 bp->b_dirtyend = eoffset;
2225 }
2226 }
2227}
2228
2229/*
2230 * getblk:
2231 *
2232 * Get a block given a specified block and offset into a file/device.
2233 * The buffers B_DONE bit will be cleared on return, making it almost
2234 * ready for an I/O initiation. B_INVAL may or may not be set on
2235 * return. The caller should clear B_INVAL prior to initiating a
2236 * READ.
2237 *
2238 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2239 * an existing buffer.
2240 *
2241 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2242 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2243 * and then cleared based on the backing VM. If the previous buffer is
2244 * non-0-sized but invalid, B_CACHE will be cleared.
2245 *
2246 * If getblk() must create a new buffer, the new buffer is returned with
2247 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2248 * case it is returned with B_INVAL clear and B_CACHE set based on the
2249 * backing VM.
2250 *
2251 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos
2252 * B_CACHE bit is clear.
2253 *
2254 * What this means, basically, is that the caller should use B_CACHE to
2255 * determine whether the buffer is fully valid or not and should clear
2256 * B_INVAL prior to issuing a read. If the caller intends to validate
2257 * the buffer by loading its data area with something, the caller needs
2258 * to clear B_INVAL. If the caller does this without issuing an I/O,
2259 * the caller should set B_CACHE ( as an optimization ), else the caller
2260 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2261 * a write attempt or if it was a successfull read. If the caller
2262 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2263 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2264 */
2265struct buf *
2266getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2267{
2268 struct buf *bp;
2269 int s;
2270#ifdef USE_BUFHASH
2271 struct bufhashhdr *bh;
2272#endif
2273
2274 if (size > MAXBSIZE)
2275 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2276
2277 s = splbio();
2278loop:
2279 /*
2280 * Block if we are low on buffers. Certain processes are allowed
2281 * to completely exhaust the buffer cache.
2282 *
2283 * If this check ever becomes a bottleneck it may be better to
2284 * move it into the else, when gbincore() fails. At the moment
2285 * it isn't a problem.
2286 *
2287 * XXX remove if 0 sections (clean this up after its proven)
2288 */
2289 if (numfreebuffers == 0) {
2290 if (curthread == PCPU_GET(idlethread))
2291 return NULL;
2292 needsbuffer |= VFS_BIO_NEED_ANY;
2293 }
2294
2295 if ((bp = gbincore(vp, blkno))) {
2296 /*
2297 * Buffer is in-core. If the buffer is not busy, it must
2298 * be on a queue.
2299 */
2300
2301 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2302 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2303 "getblk", slpflag, slptimeo) == ENOLCK)
2304 goto loop;
2305 splx(s);
2306 return (struct buf *) NULL;
2307 }
2308
2309 /*
2310 * The buffer is locked. B_CACHE is cleared if the buffer is
2311 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2312 * and for a VMIO buffer B_CACHE is adjusted according to the
2313 * backing VM cache.
2314 */
2315 if (bp->b_flags & B_INVAL)
2316 bp->b_flags &= ~B_CACHE;
2317 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2318 bp->b_flags |= B_CACHE;
2319 bremfree(bp);
2320
2321 /*
2322 * check for size inconsistancies for non-VMIO case.
2323 */
2324
2325 if (bp->b_bcount != size) {
2326 if ((bp->b_flags & B_VMIO) == 0 ||
2327 (size > bp->b_kvasize)) {
2328 if (bp->b_flags & B_DELWRI) {
2329 bp->b_flags |= B_NOCACHE;
2330 BUF_WRITE(bp);
2331 } else {
2332 if ((bp->b_flags & B_VMIO) &&
2333 (LIST_FIRST(&bp->b_dep) == NULL)) {
2334 bp->b_flags |= B_RELBUF;
2335 brelse(bp);
2336 } else {
2337 bp->b_flags |= B_NOCACHE;
2338 BUF_WRITE(bp);
2339 }
2340 }
2341 goto loop;
2342 }
2343 }
2344
2345 /*
2346 * If the size is inconsistant in the VMIO case, we can resize
2347 * the buffer. This might lead to B_CACHE getting set or
2348 * cleared. If the size has not changed, B_CACHE remains
2349 * unchanged from its previous state.
2350 */
2351
2352 if (bp->b_bcount != size)
2353 allocbuf(bp, size);
2354
2355 KASSERT(bp->b_offset != NOOFFSET,
2356 ("getblk: no buffer offset"));
2357
2358 /*
2359 * A buffer with B_DELWRI set and B_CACHE clear must
2360 * be committed before we can return the buffer in
2361 * order to prevent the caller from issuing a read
2362 * ( due to B_CACHE not being set ) and overwriting
2363 * it.
2364 *
2365 * Most callers, including NFS and FFS, need this to
2366 * operate properly either because they assume they
2367 * can issue a read if B_CACHE is not set, or because
2368 * ( for example ) an uncached B_DELWRI might loop due
2369 * to softupdates re-dirtying the buffer. In the latter
2370 * case, B_CACHE is set after the first write completes,
2371 * preventing further loops.
