"the non AFS_GLOBAL_RXLOCK_KERNEL version of
rxi_ReceiveAckPacket() wasnt counting 'hard' acks and not opening the
congestion window. using rx_Readv() directs the incoming packets into
the target iov, and sends a hard ack (skipping soft acks). "
A deadlock was detected when performing "fs flushall" if the file
server reports VNOVNODE. The scp->createBufferLock is already
held by the current thread and there is no mechanism to propagate
the knowledge. Therefore, an alternate mechanism for clearing the
cache must be developed. A new function cm_RecycleSCache(scp,flags)
has been extracted from cm_GetNewSCache(). This function performs
the task of recycling an cm_scache_t object. When called from
cm_GetNewSCache() with no flags the expectation is that there are
no associated buffers that are queued to be read or written. When
called from cm_Analyze() with the CM_SCACHE_RECYCLEFLAG_DESTROY_BUFFERS
flag, any queued buffers will be de-queued and marked as if the operations
were performed so the data can be discarded.
This patch also addresses the following issues:
* CM_ERROR codes were logged as "unknown" by cm_Analyze.
* In response to VNOVNODE, the parent is only discarded if the current
cm_scache_t is not a directory.
* In the Ioctl FlushFile and FlushVolume functions, there were no
checks to protect against flushing the Freelance SCache entries.
* In FlushFile, the wrong cm_scache_t object was being released.
* In cm_GetNewSCache, do not allow Freelance SCache entries to be
recycled. Choose a new entry if cm_RecycleSCache fails.
Speed up the performance of the cache manager by not holding the
smb_fid_t mutex across calls to cm_SyncOp and AFS RPCs.
Ensure that all smb_fid_t flag references are protected.
It is possible to have a valid callback but not to have the
required ACL info for the current user. Force acquisition of
a FetchStatus but do so without making multiple calls.
Dynamically adjust the priority of server threads based upon the age
of the cifs request that is being processed. Bump the priority one
level for each 15 seconds of age.
while investigating the cause of the delayed write errors it was observed
that all of the sleep queues are LIFO. This has the side effect of
encouraging starvation. Changing the queues to FIFOs revealed a serious
problem affecting the use of all queues which use both head and tail
pointers. The removal function osi_QRemove does not take a tail pointer
and therefore the pointer is always left hanging. If the number of elements
ever drops to zero the queue becomes corrupted.
Added osi_QRemoveHT to be used whenever head and tail pointers are used.
Updated all callers in afsd.
Do not delete an old host entry because the UUID because the UUIDs
of the clients at the old addresses are wrong. Remove the addresses.
Do not allow the same host,addr,port combination to be added into
the hash table multiple times.
====================
This delta was composed from multiple commits as part of the CVS->Git migration.
The checkin message with each commit was inconsistent.
The following are the additional commit messages.
====================
Add some more logging
Do not allow 'code' to be re-used for probing the uuid of the old host
primary address because we are relying on the value of 'code' to indicate
whether or not RCallBackConnectBack failed. If it did, we mark the
client as VENUSDOWN, if not we marked it as RESETDONE.
====================
log the old addr, not the new one
When the client changes IP addresses, in order to obtain optimal
performance it must:
(1) force the replacement of all existing rx connections
(2) ping all up and down servers
