/*
* Copyright 2000, International Business Machines Corporation and others.
* All Rights Reserved.
- *
+ *
* This software has been released under the terms of the IBM Public
* License. For details, see the LICENSE file in the top-level source
* directory or online at http://www.openafs.org/dl/license10.html
*/
#include <afsconfig.h>
-#include "../afs/param.h"
+#include "afs/param.h"
-RCSID("$Header$");
-#include "../afs/sysincludes.h"
-#include "../afs/afsincludes.h"
+#include "afs/sysincludes.h"
+#include "afsincludes.h"
#if !defined(UKERNEL)
-#include "../h/param.h"
-#include "../h/types.h"
-#include "../h/time.h"
-#if defined(AFS_AIX31_ENV) || defined(AFS_DEC_ENV)
-#include "../h/limits.h"
+#if !defined(AFS_LINUX26_ENV)
+# include "h/param.h"
+#endif
+#include "h/types.h"
+#include "h/time.h"
+#if defined(AFS_AIX31_ENV)
+#include "h/limits.h"
#endif
#if !defined(AFS_AIX_ENV) && !defined(AFS_SUN5_ENV) && !defined(AFS_SGI_ENV) && !defined(AFS_LINUX20_ENV)
-#include "../h/kernel.h" /* Doesn't needed, so it should go */
+#include "h/kernel.h" /* Doesn't needed, so it should go */
#endif
#endif /* !defined(UKERNEL) */
-#include "../afs/afs_osi.h"
-#include "../afsint/afsint.h"
-#include "../afs/lock.h"
+#include "afs/afs_osi.h"
+#include "afsint.h"
+#include "afs/lock.h"
#if !defined(UKERNEL) && !defined(AFS_LINUX20_ENV)
-#include "../h/buf.h"
+#include "h/buf.h"
#endif /* !defined(UKERNEL) */
-#include "../afs/stds.h"
-#include "../afs/volerrors.h"
-#include "../afs/exporter.h"
-#include "../afs/prs_fs.h"
-#include "../afs/afs_chunkops.h"
-#include "../afs/dir.h"
+#include "afs/stds.h"
+#include "afs/volerrors.h"
+#include "afs/exporter.h"
+#include "afs/prs_fs.h"
+#include "afs/afs_chunkops.h"
+#include "afs/dir.h"
-#include "../afs/afs_stats.h"
-#include "../afs/longc_procs.h"
-#include "../afs/afs.h"
+#include "afs/afs_stats.h"
+#include "afs/afs.h"
#ifndef BUF_TIME_MAX
#define BUF_TIME_MAX 0x7fffffff
#endif
-/* number of pages per Unix buffer, when we're using Unix buffer pool */
-#define NPB 4
+#define NPB 8 /* must be a pwer of 2 */
+static int afs_max_buffers; /* should be an integral multiple of NPB */
+
/* page size */
#define AFS_BUFFER_PAGESIZE 2048
/* log page size */
/* page hash table size - this is pretty intertwined with pHash */
#define PHSIZE (PHPAGEMASK + PHFIDMASK + 1)
/* the pHash macro */
-#define pHash(fid,page) ((((afs_int32)((fid)[0])) & PHFIDMASK) \
+#define pHash(fid,page) ((((afs_int32)(fid)) & PHFIDMASK) \
| (page & PHPAGEMASK))
#ifdef dirty
-#undef dirty /* XXX */
+#undef dirty /* XXX */
#endif
static struct buffer *Buffers = 0;
#ifdef AFS_FBSD_ENV
#define timecounter afs_timecounter
#endif
-/* The locks for individual buffer entries are now sometimes obtained while holding the
- * afs_bufferLock. Thus we now have a locking hierarchy: afs_bufferLock -> Buffers[].lock.
+
+/* A note on locking in 'struct buffer'
+ *
+ * afs_bufferLock protects the hash chain, and the 'lockers' field where that
+ * has a zero value. It must be held whenever lockers is incremented from zero.
+ *
+ * The individual buffer lock protects the contents of the structure, including
+ * the lockers field.
+ *
+ * For safety: afs_bufferLock and the individual buffer lock must be held
+ * when obtaining a reference on a structure. Only the individual buffer lock
+ * need be held when releasing a reference.