2372 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2373 * above while extending the buffer, we cannot allow the
2374 * buffer to remain with B_CACHE set after the write
2375 * completes or it will represent a corrupt state. To
2376 * deal with this we set B_NOCACHE to scrap the buffer
2377 * after the write.
2378 *
2379 * We might be able to do something fancy, like setting
2380 * B_CACHE in bwrite() except if B_DELWRI is already set,
2381 * so the below call doesn't set B_CACHE, but that gets real
2382 * confusing. This is much easier.
2383 */
2384
2385 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2386 bp->b_flags |= B_NOCACHE;
2387 BUF_WRITE(bp);
2388 goto loop;
2389 }
2390
2391 splx(s);
2392 bp->b_flags &= ~B_DONE;
2393 } else {
2394 /*
2395 * Buffer is not in-core, create new buffer. The buffer
2396 * returned by getnewbuf() is locked. Note that the returned
2397 * buffer is also considered valid (not marked B_INVAL).
2398 */
2399 int bsize, maxsize, vmio;
2400 off_t offset;
2401
2402 if (vn_isdisk(vp, NULL))
2403 bsize = DEV_BSIZE;
2404 else if (vp->v_mountedhere)
2405 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2406 else if (vp->v_mount)
2407 bsize = vp->v_mount->mnt_stat.f_iosize;
2408 else
2409 bsize = size;
2410
2411 offset = blkno * bsize;
2412 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) &&
2413 (vp->v_vflag & VV_OBJBUF);
2414 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2415 maxsize = imax(maxsize, bsize);
2416
2417 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2418 if (slpflag || slptimeo) {
2419 splx(s);
2420 return NULL;
2421 }
2422 goto loop;
2423 }
2424
2425 /*
2426 * This code is used to make sure that a buffer is not
2427 * created while the getnewbuf routine is blocked.
2428 * This can be a problem whether the vnode is locked or not.
2429 * If the buffer is created out from under us, we have to
2430 * throw away the one we just created. There is now window
2431 * race because we are safely running at splbio() from the
2432 * point of the duplicate buffer creation through to here,
2433 * and we've locked the buffer.
2434 *
2435 * Note: this must occur before we associate the buffer
2436 * with the vp especially considering limitations in
2437 * the splay tree implementation when dealing with duplicate
2438 * lblkno's.
2439 */
2440 if (gbincore(vp, blkno)) {
2441 bp->b_flags |= B_INVAL;
2442 brelse(bp);
2443 goto loop;
2444 }
2445
2446 /*
2447 * Insert the buffer into the hash, so that it can
2448 * be found by incore.
2449 */
2450 bp->b_blkno = bp->b_lblkno = blkno;
2451 bp->b_offset = offset;
2452
2453 bgetvp(vp, bp);
2454#ifdef USE_BUFHASH
2455 LIST_REMOVE(bp, b_hash);
2456 bh = bufhash(vp, blkno);
2457 LIST_INSERT_HEAD(bh, bp, b_hash);
2458#endif
2459
2460 /*
2461 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2462 * buffer size starts out as 0, B_CACHE will be set by
2463 * allocbuf() for the VMIO case prior to it testing the
2464 * backing store for validity.
2465 */
2466
2467 if (vmio) {
2468 bp->b_flags |= B_VMIO;
2469#if defined(VFS_BIO_DEBUG)
2470 if (vp->v_type != VREG)
2471 printf("getblk: vmioing file type %d???\n", vp->v_type);
2472#endif
2473 VOP_GETVOBJECT(vp, &bp->b_object);
2474 } else {
2475 bp->b_flags &= ~B_VMIO;
2476 bp->b_object = NULL;
2477 }
2478
2479 allocbuf(bp, size);
2480
2481 splx(s);
2482 bp->b_flags &= ~B_DONE;
2483 }
2484 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2485 return (bp);
2486}
2487
2488/*
2489 * Get an empty, disassociated buffer of given size. The buffer is initially
2490 * set to B_INVAL.
2491 */
2492struct buf *
2493geteblk(int size)
2494{
2495 struct buf *bp;
2496 int s;
2497 int maxsize;
2498
2499 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2500
2501 s = splbio();
2502 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2503 splx(s);
2504 allocbuf(bp, size);
2505 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2506 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2507 return (bp);
2508}
2509
2510
2511/*
2512 * This code constitutes the buffer memory from either anonymous system
2513 * memory (in the case of non-VMIO operations) or from an associated
2514 * VM object (in the case of VMIO operations). This code is able to
2515 * resize a buffer up or down.
2516 *
2517 * Note that this code is tricky, and has many complications to resolve
2518 * deadlock or inconsistant data situations. Tread lightly!!!
2519 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2520 * the caller. Calling this code willy nilly can result in the loss of data.
2521 *
2522 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2523 * B_CACHE for the non-VMIO case.
