2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 #include <afsconfig.h>
11 #include "afs/param.h"
14 #include "afs/sysincludes.h"
15 #include "afsincludes.h"
20 #if defined(AFS_AIX31_ENV)
23 #if !defined(AFS_AIX_ENV) && !defined(AFS_SUN5_ENV) && !defined(AFS_SGI_ENV) && !defined(AFS_LINUX20_ENV)
24 #include "h/kernel.h" /* Doesn't needed, so it should go */
26 #endif /* !defined(UKERNEL) */
28 #include "afs/afs_osi.h"
32 #if !defined(UKERNEL) && !defined(AFS_LINUX20_ENV)
34 #endif /* !defined(UKERNEL) */
37 #include "afs/volerrors.h"
38 #include "afs/exporter.h"
39 #include "afs/prs_fs.h"
40 #include "afs/afs_chunkops.h"
43 #include "afs/afs_stats.h"
47 #define BUF_TIME_MAX 0x7fffffff
49 #define NPB 8 /* must be a pwer of 2 */
50 static int afs_max_buffers; /* should be an integral multiple of NPB */
53 #define AFS_BUFFER_PAGESIZE 2048
56 /* If you change any of this PH stuff, make sure you don't break DZap() */
57 /* use last two bits for page */
59 /* use next five bits for fid */
61 /* page hash table size - this is pretty intertwined with pHash */
62 #define PHSIZE (PHPAGEMASK + PHFIDMASK + 1)
64 #define pHash(fid,page) ((((afs_int32)(fid)) & PHFIDMASK) \
65 | (page & PHPAGEMASK))
68 #undef dirty /* XXX */
71 static struct buffer *Buffers = 0;
72 static char *BufferData;
75 extern struct buf *geteblk();
78 #define timecounter afs_timecounter
81 /* A note on locking in 'struct buffer'
83 * afs_bufferLock protects the hash chain, and the 'lockers' field where that
84 * has a zero value. It must be held whenever lockers is incremented from zero.
86 * The individual buffer lock protects the contents of the structure, including
89 * For safety: afs_bufferLock and the individual buffer lock must be held
90 * when obtaining a reference on a structure. Only the individual buffer lock
91 * need be held when releasing a reference.
93 * The locking hierarchy is afs_bufferLock-> buffer.lock
97 static afs_lock_t afs_bufferLock;
98 static struct buffer *phTable[PHSIZE]; /* page hash table */
100 static afs_int32 timecounter;
102 /* Prototypes for static routines */
103 static struct buffer *afs_newslot(struct dcache *adc, afs_int32 apage,
106 static int dinit_flag = 0;
110 /* Initialize the venus buffer system. */
118 /* round up to next multiple of NPB, since we allocate multiple pages per chunk */
119 abuffers = ((abuffers - 1) | (NPB - 1)) + 1;
120 afs_max_buffers = abuffers << 2; /* possibly grow up to 4 times as big */
121 LOCK_INIT(&afs_bufferLock, "afs_bufferLock");
123 (struct buffer *)afs_osi_Alloc(afs_max_buffers * sizeof(struct buffer));
125 afs_stats_cmperf.bufAlloced = nbuffers = abuffers;
126 for (i = 0; i < PHSIZE; i++)
128 for (i = 0; i < abuffers; i++) {
129 if ((i & (NPB - 1)) == 0) {
130 /* time to allocate a fresh buffer */
131 BufferData = (char *) afs_osi_Alloc(AFS_BUFFER_PAGESIZE * NPB);
133 /* Fill in each buffer with an empty indication. */
136 afs_reset_inode(&tb->inode);
139 tb->data = &BufferData[AFS_BUFFER_PAGESIZE * (i & (NPB - 1))];
142 AFS_RWLOCK_INIT(&tb->lock, "buffer lock");
148 DRead(struct dcache *adc, int page)
150 /* Read a page from the disk. */
151 struct buffer *tb, *tb2;
152 struct osi_file *tfile;
156 ObtainWriteLock(&afs_bufferLock, 256);
158 #define bufmatch(tb) (tb->page == page && tb->fid == adc->index)
159 #define buf_Front(head,parent,p) {(parent)->hashNext = (p)->hashNext; (p)->hashNext= *(head);*(head)=(p);}
161 /* this apparently-complicated-looking code is simply an example of
162 * a little bit of loop unrolling, and is a standard linked-list
163 * traversal trick. It saves a few assignments at the the expense
164 * of larger code size. This could be simplified by better use of
