- 15 Nov, 2016 1 commit
-
-
Gregoire Pichon authored
This patch is the main client part of a new feature that supports multiple modify metadata RPCs in parallel. Its goal is to improve metadata operations performance of a single client, while maintening the consistency of MDT reply reconstruction and MDT recovery mechanisms. It allows to manage the number of modify RPCs in flight within the client obd structure and to assign a virtual index (the tag) to each modify RPC to help server side cleaning of reply data. The mdc component uses this feature to send multiple modify RPCs in parallel. Signed-off-by:
Gregoire Pichon <gregoire.pichon@bull.net> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5319 Reviewed-on: http://review.whamcloud.com/14374Reviewed-by:
Andreas Dilger <andreas.dilger@intel.com> Reviewed-by:
Alex Zhuravlev <alexey.zhuravlev@intel.com> Reviewed-by:
Oleg Drokin <oleg.drokin@intel.com> Signed-off-by:
James Simmons <jsimmons@infradead.org> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
- 14 Nov, 2016 39 commits
-
-
Henri Doreau authored
Implement non-blocking migration based on exclusive open instead of group lock. Implemented exclusive close operation to atomically put a lease, swap two layouts and close a file. This allows race-free migrations. Make the caller responsible for retrying on failure (EBUSY, EAGAIN) in non-blocking mode. In blocking mode, allow applications to trigger layout swaps using a grouplock they already own, to prevent race conditions between the actual data copy and the layout swap. Updated lfs accordingly. File leases are also taken in blocking mode, so that lfs migrate can issue a warning if an application attempts to open a file that is being migrated and gets blocked. Timestamps (atime/mtime) are set from userland, after the layout swap is performed, to prevent conflicts with the grouplock. lli_trunc_sem is taken/released in the vvp_io layer, under the DLM lock. This re-ordering fixes the original issue between truncate and migrate. Signed-off-by:
Henri Doreau <henri.doreau@cea.fr> Signed-off-by:
Jinshan Xiong <jinshan.xiong@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-4840 Reviewed-on: http://review.whamcloud.com/10013Reviewed-by:
John L. Hammond <john.hammond@intel.com> Reviewed-by:
frank zago <fzago@cray.com> Reviewed-by:
-
Liang Zhen authored
In current lnet selftest, both client and server side bulk have no offset and we can only test page aligned IO, this patch changed this: - user can set brw offset by lst add_test ... brw off=OFFSET ... - offset is only effective on client side so far - to simply implementation, offset needs to be eight bytes aligned Signed-off-by:
Liang Zhen <liang.zhen@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5718 Reviewed-on: http://review.whamcloud.com/12496Reviewed-by:
Doug Oucharek <doug.s.oucharek@intel.com> Reviewed-by:
James Simmons <uja.ornl@yahoo.com> Reviewed-by:
-
wang di authored
Release lmv_init_mutex once the new target is added into lmv_tgt_desc, so lmv_obd_connect will not be serialized. New target should be allowed to added to fld client lists, so FLD can always choose new added target to do the FLD lookup request, and also remove some noise error messages in this process. Signed-off-by:
wang di <di.wang@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6713 Reviewed-on: http://review.whamcloud.com/15269Reviewed-by:
Niu Yawei <yawei.niu@intel.com> Reviewed-by:
-
Liang Zhen authored
ptlrpc is using rq_xid as matchbits of bulk data, which means it has to change rq_xid for bulk resend to avoid several bulk data landing into the same buffer from different resends. This patch uses one of reserved __u64 of ptlrpc_body to transfer mbits to peer, matchbits is now separated from xid. With this change, ptlrpc can keep rq_xid unchanged on resend, it only updates matchbits for bulk data. This protocol change is only applied if both sides of connection have OBD_CONNECT_BULK_MBITS, otherwise, ptlrpc still uses old approach and update xid while resending bulk. Signed-off-by:
-
John L. Hammond authored
obd_unpackmd() is only implemented by LMV so move it from OBD operations to OBD MD operations and update the prototype to reflex the actual usage. Remove the unused function obd_free_memmd(). Signed-off-by:
-
Jinshan Xiong authored
Implement cl_req_attr_set with a cl_object operation. Get rid of cl_req and related function and data structures. Signed-off-by:
Bobi Jam <bobijam@hotmail.com> Reviewed-by:
-
Aditya Pandit authored
If you create a file, archive and release it, it keeps only a link and all information in xattr. If you tar the file with --xattr you will store the same striping information and link information in the tar. If you delete the file, the file and archive state does not make sense. Now if you restore the file using tar with xattr having the RELEASED flag turned on, then it is not correct because this is a new file. Hence ignoring the HSM xattr and masking out the "RELEASED" flag for the files, which are not archived. Signed-off-by:
