Red Hat Bugzilla – Bug 247464
soft lockup in cifs mountpoint leads to unresponsive server
Last modified: 2008-01-15 06:24:07 EST
Description of problem:
soft lockup in cifs module when doing "stat", as shown by :
kernel: BUG: soft lockup detected on CPU#1!
kernel: [<c0447f3f>] softlockup_tick+0x98/0xa6
kernel: [<c042d138>] update_process_times+0x39/0x5c
kernel: [<c04176f0>] smp_apic_timer_interrupt+0x5c/0x64
kernel: [<c04049bf>] apic_timer_interrupt+0x1f/0x24
kernel: [<c0470f5e>] generic_fillattr+0x76/0xa4
kernel: [<f8d9cce8>] cifs_getattr+0x1e/0x2b [cifs]
kernel: [<f8d9ccca>] cifs_getattr+0x0/0x2b [cifs]
kernel: [<c047132e>] vfs_getattr+0x40/0x9b
kernel: [<c0471445>] vfs_stat_fd+0x2a/0x3c
kernel: [<c04714e4>] sys_stat64+0xf/0x23
kernel: [<c04dec11>] copy_to_user+0x31/0x48
kernel: [<c0403eff>] syscall_call+0x7/0xb
Version-Release number of selected component (if applicable):
regularly (4 cases reported in the last 2 days from a universe of 336 boxes)
Steps to Reproduce:
1. create a read only mountpoint to a samba server
2. run 2 processes (better more) in an SMP server with a PAE enabled kernel that
stat files in that mount point
3. wait for it to lock (ping will work, but won't be able to ssh anymore)
system locks and requires a manual reboot (not even sysrq will respond in the
in the server the smb process is still up and connected from the defunct server,
including oplocks to the file that was being stat as exclusive.
the TCP/IP stack of the defunct server seems to be operative, but no process
that requires a fork/exec would work (rsync/sshd) and will hung TCP connections
after the 3 way handshake is completed.
probably already fixed upstream in version of cifs 1.48 as shown by the linked
bug, with the commit linked in URL
There's a pending update of the CIFS code already slated for inclusion into 5.1.
If you have a non-critical place to do so, could you test the kernels on my
people page and see if they help this issue:
No response to my query from July. I believe this is already fixed in the 5.1
release. Please reopen if that is not the case.