+ *
+ * The locking hierarchy is afs_bufferLock-> buffer.lock
+ *
*/
+
static afs_lock_t afs_bufferLock;
static struct buffer *phTable[PHSIZE]; /* page hash table */
static int nbuffers;
static afs_int32 timecounter;
/* Prototypes for static routines */
-static struct buffer *afs_newslot (afs_inode_t *afid, afs_int32 apage,register struct buffer *lp);
+static struct buffer *afs_newslot(struct dcache *adc, afs_int32 apage,
+ struct buffer *lp);
static int dinit_flag = 0;
-void DInit (int abuffers)
+void
+DInit(int abuffers)
{
/* Initialize the venus buffer system. */
- register int i;
- register struct buffer *tb;
-#if AFS_USEBUFFERS
- struct buf *tub; /* unix buffer for allocation */
-#endif
+ int i;
+ struct buffer *tb;
AFS_STATCNT(DInit);
- if (dinit_flag) return;
+ if (dinit_flag)
+ return;
dinit_flag = 1;
-#if AFS_USEBUFFERS
/* round up to next multiple of NPB, since we allocate multiple pages per chunk */
- abuffers = ((abuffers-1) | (NPB-1)) + 1;
-#endif
+ abuffers = ((abuffers - 1) | (NPB - 1)) + 1;
+ afs_max_buffers = abuffers << 2; /* possibly grow up to 4 times as big */
LOCK_INIT(&afs_bufferLock, "afs_bufferLock");
- Buffers = (struct buffer *) afs_osi_Alloc(abuffers * sizeof(struct buffer));
-#if !AFS_USEBUFFERS
- BufferData = (char *) afs_osi_Alloc(abuffers * AFS_BUFFER_PAGESIZE);
-#endif
+ Buffers = afs_osi_Alloc(afs_max_buffers * sizeof(struct buffer));
+ osi_Assert(Buffers != NULL);
timecounter = 1;
afs_stats_cmperf.bufAlloced = nbuffers = abuffers;
- for(i=0;i<PHSIZE;i++) phTable[i] = 0;
- for (i=0;i<abuffers;i++) {
-#if AFS_USEBUFFERS
- if ((i & (NPB-1)) == 0) {
+ for (i = 0; i < PHSIZE; i++)
+ phTable[i] = 0;
+ for (i = 0; i < abuffers; i++) {
+ if ((i & (NPB - 1)) == 0) {
/* time to allocate a fresh buffer */
- tub = geteblk(AFS_BUFFER_PAGESIZE*NPB);
- BufferData = (char *) tub->b_un.b_addr;
+ BufferData = afs_osi_Alloc(AFS_BUFFER_PAGESIZE * NPB);
+ osi_Assert(BufferData != NULL);
}
-#endif
- /* Fill in each buffer with an empty indication. */
+ /* Fill in each buffer with an empty indication. */
tb = &Buffers[i];
- dirp_Zap(tb->fid);
- tb->accesstime = 0;
+ tb->fid = NULLIDX;
+ afs_reset_inode(&tb->inode);
+ tb->accesstime = 0;
tb->lockers = 0;
-#if AFS_USEBUFFERS
- if ((i & (NPB-1)) == 0)
- tb->bufp = tub;
- else
- tb->bufp = 0;
- tb->data = &BufferData[AFS_BUFFER_PAGESIZE * (i&(NPB-1))];
-#else
- tb->data = &BufferData[AFS_BUFFER_PAGESIZE*i];
-#endif
+ tb->data = &BufferData[AFS_BUFFER_PAGESIZE * (i & (NPB - 1))];
tb->hashIndex = 0;
- tb->dirty = 0;
- RWLOCK_INIT(&tb->lock, "buffer lock");
+ tb->dirty = 0;
+ AFS_RWLOCK_INIT(&tb->lock, "buffer lock");
}
return;
}
-char *DRead(register afs_inode_t *fid, register int page)
+/*!
+ * Read and return the requested directory page.
