void bq_heap_t::place(item_t *item, size_t i, bool flag) throw() {
size_t j;
assert(i > 0);
for(; (j = i / 2); i = j) {
item_t *_item = get(j);
if(hcmp(_item, item)) break;
put(i, _item);
flag = false;
}
if(flag) {
for(; (j = i * 2) <= count; i = j) {
item_t *_item = get(j);
if(j < count) {
item_t *__item = get(j + 1);
if(!hcmp(_item, __item)) {
++j;
_item = __item;
}
}
if(hcmp(item, _item)) break;
put(i, _item);
}
}
put(i, item);
}
void bq_heap_t::validate() {
#ifdef DEBUG_BQ
for(size_t i = count; i > 1; --i) {
if(!hcmp(get(i / 2), get(i)))
fatal("bq_cont_heap validation failed");
}
#endif
}
void
osi_FlushText_really(struct vcache *vp)
{
afs_hyper_t fdv; /* version before which we'll flush */
AFS_STATCNT(osi_FlushText);
/* see if we've already flushed this data version */
if (hcmp(vp->f.m.DataVersion, vp->flushDV) <= 0)
return;
ObtainWriteLock(&afs_ftf, 317);
hset(fdv, vp->f.m.DataVersion);
/* why this disgusting code below?
* xuntext, called by xrele, doesn't notice when it is called
* with a freed text object. Sun continually calls xrele or xuntext
* without any locking, as long as VTEXT is set on the
* corresponding vnode.
* But, if the text object is locked when you check the VTEXT
* flag, several processes can wait in xuntext, waiting for the
* text lock; when the second one finally enters xuntext's
* critical region, the text object is already free, but the check
* was already done by xuntext's caller.
* Even worse, it turns out that xalloc locks the text object
* before reading or stating a file via the vnode layer. Thus, we
* could end up in getdcache, being asked to bring in a new
* version of a file, but the corresponding text object could be
* locked. We can't flush the text object without causing
* deadlock, so now we just don't try to lock the text object
* unless it is guaranteed to work. And we try to flush the text
* when we need to a bit more often at the vnode layer. Sun
* really blew the vm-cache flushing interface.
*/
#if defined (AFS_HPUX_ENV)
if (vp->v.v_flag & VTEXT) {
xrele(vp);
if (vp->v.v_flag & VTEXT) { /* still has a text object? */
ReleaseWriteLock(&afs_ftf);
return;
}
}
#endif
/* next do the stuff that need not check for deadlock problems */
mpurge(vp);
/* finally, record that we've done it */
hset(vp->flushDV, fdv);
ReleaseWriteLock(&afs_ftf);
}
int
afs_StoreAllSegments(struct vcache *avc, struct vrequest *areq,
int sync)
{
struct dcache *tdc;
afs_int32 code = 0;
afs_int32 index;
afs_int32 origCBs, foreign = 0;
int hash;
afs_hyper_t newDV, oldDV; /* DV when we start, and finish, respectively */
struct dcache **dcList;
unsigned int i, j, minj, moredata, high, off;
afs_size_t maxStoredLength; /* highest offset we've written to server. */
int safety, marineronce = 0;
AFS_STATCNT(afs_StoreAllSegments);
hset(oldDV, avc->f.m.DataVersion);
hset(newDV, avc->f.m.DataVersion);
hash = DVHash(&avc->f.fid);
foreign = (avc->f.states & CForeign);
dcList = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
afs_Trace2(afs_iclSetp, CM_TRACE_STOREALL, ICL_TYPE_POINTER, avc,
ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
#if !defined(AFS_AIX32_ENV) && !defined(AFS_SGI65_ENV)
/* In the aix vm implementation we need to do the vm_writep even
* on the memcache case since that's we adjust the file's size
* and finish flushing partial vm pages.
*/
if ((cacheDiskType != AFS_FCACHE_TYPE_MEM) ||
(sync & AFS_VMSYNC_INVAL) || (sync & AFS_VMSYNC) ||
(sync & AFS_LASTSTORE))
#endif /* !AFS_AIX32_ENV && !AFS_SGI65_ENV */
{
/* If we're not diskless, reading a file may stress the VM
* system enough to cause a pageout, and this vnode would be
* locked when the pageout occurs. We can prevent this problem
* by making sure all dirty pages are already flushed. We don't
* do this when diskless because reading a diskless (i.e.
* memory-resident) chunk doesn't require using new VM, and we
* also don't want to dump more dirty data into a diskless cache,
* since they're smaller, and we might exceed its available
* space.
