Re: pthread_barrier_deadlock in -rc1

[Olivier Matz <olivier.matz@6wind.com>]

Hi Jianfeng,

On Wed, May 02, 2018 at 11:05:12AM +0200, Maxime Coquelin wrote:
> 
> 
> On 05/02/2018 10:50 AM, Tan, Jianfeng wrote:
> > Hi Maxime,
> > 
> > 
> > On 5/2/2018 4:34 PM, Maxime Coquelin wrote:
> > > Hi Jiangfeng,
> > > 
> > > On 05/02/2018 10:19 AM, Tan, Jianfeng wrote:
> > > > 
> > > > 
> > > > On 5/1/2018 2:46 AM, Olivier Matz wrote:
> > > > > Hi Maxime,
> > > > > 
> > > > > Le 30 avril 2018 17:45:52 GMT+02:00, Maxime Coquelin
> > > > > <maxime.coquelin@redhat.com> a écrit :
> > > > > > Hi Olivier,
> > > > > > 
> > > > > > On 04/24/2018 04:46 PM, Olivier Matz wrote:
> > > > > > > Some parts of dpdk use their own management threads. Most of the
> > > > > > time,
> > > > > > > the affinity of the thread is not properly set: it should not be
> > > > > > scheduled
> > > > > > > on the dataplane cores, because interrupting them can cause packet
> > > > > > losses.
> > > > > > > This patchset introduces a new wrapper for thread creation that does
> > > > > > > the job automatically, avoiding code duplication.
> > > > > > > 
> > > > > > > v3:
> > > > > > > * new patch: use this API in examples when relevant.
> > > > > > > * replace pthread_kill by pthread_cancel. Note that pthread_join()
> > > > > > >     is still needed.
> > > > > > > * rebase: vfio and pdump do not have control pthreads anymore, and
> > > > > > eal
> > > > > > >     has 2 new pthreads
> > > > > > > * remove all calls to snprintf/strlcpy that truncate the thread name:
> > > > > > >     all strings lengths are already < 16.
> > > > > > > 
> > > > > > > v2:
> > > > > > > * set affinity to master core if no core is off, as suggested by
> > > > > > >     Anatoly
> > > > > > > 
> > > > > > > Olivier Matz (5):
> > > > > > >     eal: use sizeof to avoid a double use of a define
> > > > > > >     eal: new function to create control threads
> > > > > > >     eal: set name when creating a control thread
> > > > > > >     eal: set affinity for control threads
> > > > > > >     examples: use new API to create control threads
> > > > > > > 
> > > > > > >    drivers/net/kni/Makefile                     |  1 +
> > > > > > >    drivers/net/kni/rte_eth_kni.c                |  3 +-
> > > > > > >    examples/tep_termination/main.c              | 16 +++----
> > > > > > >    examples/vhost/main.c                        | 19 +++-----
> > > > > > >    lib/librte_eal/bsdapp/eal/eal.c              |  4 +-
> > > > > > >    lib/librte_eal/bsdapp/eal/eal_thread.c       |  2 +-
> > > > > > >    lib/librte_eal/common/eal_common_proc.c      | 15 ++----
> > > > > > >    lib/librte_eal/common/eal_common_thread.c    | 72
> > > > > > ++++++++++++++++++++++++++++
> > > > > > > lib/librte_eal/common/include/rte_lcore.h    | 26 ++++++++++
> > > > > > >    lib/librte_eal/linuxapp/eal/eal.c            |  4 +-
> > > > > > >    lib/librte_eal/linuxapp/eal/eal_interrupts.c | 17 ++-----
> > > > > > >    lib/librte_eal/linuxapp/eal/eal_thread.c     |  2 +-
> > > > > > >    lib/librte_eal/linuxapp/eal/eal_timer.c      | 12 +----
> > > > > > >    lib/librte_eal/rte_eal_version.map           |  1 +
> > > > > > >    lib/librte_vhost/socket.c                    | 25 ++--------
> > > > > > >    15 files changed, 135 insertions(+), 84 deletions(-)
> > > > > > > 
> > > > > > I face a deadlock issue with your series, that Jianfeng patch does not
> > > > > > resolve ("eal: fix threads block on barrier"). Reverting
> > > > > > the series and
> > > > > > Jianfeng patch makes the issue to disappear.
> > > > > > 
> > > > > > I face the problem in a VM (not seen on the host):