2524 */
2525
2526int
2527allocbuf(struct buf *bp, int size)
2528{
2529 int newbsize, mbsize;
2530 int i;
2531
2532 GIANT_REQUIRED;
2533
2534 if (BUF_REFCNT(bp) == 0)
2535 panic("allocbuf: buffer not busy");
2536
2537 if (bp->b_kvasize < size)
2538 panic("allocbuf: buffer too small");
2539
2540 if ((bp->b_flags & B_VMIO) == 0) {
2541 caddr_t origbuf;
2542 int origbufsize;
2543 /*
2544 * Just get anonymous memory from the kernel. Don't
2545 * mess with B_CACHE.
2546 */
2547 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2548 if (bp->b_flags & B_MALLOC)
2549 newbsize = mbsize;
2550 else
2551 newbsize = round_page(size);
2552
2553 if (newbsize < bp->b_bufsize) {
2554 /*
2555 * malloced buffers are not shrunk
2556 */
2557 if (bp->b_flags & B_MALLOC) {
2558 if (newbsize) {
2559 bp->b_bcount = size;
2560 } else {
2561 free(bp->b_data, M_BIOBUF);
2562 if (bp->b_bufsize) {
2563 bufmallocspace -= bp->b_bufsize;
2564 bufspacewakeup();
2565 bp->b_bufsize = 0;
2566 }
2567 bp->b_data = bp->b_kvabase;
2568 bp->b_bcount = 0;
2569 bp->b_flags &= ~B_MALLOC;
2570 }
2571 return 1;
2572 }
2573 vm_hold_free_pages(
2574 bp,
2575 (vm_offset_t) bp->b_data + newbsize,
2576 (vm_offset_t) bp->b_data + bp->b_bufsize);
2577 } else if (newbsize > bp->b_bufsize) {
2578 /*
2579 * We only use malloced memory on the first allocation.
2580 * and revert to page-allocated memory when the buffer
2581 * grows.
2582 */
2583 if ( (bufmallocspace < maxbufmallocspace) &&
2584 (bp->b_bufsize == 0) &&
2585 (mbsize <= PAGE_SIZE/2)) {
2586
2587 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2588 bp->b_bufsize = mbsize;
2589 bp->b_bcount = size;
2590 bp->b_flags |= B_MALLOC;
2591 bufmallocspace += mbsize;
2592 return 1;
2593 }
2594 origbuf = NULL;
2595 origbufsize = 0;
2596 /*
2597 * If the buffer is growing on its other-than-first allocation,
2598 * then we revert to the page-allocation scheme.
2599 */
2600 if (bp->b_flags & B_MALLOC) {
2601 origbuf = bp->b_data;
2602 origbufsize = bp->b_bufsize;
2603 bp->b_data = bp->b_kvabase;
2604 if (bp->b_bufsize) {
2605 bufmallocspace -= bp->b_bufsize;
2606 bufspacewakeup();
2607 bp->b_bufsize = 0;
2608 }
2609 bp->b_flags &= ~B_MALLOC;
2610 newbsize = round_page(newbsize);
2611 }
2612 vm_hold_load_pages(
2613 bp,
2614 (vm_offset_t) bp->b_data + bp->b_bufsize,
2615 (vm_offset_t) bp->b_data + newbsize);
2616 if (origbuf) {
2617 bcopy(origbuf, bp->b_data, origbufsize);
2618 free(origbuf, M_BIOBUF);
2619 }
2620 }
2621 } else {
2622 vm_page_t m;
2623 int desiredpages;
2624
2625 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2626 desiredpages = (size == 0) ? 0 :
2627 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2628
2629 if (bp->b_flags & B_MALLOC)
2630 panic("allocbuf: VMIO buffer can't be malloced");
2631 /*
2632 * Set B_CACHE initially if buffer is 0 length or will become
2633 * 0-length.
2634 */
2635 if (size == 0 || bp->b_bufsize == 0)
2636 bp->b_flags |= B_CACHE;
2637
2638 if (newbsize < bp->b_bufsize) {
2639 /*
2640 * DEV_BSIZE aligned new buffer size is less then the
2641 * DEV_BSIZE aligned existing buffer size. Figure out
2642 * if we have to remove any pages.
2643 */
2644 if (desiredpages < bp->b_npages) {
2645 vm_page_lock_queues();
2646 for (i = desiredpages; i < bp->b_npages; i++) {
2647 /*
2648 * the page is not freed here -- it
2649 * is the responsibility of
2650 * vnode_pager_setsize
2651 */
2652 m = bp->b_pages[i];
2653 KASSERT(m != bogus_page,
2654 ("allocbuf: bogus page found"));
2655 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2656 vm_page_lock_queues();
2657
2658 bp->b_pages[i] = NULL;
2659 vm_page_unwire(m, 0);
2660 }
2661 vm_page_unlock_queues();
2662 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2663 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2664 bp->b_npages = desiredpages;
2665 }
2666 } else if (size > bp->b_bcount) {
2667 /*
2668 * We are growing the buffer, possibly in a
2669 * byte-granular fashion.
2670 */
2671 struct vnode *vp;
2672 vm_object_t obj;
2673 vm_offset_t toff;
2674 vm_offset_t tinc;
2675
2676 /*
2677 * Step 1, bring in the VM pages from the object,
2678 * allocating them if necessary. We must clear
2679 * B_CACHE if these pages are not valid for the
2680 * range covered by the buffer.