+ *
+ * \param[in] adc pointer to directory dcache
+ * \param[in] page number of the desired directory page
+ * \param[out] entry buffer to return requested page
+ * \param[out] physerr (optional) pointer to return errno, if any
+ *
+ * \retval 0 success
+ * \retval non-zero invalid directory or internal IO error;
+ * if physerr is supplied by caller, it will be set:
+ * 0 logical error
+ * errno physical error
+ */
+int
+DReadWithErrno(struct dcache *adc, int page, struct DirBuffer *entry, int *physerr)
{
/* Read a page from the disk. */
- register struct buffer *tb, *tb2;
- void *tfile;
- register afs_int32 code, *sizep;
+ struct buffer *tb, *tb2;
+ struct osi_file *tfile;
+ int code;
AFS_STATCNT(DRead);
- MObtainWriteLock(&afs_bufferLock,256);
-/* some new code added 1/1/92 */
-#define bufmatch(tb) (tb->page == page && dirp_Eq(tb->fid, fid))
+ if (physerr != NULL)
+ *physerr = 0;
+
+ memset(entry, 0, sizeof(struct DirBuffer));
+
+ if (adc->f.chunk == 0 && adc->f.chunkBytes == 0) {
+ /* The directory blob is empty, apparently. This is not a valid dir
+ * blob, so throw an error. */
+ return EIO;
+ }
+ if (page * AFS_BUFFER_PAGESIZE >= adc->f.chunkBytes) {
+ return ENOENT; /* past the end */
+ }
+
+ ObtainWriteLock(&afs_bufferLock, 256);
+
+#define bufmatch(tb) (tb->page == page && tb->fid == adc->index)
#define buf_Front(head,parent,p) {(parent)->hashNext = (p)->hashNext; (p)->hashNext= *(head);*(head)=(p);}
/* this apparently-complicated-looking code is simply an example of
- * a little bit of loop unrolling, and is a standard linked-list
+ * a little bit of loop unrolling, and is a standard linked-list
* traversal trick. It saves a few assignments at the the expense
* of larger code size. This could be simplified by better use of
- * macros.
+ * macros.
*/
- if ((tb = phTable[pHash(fid,page)])) {
+ if ((tb = phTable[pHash(adc->index, page)])) {
if (bufmatch(tb)) {
- MObtainWriteLock(&tb->lock,257);
- ReleaseWriteLock(&afs_bufferLock);
+ ObtainWriteLock(&tb->lock, 257);
tb->lockers++;
+ ReleaseWriteLock(&afs_bufferLock);
tb->accesstime = timecounter++;
AFS_STATS(afs_stats_cmperf.bufHits++);
- MReleaseWriteLock(&tb->lock);
- return tb->data;
- }
- else {
- register struct buffer **bufhead;
- bufhead = &( phTable[pHash(fid,page)] );
- while ((tb2 = tb->hashNext)) {
- if (bufmatch(tb2)) {
- buf_Front(bufhead,tb,tb2);
- MObtainWriteLock(&tb2->lock,258);
- ReleaseWriteLock(&afs_bufferLock);
- tb2->lockers++;
- tb2->accesstime = timecounter++;
- AFS_STATS(afs_stats_cmperf.bufHits++);
- MReleaseWriteLock(&tb2->lock);
- return tb2->data;
- }
- if ((tb = tb2->hashNext)) {
- if (bufmatch(tb)) {
- buf_Front(bufhead,tb2,tb);
- MObtainWriteLock(&tb->lock,259);
- ReleaseWriteLock(&afs_bufferLock);
- tb->lockers++;
- tb->accesstime = timecounter++;
- AFS_STATS(afs_stats_cmperf.bufHits++);
- MReleaseWriteLock(&tb->lock);
- return tb->data;
- }
+ ReleaseWriteLock(&tb->lock);
+ entry->buffer = tb;
+ entry->data = tb->data;
+ return 0;
+ } else {
+ struct buffer **bufhead;
+ bufhead = &(phTable[pHash(adc->index, page)]);
+ while ((tb2 = tb->hashNext)) {
+ if (bufmatch(tb2)) {
+ buf_Front(bufhead, tb, tb2);
+ ObtainWriteLock(&tb2->lock, 258);
+ tb2->lockers++;
+ ReleaseWriteLock(&afs_bufferLock);
+ tb2->accesstime = timecounter++;
+ AFS_STATS(afs_stats_cmperf.bufHits++);
+ ReleaseWriteLock(&tb2->lock);
+ entry->buffer = tb2;
+ entry->data = tb2->data;
+ return 0;
+ }
+ if ((tb = tb2->hashNext)) {
+ if (bufmatch(tb)) {
+ buf_Front(bufhead, tb2, tb);
+ ObtainWriteLock(&tb->lock, 259);
+ tb->lockers++;
+ ReleaseWriteLock(&afs_bufferLock);
+ tb->accesstime = timecounter++;
+ AFS_STATS(afs_stats_cmperf.bufHits++);
+ ReleaseWriteLock(&tb->lock);
+ entry->buffer = tb;
+ entry->data = tb->data;
+ return 0;
+ }
+ } else
+ break;
}
- else break;
- }
}
- }
- else tb2 = NULL;
+ } else
+ tb2 = NULL;
AFS_STATS(afs_stats_cmperf.bufMisses++);
/* can't find it */
/* The last thing we looked at was either tb or tb2 (or nothing). That
- * is at least the oldest buffer on one particular hash chain, so it's
+ * is at least the oldest buffer on one particular hash chain, so it's
* a pretty good place to start looking for the truly oldest buffer.