> > > > > > # ./install/bin/testpmd -l 0,1,2 --socket-mem 1024 -n 4 --proc-type
> > > > > > auto
> > > > > > --file-prefix pg -- --portmask=3 --forward-mode=macswap
> > > > > > --port-topology=chained --disable-rss -i --rxq=1 --txq=1 --rxd=256
> > > > > > --txd=256 --nb-cores=2 --auto-start
> > > > > > EAL: Detected 3 lcore(s)
> > > > > > EAL: Detected 1 NUMA nodes
> > > > > > EAL: Auto-detected process type: PRIMARY
> > > > > > EAL: Multi-process socket /var/run/.pg_unix
> > > > > > 
> > > > > > 
> > > > > > Then it is stuck. Attaching with GDB, I get below backtrace
> > > > > > information:
> > > > > > 
> > > > > > (gdb) info threads
> > > > > >    Id   Target Id         Frame
> > > > > >    3    Thread 0x7f63e1f9f700 (LWP 8808) "rte_mp_handle"
> > > > > > 0x00007f63e2591bfd in recvmsg () at
> > > > > > ../sysdeps/unix/syscall-template.S:81
> > > > > >    2    Thread 0x7f63e179e700 (LWP 8809) "rte_mp_async"
> > > > > > pthread_barrier_wait () at
> > > > > > ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_barrier_wait.S:71
> > > > > > * 1    Thread 0x7f63e32cec00 (LWP 8807) "testpmd" pthread_barrier_wait
> > > > > > () at ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_barrier_wait.S:71
> > > > > > (gdb) bt full
> > > > > > #0  pthread_barrier_wait () at
> > > > > > ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_barrier_wait.S:71
> > > > > > No locals.
> > > > > > #1  0x0000000000520c54 in rte_ctrl_thread_create
> > > > > > (thread=thread@entry=0x7ffe5c895020, name=name@entry=0x869d86
> > > > > > "rte_mp_async", attr=attr@entry=0x0,
> > > > > > start_routine=start_routine@entry=0x521030 <async_reply_handle>,
> > > > > > arg=arg@entry=0x0)
> > > > > >      at /root/src/dpdk/lib/librte_eal/common/eal_common_thread.c:207
> > > > > >          params = 0x17b1e40
> > > > > >          lcore_id = <optimized out>
> > > > > >          cpuset = {__bits = {1, 0 <repeats 15 times>}}
> > > > > >          cpu_found = <optimized out>
> > > > > >          ret = 0
> > > > > > #2  0x00000000005220b6 in rte_mp_channel_init () at
> > > > > > /root/src/dpdk/lib/librte_eal/common/eal_common_proc.c:674
> > > > > >         path = "/var/run\000.pg_unix_*", '\000' <repeats
> > > > > > 1301 times>...
> > > > > >          dir_fd = 4
> > > > > >          mp_handle_tid = 140066969745152
> > > > > >          async_reply_handle_tid = 140066961352448
> > > > > > #3  0x000000000050c227 in rte_eal_init (argc=argc@entry=23,
> > > > > > argv=argv@entry=0x7ffe5c896378) at
> > > > > > /root/src/dpdk/lib/librte_eal/linuxapp/eal/eal.c:775
> > > > > >          i = <optimized out>
> > > > > >          fctret = 11
> > > > > >          ret = <optimized out>
> > > > > >          thread_id = 140066989861888
> > > > > >          run_once = {cnt = 1}
> > > > > >          logid = 0x17b1e00 "testpmd"
> > > > > >          cpuset = "T}\211\\\376\177", '\000' <repeats 117 times>,
> > > > > > "\020", '\000' <repeats 116 times>...
> > > > > >       thread_name =
> > > > > > "X}\211\\\376\177\000\000\226\301\036\342c\177\000"
> > > > > >          __func__ = "rte_eal_init"
> > > > > > #4  0x0000000000473214 in main (argc=23, argv=0x7ffe5c896378) at
> > > > > > /root/src/dpdk/app/test-pmd/testpmd.c:2597
> > > > > >          diag = <optimized out>
> > > > > >          port_id = <optimized out>
> > > > > >          ret = <optimized out>
> > > > > >          __func__ = "main"
> > > > > > (gdb) thread 2
> > > > > > [Switching to thread 2 (Thread 0x7f63e179e700 (LWP 8809))]
> > > > > > #0  pthread_barrier_wait () at
> > > > > > ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_barrier_wait.S:71