*/
#if defined(AFS_SUN5_ENV)
if (sync & AFS_VMSYNC_INVAL) /* invalidate VM pages */
osi_VM_TryToSmush(avc, CRED(), 1);
else
#endif
osi_VM_StoreAllSegments(avc);
}
if (AFS_IS_DISCONNECTED && !AFS_IN_SYNC) {
/* This will probably make someone sad ... */
/*printf("Net down in afs_StoreSegments\n");*/
return ENETDOWN;
}
ConvertWToSLock(&avc->lock);
/*
* Subsequent code expects a sorted list, and it expects all the
* chunks in the list to be contiguous, so we need a sort and a
* while loop in here, too - but this will work for a first pass...
* 92.10.05 - OK, there's a sort in here now. It's kind of a modified
* bin sort, I guess. Chunk numbers start with 0
*
* - Have to get a write lock on xdcache because GetDSlot might need it (if
* the chunk doesn't have a dcache struct).
* This seems like overkill in most cases.
* - I'm not sure that it's safe to do "index = .hvNextp", then unlock
* xdcache, then relock xdcache and try to use index. It is done
* a lot elsewhere in the CM, but I'm not buying that argument.
* - should be able to check IFDataMod without doing the GetDSlot (just
* hold afs_xdcache). That way, it's easy to do this without the
* writelock on afs_xdcache, and we save unneccessary disk
* operations. I don't think that works, 'cuz the next pointers
* are still on disk.
*/
origCBs = afs_allCBs;
maxStoredLength = 0;
minj = 0;
do {
memset(dcList, 0, NCHUNKSATONCE * sizeof(struct dcache *));
high = 0;
moredata = FALSE;
/* lock and start over from beginning of hash chain
* in order to avoid a race condition. */
ObtainWriteLock(&afs_xdcache, 284);
index = afs_dvhashTbl[hash];
for (j = 0; index != NULLIDX;) {
if ((afs_indexFlags[index] & IFDataMod)
&& (afs_indexUnique[index] == avc->f.fid.Fid.Unique)) {
tdc = afs_GetValidDSlot(index); /* refcount+1. */
if (!tdc) {
ReleaseWriteLock(&afs_xdcache);
code = EIO;
goto done;
}
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid) && tdc->f.chunk >= minj) {
off = tdc->f.chunk - minj;
if (off < NCHUNKSATONCE) {
if (dcList[off])
osi_Panic("dclist slot already in use!");
if (afs_mariner && !marineronce) {
/* first chunk only */
afs_MarinerLog("store$Storing", avc);
marineronce++;
}
dcList[off] = tdc;
if (off > high)
high = off;
j++;
/* DCLOCKXXX: chunkBytes is protected by tdc->lock which we
* can't grab here, due to lock ordering with afs_xdcache.
* So, disable this shortcut for now. -- kolya 2001-10-13
*/
/* shortcut: big win for little files */
/* tlen -= tdc->f.chunkBytes;
* if (tlen <= 0)
* break;
*/
} else {
moredata = TRUE;
afs_PutDCache(tdc);
if (j == NCHUNKSATONCE)
break;
}
} else {
afs_PutDCache(tdc);
}
}
index = afs_dvnextTbl[index];
}
ReleaseWriteLock(&afs_xdcache);
/* this guy writes chunks, puts back dcache structs, and bumps newDV */
/* "moredata" just says "there are more dirty chunks yet to come".
*/
if (j) {
code =
afs_CacheStoreVCache(dcList, avc, areq, sync,
minj, high, moredata,
&newDV, &maxStoredLength);
/* Release any zero-length dcache entries in our interval
* that we locked but didn't store back above.
*/
for (j = 0; j <= high; j++) {
tdc = dcList[j];
if (tdc) {
osi_Assert(tdc->f.chunkBytes == 0);
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
}
}
/* if (j) */
minj += NCHUNKSATONCE;
} while (!code && moredata);
done:
UpgradeSToWLock(&avc->lock, 29);
/* send a trivial truncation store if did nothing else */
if (code == 0) {
/*
* Call StoreMini if we haven't written enough data to extend the
* file at the fileserver to the client's notion of the file length.
*/
if ((avc->f.truncPos != AFS_NOTRUNC)
|| ((avc->f.states & CExtendedFile)
&& (maxStoredLength < avc->f.m.Length))) {
code = afs_StoreMini(avc, areq);
if (code == 0)
hadd32(newDV, 1); /* just bumped here, too */
}
avc->f.states &= ~CExtendedFile;
}
/*
* Finally, turn off DWriting, turn on DFEntryMod,
* update f.versionNo.
