- 27 Apr, 2014 40 commits
-
-
Andreas Dilger authored
The Linux VFS and Lustre OST_SYNC RPC are both capable of specifying fsync() on a sub-extent of the file {start, end} instead of the full file. This allows less than the full amount of data to be flushed, reducing or possibly eliminating the work needed before the syscall can return. However, the handling of sub-extent of the file for fsync was lost with the move to CLIO on the client and OSD API on the server. They were ignoring the passed {start, end} and using {0, OBD_OBJECT_EOF} instead. Return the ability to pass a sub-extent for fsync() from the client, to the specific stripes/OSTs that need the sync operation, and pass it down to the OSD. The ZFS OSD doesn't handle this yet, but there is room for improvement in a separate patch. Signed-off-by:Andreas Dilger <andreas.dilger@intel.com> Reviewed-on: http://review.whamcloud.com/8626 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4388Reviewed-by:
Bobi Jam <bobijam@gmail.com> Reviewed-by:
Jinshan Xiong <jinshan.xiong@intel.com> Signed-off-by:
Oleg Drokin <oleg.drokin@intel.com> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Ryan Haasken authored
In libcfs_debug_vmsg2, cdls_delay is only clamped between the minimum and the maximum when it is increased by multiplying by the backoff factor. It is not clamped when it is decreased by dividing by the backoff factor. This allows it to achieve values less than the minimum, which allows a console message to be printed that should have been skipped. This patch moves the clamping outside of the else statement, ensuring that cdls_delay is always between the min and the max after the first time through libcfs_debug_vmsg2. Signed-off-by:
Ryan Haasken <haasken@cray.com> Reviewed-on: http://review.whamcloud.com/9503 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4711Reviewed-by:
Chris Horn <hornc@cray.com> Reviewed-by:
Ann Koehler <amk@cray.com> Reviewed-by:
-
Swapnil Pimpale authored
Removed the checks for oi_lockless from osc_io_read_start() and osc_io_write_start(). This patch also removes the unnecessary call to cl_object_attr_get() in osc_io_write_start() before calling cl_object_attr_set() Signed-off-by:
Swapnil Pimpale <spimpale@ddn.com> Reviewed-on: http://review.whamcloud.com/8797 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-3868Reviewed-by:
John L. Hammond <john.hammond@intel.com> Reviewed-by:
-
Li Xi authored
spin_is_locked() is always false when the platform is uniprocessor and CONFIG_DEBUG_SPINLOCK is not enabled. This patch replaces its assertion by assert_spin_locked(). Signed-off-by:
Li Xi <lixi@ddn.com> Signed-off-by:
James Simmons <uja.ornl@gmail.com> Reviewed-on: http://review.whamcloud.com/8144 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4199Reviewed-by:
Alexey Lyashkov <alexey_lyashkov@xyratex.com> Reviewed-by:
-
John L. Hammond authored
In llite remove unused declarations, parameters, types, and unused, get-only, or set-only structure members. Add static and const qualifiers to declarations where possible. Signed-off-by:
Lai Siyao <lai.siyao@intel.com> Reviewed-by:
-
wang di authored
In some cases, cl_default_mds_easize might be zero, especially for MDC connected to non-MDT0, then mdc might pack getattr RPC with zero eadatasize. If client is trying to access remote striped directory with zero eadatasize, MDT will not return layout information of the striped direcotry, which will be mis-regarded as non-striped directory. So we should use cl_max_mds_easize if cl_default_mds_easize is zero. Signed-off-by:
wang di <di.wang@intel.com> Reviewed-on: http://review.whamcloud.com/9862 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4847Reviewed-by:
-
Jinshan Xiong authored
It's not atomic to check the last reference and state of cl_lock in cl_lock_put(). This can cause a problem that an using lock is freed, if the process is preempted between atomic_dec_and_test() and (lock->cll_state == CLS_FREEING). This problem can be solved by holding a refcount by coh_locks. In this case, it can be sure that if the lock refcount reaches zero, nobody else can have any chance to use it again. Signed-off-by:
-
Bob Glossman authored
Change the LL_IOC_LLOOP_INFO ioctl in the lustre lloop device driver to return an error instead of causing panics with LASSERT(). Signed-off-by:
Bob Glossman <bob.glossman@intel.com> Reviewed-on: http://review.whamcloud.com/9888 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4863Reviewed-by:
Nathaniel Clark <nathaniel.l.clark@intel.com> Reviewed-by:
-
Ryan Haasken authored