> > > > > > 71        cmpl    %edx, (%rdi)
> > > > > > (gdb) bt full
> > > > > > #0  pthread_barrier_wait () at
> > > > > > ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_barrier_wait.S:71
> > > > > > No locals.
> > > > > > #1  0x0000000000520777 in rte_thread_init (arg=<optimized out>) at
> > > > > > /root/src/dpdk/lib/librte_eal/common/eal_common_thread.c:156
> > > > > >          params = <optimized out>
> > > > > >          start_routine = 0x521030 <async_reply_handle>
> > > > > >          routine_arg = 0x0
> > > > > > #2  0x00007f63e258add5 in start_thread (arg=0x7f63e179e700) at
> > > > > > pthread_create.c:308
> > > > > >          __res = <optimized out>
> > > > > >          pd = 0x7f63e179e700
> > > > > >          now = <optimized out>
> > > > > >          unwind_buf = {cancel_jmp_buf = {{jmp_buf = {140066961352448,
> > > > > > 1212869169857371576, 0, 8392704, 0, 140066961352448,
> > > > > > -1291626103561052744, -1291619793368703560},
> > > > > > mask_was_saved = 0}}, priv
> > > > > > 
> > > > > > = {pad = {0x0, 0x0, 0x0, 0x0}, data = {
> > > > > >                prev = 0x0, cleanup = 0x0, canceltype = 0}}}
> > > > > >          not_first_call = <optimized out>
> > > > > >          pagesize_m1 = <optimized out>
> > > > > >          sp = <optimized out>
> > > > > >          freesize = <optimized out>
> > > > > > #3  0x00007f63e22b4b3d in clone () at
> > > > > > ../sysdeps/unix/sysv/linux/x86_64/clone.S:113
> > > > > > No locals.
> > > > > > (gdb) thread 3
> > > > > > [Switching to thread 3 (Thread 0x7f63e1f9f700 (LWP 8808))]
> > > > > > #0  0x00007f63e2591bfd in recvmsg () at
> > > > > > ../sysdeps/unix/syscall-template.S:81
> > > > > > 81    T_PSEUDO (SYSCALL_SYMBOL, SYSCALL_NAME, SYSCALL_NARGS)
> > > > > > (gdb) bt full
> > > > > > #0  0x00007f63e2591bfd in recvmsg () at
> > > > > > ../sysdeps/unix/syscall-template.S:81
> > > > > > No locals.
> > > > > > #1  0x000000000052194e in read_msg (s=0x7f63e1f9d3b0,
> > > > > > m=0x7f63e1f9d5a0)
> > > > > > 
> > > > > > at /root/src/dpdk/lib/librte_eal/common/eal_common_proc.c:258
> > > > > >          msglen = <optimized out>
> > > > > >          control =
> > > > > > "\000\000\000\000\000\000\000\000\336~\f\343c\177\000\000\005", '\000'
> > > > > > <repeats 23 times>, "\360\371\033\342c\177\000"
> > > > > >          cmsg = <optimized out>
> > > > > >          iov = {iov_base = 0x7f63e1f9d5a0, iov_len = 332}
> > > > > >        msgh = {msg_name = 0x7f63e1f9d3b0, msg_namelen = 110, msg_iov =
> > > > > > 0x7f63e1f9d370, msg_iovlen = 1, msg_control = 0x7f63e1f9d380,
> > > > > > msg_controllen = 48, msg_flags = 0}
> > > > > > #2  mp_handle (arg=<optimized out>) at
> > > > > > /root/src/dpdk/lib/librte_eal/common/eal_common_proc.c:346
> > > > > >          msg = {type = 0, msg = {name = '\000' <repeats 63 times>,
> > > > > > len_param = 0, num_fds = 0, param = '\000' <repeats 20 times>, "\002",
> > > > > > '\000' <repeats 234 times>, fds = {0, 0, 0, 0, 0, 0, 0, 0}}}
> > > > > >          sa = {sun_family = 55104,
> > > > > >            sun_path =
> > > > > > "\371\341c\177\000\000\352\372\f\343c\177\000\000\000\000\000\000\000\000\000\000\377\377\377\377\377\377\377\377\000\367\371\341c\177\000\000\030\000\000\000\000\000\000\000p\327\371\341c\177\000\000\000\367\371\341c\177\000\000\000\367\371\341c\177",
> > > > > > 
> > > > > > 
> > > > > > '\000' <repeats 34 times>, "\200\037\000\000\377\377"}
> > > > > > #3  0x00007f63e258add5 in start_thread (arg=0x7f63e1f9f700) at
> > > > > > pthread_create.c:308
> > > > > >          __res = <optimized out>
> > > > > >          pd = 0x7f63e1f9f700
