02.02.2016 10:07, Xavier Hernandez написав:
Could it be memory used by Valgrind itself to track glusterfs' memory usage ?Could you repeat the test without Valgrind and see if the memory usage after dropping caches returns to low values ?
Yup. Here are the results: === pf@server:~ » ps aux | grep volumeroot 19412 14.4 10.0 5416964 4971692 ? Ssl 10:15 36:32 /usr/sbin/glusterfs --volfile-server=server.example.com --volfile-id=volume /mnt/volume
pf@server:~ » echo 2 | sudo tee /proc/sys/vm/drop_caches 2 pf@server:~ » ps aux | grep volumeroot 19412 13.6 3.5 2336772 1740804 ? Ssl 10:15 36:53 /usr/sbin/glusterfs --volfile-server=server.example.com --volfile-id=volume /mnt/volume
===Dropped from 4.9G to 1.7G. But fresh mount consumes only 25M (megabytes):
===root 23347 0.7 0.0 698376 25124 ? Ssl 14:49 0:00 /usr/sbin/glusterfs --volfile-server=server.example.com --volfile-id=volume /mnt/volume
=== Why?
Examining statedump shows only the following snippet with high "size" value: === [mount/fuse.fuse - usage-type gf_fuse_mt_iov_base memusage] size=4234592647 num_allocs=1 max_size=4294935223 max_num_allocs=3 total_allocs=4186991 === Another leak?Grepping "gf_fuse_mt_iov_base" on GlusterFS source tree shows the following:=== $ grep -Rn gf_fuse_mt_iov_basexlators/mount/fuse/src/fuse-mem-types.h:20: gf_fuse_mt_iov_base,xlators/mount/fuse/src/fuse-bridge.c:4887: gf_fuse_mt_iov_base); === fuse-bridge.c snippet: === /* Add extra 128 byte to the first iov so that it can* accommodate "ordinary" non-write requests. It's not * guaranteed to be big enough, as SETXATTR and namespace * operations with very long names may grow behind it, * but it's good enough in most cases (and we can handle* rest via realloc). */ iov_in[0].iov_base = GF_CALLOC (1, msg0_size, gf_fuse_mt_iov_base); === Probably, some freeing missing for iov_base?This is not a real memory leak. It's only a bad accounting of memory. Note that num_allocs is 1. If you look at libglusterfs/src/mem-pool.c, you will see this: /* TBD: it would be nice to adjust the memory accounting info here, * but calling gf_mem_set_acct_info here is wrong because it bumps * up counts as though this is a new allocation - which it's not. * The consequence of doing nothing here is only that the sizes will be * wrong, but at least the counts won't be. uint32_t type = 0; xlator_t *xl = NULL; type = header->type; xl = (xlator_t *) header->xlator; gf_mem_set_acct_info (xl, &new_ptr, size, type, NULL); */ This means that memory reallocs are not correctly accounted, so the tracked size is incorrect (note that fuse_thread_proc() calls GF_REALLOC() in some cases). There are two problems here: 1. The memory is allocated with a given size S1, then reallocated with a size S2 (S2 > S1), but not accounted, so the memory accounting system still thinks that the allocated size is S1. When memory is freed, S2 is substracted from the total size used. With enough allocs/reallocs/frees, this value becomes negative. 2. statedump shows the 64-bit 'size' field representing the total memory used by a given type as an unsigned 32-bit value, loosing some information. Xavi[1] https://gist.github.com/f0cf98e8bff0c13ea38f [2] https://gist.github.com/87baa0a778ba54f0f7f7 [3] https://gist.github.com/7013b493d19c8c5fffae [4] https://gist.github.com/cc38155b57e68d7e86d5 [5] https://gist.github.com/6a24000c77760a97976a [6] https://gist.github.com/74bd7a9f734c2fd21c33 On понеділок, 1 лютого 2016 р. 14:24:22 EET Soumya Koduri wrote:On 02/01/2016 01:39 PM, Oleksandr Natalenko wrote:Wait. It seems to be my bad.Before unmounting I do drop_caches (2), and glusterfs process CPU usagegoes to 100% for a while. I haven't waited for it to drop to 0%, andinstead perform