2681 */
2682
2683 vp = bp->b_vp;
2684 obj = bp->b_object;
2685
2686 while (bp->b_npages < desiredpages) {
2687 vm_page_t m;
2688 vm_pindex_t pi;
2689
2690 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2691 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2692 /*
2693 * note: must allocate system pages
2694 * since blocking here could intefere
2695 * with paging I/O, no matter which
2696 * process we are.
2697 */
2698 m = vm_page_alloc(obj, pi,
2699 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
2700 if (m == NULL) {
2701 VM_WAIT;
2702 vm_pageout_deficit += desiredpages - bp->b_npages;
2703 } else {
2704 vm_page_lock_queues();
2705 vm_page_wakeup(m);
2706 vm_page_unlock_queues();
2707 bp->b_flags &= ~B_CACHE;
2708 bp->b_pages[bp->b_npages] = m;
2709 ++bp->b_npages;
2710 }
2711 continue;
2712 }
2713
2714 /*
2715 * We found a page. If we have to sleep on it,
2716 * retry because it might have gotten freed out
2717 * from under us.
2718 *
2719 * We can only test PG_BUSY here. Blocking on
2720 * m->busy might lead to a deadlock:
2721 *
2722 * vm_fault->getpages->cluster_read->allocbuf
2723 *
2724 */
2725 vm_page_lock_queues();
2726 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2727 continue;
2728
2729 /*
2730 * We have a good page. Should we wakeup the
2731 * page daemon?
2732 */
2733 if ((curproc != pageproc) &&
2734 ((m->queue - m->pc) == PQ_CACHE) &&
2735 ((cnt.v_free_count + cnt.v_cache_count) <
2736 (cnt.v_free_min + cnt.v_cache_min))) {
2737 pagedaemon_wakeup();
2738 }
2739 vm_page_flag_clear(m, PG_ZERO);
2740 vm_page_wire(m);
2741 vm_page_unlock_queues();
2742 bp->b_pages[bp->b_npages] = m;
2743 ++bp->b_npages;
2744 }
2745
2746 /*
2747 * Step 2. We've loaded the pages into the buffer,
2748 * we have to figure out if we can still have B_CACHE
2749 * set. Note that B_CACHE is set according to the
2750 * byte-granular range ( bcount and size ), new the
2751 * aligned range ( newbsize ).
2752 *
2753 * The VM test is against m->valid, which is DEV_BSIZE
2754 * aligned. Needless to say, the validity of the data
2755 * needs to also be DEV_BSIZE aligned. Note that this
2756 * fails with NFS if the server or some other client
2757 * extends the file's EOF. If our buffer is resized,
2758 * B_CACHE may remain set! XXX
2759 */
2760
2761 toff = bp->b_bcount;
2762 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2763
2764 while ((bp->b_flags & B_CACHE) && toff < size) {
2765 vm_pindex_t pi;
2766
2767 if (tinc > (size - toff))
2768 tinc = size - toff;
2769
2770 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2771 PAGE_SHIFT;
2772
2773 vfs_buf_test_cache(
2774 bp,
2775 bp->b_offset,
2776 toff,
2777 tinc,
2778 bp->b_pages[pi]
2779 );
2780 toff += tinc;
2781 tinc = PAGE_SIZE;
2782 }
2783
2784 /*
2785 * Step 3, fixup the KVM pmap. Remember that
2786 * bp->b_data is relative to bp->b_offset, but
2787 * bp->b_offset may be offset into the first page.
2788 */
2789
2790 bp->b_data = (caddr_t)
2791 trunc_page((vm_offset_t)bp->b_data);
2792 pmap_qenter(
2793 (vm_offset_t)bp->b_data,
2794 bp->b_pages,
2795 bp->b_npages
2796 );
2797
2798 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2799 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2800 }
2801 }
2802 if (newbsize < bp->b_bufsize)
2803 bufspacewakeup();
2804 bp->b_bufsize = newbsize; /* actual buffer allocation */
2805 bp->b_bcount = size; /* requested buffer size */
2806 return 1;
2807}
2808
2809void
2810biodone(struct bio *bp)
2811{
2812 bp->bio_flags |= BIO_DONE;
2813 if (bp->bio_done != NULL)
2814 bp->bio_done(bp);
2815 else
2816 wakeup(bp);
2817}
2818
2819/*
2820 * Wait for a BIO to finish.
2821 * XXX: For now resort to a timeout, the optimal locking (if any) for this
2822 * case is not at this point obvious.
2823 */
2824int
2825biowait(struct bio *bp, const char *wchan)
2826{
2827
2828 while ((bp->bio_flags & BIO_DONE) == 0)
2829 msleep(bp, NULL, 0, wchan, hz);
2830 if (!(bp->bio_flags & BIO_ERROR))
2831 return (0);
2832 if (bp->bio_error)
2833 return (bp->bio_error);
2834 return (EIO);
2835}
2836
2837/*
2838 * bufwait:
2839 *
2840 * Wait for buffer I/O completion, returning error status. The buffer
2841 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR
2842 * error and cleared.