167 if ((tb = phTable[pHash(adc->index, page)])) {
169 ObtainWriteLock(&tb->lock, 257);
171 ReleaseWriteLock(&afs_bufferLock);
172 tb->accesstime = timecounter++;
173 AFS_STATS(afs_stats_cmperf.bufHits++);
174 ReleaseWriteLock(&tb->lock);
177 struct buffer **bufhead;
178 bufhead = &(phTable[pHash(adc->index, page)]);
179 while ((tb2 = tb->hashNext)) {
181 buf_Front(bufhead, tb, tb2);
182 ObtainWriteLock(&tb2->lock, 258);
184 ReleaseWriteLock(&afs_bufferLock);
185 tb2->accesstime = timecounter++;
186 AFS_STATS(afs_stats_cmperf.bufHits++);
187 ReleaseWriteLock(&tb2->lock);
190 if ((tb = tb2->hashNext)) {
192 buf_Front(bufhead, tb2, tb);
193 ObtainWriteLock(&tb->lock, 259);
195 ReleaseWriteLock(&afs_bufferLock);
196 tb->accesstime = timecounter++;
197 AFS_STATS(afs_stats_cmperf.bufHits++);
198 ReleaseWriteLock(&tb->lock);
208 AFS_STATS(afs_stats_cmperf.bufMisses++);
210 /* The last thing we looked at was either tb or tb2 (or nothing). That
211 * is at least the oldest buffer on one particular hash chain, so it's
212 * a pretty good place to start looking for the truly oldest buffer.
214 tb = afs_newslot(adc, page, (tb ? tb : tb2));
216 ReleaseWriteLock(&afs_bufferLock);
219 ObtainWriteLock(&tb->lock, 260);
221 ReleaseWriteLock(&afs_bufferLock);
222 if (page * AFS_BUFFER_PAGESIZE >= adc->f.chunkBytes) {
224 afs_reset_inode(&tb->inode);
226 ReleaseWriteLock(&tb->lock);
229 tfile = afs_CFileOpen(&adc->f.inode);
231 afs_CFileRead(tfile, tb->page * AFS_BUFFER_PAGESIZE, tb->data,
232 AFS_BUFFER_PAGESIZE);
233 afs_CFileClose(tfile);
234 if (code < AFS_BUFFER_PAGESIZE) {
236 afs_reset_inode(&tb->inode);
238 ReleaseWriteLock(&tb->lock);
241 /* Note that findslot sets the page field in the buffer equal to
242 * what it is searching for. */
243 ReleaseWriteLock(&tb->lock);
248 FixupBucket(struct buffer *ap)
250 struct buffer **lp, *tp;
252 /* first try to get it out of its current hash bucket, in which it
254 AFS_STATCNT(FixupBucket);
257 for (tp = *lp; tp; tp = tp->hashNext) {
264 /* now figure the new hash bucket */
265 i = pHash(ap->fid, ap->page);
266 ap->hashIndex = i; /* remember where we are for deletion */
267 ap->hashNext = phTable[i]; /* add us to the list */
268 phTable[i] = ap; /* at the front, since it's LRU */
271 /* lp is pointer to a fairly-old buffer */
272 static struct buffer *
273 afs_newslot(struct dcache *adc, afs_int32 apage, struct buffer *lp)
275 /* Find a usable buffer slot */
279 struct osi_file *tfile;
281 AFS_STATCNT(afs_newslot);
282 /* we take a pointer here to a buffer which was at the end of an
283 * LRU hash chain. Odds are, it's one of the older buffers, not
284 * one of the newer. Having an older buffer to start with may
285 * permit us to avoid a few of the assignments in the "typical
286 * case" for loop below.
288 if (lp && (lp->lockers == 0)) {
294 /* timecounter might have wrapped, if machine is very very busy
295 * and stays up for a long time. Timecounter mustn't wrap twice
296 * (positive->negative->positive) before calling newslot, but that
297 * would require 2 billion consecutive cache hits... Anyway, the
298 * penalty is only that the cache replacement policy will be
299 * almost MRU for the next ~2 billion DReads... newslot doesn't
300 * get called nearly as often as DRead, so in order to avoid the
301 * performance penalty of using the hypers, it's worth doing the
302 * extra check here every time. It's probably cheaper than doing
303 * hcmp, anyway. There is a little performance hit resulting from
304 * resetting all the access times to 0, but it only happens once
305 * every month or so, and the access times will rapidly sort
306 * themselves back out after just a few more DReads.
