Problem: Before executing a fop in POSIX xlator it builds an internal
path based on GFID.To validate the path it call's (l)stat
system call and while .glusterfs is heavily loaded kernel takes
time to lookup inode and due to that performance drops
Solution: In this patch we followed two ways to improve the performance.
1) Keep open fd specific to first level directory(gfid[0])
in .glusterfs, it would force to kernel keep the inodes
from all those files in cache. In case of memory pressure
kernel won't uncache first level inodes. We need to open
256 fd's per brick to access the entry faster.
2) Use at based call's to access relative path to reduce
path based lookup time.
Note: To verify the patch we have executed kernel untar 100 times on 6
different clients after enabling metadata group-cache and some
other option.We were getting more than 20 percent improvement in
kenel untar after applying the patch.
Credits: Xavi Hernandez <xhernandez@redhat.com>
Change-Id: I1643e6b01ed669b2bb148d02f4e6a8e08da45343
updates: #891
Signed-off-by: Mohit Agrawal <moagrawal@redhat.com>
Following tests are done - 1 - After finishing reset-brick all the bricks should be up. 2 - Heal should be completed. 3 - Check number of entries present on brick which was reset. Change-Id: I9314bed180293a99d400d94bb8cc7ece999da29e Updates: #1144
fixes: bz#1760189 Change-Id: Iffbf8d6f4c50b8e2de8364658697bdbe96549f5d Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
fixes: #725 Change-Id: Iaaefe6f49c8193c476b987b92df6bab3e2f62601 Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
We've found perf xlators io-cache and read-ahead not adding any performance improvement. At best read-ahead is redundant due to kernel read-ahead and at worst io-cache is degrading the performance for workloads that doesn't involve re-read. Given that VFS already have both these functionalities, this patch makes these two translators turned off by default for native fuse mounts. For non-native fuse mounts like gfapi (NFS-ganesha/samba) we can have these xlators on by having custom profiles. Change-Id: Ie7535788909d4c741844473696f001274dc0bb60 Signed-off-by: Raghavendra Gowdappa <rgowdapp@redhat.com> fixes: bz#1676479
fixes: #721 Change-Id: I5333540e3c635ccf441cf1f4696e4c8986e38ea8 Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
Problem: If update size/version is not successful on the file, updates on the same stripe could lead to data corruptions if the earlier un-aligned write is not successful on all the bricks. Application won't have any knowledge of this because update size/version happens in the background. Fix: Fail fsync/flush on fds that are opened before update-size-version went bad. fixes: bz#1748836 Change-Id: I9d323eddcda703bd27d55f340c4079d76e06e492 Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
EC doesn't allow concurrent writes on overlapping areas, they are serialized. However non-overlapping writes are serviced in parallel. When a write is not aligned, EC first needs to read the entire chunk from disk, apply the modified fragment and write it again. The problem appears on sparse files because a write to an offset implicitly creates data on offsets below it (so, in some way, they are overlapping). For example, if a file is empty and we read 10 bytes from offset 10, read() will return 0 bytes. Now, if we write one byte at offset 1M and retry the same read, the system call will return 10 bytes (all containing 0's). So if we have two writes, the first one at offset 10 and the second one at offset 1M, EC will send both in parallel because they do not overlap. However, the first one will try to read missing data from the first chunk (i.e. offsets 0 to 9) to recombine the entire chunk and do the final write. This read will happen in parallel with the write to 1M. What could happen is that half of the bricks process the write before the read, and the half do the read before the write. Some bricks will return 10 bytes of data while the otherw will return 0 bytes (because the file on the brick has not been expanded yet). When EC tries to recombine the answers from the bricks, it can't, because it needs more than half consistent answers to recover the data. So this read fails with EIO error. This error is propagated to the parent write, which is aborted and EIO is returned to the application. The issue happened because EC assumed that a write to a given offset implies that offsets below it exist. This fix prevents the read of the chunk from bricks if the current size of the file is smaller than the read chunk offset. This size is correctly tracked, so this fixes the issue. Also modifying ec-stripe.t file for Test #13 within it. In this patch, if a file size is less than the offset we are writing, we fill zeros in head and tail and do not consider it strip cache miss. That actually make sense as we know what data that part holds and there is no need of reading it from bricks. Change-Id: Ic342e8c35c555b8534109e9314c9a0710b6225d6 Fixes: bz#1730715 Signed-off-by: Xavi Hernandez <xhernandez@redhat.com>
Problem:
Race:
Thread-1 Thread-2
1) Does ec_get_size_version() to perform
pre-op fxattrop as part of write-1
2) Calls ec_set_dirty_flag() in
ec_get_size_version() for write-2.
This sets dirty[] to 1
3) Completes executing
ec_prepare_update_cbk leading to
ctx->dirty[] = '1'
4) Takes LOCK(inode->lock) to check if there are
any flags and sets dirty-flag because
lock->waiting_flag is 0 now. This leads to
fxattrop to increment on-disk dirty[] to '2'
At the end of the writes the file will be marked for heal even when it doesn't need heal.
Fix:
Perform ec_set_dirty_flag() and other checks inside LOCK() to prevent dirty[] to be marked
as '1' in step 2) above
Updates bz#1593224
Change-Id: Icac2ab39c0b1e7e154387800fbededc561612865
Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
Also fixed some issues on test ec-1468261.t. Change-Id: If156f86af986d9eed13cdd1f15c5a7214cd11706 Updates: bz#1193929 Signed-off-by: Xavier Hernandez <jahernan@redhat.com>