*/
- tb = afs_newslot(fid, page, (tb ? tb : tb2));
+ tb = afs_newslot(adc, page, (tb ? tb : tb2));
if (!tb) {
- MReleaseWriteLock(&afs_bufferLock);
- return 0;
+ ReleaseWriteLock(&afs_bufferLock);
+ return EIO;
}
- MObtainWriteLock(&tb->lock,260);
- MReleaseWriteLock(&afs_bufferLock);
+ ObtainWriteLock(&tb->lock, 260);
tb->lockers++;
- tfile = afs_CFileOpen(fid[0]);
- sizep = (afs_int32 *)tfile;
- if (page * AFS_BUFFER_PAGESIZE >= *sizep) {
- dirp_Zap(tb->fid);
- tb->lockers--;
- MReleaseWriteLock(&tb->lock);
- afs_CFileClose(tfile);
- return 0;
+ ReleaseWriteLock(&afs_bufferLock);
+ tfile = afs_CFileOpen(&adc->f.inode);
+ if (!tfile) {
+ code = EIO;
+ goto error;
}
- code = afs_CFileRead(tfile, tb->page * AFS_BUFFER_PAGESIZE,
- tb->data, AFS_BUFFER_PAGESIZE);
+ code =
+ afs_CFileRead(tfile, tb->page * AFS_BUFFER_PAGESIZE, tb->data,
+ AFS_BUFFER_PAGESIZE);
afs_CFileClose(tfile);
if (code < AFS_BUFFER_PAGESIZE) {
- dirp_Zap(tb->fid);
- tb->lockers--;
- MReleaseWriteLock(&tb->lock);
- return 0;
+ if (code < 0 && physerr != NULL)
+ *physerr = -code;
+ code = EIO;
+ goto error;
}
/* Note that findslot sets the page field in the buffer equal to
* what it is searching for. */
- MReleaseWriteLock(&tb->lock);
- return tb->data;
+ ReleaseWriteLock(&tb->lock);
+ entry->buffer = tb;
+ entry->data = tb->data;
+ return 0;
+
+ error:
+ tb->fid = NULLIDX;
+ afs_reset_inode(&tb->inode);
+ tb->lockers--;
+ ReleaseWriteLock(&tb->lock);
+ return code;
+}
+
+/*!
+ * Read and return the requested directory page.
+ *
+ * \param[in] adc pointer to directory dcache
+ * \param[in] page number of the desired directory page
+ * \param[out] entry buffer to return requested page
+ *
+ * \retval 0 success
+ * \retval non-zero invalid directory or internal IO error;
+ */
+int
+DRead(struct dcache *adc, int page, struct DirBuffer *entry)
+{
+ return DReadWithErrno(adc, page, entry, NULL);
}
-static void FixupBucket(register struct buffer *ap)
+static void
+FixupBucket(struct buffer *ap)
{
- register struct buffer **lp, *tp;
- register int i;
+ struct buffer **lp, *tp;
+ int i;
/* first try to get it out of its current hash bucket, in which it
* might not be */
AFS_STATCNT(FixupBucket);
i = ap->hashIndex;
lp = &phTable[i];
- for(tp = *lp; tp; tp=tp->hashNext) {
+ for (tp = *lp; tp; tp = tp->hashNext) {
if (tp == ap) {
*lp = tp->hashNext;
break;
lp = &tp->hashNext;
}
/* now figure the new hash bucket */
- i = pHash(ap->fid,ap->page);
+ i = pHash(ap->fid, ap->page);
ap->hashIndex = i; /* remember where we are for deletion */
ap->hashNext = phTable[i]; /* add us to the list */
- phTable[i] = ap; /* at the front, since it's LRU */
+ phTable[i] = ap; /* at the front, since it's LRU */
}
/* lp is pointer to a fairly-old buffer */
-static struct buffer *afs_newslot (afs_inode_t *afid, afs_int32 apage,register struct buffer *lp)
+static struct buffer *
+afs_newslot(struct dcache *adc, afs_int32 apage, struct buffer *lp)
{
/* Find a usable buffer slot */
- register afs_int32 i;
- afs_int32 lt;
- register struct buffer *tp;
- void *tfile;
+ afs_int32 i;
+ afs_int32 lt = 0;
+ struct buffer *tp;
+ struct osi_file *tfile;
AFS_STATCNT(afs_newslot);
/* we take a pointer here to a buffer which was at the end of an
* case" for loop below.