2843 */
2844int
2845bufwait(register struct buf * bp)
2846{
2847 int s;
2848
2849 s = splbio();
2850 while ((bp->b_flags & B_DONE) == 0) {
2851 if (bp->b_iocmd == BIO_READ)
2852 tsleep(bp, PRIBIO, "biord", 0);
2853 else
2854 tsleep(bp, PRIBIO, "biowr", 0);
2855 }
2856 splx(s);
2857 if (bp->b_flags & B_EINTR) {
2858 bp->b_flags &= ~B_EINTR;
2859 return (EINTR);
2860 }
2861 if (bp->b_ioflags & BIO_ERROR) {
2862 return (bp->b_error ? bp->b_error : EIO);
2863 } else {
2864 return (0);
2865 }
2866}
2867
2868 /*
2869 * Call back function from struct bio back up to struct buf.
2870 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY().
2871 */
2872void
2873bufdonebio(struct bio *bp)
2874{
2875 bufdone(bp->bio_caller2);
2876}
2877
2878/*
2879 * bufdone:
2880 *
2881 * Finish I/O on a buffer, optionally calling a completion function.
2882 * This is usually called from an interrupt so process blocking is
2883 * not allowed.
2884 *
2885 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2886 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2887 * assuming B_INVAL is clear.
2888 *
2889 * For the VMIO case, we set B_CACHE if the op was a read and no
2890 * read error occured, or if the op was a write. B_CACHE is never
2891 * set if the buffer is invalid or otherwise uncacheable.
2892 *
2893 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2894 * initiator to leave B_INVAL set to brelse the buffer out of existance
2895 * in the biodone routine.
2896 */
2897void
2898bufdone(struct buf *bp)
2899{
2900 int s;
2901 void (*biodone)(struct buf *);
2902
2903 GIANT_REQUIRED;
2904
2905 s = splbio();
2906
2907 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
2908 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2909
2910 bp->b_flags |= B_DONE;
2911 runningbufwakeup(bp);
2912
2913 if (bp->b_iocmd == BIO_DELETE) {
2914 brelse(bp);
2915 splx(s);
2916 return;
2917 }
2918
2919 if (bp->b_iocmd == BIO_WRITE) {
2920 vwakeup(bp);
2921 }
2922
2923 /* call optional completion function if requested */
2924 if (bp->b_iodone != NULL) {
2925 biodone = bp->b_iodone;
2926 bp->b_iodone = NULL;
2927 (*biodone) (bp);
2928 splx(s);
2929 return;
2930 }
2931 if (LIST_FIRST(&bp->b_dep) != NULL)
2932 buf_complete(bp);
2933
2934 if (bp->b_flags & B_VMIO) {
2935 int i;
2936 vm_ooffset_t foff;
2937 vm_page_t m;
2938 vm_object_t obj;
2939 int iosize;
2940 struct vnode *vp = bp->b_vp;
2941
2942 obj = bp->b_object;
2943
2944#if defined(VFS_BIO_DEBUG)
2945 mp_fixme("usecount and vflag accessed without locks.");
2946 if (vp->v_usecount == 0) {
2947 panic("biodone: zero vnode ref count");
2948 }
2949
2950 if ((vp->v_vflag & VV_OBJBUF) == 0) {
2951 panic("biodone: vnode is not setup for merged cache");
2952 }
2953#endif
2954
2955 foff = bp->b_offset;
2956 KASSERT(bp->b_offset != NOOFFSET,
2957 ("biodone: no buffer offset"));
2958
2959#if defined(VFS_BIO_DEBUG)
2960 if (obj->paging_in_progress < bp->b_npages) {
2961 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
2962 obj->paging_in_progress, bp->b_npages);
2963 }
2964#endif
2965
2966 /*
2967 * Set B_CACHE if the op was a normal read and no error
2968 * occured. B_CACHE is set for writes in the b*write()
2969 * routines.
2970 */
2971 iosize = bp->b_bcount - bp->b_resid;
2972 if (bp->b_iocmd == BIO_READ &&
2973 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
2974 !(bp->b_ioflags & BIO_ERROR)) {
2975 bp->b_flags |= B_CACHE;
2976 }
2977 vm_page_lock_queues();
2978 for (i = 0; i < bp->b_npages; i++) {
2979 int bogusflag = 0;
2980 int resid;
2981
2982 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2983 if (resid > iosize)
2984 resid = iosize;
2985
2986 /*
2987 * cleanup bogus pages, restoring the originals
2988 */
2989 m = bp->b_pages[i];
2990 if (m == bogus_page) {
2991 bogusflag = 1;
2992 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2993 if (m == NULL)
2994 panic("biodone: page disappeared!");
2995 bp->b_pages[i] = m;
2996 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2997 }
2998#if defined(VFS_BIO_DEBUG)
2999 if (OFF_TO_IDX(foff) != m->pindex) {
3000 printf(
3001"biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3002 (intmax_t)foff, (uintmax_t)m->pindex);
3003 }
3004#endif
3005
3006 /*
3007 * In the write case, the valid and clean bits are
3008 * already changed correctly ( see bdwrite() ), so we
3009 * only need to do this here in the read case.