308 if (timecounter < 0) {
311 for (i = 0; i < nbuffers; i++, tp++) {
313 if (!lp && !tp->lockers) /* one is as good as the rest, I guess */
317 /* this is the typical case */
319 for (i = 0; i < nbuffers; i++, tp++) {
320 if (tp->lockers == 0) {
321 if (!lp || tp->accesstime < lt) {
330 /* No unlocked buffers. If still possible, allocate a new increment */
331 if (nbuffers + NPB > afs_max_buffers) {
332 /* There are no unlocked buffers -- this used to panic, but that
333 * seems extreme. To the best of my knowledge, all the callers
334 * of DRead are prepared to handle a zero return. Some of them
335 * just panic directly, but not all of them. */
336 afs_warn("afs: all buffers locked\n");
340 BufferData = (char *) afs_osi_Alloc(AFS_BUFFER_PAGESIZE * NPB);
341 for (i = 0; i< NPB; i++) {
342 /* Fill in each buffer with an empty indication. */
343 tp = &Buffers[i + nbuffers];
345 afs_reset_inode(&tp->inode);
348 tp->data = &BufferData[AFS_BUFFER_PAGESIZE * i];
351 AFS_RWLOCK_INIT(&tp->lock, "buffer lock");
353 lp = &Buffers[nbuffers];
358 /* see DFlush for rationale for not getting and locking the dcache */
359 tfile = afs_CFileOpen(&lp->inode);
360 afs_CFileWrite(tfile, lp->page * AFS_BUFFER_PAGESIZE, lp->data,
361 AFS_BUFFER_PAGESIZE);
363 afs_CFileClose(tfile);
364 AFS_STATS(afs_stats_cmperf.bufFlushDirty++);
367 /* Now fill in the header. */
368 lp->fid = adc->index;
369 afs_copy_inode(&lp->inode, &adc->f.inode);
371 lp->accesstime = timecounter++;
372 FixupBucket(lp); /* move to the right hash bucket */
378 DRelease(void *loc, int flag)
380 /* Release a buffer, specifying whether or not the buffer has been
381 * modified by the locker. */
382 struct buffer *bp = (struct buffer *)loc;
386 AFS_STATCNT(DRelease);
389 /* look for buffer by scanning Unix buffers for appropriate address */
390 /* careful: despite the declaration above at this point bp is still simply
391 * an address inside the buffer, not a pointer to the buffer header */
393 for (index = 0; index < nbuffers; index += NPB, tp += NPB) {
394 if ( (char *) bp >= (char *) tp->data &&
395 (char *) bp < (char *) tp->data + AFS_BUFFER_PAGESIZE * NPB) {
396 /* we found the right range */
397 index += ((char *) bp - (char *) tp->data) >> LOGPS;
401 tp = &(Buffers[index]);
402 ObtainWriteLock(&tp->lock, 261);
406 ReleaseWriteLock(&tp->lock);
412 /* Return the byte within a file represented by a buffer pointer. */
415 AFS_STATCNT(DVOffset);
416 /* look for buffer by scanning Unix buffers for appropriate address */
417 /* see comment in DRelease about the meaning of ap/bp */
419 for (index = 0; index < nbuffers; index += NPB, tp += NPB) {
420 if ( (char *) ap >= (char *) tp->data &&
421 (char *) ap < (char *) tp->data + AFS_BUFFER_PAGESIZE * NPB) {
422 /* we found the right range */
423 index += ((char *) ap - (char *) tp->data) >> LOGPS;
427 if (index < 0 || index >= nbuffers)
429 tp = &(Buffers[index]);
430 return AFS_BUFFER_PAGESIZE * tp->page + (int)(((char *)ap) - tp->data);
434 * Zap one dcache entry: destroy one FID's buffers.
436 * 1/1/91 - I've modified the hash function to take the page as well
437 * as the *fid, so that lookup will be a bit faster. That presents some
438 * difficulties for Zap, which now has to have some knowledge of the nature
439 * of the hash function. Oh well. This should use the list traversal
442 * \param adc The dcache entry to be zapped.