*/
if (lp && (lp->lockers == 0)) {
- lt = lp->accesstime;
- }
- else {
- lp = 0;
- lt = BUF_TIME_MAX;
+ lt = lp->accesstime;
+ } else {
+ lp = NULL;
}
/* timecounter might have wrapped, if machine is very very busy
* themselves back out after just a few more DReads.
*/
if (timecounter < 0) {
- timecounter = 1;
- tp = Buffers;
- for (i=0;i<nbuffers;i++,tp++) {
- tp->accesstime = 0;
- if (!lp && !tp->lockers) /* one is as good as the rest, I guess */
- lp = tp;
- }
- }
- else {
- /* this is the typical case */
- tp = Buffers;
- for (i=0;i<nbuffers;i++,tp++) {
- if (tp->lockers == 0) {
- if (tp->accesstime < lt) {
- lp = tp;
- lt = tp->accesstime;
- }
+ timecounter = 1;
+ tp = Buffers;
+ for (i = 0; i < nbuffers; i++, tp++) {
+ tp->accesstime = 0;
+ if (!lp && !tp->lockers) /* one is as good as the rest, I guess */
+ lp = tp;
+ }
+ } else {
+ /* this is the typical case */
+ tp = Buffers;
+ for (i = 0; i < nbuffers; i++, tp++) {
+ if (tp->lockers == 0) {
+ if (!lp || tp->accesstime < lt) {
+ lp = tp;
+ lt = tp->accesstime;
+ }
+ }
}
- }
}
if (lp == 0) {
- /* There are no unlocked buffers -- this used to panic, but that
- * seems extreme. To the best of my knowledge, all the callers
- * of DRead are prepared to handle a zero return. Some of them
- * just panic directly, but not all of them. */
- afs_warn ("all buffers locked");
- return 0;
+ /* No unlocked buffers. If still possible, allocate a new increment */
+ if (nbuffers + NPB > afs_max_buffers) {
+ /* There are no unlocked buffers -- this used to panic, but that
+ * seems extreme. To the best of my knowledge, all the callers
+ * of DRead are prepared to handle a zero return. Some of them
+ * just panic directly, but not all of them. */
+ afs_warn("afs: all buffers locked\n");
+ return 0;
+ }
+
+ BufferData = afs_osi_Alloc(AFS_BUFFER_PAGESIZE * NPB);
+ osi_Assert(BufferData != NULL);
+ for (i = 0; i< NPB; i++) {
+ /* Fill in each buffer with an empty indication. */
+ tp = &Buffers[i + nbuffers];
+ tp->fid = NULLIDX;
+ afs_reset_inode(&tp->inode);
+ tp->accesstime = 0;
+ tp->lockers = 0;
+ tp->data = &BufferData[AFS_BUFFER_PAGESIZE * i];
+ tp->hashIndex = 0;
+ tp->dirty = 0;
+ AFS_RWLOCK_INIT(&tp->lock, "buffer lock");
+ }
+ lp = &Buffers[nbuffers];
+ nbuffers += NPB;
}
if (lp->dirty) {
- tfile = afs_CFileOpen(lp->fid[0]);
- afs_CFileWrite(tfile, lp->page * AFS_BUFFER_PAGESIZE,
- lp->data, AFS_BUFFER_PAGESIZE);
- lp->dirty = 0;
+ /* see DFlush for rationale for not getting and locking the dcache */
+ tfile = afs_CFileOpen(&lp->inode);
+ if (!tfile)
+ return NULL; /* Callers will flag as EIO */
+
+ afs_CFileWrite(tfile, lp->page * AFS_BUFFER_PAGESIZE, lp->data,
+ AFS_BUFFER_PAGESIZE);
+ lp->dirty = 0;
afs_CFileClose(tfile);
AFS_STATS(afs_stats_cmperf.bufFlushDirty++);
- }
+ }
+ /* Zero out the data so we don't leak something we shouldn't. */
+ memset(lp->data, 0, AFS_BUFFER_PAGESIZE);
/* Now fill in the header. */