3010 */
3011 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3012 vfs_page_set_valid(bp, foff, i, m);
3013 }
3014 vm_page_flag_clear(m, PG_ZERO);
3015
3016 /*
3017 * when debugging new filesystems or buffer I/O methods, this
3018 * is the most common error that pops up. if you see this, you
3019 * have not set the page busy flag correctly!!!
3020 */
3021 if (m->busy == 0) {
3022 printf("biodone: page busy < 0, "
3023 "pindex: %d, foff: 0x(%x,%x), "
3024 "resid: %d, index: %d\n",
3025 (int) m->pindex, (int)(foff >> 32),
3026 (int) foff & 0xffffffff, resid, i);
3027 if (!vn_isdisk(vp, NULL))
3028 printf(" iosize: %ld, lblkno: %jd, flags: 0x%lx, npages: %d\n",
3029 bp->b_vp->v_mount->mnt_stat.f_iosize,
3030 (intmax_t) bp->b_lblkno,
3031 bp->b_flags, bp->b_npages);
3032 else
3033 printf(" VDEV, lblkno: %jd, flags: 0x%lx, npages: %d\n",
3034 (intmax_t) bp->b_lblkno,
3035 bp->b_flags, bp->b_npages);
3036 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3037 m->valid, m->dirty, m->wire_count);
3038 panic("biodone: page busy < 0\n");
3039 }
3040 vm_page_io_finish(m);
3041 vm_object_pip_subtract(obj, 1);
3042 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3043 iosize -= resid;
3044 }
3045 vm_page_unlock_queues();
3046 if (obj)
3047 vm_object_pip_wakeupn(obj, 0);
3048 }
3049
3050 /*
3051 * For asynchronous completions, release the buffer now. The brelse
3052 * will do a wakeup there if necessary - so no need to do a wakeup
3053 * here in the async case. The sync case always needs to do a wakeup.
3054 */
3055
3056 if (bp->b_flags & B_ASYNC) {
3057 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3058 brelse(bp);
3059 else
3060 bqrelse(bp);
3061 } else {
3062 wakeup(bp);
3063 }
3064 splx(s);
3065}
3066
3067/*
3068 * This routine is called in lieu of iodone in the case of
3069 * incomplete I/O. This keeps the busy status for pages
3070 * consistant.
3071 */
3072void
3073vfs_unbusy_pages(struct buf * bp)
3074{
3075 int i;
3076
3077 GIANT_REQUIRED;
3078
3079 runningbufwakeup(bp);
3080 if (bp->b_flags & B_VMIO) {
3081 vm_object_t obj;
3082
3083 obj = bp->b_object;
3084 vm_page_lock_queues();
3085 for (i = 0; i < bp->b_npages; i++) {
3086 vm_page_t m = bp->b_pages[i];
3087
3088 if (m == bogus_page) {
3089 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3090 if (!m) {
3091 panic("vfs_unbusy_pages: page missing\n");
3092 }
3093 bp->b_pages[i] = m;
3094 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3095 }
3096 vm_object_pip_subtract(obj, 1);
3097 vm_page_flag_clear(m, PG_ZERO);
3098 vm_page_io_finish(m);
3099 }
3100 vm_page_unlock_queues();
3101 vm_object_pip_wakeupn(obj, 0);
3102 }
3103}
3104
3105/*
3106 * vfs_page_set_valid:
3107 *
3108 * Set the valid bits in a page based on the supplied offset. The
3109 * range is restricted to the buffer's size.
3110 *
3111 * This routine is typically called after a read completes.
3112 */
3113static void
3114vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3115{
3116 vm_ooffset_t soff, eoff;
3117
3118 GIANT_REQUIRED;
3119 /*
3120 * Start and end offsets in buffer. eoff - soff may not cross a
3121 * page boundry or cross the end of the buffer. The end of the
3122 * buffer, in this case, is our file EOF, not the allocation size
3123 * of the buffer.
3124 */
3125 soff = off;
3126 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3127 if (eoff > bp->b_offset + bp->b_bcount)
3128 eoff = bp->b_offset + bp->b_bcount;
3129
3130 /*
3131 * Set valid range. This is typically the entire buffer and thus the
3132 * entire page.
3133 */
3134 if (eoff > soff) {
3135 vm_page_set_validclean(
3136 m,
3137 (vm_offset_t) (soff & PAGE_MASK),
3138 (vm_offset_t) (eoff - soff)
3139 );
3140 }
3141}
3142
3143/*
3144 * This routine is called before a device strategy routine.
3145 * It is used to tell the VM system that paging I/O is in
3146 * progress, and treat the pages associated with the buffer
3147 * almost as being PG_BUSY. Also the object paging_in_progress
3148 * flag is handled to make sure that the object doesn't become
3149 * inconsistant.