Aditya Pandit <panditadityashreesh@yahoo.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6214 Reviewed-on: http://review.whamcloud.com/16060Reviewed-by:
-
wang di authored
If there are bad stripe during striped dir revalidation, most likely due the race between close(unlink) and getattr, then let's revalidate the dentry, instead of return error, like normal directory. Signed-off-by:
-
Andreas Dilger authored
Fix the alignment of fields in commonly-used structures to reduce memory usage on the client and server. Structures fixed: ptlrpc_reply_state: reduced by 8 bytes obd_device: reduced by 16 bytes niobuf_local: reduced by 8 bytes Signed-off-by:
Dmitry Eremin <dmitry.eremin@intel.com> Reviewed-by:
-
Henri Doreau authored
There used to be several pre-cleanup phases, but only OBD_CLEANUP_EXPORTS is actually used. Thus remove the whole notion of precleanup phases. Signed-off-by:
-
Ben Evans authored
Put IT_* definitions into an enum, as they're sent over the wire, adjust calls, print statements, etc. to use the new enum. Signed-off-by:
Ben Evans <bevans@cray.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6746 Reviewed-on: http://review.whamcloud.com/16228Reviewed-by:
-
Andreas Dilger authored
Improve the error messages related to DFID output and parsing for usage in userspace. Signed-off-by:
-
Oleg Drokin authored
osc_ast_guard has been removed by the clio simplification. Remove the last lock class definition. Signed-off-by:
-
Andreas Dilger authored
Remove unused file attribute flag LUSTRE_BFLAG_UNCOMMITTED_WRITES that was used internally on the client at one point. Add flags from the kernel which may be useful in the near future. Signed-off-by:
-
wang di authored
Do not lookup master inode by ilookup5, instead it should use ilookup5_nowait, otherwise it will cause dead lock, 1. Client1 send chmod req to the MDT0, then on MDT0, it enqueues master and all of its slaves lock, (mdt_attr_set() ->mdt_lock_slaves()), after gets master and stripe0 lock, it will send the enqueue request(for stripe1) to MDT1, then MDT1 finds the lock has been granted to client2. Then MDT1 sends blocking ast to client2. 2. At the same time, client2 tries to unlink the striped dir (rm -rf striped_dir), and during lookup, it will hold the master inode of the striped directory, whose inode state is NEW, then tries to revalidate all of its slaves, (ll_prep_inode()->ll_iget()->ll_read_inode2()-> ll_update_inode().). And it will be blocked on the server side because of 1. 3. Then the client get the blocking_ast request, cancel the lock, but being blocked by ilookup5 in ll_md_blocking_ast(), because the inode state is still NEW. Signed-off-by:
Lai Siyao <lai.siyao@intel.com> Reviewed-by:
-
Alexander Boyko authored
Fix: 5c689e68 ("staging/lustre/ptlrpc: race at req processing") decreased the race window, but does not remove it. Disable rq_resend right after MSG_REPLAY flag set. Import lock protects two threads from race between set/clear MSG_REPLAY and rq_resend flags. Signed-off-by:
Alexander Boyko <alexander_boyko@xyratex.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5554 Xyratex-bug-id: MRP-1888 Reviewed-on: http://review.whamcloud.com/10735Reviewed-by:
Mike Pershin <mike.pershin@intel.com> Reviewed-by:
-
Henri Doreau authored
Do not release nrs_lock in nrs_policy_stop0 to prevent op_policy_stop() from being executed concurrently. Signed-off-by:
Emoly Liu <emoly.liu@intel.com> Reviewed-by:
-
Jinshan Xiong authored
To avoid a race that osc_extent and osc_object destroy happens on the same time, which causes kernel crash. Signed-off-by:
Patrick Farrell <paf@cray.com> Reviewed-by:
-
wang di authored
If ll_iget fails during inode initialization, especially during striped directory lookup after creation failed, then it should clear stripe MD before make_bad_inode(), because make_bad_inode() will reset the i_mode, which can cause ll_clear_inode() skip freeing those stripe MD. Signed-off-by:
-
wang di authored
If clients or other targets can not get IR config lock or lock, the mount should continue, instead of failing. Because timeout mechanism will handle the recovery anyway. Signed-off-by:
-
John L. Hammond authored
In ldlm_expired_completion_wait() remove the useless LCONSOLE_WARN() message and upgrade the LDLM_DEBUG() statement to LDLM_ERROR(). Signed-off-by:
-
Hongchao Zhang authored
At client side, the replay cursor using to speed up the lookup of committed open requests in its obd_import should be resetted for normal connection (not reconnection) during recovery. Signed-off-by:
Hongchao Zhang <hongchao.zhang@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6802 Reviewed-on: http://review.whamcloud.com/17351Reviewed-by:
-
Hiroya Nozaki authored
Because of a implementation of generic_perform_write(), write(2) may return 0 with no errno even if EDQUOT or ENOSPC actually happened in it. This patch fixes the issue with setting a proper errno to ci_result. Signed-off-by:
Hiroya Nozaki <nozaki.hiroya@jp.fujitsu.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6732 Reviewed-on: http://review.whamcloud.com/15302Reviewed-by:
-
Andreas Dilger authored
