The cm_scache structure has included a reference to the associated cm_volume_t,
volp. By doing so the reference count on the cm_volume_t objects never hits
zero unless all of the cm_scache_t objects in that volume have been reused.
This prevents cm_volume object recycling.
This commit removes the hard reference and instead adds a function
cm_GetVolumeByFID which obtains a reference to the cm_volume that matches
the cm_scache fid value as needed.
A new "Volumes" registry value is added to permit explicit specification
of the number of volume objects to be allocated.
Update the fileserver documentation for demand-attach and add documentation
of other missing options and notes where some options are only applicable
with particular builds.
Introduce a new set of macros that can be used to permit either mutex
based protection or Interlocked operation protection increments,
decrements, or additions.
adapted from marc dionne's patch, but works more directly like the darwin port,
further, i'm guessing iget_locked will go away evewntually. maybe this is
too conservative?
If we want to enforce the open for read flag in smb_ReadData we must
also check for execute access. We map execute access to read access
because AFS doesn't have a separate permission for that.
change the RxMaxMtu setting to 0 to restore default
behavior (aka performance). The Cisco VPN 5.0.01.0600
client no longer requires a reduced mtu size for compatibility
with AFS.
syscfg_GetIFInfo() obtains the current list of IP addresses.
Its a really expensive operation. Call the function once
and cache the data until the next IP address change instead
of calling it everytime we receive a WhoAreYou or construct a
new server object. Adds a new global rw lock, cm_syscfgLock
This is a first cut at a cache manager statistics monitor
that can be used to determine the necessary cache parameters
to support the working set.
Off by default the performance package can be activated
by setting "daemonPerformanceTuningInterval" in the service
Parameters key. As with the other daemon interval values
the unit is in seconds.
At service start and each succeeding interval the cache
manager will write statistics to %TEMP%\afsd_performance.txt
showing the relative usage of cm_scache_t, cm_volume_t and
cm_buf_t objects. The FID statistics keep track of all FIDs
seen by the cache manager during the service session whether
or not they are backed by any live object in the cache.
These statistics are not stored in the cache file.
Convert cm_conn_t reference counts to use Interlocked operations.
This permits several cm_serverLock holds to be converted to read
locks.
Add string translation for VL errors to cm_Analyze().
Permit RXKAD errors other than RXKADEXPIRED to be treated as
non-fatal. Instead immediately retry an alternate server if
there is one. This will permit the client to continue to
access replicated data from an alternate site if one of the
file servers is misconfigured.
Prior to looking up a volume in the vldb the cm_volume_t
object is constructed. If the response is VL_NOENT the
cm_volume_t object should be removed and place at the
head of the recycling list so that we do not recycle
something we might actually care about.
Convert more cm_volumeLock holds to read locks now
that we are using Interlocked operations to manipulate
the reference counts.
If an application opens a file for read only and then attempts to
write to it, we would attempt to release an scp mutex without having
obtained either the scp or the mutex within smb_WriteData
There was no check in smb_ReadData to ensure that the file was opened
for reading. Add one.
on VSALVAGE the client will correctly resort its list and fail over to
other replicas, but only for RO where there is more than one replica,
obviously; since in dafs we expect to only delay on salvage this is the most
useful course of action
(1) an attempt to make better use of bandwidth from the BkgDaemon threads
by preventing the thread from blocking on a vnode that is already
storing data in another thread
(2) prevents CM_SCACHEFLAG_ASYNCSTORE from being reset on a write failure.
(3) fixes cm_EvaluateSysName to avoid accessing uninitialized memory
(4) prevents a lock leak if the symlink's mountpointstring is too long.
(This could never actually happen but better to correct the code.)
Replace the cm_scache_t mutex with a rwlock permitting a small amount
of additional parallelization in places where it is safe to use read
locks instead of write locks. All functions that eventually call
cm_SyncOp must use write locks.
when support for multiple valid buffer data versions was added
forget to remove the force dv change from the freelance code.
this broken automatic additions of new mount points.
somewhere along the way Microsoft stopped opening Event Message files
explicitly. Instead they search the PATH environment variable for
the specified file. I think this is broken but simply registering
"afsd_service.exe" instead of the fully qualified path works.
(1) Add an undocumented store behind mode for use in testing. Set
EnableSMBAsyncStore to 2. When set all smb_WriteData calls are
background writes, all calls to cm_FSync are skipped and file close
operations do not block for dirty buffers to be written. This
permits all writes to be performed in the buf_IncrSyncer thread.
(2) Do not use I64 in osi_Log() format strings as all parameters are
converted to size_t which is 32-bit on 32-bit Windows.