3150 *
3151 * Since I/O has not been initiated yet, certain buffer flags
3152 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3153 * and should be ignored.
3154 */
3155void
3156vfs_busy_pages(struct buf * bp, int clear_modify)
3157{
3158 int i, bogus;
3159
3160 if (bp->b_flags & B_VMIO) {
3161 vm_object_t obj;
3162 vm_ooffset_t foff;
3163
3164 obj = bp->b_object;
3165 foff = bp->b_offset;
3166 KASSERT(bp->b_offset != NOOFFSET,
3167 ("vfs_busy_pages: no buffer offset"));
3168 vfs_setdirty(bp);
3169retry:
3170 vm_page_lock_queues();
3171 for (i = 0; i < bp->b_npages; i++) {
3172 vm_page_t m = bp->b_pages[i];
3173
3174 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3175 goto retry;
3176 }
3177 bogus = 0;
3178 for (i = 0; i < bp->b_npages; i++) {
3179 vm_page_t m = bp->b_pages[i];
3180
3181 vm_page_flag_clear(m, PG_ZERO);
3182 if ((bp->b_flags & B_CLUSTER) == 0) {
3183 vm_object_pip_add(obj, 1);
3184 vm_page_io_start(m);
3185 }
3186 /*
3187 * When readying a buffer for a read ( i.e
3188 * clear_modify == 0 ), it is important to do
3189 * bogus_page replacement for valid pages in
3190 * partially instantiated buffers. Partially
3191 * instantiated buffers can, in turn, occur when
3192 * reconstituting a buffer from its VM backing store
3193 * base. We only have to do this if B_CACHE is
3194 * clear ( which causes the I/O to occur in the
3195 * first place ). The replacement prevents the read
3196 * I/O from overwriting potentially dirty VM-backed
3197 * pages. XXX bogus page replacement is, uh, bogus.
3198 * It may not work properly with small-block devices.
3199 * We need to find a better way.
3200 */
3201 vm_page_protect(m, VM_PROT_NONE);
3202 if (clear_modify)
3203 vfs_page_set_valid(bp, foff, i, m);
3204 else if (m->valid == VM_PAGE_BITS_ALL &&
3205 (bp->b_flags & B_CACHE) == 0) {
3206 bp->b_pages[i] = bogus_page;
3207 bogus++;
3208 }
3209 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3210 }
3211 vm_page_unlock_queues();
3212 if (bogus)
3213 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3214 }
3215}
3216
3217/*
3218 * Tell the VM system that the pages associated with this buffer
3219 * are clean. This is used for delayed writes where the data is
3220 * going to go to disk eventually without additional VM intevention.
3221 *
3222 * Note that while we only really need to clean through to b_bcount, we
3223 * just go ahead and clean through to b_bufsize.
3224 */
3225static void
3226vfs_clean_pages(struct buf * bp)
3227{
3228 int i;
3229
3230 GIANT_REQUIRED;
3231
3232 if (bp->b_flags & B_VMIO) {
3233 vm_ooffset_t foff;
3234
3235 foff = bp->b_offset;
3236 KASSERT(bp->b_offset != NOOFFSET,
3237 ("vfs_clean_pages: no buffer offset"));
3238 for (i = 0; i < bp->b_npages; i++) {
3239 vm_page_t m = bp->b_pages[i];
3240 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3241 vm_ooffset_t eoff = noff;
3242
3243 if (eoff > bp->b_offset + bp->b_bufsize)
3244 eoff = bp->b_offset + bp->b_bufsize;
3245 vfs_page_set_valid(bp, foff, i, m);
3246 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3247 foff = noff;
3248 }
3249 }
3250}
3251
3252/*
3253 * vfs_bio_set_validclean:
3254 *
3255 * Set the range within the buffer to valid and clean. The range is
3256 * relative to the beginning of the buffer, b_offset. Note that b_offset
3257 * itself may be offset from the beginning of the first page.
3258 *
3259 */
3260
3261void
3262vfs_bio_set_validclean(struct buf *bp, int base, int size)
3263{
3264 if (bp->b_flags & B_VMIO) {
3265 int i;
3266 int n;
3267
3268 /*
3269 * Fixup base to be relative to beginning of first page.
3270 * Set initial n to be the maximum number of bytes in the
3271 * first page that can be validated.
3272 */
3273
3274 base += (bp->b_offset & PAGE_MASK);
3275 n = PAGE_SIZE - (base & PAGE_MASK);
3276
3277 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3278 vm_page_t m = bp->b_pages[i];
3279
3280 if (n > size)
3281 n = size;
3282
3283 vm_page_set_validclean(m, base & PAGE_MASK, n);
3284 base += n;
3285 size -= n;
3286 n = PAGE_SIZE;
3287 }
3288 }
3289}
3290
3291/*
3292 * vfs_bio_clrbuf:
3293 *
3294 * clear a buffer. This routine essentially fakes an I/O, so we need
3295 * to clear BIO_ERROR and B_INVAL.
3296 *
3297 * Note that while we only theoretically need to clear through b_bcount,
3298 * we go ahead and clear through b_bufsize.
