Since this works fine with krbd, it sounds like the bug may be in
librbd. Could you install debug symbols (the librbd1-dbg package) and
when this occurs, attach to the qemu process with gdb and get a
backtrace of all threads (there will be a lot of them) via 'gdb -p $pid'
and in gdb 'thread apply a
Here are gdb oupout with librbd1-dbg and librados2-dbg.
-
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
0x7ff8cf8dddff in ppoll () from /lib/x86_64-linux-gnu/libc.
Can you run 'perf top' against just the QEMU process? There was an
email chain from nearly a year ago about tcmalloc causing extremely high
'_raw_spin_lock' calls under high IOPS scenarios.
--
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to Q
Here are 'perf top -p `pgrep qemu` -a` output; I met tcmalloc problem on the
osd host and fix it with a larger TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES. The
perf top of tcmalloc problem is a little bit different of my problem.
--
Sa
Some more test I have do:
1. running qemu with TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES 256MB, still got
problem
2. prevent cpu go into C3 and C6 state, still got problem
3. running qemu with aio=native, still got problem
--
You received this bug notification because you are a member of qemu-
devel
Any chance you can re-test with a more recent kernel on the hypervisor
host? If the spin-lock was coming from user-space, I would expect
futex_wait_setup and futex_wake to be much higher.
--
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to QEMU