445 DZap(struct dcache *adc)
448 /* Destroy all buffers pertaining to a particular fid. */
452 ObtainReadLock(&afs_bufferLock);
454 for (i = 0; i <= PHPAGEMASK; i++)
455 for (tb = phTable[pHash(adc->index, i)]; tb; tb = tb->hashNext)
456 if (tb->fid == adc->index) {
457 ObtainWriteLock(&tb->lock, 262);
459 afs_reset_inode(&tb->inode);
461 ReleaseWriteLock(&tb->lock);
463 ReleaseReadLock(&afs_bufferLock);
467 DFlushBuffer(struct buffer *ab)
469 struct osi_file *tfile;
471 tfile = afs_CFileOpen(&ab->inode);
472 afs_CFileWrite(tfile, ab->page * AFS_BUFFER_PAGESIZE,
473 ab->data, AFS_BUFFER_PAGESIZE);
474 ab->dirty = 0; /* Clear the dirty flag */
475 afs_CFileClose(tfile);
479 DFlushDCache(struct dcache *adc)
484 ObtainReadLock(&afs_bufferLock);
486 for (i = 0; i <= PHPAGEMASK; i++)
487 for (tb = phTable[pHash(adc->index, i)]; tb; tb = tb->hashNext)
488 if (tb->fid == adc->index) {
489 ObtainWriteLock(&tb->lock, 701);
491 ReleaseReadLock(&afs_bufferLock);
496 ReleaseWriteLock(&tb->lock);
497 ObtainReadLock(&afs_bufferLock);
500 ReleaseReadLock(&afs_bufferLock);
506 /* Flush all the modified buffers. */
512 ObtainReadLock(&afs_bufferLock);
513 for (i = 0; i < nbuffers; i++, tb++) {
515 ObtainWriteLock(&tb->lock, 263);
517 ReleaseReadLock(&afs_bufferLock);
519 /* it seems safe to do this I/O without having the dcache
520 * locked, since the only things that will update the data in
521 * a directory are the buffer package, which holds the relevant
522 * tb->lock while doing the write, or afs_GetDCache, which
523 * DZap's the directory while holding the dcache lock.
524 * It is not possible to lock the dcache or even call
525 * afs_GetDSlot to map the index to the dcache since the dir
526 * package's caller has some dcache object locked already (so
527 * we cannot lock afs_xdcache). In addition, we cannot obtain
528 * a dcache lock while holding the tb->lock of the same file
529 * since that can deadlock with DRead/DNew */
533 ReleaseWriteLock(&tb->lock);
534 ObtainReadLock(&afs_bufferLock);
537 ReleaseReadLock(&afs_bufferLock);
541 DNew(struct dcache *adc, int page)
543 /* Same as read, only do *not* even try to read the page, since it probably doesn't exist. */
546 ObtainWriteLock(&afs_bufferLock, 264);
547 if ((tb = afs_newslot(adc, page, NULL)) == 0) {
548 ReleaseWriteLock(&afs_bufferLock);
551 /* extend the chunk, if needed */
552 /* Do it now, not in DFlush or afs_newslot when the data is written out,
553 * since now our caller has adc->lock writelocked, and we can't acquire
554 * that lock (or even map from a fid to a dcache) in afs_newslot or
555 * DFlush due to lock hierarchy issues */
556 if ((page + 1) * AFS_BUFFER_PAGESIZE > adc->f.chunkBytes) {
557 afs_AdjustSize(adc, (page + 1) * AFS_BUFFER_PAGESIZE);
558 afs_WriteDCache(adc, 1);
560 ObtainWriteLock(&tb->lock, 265);
562 ReleaseWriteLock(&afs_bufferLock);
563 ReleaseWriteLock(&tb->lock);
568 shutdown_bufferpackage(void)
573 AFS_STATCNT(shutdown_bufferpackage);
574 /* Free all allocated Buffers and associated buffer pages */
576 if (afs_cold_shutdown) {
579 for (i = 0; i < nbuffers; i += NPB, tp += NPB) {
580 afs_osi_Free(tp->data, NPB * AFS_BUFFER_PAGESIZE);
582 afs_osi_Free(Buffers, nbuffers * sizeof(struct buffer));
585 for (i = 0; i < PHSIZE; i++)
587 memset(&afs_bufferLock, 0, sizeof(afs_lock_t));