(3) Reduce the number of times the cm_buf_t mutex is obtained, dropped,
obtained, dropped in buf_IncrSyncer
(4) In buf_CleanAsyncLocked, request that a full chunk be written instead
of just the current buffer. cm_SetupStoreBIOD will stop at the
first clean buffer. This reduces the overall number of RPCs that
must be performed.
(5) Define CM_BUF_VERSION_BAD and use it instead of -1.
avoid deadlock in buf_FlushCleanPages().
cannot obtain buffer mutex after a successful
Stabilize call because the scp will be be locked
and obtaining buffer mutex after scache mutex
is a lock order violation.
Attempts to open files which are already write-locked by another
client took forever to return a lock not granted error. This
was because cm_Analyze() would retry the lock request for up to
the RDRtimeout in response to the EAGAIN error. The problem
was that cm_IntSetLock() was not setting the CM_REQ_NORETRY flag.
While examining this issue, discovered two other things:
(1) the infinite wait logic on lock request processing was broken
(2) the cancel outstanding lock request logic wasn't implemented
(3) cm_Analyze() would put the thread to sleep even when retries
were not permitted.
Add lock_convertRToW which permits a read-lock to be upgraded to a
write lock. If the caller is the only reader it permits a fast
transition otherwise it adds the caller to the waiters queue.
In the osi_Log macros, check to see if the log is enabled before making
the function call. This avoids significant function call overhead.
Remove race conditions in the statistics code by switching to
Interlocked functions.
Speed up cm_dnlcEnter() by searching first with a read lock and
then switching to a write lock if not found.
Add HKLM\SOFTWARE\OpenAFS\Client registry values "UseDNLC" and "Debug"DNLC".
I suspect with the B+tree code that the DNLC is unnecessary overhead but
leave it on by default for now.
correctly set and unset HWHO_INPROGRESS as we pass through host probing package.
also, do not block on lock checking for HOSTDELETED. instead, return a null host
and VBUSY; restructure code to allow returning a null host with prejudice.
Further testing on the 115KB AT&T Edge connection shows that 128KB
provides reasonably robust performance with six simultaneous copies of
multi-MB files to AFS.
Instead of tying the async store size to either the chunksize (too large)
or the buffer block size (too small) provide an intermediate value that
can be independently controlled.
In the future it would be desireable for the async store size to be
dynamically determined based upon measurable characteristics of the
network. In the meantime, 32KB is an acceptable performance compromise
that should work well on 1Gbit networks and low-speed cellular networks.
Avoid the need to update the data version number on each buffer associated
with a scache when MergeStatus is called after a StoreData by maintaining
a range of valid data versions as part of the cm_scache_t object.
Improve parallelism by using InterlockedIncrement and InterlockedDecrement
to adjust the refCount fields of cm_buf_t and cm_scache_t objects. This
permits read locks to be used for increments and decrements but requires
a write-lock to perform any actions based upon the refCount hitting zero.
2. Add a 'hash' value to the cm_fid_t structure in order to speed up cm_FidCmp().
Add cm_SetFid() for use in constructing a fid complete with the hash.
3. Redefine the BUF_HASH and BUF_FILEHASH in terms of the cm_fid_t hash which
has a better distribution
4. Modify cm_ConsiderPrefetch to evaluate the amount of data in the most recent
read request instead of the next chunkSize. cm_chunkSize can be dozens or
hundreds of buffers. As a result too much time is spent performing the
evaluation.
5. Fix the usage of cm_scache_t bufCreateLock. The purpose of this lock is
to prevent the creation of new buffers while a truncation is being performed.
All references to bufCreateLock have been removed except in two places:
i. a write-lock surrounding the function that calls buf_Truncate()
ii. a read-lock within buf_GetNewLocked() that actually allocates new
buffers
6. Modify the CM_CONFIG_DATA_MAGIC value to include a CM_CONFIG_DATA_VERSION
number which value be used to force the replacement of the cache file
contents when incompatible changes are made between releases.
7. CM_SCACHESYNC_ASYNCSTORE should not be ordered by cm_SyncOpCheckContinue
8. Avoid calls to multi_Rx if nconns == 0
9. Modify smb_WriteData to perform background writes based upon the crossing
of buf_blockSize boundaries instead of cm_chunkSize boundaries. This will
slow down writes from the SMB interface but will avoid the risk of
the CIFS client disconnecting from the AFS client SMB server.
Add RxEnableHotThread registry option to permit rx_enable_hot_thread
to be set. The default is off for the moment but turning it on does
provide significant performance benefits.
projects / packages / o / openafs.git / log