Hide Forgot
Reported by Mark Mielke on gluster-devel: Had a client crash with bonnie++ running on one client, and the other client listing a directory that was removed by Bonnie++ before or during the list request: frame : type(1) op(GETXATTR) patchset: git://git.sv.gnu.org/gluster.git signal received: 11 time of crash: 2009-09-07 02:31:30 configuration details: argp 1 backtrace 1 bdb->cursor->get 1 db.h 1 dlfcn 1 fdatasync 1 libpthread 1 llistxattr 1 setfsid 1 spinlock 1 epoll.h 1 xattr.h 1 st_atim.tv_nsec 1 package-string: glusterfs 2.1.0git /lib64/libc.so.6[0x3133a33370] /opt/glusterfs/lib/libglusterfs.so.0(dict_foreach+0x18)[0x7f66c043c858] /opt/glusterfs/lib/glusterfs/2.1.0git/xlator/cluster/replicate.so(__filter_xattrs+0x2a)[0x7f66bf771d3a] /opt/glusterfs/lib/glusterfs/2.1.0git/xlator/cluster/replicate.so(afr_getxattr_cbk+0x54)[0x7f66bf771e04] /opt/glusterfs/lib/glusterfs/2.1.0git/xlator/protocol/client.so(client_getxattr_cbk+0x145)[0x7f66bf99f9a5] /opt/glusterfs/lib/glusterfs/2.1.0git/xlator/protocol/client.so(protocol_client_pollin+0xca)[0x7f66bf996dba] /opt/glusterfs/lib/glusterfs/2.1.0git/xlator/protocol/client.so(notify+0xe8)[0x7f66bf9a0ae8] /opt/glusterfs/lib/libglusterfs.so.0(xlator_notify+0x43)[0x7f66c04441c3] /opt/glusterfs/lib/glusterfs/2.1.0git/transport/socket.so(socket_event_handler+0xc8)[0x7f66be318be8] /opt/glusterfs/lib/libglusterfs.so.0[0x7f66c045dd5d] /opt/glusterfs/sbin/glusterfs(main+0x75d)[0x403bbd] /lib64/libc.so.6(__libc_start_main+0xfd)[0x3133a1ea2d] /opt/glusterfs/sbin/glusterfs[0x402579] This is for a 3-node cluster/replicate cluster. Hope the info helps...
Here is the backtrace from the core dump - sorry, no debug symbols enabled although I will try to reproduce once I recompile with debugging enabled. #0 dict_foreach (dict=0x0, fn=0x7f66bf770be0 <__gather_xattr_keys>, data=0x7ffff50bccf0) at dict.c:1182 #1 0x00007f66bf771d3a in __filter_xattrs (dict=0x0) at afr-inode-read.c:586 #2 0x00007f66bf771e04 in afr_getxattr_cbk (frame=0x7f66b00022e0, cookie=0x7f66bf770be0, this=<value optimized out>, op_ret=-1, op_errno=<value optimized out>, dict=<value optimized out>) at afr-inode-read.c:645 #3 0x00007f66bf99f9a5 in client_getxattr_cbk (frame=0x23ca860, hdr=<value optimized out>, hdrlen=<value optimized out>, iobuf=<value optimized out>) at client-protocol.c:4950 #4 0x00007f66bf996dba in protocol_client_pollin (this=0x23bec00, trans=0x23c3a80) at client-protocol.c:6082 #5 0x00007f66bf9a0ae8 in notify (this=0x0, event=<value optimized out>, data=0x23c3a80) at client-protocol.c:6125 #6 0x00007f66c04441c3 in xlator_notify (xl=0x23bec00, event=2, data=0x23c3a80) at xlator.c:930 #7 0x00007f66be318be8 in socket_event_handler (fd=<value optimized out>, idx=3, data=0x23c3a80, poll_in=1, poll_out=0, poll_err=<value optimized out>) at socket.c:813 #8 0x00007f66c045dd5d in event_dispatch_epoll_handler (i=<value optimized out>, events=<value optimized out>, event_pool=<value optimized out>) at event.c:804 #9 event_dispatch_epoll (i=<value optimized out>, events=<value optimized out>, event_pool=<value optimized out>) at event.c:867 #10 0x0000000000403bbd in main (argc=<value optimized out>, argv=<value optimized out>) at glusterfsd.c:1294
I can reproduce it fairly quickly now - Three-node cluster/replication cluster. The servers each have storage/posix, features/locks, performance/io-threads(4), protocol/server(tcp). The clients each have 3 x protocol/client(tcp) (1 for each server), cluster/replicate (read-subvolume each set to themselves), performance/io-threads(4), performance/write-behind(1MB), performance/io-cache(256MB). The clients and servers are the same machines, connected via 100BaseTX network. On the first client/server, I run: $ cd /gluster/volume $ perl -e 'while (1) { open(f, "> abc"); close(f); unlink("abc") }' ... let it run ... On the third client/server, I run: $ cd /gluster/volume $ while true; do /bin/ls; done ... let it run ... On the third client/server, this eventually (15 seconds?) results in: ls: cannot open directory .: Transport endpoint is not connected The /core.N file has identical contents to my earlier comment, with the failure being listed as line 1182 of dict.c. The mount needs to be unmounting and remounted to clear this condition. I think this is enough info for you to reproduce, so unless you cannot reproduce the problem - I'm not going to figure out how to re-install with debugging yet.
More information: I simplified the client configuration to take out all the performance/ modules. Not it only has the three protocol/client entries, and the cluster/replication entry. I also adjusted the Perl script to increase the frequency of errors: [markm@wcarh033]/gluster/tools/markm% perl -e 'while (1) { for (0 .. 9) { open(f, "> $_"); close(f) } for (0 .. 9) { unlink($_) } }' By creating 10 and then removing 10, it makes the /bin/ls output more chance of having a failure than with just one. Also, I found that my: while true; do /bin/ls; done Does not actually work. For some reason, the continuous execution causes it to *succeed* continuously. No errors. But, as soon as I start typing 'ls' a few times on the command line randomly instead, I start to get: [root@wcarh035 markm]# ls ls: cannot access 1: No such file or directory ls: cannot access 2: No such file or directory ls: cannot access 3: No such file or directory 1 2 3 4 5 6 7 8 9 abc And then: [root@wcarh035 markm]# ls ls: cannot open directory .: Transport endpoint is not connected
Mark, Thank you for the excellent diagnostics. A patch to fix this is under review: http://patches.gluster.com/patch/1259/
Oops, typed the wrong bug number for the previous patch. http://patches.gluster.com/patch/1261/
PATCH: http://patches.gluster.com/patch/1263 in master (cluster/afr: Check op_ret in afr_getxattr_cbk before accessing dict)
PATCH: http://patches.gluster.com/patch/1264 in release-2.0 (cluster/afr: Check op_ret in afr_getxattr_cbk before accessing dict)