* A lot of this could be integrated into the loop above
*/
if (!code) {
afs_hyper_t h_unset;
hones(h_unset);
minj = 0;
do {
moredata = FALSE;
memset(dcList, 0,
NCHUNKSATONCE * sizeof(struct dcache *));
/* overkill, but it gets the lock in case GetDSlot needs it */
ObtainWriteLock(&afs_xdcache, 285);
for (j = 0, safety = 0, index = afs_dvhashTbl[hash];
index != NULLIDX && safety < afs_cacheFiles + 2;
index = afs_dvnextTbl[index]) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
tdc = afs_GetValidDSlot(index);
if (!tdc) {
/* This is okay; since manipulating the dcaches at this
* point is best-effort. We only get a dcache here to
* increment the dv and turn off DWriting. If we were
* supposed to do that for a dcache, but could not
* due to an I/O error, it just means the dv won't
* be updated so we don't be able to use that cached
* chunk in the future. That's inefficient, but not
* an error. */
continue;
}
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid)
&& tdc->f.chunk >= minj) {
off = tdc->f.chunk - minj;
if (off < NCHUNKSATONCE) {
/* this is the file, and the correct chunk range */
if (j >= NCHUNKSATONCE)
osi_Panic
("Too many dcache entries in range\n");
dcList[j++] = tdc;
} else {
moredata = TRUE;
afs_PutDCache(tdc);
if (j == NCHUNKSATONCE)
break;
}
} else {
afs_PutDCache(tdc);
}
}
}
ReleaseWriteLock(&afs_xdcache);
for (i = 0; i < j; i++) {
/* Iterate over the dcache entries we collected above */
tdc = dcList[i];
ObtainSharedLock(&tdc->lock, 677);
/* was code here to clear IFDataMod, but it should only be done
* in storedcache and storealldcache.
*/
/* Only increase DV if we had up-to-date data to start with.
* Otherwise, we could be falsely upgrading an old chunk
* (that we never read) into one labelled with the current
* DV #. Also note that we check that no intervening stores
* occurred, otherwise we might mislabel cache information
* for a chunk that we didn't store this time
*/
/* Don't update the version number if it's not yet set. */
if (!hsame(tdc->f.versionNo, h_unset)
&& hcmp(tdc->f.versionNo, oldDV) >= 0) {
if ((!(afs_dvhack || foreign)
&& hsame(avc->f.m.DataVersion, newDV))
|| ((afs_dvhack || foreign)
&& (origCBs == afs_allCBs))) {
/* no error, this is the DV */
UpgradeSToWLock(&tdc->lock, 678);
hset(tdc->f.versionNo, avc->f.m.DataVersion);
tdc->dflags |= DFEntryMod;
/* DWriting may not have gotten cleared above, if all
* we did was a StoreMini */
tdc->f.states &= ~DWriting;
ConvertWToSLock(&tdc->lock);
}
}
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
minj += NCHUNKSATONCE;
} while (moredata);
}
if (code) {
/*
* Invalidate chunks after an error for ccores files since
* afs_inactive won't be called for these and they won't be
* invalidated. Also discard data if it's a permanent error from the
* fileserver.
*/
if (areq->permWriteError || (avc->f.states & CCore)) {
afs_InvalidateAllSegments(avc);
}
}
afs_Trace3(afs_iclSetp, CM_TRACE_STOREALLDONE, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, avc->f.m.Length, ICL_TYPE_INT32, code);
/* would like a Trace5, but it doesn't exist... */
afs_Trace3(afs_iclSetp, CM_TRACE_AVCLOCKER, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, avc->lock.wait_states, ICL_TYPE_INT32,
avc->lock.excl_locked);
afs_Trace4(afs_iclSetp, CM_TRACE_AVCLOCKEE, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, avc->lock.wait_states, ICL_TYPE_INT32,
avc->lock.readers_reading, ICL_TYPE_INT32,
avc->lock.num_waiting);
/*
* Finally, if updated DataVersion matches newDV, we did all of the
* stores. If mapDV indicates that the page cache was flushed up
* to when we started the store, then we can relabel them as flushed
* as recently as newDV.
* Turn off CDirty bit because the stored data is now in sync with server.
*/
if (code == 0 && hcmp(avc->mapDV, oldDV) >= 0) {
if ((!(afs_dvhack || foreign) && hsame(avc->f.m.DataVersion, newDV))
|| ((afs_dvhack || foreign) && (origCBs == afs_allCBs))) {
hset(avc->mapDV, newDV);
avc->f.states &= ~CDirty;
}
}
osi_FreeLargeSpace(dcList);
/* If not the final write a temporary error is ok. */
if (code && !areq->permWriteError && !(sync & AFS_LASTSTORE))
code = 0;
return code;
} /*afs_StoreAllSegments (new 03/02/94) */
/* this call, unlike osi_FlushText, is supposed to discard caches that may
contain invalid information if a file is written remotely, but that may
contain valid information that needs to be written back if the file is
being written locally. It doesn't subsume osi_FlushText, since the latter
function may be needed to flush caches that are invalidated by local writes.
avc->pvnLock is already held, avc->lock is guaranteed not to be held (by
us, of course).