Messages which are printed by ll_dirty_page_discard_warn() should not be rate limited. If they are rate limited, some files which may be corrupted on client eviction will not be reported to the user. This patch changes the CWARN to a CDEBUG to disable console message rate limiting for this message. The dirty page discard warnings are already limited on a per-file basis by the function vvp_vmpage_error which calls ll_dirty_page_discard_warn only if the ccc_object's cob_discard_page_warned == 0. Signed-off-by:
Cory Spitz <spitzcor@cray.com> Reviewed-by:
-
Dmitry Eremin authored
According https://www.kernel.org/doc/Documentation/mutex-design.txt: - the mutex subsystem is slightly faster and has better scalability for contended workloads. In terms of 'ops per CPU cycle', the semaphore kernel performed 551 ops/sec per 1% of CPU time used, while the mutex kernel performed 3825 ops/sec per 1% of CPU time used - it was 6.9 times more efficient. - there are no fastpath tradeoffs, the mutex fastpath is just as tight as the semaphore fastpath. On x86, the locking fastpath is 2 instructions. - 'struct mutex' semantics are well-defined and are enforced if CONFIG_DEBUG_MUTEXES is turned on. Semaphores on the other hand have virtually no debugging code or instrumentation. One more benefit of mutex is optimistic spinning. It try to spin for acquisition when there are no pending waiters and the lock owner is currently running on a (different) CPU. The rationale is that if the lock owner is running, it is likely to release the lock soon. This significantly reduce amount of context switches when locked region is small and we have high contention. Signed-off-by:
Dmitry Eremin <dmitry.eremin@intel.com> Reviewed-on: http://review.whamcloud.com/9095 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4257Reviewed-by:
-
Dmitry Eremin authored
It's just optimization. The mutex subsystem is slightly faster and has better scalability for contended workloads. Remove the lustre_lock and it's accessor functions l_lock(), l_unlock(), l_lock_init(), and l_has_lock() since they have not been used by the code since Lustre 1.6. Signed-off-by:
Alex Zhuravlev <alexey.zhuravlev@intel.com> Signed-off-by:
-
Li Xi authored
Some CPU table functions for uniprocessor architecture is missing. Signed-off-by:
Liang Zhen <liang.zhen@intel.com> Reviewed-by:
-
John L. Hammond authored
Remove the nowhere included header lustre/include/ioctl.h. Signed-off-by:
-
Liang Zhen authored
lnet_shutdown_lndnis() may enter endless loop if there is a busy NI, this is injected by LNet SMP improvements. It's fixed in this patch. Signed-off-by:
Isaac Huang <he.huang@intel.com> Reviewed-by:
-
Dmitry Eremin authored
sscanf format specification '%u' expects type 'unsigned int *' for 'u', but parameter 3 has a different type 'int*'. Signed-off-by:
-
Dmitry Eremin authored
Pointer 'mod' checked for NULL at line 160 may be dereferenced at line 208. Signed-off-by:
Fan Yong <fan.yong@intel.com> Signed-off-by:
-
Dmitry Eremin authored
Pointer '*exp' returned from call to function 'class_conn2export' at line 523 may be NULL and may be dereferenced at line 543. Signed-off-by:
-
Andreas Dilger authored
Locate the loh_flags and loh_ref fields together in lu_object_header to avoid holes and shrink the structure by 8 bytes. Signed-off-by:
-
Jinshan Xiong authored
We used to access layout version under the protection of ldlm lock, this introduces extra overhead for dlm lock matching. In this patch, lli_layout_lock is introduced to access the layout version. Also, when a layout lock is losing, we should tear down mmap of the correspoding inode to avoid stale data accessing in the future. This is part of technical verification of replication. Signed-off-by:
-
Christopher J. Morrone authored
The statahead debug messages include the pid of the current process in their body. This is both redudant (because all lustre log messages contain the pid), and sometimes downright misleading. For instance the messages would say something like "stopping statahead thread 3446". One would probably think that 3446 is the pid of the process that is being stopped, but in fact it was the pid of the caller issuing the stop signal. We remove all superfluous pids from the messages. Next we have the ll_statahead_thread() and the ll_agl_thread() record their respective pids in their respective ptlrpc_thread structures. This allows to print the pid of the thread that we are trying to stop (which is actually useful info) from other threads, such as those calling ll_stop_statahead(). Signed-off-by:
Christopher J. Morrone <morrone2@llnl.gov> Reviewed-on: http://review.whamcloud.com/9360 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4624Reviewed-by:
-
Christopher J. Morrone authored
