- 31 Mar, 2016 40 commits
-
-
Vitaly Fertman authored
after LU-4300 enqueue does not ELC anymore, however if enqueue is agressive (ls -la of a large dir) we may exceed lru-resize limit quickly because LRUR shrinker and recalc are called not so often. ELC is to be restored in enqueue. ELC also should check for the lock weight, in addition to LRUR. ELC can also keep "skipped" locks, i.e. once checked for the weight and left in the lru - let LRUR take care about them later. LRUR is to be left untouched, no weight logic, otherwise LU-5727 appears and OPEN locks do not get canceled. Xyratex-bug-id: MRP-2550 Signed-off-by:
Vitaly Fertman <vitaly.fertman@seagate.com> Reviewed-on: http://review.whamcloud.com/14342 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6390Reviewed-by:
Jinshan Xiong <jinshan.xiong@intel.com> Reviewed-by:
Niu Yawei <yawei.niu@intel.com> Reviewed-by:
Andreas Dilger <andreas.dilger@intel.com> Signed-off-by:
Oleg Drokin <green@linuxhacker.ru> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Niu Yawei authored
The changes to ldlm_cancel_aged_policy() introduced from LU-4300 was incorrect. This patch revert this part of changes. Signed-off-by:
Bobi Jam <bobijam@gmail.com> Reviewed-by:
-
Jinshan Xiong authored
Change the policy of ELC to pick locks that have no dirty pages, no page in writeback state, and no locked pages. Signed-off-by:
-
Oleg Drokin authored
vvp_io_setattr_lock is the only user that sets it, but it's never checked anywhere, so could go away. Also get rid of enum ccc_setattr_lock_type that becomes unused. Signed-off-by:
-
John Hammond authored
Move several declarations between llite_internal.h and vvp_internal.h with the goal of reserving the latter header for functions that pertain to vvp_{device,object,page,...}. Signed-off-by:John L. Hammond <john.hammond@intel.com> Signed-off-by:
Bobi Jam <bobijam@hotmail.com> Reviewed-by:
James Simmons <uja.ornl@gmail.com> Signed-off-by:
-
Oleg Drokin authored
ll_dirent_type_get is only used in one place in llite/dir.c, so move it there. Signed-off-by:
-
John Hammond authored
Merge their contents into vvp_global_{init,fini}() and {init,exit}_lustre_lite(). Rename ccc_inode_fini_* to cl_inode_fini_*. Signed-off-by: -
John Hammond authored
struct ccc_thread_info is used in the VVP parts of llite so rename it struct vvp_thread_info. Rename supporting functions accordingly. Move init code from lcommon_cl.c to vvp_dev.c Signed-off-by:
-
John Hammond authored
struct vvp_thread_info is used in the non-VVP parts of llite so rename it struct ll_thread_info. Rename supporting functions accordingly. Signed-off-by:
-
John Hammond authored
And move the definition from vvp_internal.h to llite_internal.h. Signed-off-by:
-
John L. Hammond authored
Rename struct ccc_req to struct vvp_req and move related functions from lustre/llite/lcommon_cl.c to the new file lustre/llite/vvp_req.c. Signed-off-by:
-
John L. Hammond authored
Move all vvp_io related functions from lustre/llite/lcommon_cl.c to the sole file where they are used lustre/llite/vvp_io.c. Signed-off-by:
-
John L. Hammond authored
Rename members of struct vvp_io to used to start with vui_ rather than cui_. Rename several instances of struct vvp_io * from cio to vio. Signed-off-by:
Lai Siyao <lai.siyao@intel.com> Reviewed-by:
-
John L. Hammond authored
Move the contents of struct vvp_io into struct ccc_io, delete the former, and rename the latter to struct vvp_io. Rename various ccc_io related functions to use vvp rather than ccc. Signed-off-by:
-
John L. Hammond authored
Ever since removal of the the unused function ll_ra_read_get(), the struct ll_ra_read members lrr_reader and lrr_linkage and the struct ll_readahead_state member ras_read_beads unnecessary so remove them. In struct vvp_io replace the struct ll_ra_read cui_bead member with cui_ra_start and cui_ra_count. Signed-off-by:
-
John L. Hammond authored
Rename struct ccc_lock to struct vvp_lock and merge the CCC lock methods into the VVP lock methods. Signed-off-by:
-
John L. Hammond authored
Rename struct ccc_page to struct vvp_page and remove obsolete CCC page methods. Signed-off-by:
-
John L. Hammond authored
Rename struct ccc_object to struct vvp_object and merge the CCC object methods into the VVP object methods. Signed-off-by:
Dmitry Eremin <dmitry.eremin@intel.com> Signed-off-by:
-
John L. Hammond authored
Rename struct ccc_device to struct vvp_device and merge the CCC device methods into the VVP device methods. Signed-off-by:
-
John L. Hammond authored
Move the definition of struct cl_client_cache to lustre/include/cl_object.h and move the rest of lustre/include/lclient.h in to lustre/llite/vvp_internal.h. Signed-off-by:
-
Li Dongyang authored
When we are doing a partial IO on both first and last page, the logic currently only call cl_page_clip on the first page, which will end up with a incorrect i_size. Signed-off-by:
Li Dongyang <dongyang.li@anu.edu.au> Reviewed-on: http://review.whamcloud.com/11630 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5552Reviewed-by:
Ian Costello <costello.ian@gmail.com> Reviewed-by:
Li Xi <pkuelelixi@gmail.com> Reviewed-by:
-
Jinshan Xiong authored
Writing thread already locked page #1, and then wait for the Writeback bit of page #2; Ptlrpc thread is composing a write RPC, so it sets Writeback on page #2 and tries to lock page #1 to make it ready. Deadlocked. Signed-off-by:
wangdi <di.wang@intel.com> Signed-off-by:
-
Jinshan Xiong authored
The original code was wrong which clipped page incorrectly for partial pages started with zero. Signed-off-by:
-
Bobi Jam authored
Update comments to reflect current cl_lock situations. Signed-off-by:
Bobi Jam <bobijam.xu@intel.com> Reviewed-on: http://review.whamcloud.com/13137 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6046Reviewed-by:
-
Jinshan Xiong authored
In this patch, the cl_lock cache is eliminated. cl_lock is turned into a cacheless data container for the requirements of locks to complete the IO. cl_lock is created before I/O starts and destroyed when the I/O is complete. cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached to cl_lock at OSC layer. LDLM lock is still cacheable. Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue() methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock consists of multiple sub cl_locks, each sub locks will be enqueued correspondingly. At OSC layer, the lock enqueue request will tend to reuse cached LDLM lock; otherwise a new LDLM lock will have to be requested from OST side. cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel() method will be called for each layer to release the resource held by this lock. At OSC layer, the reference count of LDLM lock, which is held at clo_enqueue time, is released. LDLM lock can only be canceled if there is no cl_lock using it. Signed-off-by:
-
Jinshan Xiong authored
To make cl_sync_io interfaces not just wait for pages, but to be a generic synchronization mechanism. Also remove cl_io_cancel that became not used. Signed-off-by:
-
Oleg Drokin authored
This was a compat code from the time it had ia_attr_flags. Instead convert all the cryptic callers that did ((struct ll_iattr *)&op_data->op_attr)->ia_attr_flags into direct access to op_data->op_attr_flags This also makes lustre/include/linux/obd.h not needed anymore, so remove it. Signed-off-by:
-
John L. Hammond authored
In llite remove the wrapper functions and macros: cl_i2info() cl_i2sbi() cl_iattr2fd() cl_inode_info cl_inode_mode() cl_inode_{a,m,c}time() cl_isize_{read,write,write_nolock}() Signed-off-by: -
John L. Hammond authored
Remove the definition of struct client_obd_lock and the functions client_obd_list_{init,lock,unlock,done}(). Use spinlock_t for the cl_{loi,lru}_list_lock members of struct client_obd and call spin_{lock,unlock}() directly. Signed-off-by: -
John L. Hammond authored
In struct lmv_obd rename the init_mutex member to lmv_init_mutex. Remove the compat macros lmv_init_{lock,unlock}() and use mutex_{lock,unlock}(&lmv->lmv_init_mutex) instead. Signed-off-by:Bob Glossman <bob.glossman@intel.com> Reviewed-by:
-
John L. Hammond authored
In struct ll_inode_info remove the struct ost_lvb lli_lvb member and replace it with obd_time lli_{a,m,c}time. Rename ll_merge_lvb() to ll_merge_attr(). Remove cl_merge_lvb() and replace calls to it with calls to ll_merge_attr(). Signed-off-by: -
Jinshan Xiong authored
It used to check each page in the readahead window is covered by a lock underneath, now cpo_page_is_under_lock() provides @max_index to help decide the maximum ra window. @max_index can be modified by OSC to extend the maximum lock region, to align stripe boundary at LOV, and to make sure the readahead region at least covers read region at LLITE layer. After this is done, usually readahead code calls cpo_page_is_under_lock() for each stripe. Signed-off-by:
-
Jinshan Xiong authored
>From now on, cl_page becomes one to one mapping of vmpage. Signed-off-by:
-
Jinshan Xiong authored
Use weigh_ast to decide if a lock covers any pages. In recovery, weigh_ast will be used to decide if a DLM read lock covers any locked pages, or it will be canceled instead being recovered. The problem with the original implementation is that it attached each osc_page to an osc_lock also changed lock state to add every pages for readahead. Signed-off-by:
-
Jinshan Xiong authored
in ll_write_end(), instead of adding the page into writeback cache directly, it will be held in a page list. After enough pages have been collected, issue them all with cio_commit_async(). Signed-off-by:
-
Jinshan Xiong authored
This change adds support for a single preallocated cl_env per CPU which can be used in circumstances where reschedule is not possible. Currently this interface is only used by the ll_releasepage function. Signed-off-by:
Prakash Surya <surya1@llnl.gov> Reviewed-on: http://review.whamcloud.com/8174 Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-3321Reviewed-by:
-
Jinshan Xiong authored
Move radix tree to osc layer to for performance improvement. Signed-off-by:
-
Jinshan Xiong authored
This way we can drop it async. Signed-off-by:
-
Jinshan Xiong authored
Add and discard pages from LRU in batch. Signed-off-by:
-
Jinshan Xiong authored
This reverts a patch that was merged before lustre client was introduced into the stagign tree, so it's not in the history. The performance dropped a lot when memory reclaim process kicked in as ll_releasepage() was called to destroy lustre pages. It turned out that big overhead to allocate cl_env and keys on the fly so we have to revert this patch. The original problem for the reverted patch would be solved in a follow on patch instead. Signed-off-by:
-