- dirp_Cpy(lp->fid, afid); /* set this */
+ lp->fid = adc->index;
+ afs_copy_inode(&lp->inode, &adc->f.inode);
lp->page = apage;
lp->accesstime = timecounter++;
FixupBucket(lp); /* move to the right hash bucket */
return lp;
}
-void DRelease (register struct buffer *bp, int flag)
+void
+DRelease(struct DirBuffer *entry, int flag)
{
- /* Release a buffer, specifying whether or not the buffer has been
- * modified by the locker. */
- register int index;
-#if AFS_USEBUFFERS
- register struct buffer *tp;
-#endif
+ struct buffer *tp;
AFS_STATCNT(DRelease);
- if (!bp) return;
-#if AFS_USEBUFFERS
- /* look for buffer by scanning Unix buffers for appropriate address */
- tp = Buffers;
- for(index = 0; index < nbuffers; index += NPB, tp += NPB) {
- if ((afs_int32)bp >= (afs_int32)tp->data
- && (afs_int32)bp < (afs_int32)tp->data + AFS_BUFFER_PAGESIZE*NPB) {
- /* we found the right range */
- index += ((afs_int32)bp - (afs_int32)tp->data) >> LOGPS;
- break;
- }
- }
-#else
- index = (((char *)bp)-((char *)BufferData))>>LOGPS;
-#endif
- bp = &(Buffers[index]);
- MObtainWriteLock(&bp->lock,261);
- bp->lockers--;
- if (flag) bp->dirty=1;
- MReleaseWriteLock(&bp->lock);
+
+ tp = entry->buffer;
+ if (tp == NULL)
+ return;
+
+ ObtainWriteLock(&tp->lock, 261);
+ tp->lockers--;
+ if (flag)
+ tp->dirty = 1;
+ ReleaseWriteLock(&tp->lock);
}
-int DVOffset (register void *ap)
+int
+DVOffset(struct DirBuffer *entry)
{
- /* Return the byte within a file represented by a buffer pointer. */
- register struct buffer *bp;
- register int index;
-#if AFS_USEBUFFERS
- register struct buffer *tp;
-#endif
+ struct buffer *bp;
+
AFS_STATCNT(DVOffset);
- bp=ap;
-#if AFS_USEBUFFERS
- /* look for buffer by scanning Unix buffers for appropriate address */
- tp = Buffers;
- for(index = 0; index < nbuffers; index += NPB, tp += NPB) {
- if ((afs_int32)bp >= (afs_int32)tp->data && (afs_int32)bp < (afs_int32)tp->data + AFS_BUFFER_PAGESIZE*NPB) {
- /* we found the right range */
- index += ((afs_int32)bp - (afs_int32)tp->data) >> LOGPS;
- break;
- }
- }
-#else
- index = (((char *)bp)-((char *)BufferData))>>LOGPS;
-#endif
- if (index<0 || index >= nbuffers) return -1;
- bp = &(Buffers[index]);
- return AFS_BUFFER_PAGESIZE*bp->page+(int)(((char *)ap)-bp->data);
+
+ bp = entry->buffer;
+ return AFS_BUFFER_PAGESIZE * bp->page
+ + (char *)entry->data - (char *)bp->data;
}
-/* 1/1/91 - I've modified the hash function to take the page as well
+/*!
+ * Zap one dcache entry: destroy one FID's buffers.
+ *
+ * 1/1/91 - I've modified the hash function to take the page as well
* as the *fid, so that lookup will be a bit faster. That presents some
* difficulties for Zap, which now has to have some knowledge of the nature
- * of the hash function. Oh well. This should use the list traversal
+ * of the hash function. Oh well. This should use the list traversal
* method of DRead...
+ *
+ * \param adc The dcache entry to be zapped.