The statahead and statahead agl threads blindly set their thread state to SVC_RUNNING without checking the state first. If, for instance, another thread sets the state to SVC_STOPPING that stop signal will now have been lost. Deadlock ensues. We also partly improve the sai reference counting, because a race exists where the ll_stop_statahead thread can drop the default reference, and the statahead thread can exit and drop its reference as well. With no references on the sai, the final put will poison and free the buffer. The original do_statahead_enter() function may then continue to access the buffer after it is freed because it did not take a reference of its own. We add a local reference to address that. Signed-off-by:
-
Brian Behlendorf authored
When allocating a reply buffer for the striping information don't assume the unlikely worst case. Instead, assume the common case and size the buffer based on the observed default ea/cookie size. The default size is initialized to a single stripe and allowed to grow up to an entire page if needed. This means that for smallish filesystems (less than ~21 OSTs) where the worst case striping information can fit in a single page there is effectively no change. Only for larger filesystem will the default be less than the maximum. This has a number of advantages. * By limiting the default reply buffer size we avoid always vmalloc()'ing the buffer because it exceeds four pages in size and instead kmalloc() it. This prevents the client from thrashing on the global vmalloc() spin lock. * A reply buffer of exactly the right size (no larger) is allocated in the overflow case. These larger reply buffers are still unlikely to exceed the 16k limit where a vmalloc() will occur. * Saves memory in the common case. Wide striped files exceeded the default are expected to be the exception. The reason this patch works is because the ptlrpc layer is smart enough to reallocate the reply buffer when an overflow occurs. Therefore the client doesn't have to drop the incoming reply and send a new request with a larger reply buffer. It's also worth mentioning that the reply buffer always contains a significant amount of extra padding because they are rounded up to the nearest power of two. This means that even files striped wider than the default have a good chance of fitting in the allocated reply buffer. Also remove client eadatasize check in mdt xattr packing because as said above client can handle -EOVERFLOW. Signed-off-by:
Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by:
-
Jinshan Xiong authored
Otherwise, it will cause deadlock because it essentially holds some sub locks and then to request others in an arbitrary order. Signed-off-by:
-
Bobi Jam authored
File write before io loop will take lli_trun_sem read semaphore to protect osc_extent, while after generic_file_aio_write() done, it could possible need to kill suid or sgid, which will call ll_setattr_raw() to change the inode's attribute, and it does not involve size. So the ll_truc_sem write semaphore should be constrained around ll_setattr_ost() to not come across the lli_trunc_sem read semaphore get from the normal file write path. Signed-off-by:
Bobi Jam <bobijam.xu@intel.com> Reviewed-on: http://review.whamcloud.com/9267 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4627Reviewed-by:
-
Bobi Jam authored
The last parameter @datasync of fsync() has following indication: * if datasync=0, we'd always flush data and metadata * if datasync=1, we'd always flush data while does not flush modifed metadata unless that metadata is needed in order to allow a subsequent data retrieval to be correctly handled. For example, a change to the file size would require a metadata flush. Lustre client can not tell the difference easily, and would issue MDS_SYNC and OST_SYNC in all cases. Signed-off-by:
-
Li Xi authored
Setxattr does not check the permission when setting ACL xattrs. This will cause security problem because any user can walk around permission checking by changing ACL rules. Signed-off-by:
-
James Nunez authored
The total size of an HSM archive request may exceed the desired (LNET) message. When this happens, it can hang the client and not allow the archive request to succeed. Before we know the total size of the hsm_action_items, we need to limit the size of the reguest. Doing this limits the number of items that can be sent in one archive request. We'e reduced the size allowed for the user archive request to MDS_MAXREQSIZE/3. Signed-off-by:
James Nunez <james.a.nunez@intel.com> Reviewed-on: http://review.whamcloud.com/9393 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4639Reviewed-by:
-
Peng Tao authored
If fsname is 8-byte aligned, hai_zero fails to count the ending NULL terminator causing hai to directly attached after fsname and future hai_zero will return a different position for first hai. Signed-off-by:
Peng Tao <bergwolf@gmail.com> Reviewed-on: http://review.whamcloud.com/9431 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4689Reviewed-by:
-
John L. Hammond authored
Remove the unused function llog_obd_add(). Remove the unused count and parameters from llog_cancel(). Move dump_lsm() from obdclass to the only module that uses it (lov). Remove obd_lov.h. Signed-off-by:
Mike Pershin <mike.pershin@intel.com> Signed-off-by:
-
John L. Hammond authored
Remove the unused functions lov_llog_init(), lov_llog_finish(), their supporting functions, and the file lov_log.c. Signed-off-by:
-
Andreas Dilger authored
Quiet some common console error messages for permission errors that can be hit in common cases. Signed-off-by:
Faccini Bruno <bruno.faccini@intel.com> Reviewed-by:
-
Ann Koehler authored
In Lustre 2.4, the flags passed to the memory allocation functions are translated from CFS enumeration values types to the kernel GFP values by calling cfs_alloc_flags_to_gfp(). This function adds __GFP_WAIT to all flags except CFS_ALLOC_ATOMIC. In 2.5, when the cfs wrappers were dropped, cfs_alloc_flags_to_gfp() was removed and the CFS_ALLOC_xxxx was simply replaced with __GFP_xxxx. This means that most memory allocation calls are missing the __GFP_WAIT flag. The result is that Lustre experiences more ENOMEM errors, many of which the higher levels of Lustre do not handle robustly. Notes GFP_NOFS = __GFP_WAIT | __GFP_IO. So the patch replaces __GFP_IO with GFP_NOFS. Patch does not add __GFP_WAIT to GFP_IOFS. GFP_IOFS was not used in Lustre 2.4 so it has never been used with __GFP_WAIT. Signed-off-by:
Emoly Liu <emoly.liu@intel.com> Reviewed-on: http://review.whamcloud.com/9223 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4357Reviewed-by:
-
Andriy Skulysh authored
Add rpc_cache for allocating ptlrpc_requests. Xyratex-bug-id: MRP-689 Signed-off-by:
Andriy Skulysh <Andriy_Skulysh@xyratex.com> Signed-off-by:
Niu Yawei <yawei.niu@intel.com> Reviewed-on: http://review.whamcloud.com/6874 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2424Reviewed-by:
-
Dmitry Eremin authored
Array 'message_buf' of size 500 may use index value(s) -1 Object 'enc_key.data' was freed at line 164 after being freed by calling 'free' at line 150. Also there are 3 similar errors on line(s) 164. Suspicious dereference of pointer 'vmsg' before NULL check at line 187. Also there are 2 similar errors on line(s) 196, 205. Suspicious dereference of pointer 'rmsg' before NULL check at line 191. Also there are 2 similar errors on line(s) 200, 209. Signed-off-by:
-
Andreas Dilger authored
If a connection has been stopped with ptlrpc_pinger_del_import() and marked obd_no_recov, don't reconnect in ptlrpc_disconnect_import() if the import is already disconnected. Otherwise, without the pinger it will just wait there indefinitely for the reconnection that will never happen. Put the obd_no_recov check inside ptlrpc_import_in_recovery() so that any threads waiting on the connection to recover would also be broken out of their sleep if obd_no_recov is set. Signed-off-by:
-
Niu Yawei authored
The LDLM_FL_BLOCK_NOWAIT flag should be cleared when re-enqueue the agl lock as normal glimpse, otherwise, it won't get size back if there is conflicting locks on other client. Signed-off-by:
Ned Bass <bass6@llnl.gov> Reviewed-by:
-
Li Xi authored
This patch moves lock's skip flag clearing from lru-delete to lru-add code to prevent clearing lock's flag without resource lock protection. Signed-off-by:
-
Matt Ezell authored
LNET messages that are dropped are not accounted for correctly in /proc/sys/lnet/stats. What I assume to be a simple typo is causing drop_length to be double-counted and drop_count to never be incremented. Signed-off-by:
Matt Ezell <ezellma@ornl.gov> Reviewed-on: http://review.whamcloud.com/9096 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4577Reviewed-by:
-
Hongchao Zhang authored
In ptlrpc_activate_import(), obd_import->imp_deactive should be checked if it is deactivated, otherwise it will trigger an LBUG in ptlrpc_invalidate_import(): ptlrpc_invalidate_import() ASSERTION(imp->imp_invalid) failed Signed-off-by:
Hongchao Zhang <hongchao.zhang@intel.com> Reviewed-on: http://review.whamcloud.com/8747 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4386Reviewed-by:
-
Cheng Shao authored
Function ptlrpc_update_export_timer generates lots of D_HA level log messages whenever the export timer gets updated. Those log messages are found little use for issue investigations, and it will take space in the Lustre log buffer. We are removing it now. Xyratex-bug-id: MRP-733 Signed-off-by:
Cheng Shao <cheng_shao@xyratex.com> Reviewed-on: http://review.whamcloud.com/9147 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4590Reviewed-by:
-