3299 */
3300
3301void
3302vfs_bio_clrbuf(struct buf *bp)
3303{
3304 int i, mask = 0;
3305 caddr_t sa, ea;
3306
3307 GIANT_REQUIRED;
3308
3309 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3310 bp->b_flags &= ~B_INVAL;
3311 bp->b_ioflags &= ~BIO_ERROR;
3312 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3313 (bp->b_offset & PAGE_MASK) == 0) {
3314 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3315 if ((bp->b_pages[0]->valid & mask) == mask) {
3316 bp->b_resid = 0;
3317 return;
3318 }
3319 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3320 ((bp->b_pages[0]->valid & mask) == 0)) {
3321 bzero(bp->b_data, bp->b_bufsize);
3322 bp->b_pages[0]->valid |= mask;
3323 bp->b_resid = 0;
3324 return;
3325 }
3326 }
3327 ea = sa = bp->b_data;
3328 for(i=0;i<bp->b_npages;i++,sa=ea) {
3329 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3330 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3331 ea = (caddr_t)(vm_offset_t)ulmin(
3332 (u_long)(vm_offset_t)ea,
3333 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3334 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3335 if ((bp->b_pages[i]->valid & mask) == mask)
3336 continue;
3337 if ((bp->b_pages[i]->valid & mask) == 0) {
3338 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3339 bzero(sa, ea - sa);
3340 }
3341 } else {
3342 for (; sa < ea; sa += DEV_BSIZE, j++) {
3343 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3344 (bp->b_pages[i]->valid & (1<<j)) == 0)
3345 bzero(sa, DEV_BSIZE);
3346 }
3347 }
3348 bp->b_pages[i]->valid |= mask;
3349 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3350 }
3351 bp->b_resid = 0;
3352 } else {
3353 clrbuf(bp);
3354 }
3355}
3356
3357/*
3358 * vm_hold_load_pages and vm_hold_free_pages get pages into
3359 * a buffers address space. The pages are anonymous and are
3360 * not associated with a file object.
3361 */
3362static void
3363vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3364{
3365 vm_offset_t pg;
3366 vm_page_t p;
3367 int index;
3368
3369 GIANT_REQUIRED;
3370
3371 to = round_page(to);
3372 from = round_page(from);
3373 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3374
3375 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3376tryagain:
3377 /*
3378 * note: must allocate system pages since blocking here
3379 * could intefere with paging I/O, no matter which
3380 * process we are.
3381 */
3382 p = vm_page_alloc(kernel_object,
3383 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3384 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3385 if (!p) {
3386 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3387 VM_WAIT;
3388 goto tryagain;
3389 }
3390 vm_page_lock_queues();
3391 p->valid = VM_PAGE_BITS_ALL;
3392 vm_page_flag_clear(p, PG_ZERO);
3393 vm_page_unlock_queues();
3394 pmap_qenter(pg, &p, 1);
3395 bp->b_pages[index] = p;
3396 vm_page_wakeup(p);
3397 }
3398 bp->b_npages = index;
3399}
3400
3401/* Return pages associated with this buf to the vm system */
3402void
3403vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3404{
3405 vm_offset_t pg;
3406 vm_page_t p;
3407 int index, newnpages;
3408
3409 GIANT_REQUIRED;
3410
3411 from = round_page(from);
3412 to = round_page(to);
3413 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3414
3415 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3416 p = bp->b_pages[index];
3417 if (p && (index < bp->b_npages)) {
3418 if (p->busy) {
3419 printf(
3420 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3421 (intmax_t)bp->b_blkno,
3422 (intmax_t)bp->b_lblkno);
3423 }
3424 bp->b_pages[index] = NULL;
3425 pmap_qremove(pg, 1);
3426 vm_page_lock_queues();
3427 vm_page_busy(p);
3428 vm_page_unwire(p, 0);
3429 vm_page_free(p);
3430 vm_page_unlock_queues();
3431 }
3432 }
3433 bp->b_npages = newnpages;
3434}
3435
3436
3437#include "opt_ddb.h"
3438#ifdef DDB
3439#include <ddb/ddb.h>
3440
3441/* DDB command to show buffer data */
3442DB_SHOW_COMMAND(buffer, db_show_buffer)
3443{
3444 /* get args */
3445 struct buf *bp = (struct buf *)addr;
3446
3447 if (!have_addr) {
3448 db_printf("usage: show buffer <addr>\n");
3449 return;
3450 }
3451
3452 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3453 db_printf(
3454 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3455 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n",
3456 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3457 major(bp->b_dev), minor(bp->b_dev), bp->b_data,
3458 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno);
3459 if (bp->b_npages) {
3460 int i;
3461 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3462 for (i = 0; i < bp->b_npages; i++) {
3463 vm_page_t m;
3464 m = bp->b_pages[i];
3465 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3466 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3467 if ((i + 1) < bp->b_npages)
3468 db_printf(",");
3469 }
3470 db_printf("\n");
3471 }
3472}
3473#endif /* DDB */
|