- 03 Feb, 2017 40 commits
-
-
James Simmons authored
Remove the inline function socklnd_init_msg. Its only used by the kernel code so no point keeping it in an UAPI header. Signed-off-by:
James Simmons <uja.ornl@yahoo.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6142 Reviewed-on: https://review.whamcloud.com/18506Reviewed-by:
Dmitry Eremin <dmitry.eremin@intel.com> Reviewed-by:
Doug Oucharek <doug.s.oucharek@intel.com> Reviewed-by:
John L. Hammond <john.hammond@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>
-
Liang Zhen authored
ksocklnd reaper thread always tries to close the connection for the first timedout zero-copy TX. This is wrong if this connection is already being closed, because the reaper will see the same TX again and again and cannot find out other timedout zero-copy TXs and close connections for them. Signed-off-by:
Liang Zhen <liang.zhen@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8867 Reviewed-on: https://review.whamcloud.com/23973Reviewed-by:
-
John L. Hammond authored
In mdc_close() if ptlrpc_request_pack() fails then set req to NULL so that an already freed request is not returned in *request. Signed-off-by:
Patrick Farrell <paf@cray.com> Reviewed-by:
Andreas Dilger <andreas.dilger@intel.com> Signed-off-by:
-
Patrick Farrell authored
An earlier commit accidentally changed handling of IT_OPEN, making it take the MDS_INODELOCK_UPDATE bits lock instead of MDS_INODELOCK_LOOKUP. This does not cause any known bugs. Signed-off-by:
Lai Siyao <lai.siyao@intel.com> Reviewed-by:
-
Steve Guminski authored
This patch makes no functional changes. Struct initializers in the libcfs directory that use C89 or GCC-only syntax are updated to C99 syntax. The C99 syntax prevents incorrect initialization if values are accidently placed in the wrong position, allows changes in the struct definition, and clears any members that are not given an explicit value. The following struct initializers have been updated: libcfs/include/libcfs/libcfs_crypto.h: static struct cfs_crypto_hash_type hash_types[] Signed-off-by:Steve Guminski <stephenx.guminski@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6210 Reviewed-on: https://review.whamcloud.com/23332Reviewed-by:
Frank Zago <fzago@cray.com> Reviewed-by:
Nathaniel Clark <nathaniel.l.clark@intel.com> Reviewed-by:
-
Dmitry Eremin authored
Don't treat unability to set CPU partition affinity as error. Improve those warning messages. Signed-off-by:
-
John L. Hammond authored
None of the obd_notify() handlers listen for the OBD_NOTIFY_CREATE event, so remove it and its sole use in lov_add_target(). Signed-off-by:
Ben Evans <bevans@cray.com> Reviewed-by:
-
Vladimir Saveliev authored
ptlrpc_import_delay_req() refuses to delay blocking asts when import is not in LUSTRE_IMP_FULL yet. That leads to client eviction assuming that it failed to respond. Allow delays for blocking asts being resent. Signed-off-by:
Vladimir Saveliev <vladimir.saveliev@seagate.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8351 Seagate-bug-id: MRP-3500 Reviewed-on: https://review.whamcloud.com/21065Reviewed-by:
Bobi Jam <bobijam@hotmail.com> Reviewed-by:
-
Steve Guminski authored
The patch updates the prototype in osc_internal.h to match the enums used in the declaration. The osc_match_base declaration in lustre/osc/osc_request.c uses enums for stricter checking on the type and mode parameters: int osc_match_base(struct obd_export *exp, ... --> enum ldlm_type type, union ldlm_policy_data *policy, --> enum ldlm_mode mode, ... int unref) The prototype in lustre/osc/osc_internal.h instead used unsigned ints: int osc_match_base(struct obd_export *exp, ... --> __u32 type, union ldlm_policy_data *policy, --> __u32 mode, ... int unref); Signed-off-by:Bob Glossman <bob.glossman@intel.com> Reviewed-by:
-
Niu Yawei authored
reply_out_callback() should call ptlrpc_schedule_difficult_reply() to finalize the rs if it's already not on uncommitted list, otherwise, the rs and the export held by rs could be leaked: - target_send_reply() sends a difficult reply before the transaction committed, the reply is linked to scp_rep_active; - export gets disconnected by umount or whatever reason, server_disconnect_export() is called to complete all outstanding replies, which will calls into ptlrpc_handle_rs() to dispose of the rs, so the rs is removed from the uncommitted list and LNetMDUnlink() is called to unlink the reply buffer and generate an unlink event; - reply_out_callback() is called to process above unlink event, ptlrpc_schedule_difficult_reply() is supposed to be called to dispose of the rs finally. However, it could be skipped because of following flawed code snippet: if (!rs->rs_no_ack || rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed) ptlrpc_schedule_difficult_reply(rs); The intention of above code is: if rs_no_ack is true (COS enabled), and transaction is not committed, we should rely on commit callback to release the rs. However, it overlooked the situation that rs could have been removed from the uncommitted list by disconnecting export. Signed-off-by:Niu Yawei <yawei.niu@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-7903 Reviewed-on: http://review.whamcloud.com/22696Reviewed-by:
-
James Simmons authored
Replace the ulong_ptr_t and long_ptr_t with standard kernel types. Signed-off-by:
-
Rahul Deshmukh authored
There exists timing race between umount and other thread which will increment the reference count on mnt e.g. getattr. If umount thread lose the race then umount fails with EBUSY error. To avoid this timed wait is added so that umount thread will wait for user to decrement the mnt reference count. Signed-off-by:
Rahul Deshmukh <rahul.deshmukh@seagate.com> Signed-off-by:
Lokesh Nagappa Jaliminche <lokesh.jaliminche@seagate.com> Signed-off-by:
Jian Yu <jian.yu@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-1882 Seagate-bug-id: MRP-1192 Reviewed-on: http://review.whamcloud.com/20061Reviewed-by:
-
Bobi Jam authored
So that debug log only contains relevant messages for debugging purpose. Signed-off-by:
Bobi Jam <bobijam.xu@intel.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8413 Reviewed-on: http://review.whamcloud.com/22753Reviewed-by:
Fan Yong <fan.yong@intel.com> Reviewed-by:
-
James Simmons authored
In the case of interval_tree.h only interval_set() uses LASSERT which is removed in this patch and interval_set() instead reports a real error. The header libcfs.h for interval_tree.h is not needed anymore so we can just use the standard linux kernel headers instead.h Signed-off-by:
-
Bobi Jam authored
lov_device::ld_target[ost_idx] could be NULL if the OST target is not filled in lov_device::ld_lov::lov_tgt_desc[ost_idx] yet. Signed-off-by:
Jinshan Xiong <jinshan.xiong@intel.com> Reviewed-by:
-
Dmitry Eremin authored
Change default value of CPT pattern and make it match NUMA topology Signed-off-by:
Olaf Weber <olaf@sgi.com> Reviewed-by:
-
Dmitry Eremin authored
The function cfs_cpt_table_create_pattern() alters the string passed to it. Currently we are passing in the module parameter string cpu_pattern which is incorrect. Instead lets duplicate the module parameter string and pass that to the function cfs_cpt_table_create_pattern(). Signed-off-by:
-
Jinshan Xiong authored
OSC has to make sure that it won't issue write RPCs with too many chunks otherwise it will casue ZFS to create transactions much bigger than DMU_MAX_ACCESS in size, which will end up with write failure. Signed-off-by:
-
Niu Yawei authored
Sync write should update m/ctime promptly, otherwise, stale m/ctime could be updated on the OST object by the sync write RPC. Signed-off-by:
-
Fan Yong authored
The 'fld_read_server' uses 'RMF_GENERIC_DATA' to hold the 'FLD_QUERY' RPC reply that is composed of 'struct lu_seq_range_array'. But there is not registered swabber function for 'RMF_GENERIC_DATA'. So the RPC peers need to handle the RPC reply with fixed little-endian format. In theory, we can define new structure with some swabber registered to handle the 'FLD_QUERY' RPC reply result automatically. But from the implementation view, it is not easy to be done within current 'struct req_msg_field' framework. Because the sequence range array in the RPC reply is not fixed length, instead, its length depends on 'lu_seq_range' count, that is unknown when prepare the RPC buffer. Generally, for such flexible length RPC usage, there will be a field in the RPC layout to indicate the data length. But for the 'FLD_READ' RPC, we have no way to do that unless we add new length filed that will broken the on-wire RPC protocol and cause interoperability trouble with old peer. Signed-off-by:
-
Dmitry Eremin authored
Setting extended attributes permissions are properly checked with and without ACLs. In user.* namespace, only regular files and directories can have extended attributes. For sticky directories, only the owner and privileged user can write attributes. Signed-off-by:
Alex Zhuravlev <alexey.zhuravlev@intel.com> Reviewed-by:
-
Andriy Skulysh authored
Whole policy structure was zeroed twice. Once during enqueue and second time during resend or replay. Policy structure should be initialized with default values only in ldlm_lock_new(). Signed-off-by:
Andriy Skulysh <andriy.skulysh@seagate.com> Signed-off-by:
Alexander Boyko <alexander.boyko@seagate.com> Reviewed-by:
Vitaly Fertman <vitaly.fertman@seagate.com> Reviewed-by:
Mike Pershin <mike.pershin@intel.com> Reviewed-by:
-
Fan Yong authored
Originally, the logic of handling config_llog_data::cld_refcount is some confusing, it may cause the cld_refcount to be leaked or trigger "LASSERT(atomic_read(&cld->cld_refcount) > 0);" when put the reference. This patch clean related logic as following: 1) When the 'cld' is created, its reference is set as 1. 2) No need additional reference when add the 'cld' into the list 'config_llog_list'. 3) Inrease 'cld_refcount' when set lock data after mgc_enqueue() done successfully by mgc_process_log(). 4) When mgc_requeue_thread() traversals the 'config_llog_list', it needs to take additional reference on each 'cld' to avoid being freed during subsequent processing. The reference also prevents the 'cld' to be dropped from the 'config_llog_list', then the mgc_requeue_thread() can safely locate next 'cld', and then decrease the 'cld_refcount' for previous one. 5) mgc_blocking_ast() will drop the reference of 'cld_refcount' that is taken in mgc_process_log(). 6) The others need to call config_log_find() to find the 'cld' if want to access related config log data. That will increase the 'cld_refcount' to avoid being freed during accessing. The sponsor needs to call config_log_put() after using the 'cld'. 7) Other confused or redundant logic are dropped. On the other hand, the patch also enhances the protection for 'config_llog_data' flags, such as 'cld_stopping'/'cld_lostlock' as following. a) Use 'config_list_lock' (spinlock) to handle the possible parallel accessing of these flags among mgc_requeue_thread() and others config llog data visitors, such as mount/umount, blocking_ast, and so on. b) Use 'config_llog_data::cld_lock' (mutex) to pretect other parallel accessing of these flags among kinds of blockable operations, such as mount, umount, and blocking ast. The 'config_llog_data::cld_lock' is also used for protecting the sub-cld members, such as 'cld_sptlrpc'/'cld_params', and so on. Signed-off-by:
Hongchao Zhang <hongchao.zhang@intel.com> Reviewed-by:
-
Oleg Drokin authored
By forcing creates to always go via lookup we lose some important caching benefits too. Instead let's trust creates with positive cached entries. Then we have 3 possible outcomes: 1. Negative dentry - we go via atomic_open and do the create by name there. 2. Positive dentry, no contention - we just go straight to ll_intent_file_open and open by fid. 3. positive dentry, contention - by the time we reach the server, the inode is gone. We get ENOENT which is unacceptable to return from create. But since we know it's a create, we substitute it with ESTALE and VFS retries again with LOOKUP_REVAL set, we catch that in revalidate and force a lookup (same path as before this patch). Signed-off-by:
-
Lai Siyao authored
Current client doesn't check permission before updating filesystem default stripe on MGS, which isn't secure and obvious. Since we setattr on MDS first, and then set default stripe on MGS, we can just return error upon setattr failure. Now filesystem default stripe is stored in ROOT in MDT, so saving it in system config is for compatibility with old servers, this will be removed in the future. Signed-off-by:
-
Ben Evans authored
These #if 0 blocks have been in place for years. Assume they are not used and remove them Signed-off-by:
-
John L. Hammond authored
Remove the unused lmv.*.placement parameter along with supporting functions and struct members. Signed-off-by:
-
John L. Hammond authored
Remove obsolete comments about the behavior of ll_unlink() Signed-off-by:
Andrew Perepechko <andrew.perepechko@seagate.com> Reviewed-by:
-
Andreas Dilger authored
Change the order of checks for inactive OSCs in lov_prep_statfs_set() so that administratively disabled OSTs do not generate any output in "lfs df" at all, to avoid needlessly cluttering the output. Enable the lazystatfs mount option by default, so that "df" does not hang when an OST is temporarily offline. Signed-off-by:
-
Alexander Boyko authored
This patch resolves IO vs eviction race. After eviction failed export stayed at stale list, a client had IO processing and reconnected during it. A client sent brw rpc with last lock cookie and new connection. The lock with failed export was found and assert was happened. (ost_handler.c:1812:ost_prolong_lock_one()) ASSERTION( lock->l_export == opd->opd_exp ) failed: 1. Skip the lock at ldlm_handle2lock if lock export failed. 2. Validation of lock for IO was added at hpreq_check(). The lock searching is based on granted interval tree. If server doesn`t have a valid lock, it reply to client with ESTALE. Signed-off-by: -