> > > > > >          now = <optimized out>
> > > > > >          unwind_buf = {cancel_jmp_buf = {{jmp_buf = {140066969745152,
> > > > > > 1212869169857371576, 0, 8392704, 0, 140066969745152,
> > > > > > -1291625004586295880, -1291619793368703560},
> > > > > > mask_was_saved = 0}}, priv
> > > > > > 
> > > > > > = {pad = {0x0, 0x0, 0x0, 0x0}, data = {
> > > > > >                prev = 0x0, cleanup = 0x0, canceltype = 0}}}
> > > > > >          not_first_call = <optimized out>
> > > > > >          pagesize_m1 = <optimized out>
> > > > > >          sp = <optimized out>
> > > > > >          freesize = <optimized out>
> > > > > > #4  0x00007f63e22b4b3d in clone () at
> > > > > > ../sysdeps/unix/sysv/linux/x86_64/clone.S:113
> > > > > > No locals.
> > > > > > 
> > > > > > I don't have more info for now.
> > > > > > 
> > > > > 
> > > > > Thanks for the feedback on this issue. I don't see obvious
> > > > > reason for this deadlock yet.
> > > > > 
> > > > > I'll investigate it asap (not tomorrow, but wednesday). In
> > > > > the worst case, we can revert the series if I cannot find
> > > > > the root cause rapidly.
> > > > 
> > > > I might think that the suggestion from Stephen of destroying the
> > > > barrier can help this issue. I'll try to reproduce it and test
> > > > it before sending a patch to fix it.
> > > 
> > > In case you don't reproduce, feel free to send me the patch to test it.
> > 
> > Below patch can fix another strange sigsegv issue in my VM. Please check
> > if it works for you. I doubt it's use-after-free problem which could
> > lead to different issues in different env. Please have a try.
> > 
> > 
> > diff --git a/lib/librte_eal/common/eal_common_thread.c
> > b/lib/librte_eal/common/eal_common_thread.c
> > index de69452..d91b67d 100644
> > --- a/lib/librte_eal/common/eal_common_thread.c
> > +++ b/lib/librte_eal/common/eal_common_thread.c
> > @@ -205,6 +205,7 @@ rte_ctrl_thread_create(pthread_t *thread, const char
> > *name,
> >                  goto fail;
> > 
> >          pthread_barrier_wait(&params->configured);
> > +       pthread_barrier_destroy(&params->configured);
> 
> Thanks Jianfeng, that fixes my issue.
> For correctness, I wonder whether we should check pthread_barrier_wait
> return, and only call destroy() if PTHREAD_BARRIER_SERIAL_THREAD?
> And so also do same the same thing in rte_thread_init().
> 
> What do you think?
> Thanks,
> Maxime


Thanks for the update. I also have a patch that replaces the barrier by
a lock which could also work, but if Jianfeng's one fixes the issue, I
think it is better.

About the PTHREAD_BARRIER_SERIAL_THREAD, not sure it will change
something:

       Upon successful completion, the pthread_barrier_wait() function
       shall return PTHREAD_BARRIER_SERIAL_THREAD for a single
       (arbitrary) thread synchronized at the barrier and zero for each
       of the other threads. Otherwise, an error number shall be
       returned to indicate the error.

I understand that it will ensure that only one barrier will return
PTHREAD_BARRIER_SERIAL_THREAD, but not necessarily the last one. So
if destroy() is called in the parent thread, it should be the same, no?

By the way, there is also a small memory leak that was introduced by
the previous patch, maybe you can add the fix too:

-       if (ret != 0)
+       if (ret != 0) {
+               free(params);
                return ret;
+       }


Olivier



> 
> >          free(params);
> > 
> >          return 0;
> > 
> > Thanks,
> > Jianfeng
> > 
> > > 
> > > Thanks,
> > > Maxime
> > > 
> > > > Thanks,
> > > > Jianfeng
> > > > 
> > > > > 
> > > > > Olivier
> > > > > 
> > > > 
> >