*/
void
osi_FlushPages(struct vcache *avc, afs_ucred_t *credp)
{
afs_hyper_t origDV;
#if defined(AFS_CACHE_BYPASS)
/* The optimization to check DV under read lock below is identical a
* change in CITI cache bypass work. The problem CITI found in 1999
* was that this code and background daemon doing prefetching competed
* for the vcache entry shared lock. It's not clear to me from the
* tech report, but it looks like CITI fixed the general prefetch code
* path as a bonus when experimenting on prefetch for cache bypass, see
* citi-tr-01-3.
*/
#endif
if (vType(avc) == VDIR) {
/* not applicable to directories; they're never mapped or stored in
* pages */
return;
}
ObtainReadLock(&avc->lock);
/* If we've already purged this version, or if we're the ones
* writing this version, don't flush it (could lose the
* data we're writing). */
if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0)
|| ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) {
ReleaseReadLock(&avc->lock);
return;
}
ReleaseReadLock(&avc->lock);
ObtainWriteLock(&avc->lock, 10);
/* Check again */
if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0)
|| ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) {
ReleaseWriteLock(&avc->lock);
return;
}
/* At this point, you might think that we can skip trying to flush pages
* if mapDV is zero, since a file with a zero DV will not have any data in
* it. However, some platforms (notably Linux 2.6.22+) will keep a page
* full of zeroes around for an empty file. So play it safe and always
* flush pages. */
AFS_STATCNT(osi_FlushPages);
hset(origDV, avc->f.m.DataVersion);
afs_Trace3(afs_iclSetp, CM_TRACE_FLUSHPAGES, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, origDV.low, ICL_TYPE_INT32, avc->f.m.Length);
ReleaseWriteLock(&avc->lock);
#ifndef AFS_FBSD70_ENV
AFS_GUNLOCK();
#endif
osi_VM_FlushPages(avc, credp);
#ifndef AFS_FBSD70_ENV
AFS_GLOCK();
#endif
ObtainWriteLock(&avc->lock, 88);
/* do this last, and to original version, since stores may occur
* while executing above PUTPAGE call */
hset(avc->mapDV, origDV);
ReleaseWriteLock(&avc->lock);
}
/* this call, unlike osi_FlushText, is supposed to discard caches that may
contain invalid information if a file is written remotely, but that may
contain valid information that needs to be written back if the file is
being written locally. It doesn't subsume osi_FlushText, since the latter
function may be needed to flush caches that are invalidated by local writes.
avc->pvnLock is already held, avc->lock is guaranteed not to be held (by
us, of course).
*/
void
osi_FlushPages(struct vcache *avc, afs_ucred_t *credp)
{
#ifdef AFS_FBSD70_ENV
int vfslocked;
#endif
afs_hyper_t origDV;
#if defined(AFS_CACHE_BYPASS)
/* The optimization to check DV under read lock below is identical a
* change in CITI cache bypass work. The problem CITI found in 1999
* was that this code and background daemon doing prefetching competed
* for the vcache entry shared lock. It's not clear to me from the
* tech report, but it looks like CITI fixed the general prefetch code
* path as a bonus when experimenting on prefetch for cache bypass, see
* citi-tr-01-3.
*/
#endif
ObtainReadLock(&avc->lock);
/* If we've already purged this version, or if we're the ones
* writing this version, don't flush it (could lose the
* data we're writing). */
if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0)
|| ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) {
ReleaseReadLock(&avc->lock);
return;
}
ReleaseReadLock(&avc->lock);
ObtainWriteLock(&avc->lock, 10);
/* Check again */
if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0)
|| ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) {
ReleaseWriteLock(&avc->lock);
return;
}
if (hiszero(avc->mapDV)) {
hset(avc->mapDV, avc->f.m.DataVersion);
ReleaseWriteLock(&avc->lock);
return;
}
AFS_STATCNT(osi_FlushPages);
hset(origDV, avc->f.m.DataVersion);
afs_Trace3(afs_iclSetp, CM_TRACE_FLUSHPAGES, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, origDV.low, ICL_TYPE_INT32, avc->f.m.Length);
ReleaseWriteLock(&avc->lock);
#ifdef AFS_FBSD70_ENV
vfslocked = VFS_LOCK_GIANT(AFSTOV(avc)->v_mount);
#endif
#ifndef AFS_FBSD70_ENV
AFS_GUNLOCK();
#endif
osi_VM_FlushPages(avc, credp);
#ifndef AFS_FBSD70_ENV
AFS_GLOCK();
#endif
#ifdef AFS_FBSD70_ENV
VFS_UNLOCK_GIANT(vfslocked);
#endif
ObtainWriteLock(&avc->lock, 88);
/* do this last, and to original version, since stores may occur
* while executing above PUTPAGE call */
hset(avc->mapDV, origDV);
ReleaseWriteLock(&avc->lock);
}