*/
-void DZap (afs_inode_t *fid)
+void
+DZap(struct dcache *adc)
{
- register int i;
+ int i;
/* Destroy all buffers pertaining to a particular fid. */
- register struct buffer *tb;
-
+ struct buffer *tb;
+
AFS_STATCNT(DZap);
- MObtainReadLock(&afs_bufferLock);
-
- for (i=0;i<=PHPAGEMASK;i++)
- for(tb=phTable[pHash(fid,i)]; tb; tb=tb->hashNext)
- if (dirp_Eq(tb->fid,fid)) {
- MObtainWriteLock(&tb->lock,262);
- dirp_Zap(tb->fid);
- tb->dirty = 0;
- MReleaseWriteLock(&tb->lock);
- }
- MReleaseReadLock(&afs_bufferLock);
+ ObtainReadLock(&afs_bufferLock);
+
+ for (i = 0; i <= PHPAGEMASK; i++)
+ for (tb = phTable[pHash(adc->index, i)]; tb; tb = tb->hashNext)
+ if (tb->fid == adc->index) {
+ ObtainWriteLock(&tb->lock, 262);
+ tb->fid = NULLIDX;
+ afs_reset_inode(&tb->inode);
+ tb->dirty = 0;
+ ReleaseWriteLock(&tb->lock);
+ }
+ ReleaseReadLock(&afs_bufferLock);
+}
+
+static void
+DFlushBuffer(struct buffer *ab)
+{
+ struct osi_file *tfile;
+
+ tfile = afs_CFileOpen(&ab->inode);
+ osi_Assert(tfile);
+ afs_CFileWrite(tfile, ab->page * AFS_BUFFER_PAGESIZE,
+ ab->data, AFS_BUFFER_PAGESIZE);
+ ab->dirty = 0; /* Clear the dirty flag */
+ afs_CFileClose(tfile);
+}
+
+void
+DFlushDCache(struct dcache *adc)
+{
+ int i;
+ struct buffer *tb;
+
+ ObtainReadLock(&afs_bufferLock);
+
+ for (i = 0; i <= PHPAGEMASK; i++)
+ for (tb = phTable[pHash(adc->index, i)]; tb; tb = tb->hashNext)
+ if (tb->fid == adc->index) {
+ ObtainWriteLock(&tb->lock, 701);
+ tb->lockers++;
+ ReleaseReadLock(&afs_bufferLock);
+ if (tb->dirty) {
+ DFlushBuffer(tb);
+ }
+ tb->lockers--;
+ ReleaseWriteLock(&tb->lock);
+ ObtainReadLock(&afs_bufferLock);
+ }
+
+ ReleaseReadLock(&afs_bufferLock);
}
-void DFlush (void)
+int
+DFlush(void)
{
/* Flush all the modified buffers. */
- register int i;
- register struct buffer *tb;
- void *tfile;
+ int i;
+ struct buffer *tb;
AFS_STATCNT(DFlush);
tb = Buffers;
- MObtainReadLock(&afs_bufferLock);
- for(i=0;i<nbuffers;i++,tb++) {
- if (tb->dirty) {
- MObtainWriteLock(&tb->lock,263);
+ ObtainReadLock(&afs_bufferLock);
+ for (i = 0; i < nbuffers; i++, tb++) {
+ if (tb->dirty) {
+ ObtainWriteLock(&tb->lock, 263);
tb->lockers++;
- MReleaseReadLock(&afs_bufferLock);
+ ReleaseReadLock(&afs_bufferLock);
if (tb->dirty) {
- tfile = afs_CFileOpen(tb->fid[0]);
- afs_CFileWrite(tfile, tb->page * AFS_BUFFER_PAGESIZE,
- tb->data, AFS_BUFFER_PAGESIZE);
- tb->dirty = 0; /* Clear the dirty flag */
- afs_CFileClose(tfile);
+ /* it seems safe to do this I/O without having the dcache
+ * locked, since the only things that will update the data in
+ * a directory are the buffer package, which holds the relevant
+ * tb->lock while doing the write, or afs_GetDCache, which
+ * DZap's the directory while holding the dcache lock.