Alex Zhuravlev authored
First of all, this is expensive procedure including a global mutex and per-bucket spinlocks. also, all the threads observed exceed will be calling lu_site_purge() and essentially serialized on that. instead we can let other threads to skip the whole procedure. Signed-off-by:
-
Jinshan Xiong authored
ignore_layout can be set for operations that layout won't be changed, typically page operations. Ignoring layout change in group lock request will confuse layout change code at LOV layer and hit assertion. Signed-off-by:
Henri Doreau <henri.doreau@cea.fr> Signed-off-by:
-
frank zago authored
The function hai_dump_data_field will do a stack buffer overrun when cat'ing /sys/fs/lustre/.../hsm/actions if an action has some data in it. hai_dump_data_field uses snprintf. But there is no check for truncation, and the value returned by snprintf is used as-is. The coordinator code calls hai_dump_data_field with 12 bytes in the buffer. The 6th byte of data is printed incompletely to make room for the terminating NUL. However snprintf still returns 2, so when hai_dump_data_field writes the final NUL, it does it outside the reserved buffer, in the 13th byte of the buffer. This stack buffer overrun hangs my VM. Fix by checking that there is enough room for the next 2 characters plus the NUL terminator. Don't print half bytes. Change the format to 02X instead of .2X, which makes more sense. Signed-off-by:
Jean-Baptiste Riaux <riaux.jb@intel.com> Reviewed-by:
-
Ulka Vaze authored
This issue is found by smatch; has been reported as- Unchecked usage of potential ERR_PTR result in lmv_hsm_req_count and lmv_hsm_req_build. Added ERR_PTR in both functions and also return value check added. Signed-off-by:
Ulka Vaze <ulka.vaze@yahoo.in> Signed-off-by:
Aditya Pandit <panditadityashreesh@yahoo.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6523 Reviewed-on: http://review.whamcloud.com/14918Reviewed-by:
-
Ann Koehler authored
The algorithm for counting freeable objects in the lu_cache shrinker does not scale with the number of cpus. The LU_SS_LRU_LEN counter for each cpu is read and summed at shrink time while holding the lu_sites_guard mutex. With a large number of cpus and low memory conditions, processes bottleneck on the mutex. This mod reduces the time spent counting by using the kernel's percpu counter functions to maintain the length of a site's lru. The summing occurs when a percpu value is incremented or decremented and a threshold is exceeded. lu_cache_shrink_count() simply returns the last such computed sum. This mod also replaces the lu_sites_guard mutex with a rw semaphore. The lock protects the lu_site list, which is modified when a file system is mounted/umounted or when the lu_site is purged. lu_cache_shrink_count simply reads data so it does not need to wait for other readers. lu_cache_shrink_scan, which actually frees the unused objects, is still serialized. Signed-off-by:
Ann Koehler <amk@cray.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-7997 Reviewed-on: http://review.whamcloud.com/19390Reviewed-by:
-
Giuseppe Di Natale authored
The target_obd debugfs file was not being generated correctly in cases where nonconsecutive MDT indices were used when generating a filesystem. Signed-off-by:
Giuseppe Di Natale <dinatale2@llnl.gov> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8100 Reviewed-on: http://review.whamcloud.com/20336Reviewed-by:
-
Jeremy Filizetti authored
Restore connect flags on failure of ptlrpc_connect_import() to prevent an LBUG due to flags mismatch. Signed-off-by:
Jeremy Filizetti <jeremy.filizetti@gmail.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-7185 Reviewed-on: http://review.whamcloud.com/16950Reviewed-by:
Sebastien Buisson <sebastien.buisson@bull.net> Reviewed-by:
-
Niu Yawei authored
Set mbits for EINPROGRESS resend in ptl_send_rpc(). Signed-off-by:
-
Jinshan Xiong authored
Read ahead currently doesn't handle 16MB RPC packets correctly by assuming the packets are a default size instead of querying the size. This work adjust the read ahead policy to issue read ahead RPC by the underlying RPC size. Signed-off-by:
Gu Zheng <gzheng@ddn.com> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-7990 Reviewed-on: http://review.whamcloud.com/19368Reviewed-by:
Li Xi <lixi@ddn.com> Reviewed-by:
-
Andreas Dilger authored
If a user process is waiting for MDS recovery during close, but the process is interrupted, the file is still closed but it prints a message on the console. Quiet the console message for -EINTR, since this is expected behaviour. Signed-off-by:
Emoly Liu <emoly.liu@intel.com> Reviewed-by:
-