In some cases with a very long pathname, such as with sanity.sh test_154c, mdc_ioc_fid2path() would spew long debug messages to the log, because libcfs_debug_vmsg2() refuses to log messages over one page in size. Truncate the debug message to only log the last 512 characters of the pathname, which is sufficient for most debugging, saves a bit of space in the debug log, and will prevent the debug logging from printing to the console in the first place. Signed-off-by:
Jian Yu <jian.yu@intel.com> Reviewed-by:
-
John L. Hammond authored
The reference copytool cannot handle migration of HSM archive files. In the MDT migration path check for HSM attributes and fail if they are present. In the LMV layer allow creation of volatile files with any MDT index. Add a test to sanity-hsm to ensure that attempting to migrate an HSM archive file is handled safely. Signed-off-by:
-
Andrew Wellington authored
SELinux contexts are applied by the kernel if mount options are not binary. As we don't use any binary mount options in Lustre, remove the binary mount option flag. Signed-off-by:
Andrew Wellington <andrew.wellington@anu.edu.au> Signed-off-by:
Sebastien Buisson <sebastien.buisson@bull.net> Reviewed-by:
-
Jinshan Xiong authored
After range lock is introduced to Lustre, it's possible for multiple threads to submit osc_extents with partial pages, and finally I/O engine may try to merge these extents, which will end up with assert in osc_build_rpc(). In this patch, osc_extent::oe_no_merge is introduced, and this flag is set if osc_extent submitted via osc_io_submit() includes partial pages. This flag is used by I/O engine to stop merging this kind of extents. Signed-off-by:
-
Lai Siyao authored
During statahead file may be recreated, though this is rare case, current code will leak the lock, this patch will release lock in this case. Signed-off-by:
Fan Yong <fan.yong@intel.com> Reviewed-by:
-
Gregoire Pichon authored
The OSC tunable max_dirty_mb must be set to a value strictly lower than 2048, as it is assumed by OSS in ofd_grant_alloc() routine. Signed-off-by:
-
Arnd Bergmann authored
James Simmons reports: > The ldlm_pool field pl_recalc_time is set to the current > monotonic clock value but the interval period is calculated > with the wall clock. This means the interval period will > always be far larger than the pl_recalc_period, which is > just a small interval time period. The correct thing to > do is to use monotomic clock current value instead of the > wall clocks value when calculating recalc_interval_sec. This broke when I converted the 32-bit get_seconds() into ktime_get_{real_,}seconds() inconsistently. Either one of those two would have worked, but mixing them does not. Staying with the original intention of the patch, this changes the ktime_get_seconds() calls into ktime_get_real_seconds(), using real time instead of mononic time. Cc: stable@vger.kernel.org # v4.4+ Fixes: 8f83409c ("staging/lustre: use 64-bit time for pl_recalc") Reported-by:Arnd Bergmann <arnd@arndb.de> Reviewed-by:
-
Dan Carpenter authored
We accidentally removed a tab here. Let's add it back, and some curly braces as well since this is a muti-line indent. Signed-off-by:
Dan Carpenter <dan.carpenter@oracle.com> Reviewed-by:
-
Bobi Jam authored
When lu_object_alloc() reaches to LOV object init, we need initialize its stripe type beforehand, so that if something wrong in the conf buffer, the object chain need to be traversed to free what has been allocated, with LOV object type be set as LLT_EMPTY, and when the LOV part is reached, it won't panic without knowing what stripe type it is. This patch also improves debug messages in lsm_unpackmd_common(), and does not return error if the LOV device is still processing config log while trying to verify a layout buffer. Signed-off-by:
Bobi Jam <bobijam.xu@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6744 Reviewed-on: http://review.whamcloud.com/15362Reviewed-by:
-
John L. Hammond authored
In lsm_alloc_plain() use a signed loop index to avoid an infinite loop in the error path. Signed-off-by:
-
Shiva Kerdel authored
Follow the kernel type preferrences of using 's32' over 'int32_t'. Signed-off-by:
Shiva Kerdel <shiva@exdev.nl> Acked-by:
Stuart Yoder <stuart.yoder@nxp.com> Signed-off-by:
-
Shiva Kerdel authored
Follow the kernel type preferrences of using 's16' over 'int16_t'. Signed-off-by:
-
Punit Vara authored
Replace all occurences of (1<<x) by BIT(x) in the file ks7010_sdio.h to get rid of checkpatch.pl "CHECK" output "Prefer using BIT macro". Signed-off-by:
Punit Vara <punitvara@gmail.com> Signed-off-by:
-
Daeseok Youn authored
Re-arrange the functions for removing forward declarations in dgnc_driver.c Signed-off-by:
Daeseok Youn <daeseok.youn@gmail.com> Signed-off-by:
-
Greg Kroah-Hartman authored
This reverts commit d4ef1313. It's wrong :( Reported-by:
-
Sergio Paracuellos authored
This patch replace actual mask stuff using BIT macros with or operators to make use of GENMASK macro which simplifies code clearity and readibility. It applies for defines included in hfa384x.h header file. Signed-off-by:
Sergio Paracuellos <sergio.paracuellos@gmail.com> Signed-off-by:
-