+ * It is not possible to lock the dcache or even call
+ * afs_GetDSlot to map the index to the dcache since the dir
+ * package's caller has some dcache object locked already (so
+ * we cannot lock afs_xdcache). In addition, we cannot obtain
+ * a dcache lock while holding the tb->lock of the same file
+ * since that can deadlock with DRead/DNew */
+ DFlushBuffer(tb);
}
tb->lockers--;
- MReleaseWriteLock(&tb->lock);
- MObtainReadLock(&afs_bufferLock);
+ ReleaseWriteLock(&tb->lock);
+ ObtainReadLock(&afs_bufferLock);
}
}
- MReleaseReadLock(&afs_bufferLock);
+ ReleaseReadLock(&afs_bufferLock);
+
+ return 0;
}
-char *DNew (register afs_inode_t *fid, register int page)
+int
+DNew(struct dcache *adc, int page, struct DirBuffer *entry)
{
- /* Same as read, only do *not* even try to read the page, since it probably doesn't exist. */
- register struct buffer *tb;
+ /* Same as read, only do *not* even try to read the page, since it
+ * probably doesn't exist. */
+ struct buffer *tb;
AFS_STATCNT(DNew);
- MObtainWriteLock(&afs_bufferLock,264);
- if ((tb = afs_newslot(fid,page,NULL)) == 0) {
- MReleaseWriteLock(&afs_bufferLock);
- return 0;
+
+ ObtainWriteLock(&afs_bufferLock, 264);
+ if ((tb = afs_newslot(adc, page, NULL)) == 0) {
+ ReleaseWriteLock(&afs_bufferLock);
+ return EIO;
}
- MObtainWriteLock(&tb->lock,265);
- MReleaseWriteLock(&afs_bufferLock);
+ /* extend the chunk, if needed */
+ /* Do it now, not in DFlush or afs_newslot when the data is written out,
+ * since now our caller has adc->lock writelocked, and we can't acquire
+ * that lock (or even map from a fid to a dcache) in afs_newslot or
+ * DFlush due to lock hierarchy issues */
+ if ((page + 1) * AFS_BUFFER_PAGESIZE > adc->f.chunkBytes) {
+ afs_AdjustSize(adc, (page + 1) * AFS_BUFFER_PAGESIZE);
+ osi_Assert(afs_WriteDCache(adc, 1) == 0);
+ }
+ ObtainWriteLock(&tb->lock, 265);
tb->lockers++;
- MReleaseWriteLock(&tb->lock);
- return tb->data;
+ ReleaseWriteLock(&afs_bufferLock);
+ ReleaseWriteLock(&tb->lock);
+ entry->buffer = tb;
+ entry->data = tb->data;
+
+ return 0;
}
-void shutdown_bufferpackage(void)
+void
+shutdown_bufferpackage(void)
{
-#if AFS_USEBUFFERS
- register struct buffer *tp;
-#endif
+ struct buffer *tp;
int i;
- extern int afs_cold_shutdown;
AFS_STATCNT(shutdown_bufferpackage);
/* Free all allocated Buffers and associated buffer pages */
DFlush();
+
+ dinit_flag = 0;
+ tp = Buffers;
+ for (i = 0; i < nbuffers; i += NPB, tp += NPB) {
+ afs_osi_Free(tp->data, NPB * AFS_BUFFER_PAGESIZE);
+ }
+ afs_osi_Free(Buffers, nbuffers * sizeof(struct buffer));
+ Buffers = NULL;
+ nbuffers = 0;
+ timecounter = 1;
+ for (i = 0; i < PHSIZE; i++)
+ phTable[i] = NULL;
+
if (afs_cold_shutdown) {
- dinit_flag = 0;
-#if !AFS_USEBUFFERS
- afs_osi_Free(BufferData, nbuffers * AFS_BUFFER_PAGESIZE);
-#else
- tp = Buffers;
- for (i=0; i < nbuffers; i+= NPB, tp += NPB) {
- /* The following check shouldn't be necessary and it will be removed soon */
- if (!tp->bufp)
- afs_warn("shutdown_bufferpackage: bufp == 0!! Shouldn't happen\n");
- else {
- brelse(tp->bufp);
- tp->bufp = 0;
- }
- }
-#endif
- afs_osi_Free(Buffers, nbuffers * sizeof(struct buffer));
- nbuffers = 0;
- timecounter = 1;
- for(i=0;i<PHSIZE;i++) phTable[i] = 0;
- memset((char *)&afs_bufferLock, 0, sizeof(afs_lock_t));
- }
-}
+ memset(&afs_bufferLock, 0, sizeof(afs_lock_t));
+ }
+}
git://git.openafs.org / openafs.git / blobdiff