unmount. It seems glusterfs is purging inodes and that'swhy it uses 100% of CPU. I've re-tested it, waiting for CPU usage to become normal, and got no leaks. Will verify this once again and report more.BTW, if that works, how could I limit inode cache for FUSE client? I do not want it to go beyond 1G, for example, even if I have 48G of RAM onmy server.Its hard-coded for now. For fuse the lru limit (of the inodes which arenot active) is (32*1024).One of the ways to address this (which we were discussing earlier) is to have an option to configure inode cache limit. If that sounds good, we can then check on if it has to be global/volume-level, client/server/both.Thanks, Soumya01.02.2016 09:54, Soumya Koduri написав:On 01/31/2016 03:05 PM, Oleksandr Natalenko wrote:Unfortunately, this patch doesn't help. RAM usage on "find" finish is ~9G. Here is statedump before drop_caches: https://gist.github.com/ fc1647de0982ab447e20[mount/fuse.fuse - usage-type gf_common_mt_inode_ctx memusage] size=706766688 num_allocs=2454051And after drop_caches: https://gist.github.com/5eab63bc13f78787ed19[mount/fuse.fuse - usage-type gf_common_mt_inode_ctx memusage] size=550996416 num_allocs=1913182 There isn't much significant drop in inode contexts. One of thereasons could be because of dentrys holding a refcount on the inodeswhich shall result in inodes not getting purged even after fuse_forget. pool-name=fuse:dentry_t hot-count=32761if '32761' is the current active dentry count, it still doesn't seemto match up to inode count. Thanks, SoumyaAnd here is Valgrind output: https://gist.github.com/2490aeac448320d98596 On субота, 30 січня 2016 р. 22:56:37 EET Xavier Hernandez wrote:There's another inode leak caused by an incorrect counting of lookups on directory reads. Here's a patch that solves the problem for 3.7: http://review.gluster.org/13324 Hopefully with this patch the memory leaks should disapear. Xavi On 29.01.2016 19:09, Oleksandr Natalenko wrote:Here is intermediate summary of current memoryleaks in FUSE clientinvestigation. I use GlusterFS v3.7.6release with the following patches:===Kaleb S KEITHLEY (1):fuse: use-after-free fix in fuse-bridge, revisitedPranith Kumar K(1):mount/fuse: Fix use-after-free crashSoumya Koduri (3):gfapi: Fix inode nlookup countsinode: Retire the inodes from the lrulist in inode_table_destroyupcall: free the xdr* allocations ===With those patches we got API leaks fixed (I hope, brief tests showthat) andgot rid of "kernel notifier loop terminated" message.Nevertheless, FUSEclient still leaks. I have several testvolumes with several million of small files (100K…2M inaverage). Ido 2 types of FUSE client testing:1) find /mnt/volume -type d 2)rsync -av -H /mnt/source_volume/* /mnt/target_volume/And mostup-to-date results are shown below:=== find /mnt/volume -type d===Memory consumption: ~4GStatedump:https://gist.github.com/10cde83c63f1b4f1dd7aValgrind:https://gist.github.com/097afb01ebb2c5e9e78dI guess,fuse-bridge/fuse-resolve. related.=== rsync -av -H/mnt/source_volume/* /mnt/target_volume/ ===Memory consumption:~3.3...4GStatedump (target volume):https://gist.github.com/31e43110eaa4da663435Valgrind (target volume):https://gist.github.com/f8e0151a6878cacc9b1aI guess,DHT-related.Give me more patches to test :)._______________________________________________Gluster-devel mailinglistGluster-devel@xxxxxxxxxxxhttp://www.gluster.org/mailman/listinfo/gluster-devel_______________________________________________ Gluster-devel mailing list Gluster-devel@xxxxxxxxxxx http://www.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________ Gluster-users mailing list Gluster-users@xxxxxxxxxxx http://www.gluster.org/mailman/listinfo/gluster-users
