[lttng-dev] Relayd trace drops
Aravind HT
aravind.ht at gmail.com
Wed Dec 16 08:02:25 EST 2015
Hi Mathieu,
Ive been able to analyze this sessiond - consumerd hang scenario and I see
that the fix is not available on 2.6.1 or on 2.7.0.
Root Cause :
The hang is mainly caused because of a race condition for the
consumer_metadata_socket between the following two flows of code
Flow 1. do_sync_metadata() -> lttng_ustconsumer_sync_metadat() ->
lttng_ustconsumer_request_metadata() -> lttncomm_recv_unix_sock() ->
recvmsg()
and
Flow 2. lttng_consumer_recv_cmd() -> lttng_ust_consumer_recv_cmd() ->
lttng_ustconsumer_recv_metadata()
Flow 1 above initiates a metadata requests and recvmsg()s for the reply
from sessiond on the consumer_metadata_socket ( it has taken the
metadata_socket_lock and the stream->lock ) , but simultaneously this
socket is being polled from Flow 2. Once in a while Flow 2 gets to receive
while Flow 1 just hangs on recvmsg() . Since Flow 1 has taken the
stream->lock, this then introduces a deadlock at a later stage
(on_buffer_ready() when the metadata is being flushed. )
Think using the metadata_socket_lock in Flow 2 before attempting to receive
may address this issue, but is a bit complicated as the message would have
already been received at the point of using that lock. The other
alternative is to remove the complete recv() from Flow 1 and let that be
taken care by Flow 2.
Please let me know if this has been fixed in the later builds, may be I
searched it wrong.
Regards,
Aravind.
On Sat, Dec 12, 2015 at 4:13 AM, Mathieu Desnoyers <
mathieu.desnoyers at efficios.com> wrote:
> Hi,
>
> Please upgrade to 2.6.1 or 2.7.0, those are old issues.
>
> Thanks,
>
> Mathieu
>
>
> ----- On Dec 11, 2015, at 1:42 PM, Aravind HT <aravind.ht at gmail.com>
> wrote:
>
>
> NPBS understand it, I will continue to investigate this, got a bit side
> tracked with the sessiond hang issue that I mentioned earlier.
> Making some progress on that, will update as soon as I find the root
> cause, it looks like a mutex deadlock between
> the following threads in the consumerd daemon
>
> #1 0x0000000000410b5d in lttcomm_recv_unix_sock (sock=26, buf=buf at entry=0x7f16acca57f0,
> len=len at entry=4476) at unix.c:175
> #2 0x000000000041dc96 in lttng_ustconsumer_request_metadata (ctx=ctx at entry=0x640d10,
> channel=0x7f16980192b0, timer=timer at entry=0, wait=wait at entry=0) at
> ust-consumer.c:2331
> #3 0x000000000041ff3c in lttng_ustconsumer_sync_metadata (ctx=ctx at entry=0x640d10,
> metadata=metadata at entry=0x7f169801aa40) at ust-consumer.c:1829
> #4 0x000000000040eabf in do_sync_metadata (ctx=<optimized out>,
> metadata=<optimized out>) at consumer-stream.c:468
> #5 consumer_stream_sync_metadata (ctx=ctx at entry=0x640d10, session_id=8)
> at consumer-stream.c:545
> #6 0x000000000041d316 in lttng_ustconsumer_read_subbuffer
> (stream=stream at entry=0x7f1698018fa0, ctx=ctx at entry=0x640d10) at
> ust-consumer.c:2063
> #7 0x000000000040b613 in lttng_consumer_read_subbuffer
> (stream=0x7f1698018fa0, ctx=0x640d10) at consumer.c:3203
> #8 0x0000000000409608 in consumer_thread_data_poll (data=0x640d10) at
> consumer.c:2557
> #9 0x00007f16af6a6fe3 in start_thread (arg=0x7f16acca7700) at
> pthread_create.c:312
> #10 0x00007f16af3dcafd in clone () at
> ../sysdeps/unix/sysv/linux/x86_64/clone.S:111
>
> and
>
> consumer_thread_metadata_poll() calling lttng_consumer_read_subbuffer() /
> on_buffer_ready() calling pthread_mutex_lock(stream->lock) and then hanging
>
>
> This again can be reproduced by the same script, just running consecutive
> sessions of short duration (<10 seconds) with traces getting captured on a
> remote node.
> I hit this every time I run the script.
>
> Regards,
> Aravind.
>
>
>
> On Fri, Dec 11, 2015 at 11:33 PM, Jonathan Rajotte Julien <
> Jonathan.rajotte-julien at efficios.com> wrote:
>
>> Hi Aravind,
>>
>> On 2015-12-11 12:16 PM, Aravind HT wrote:
>>
>>> Hi,
>>>
>>> Did you get a chance to reproduce the problem with the scripts ? Let me
>>> know if you need any help running it.
>>>
>>
>> Time is a scarce resource unfortunately. But until I have the time to
>> look at it you might want to investigate if you are able to reproduce it on
>> 2.6.1/2.7.0/master on 2 regular (x86_64, ubuntu/debian) VM this would
>> really help.
>>
>> Cheers!
>>
>>
>>> Regards,
>>> Aravind.
>>>
>>>
>>> On Wed, Dec 9, 2015 at 3:04 PM, Aravind HT <aravind.ht at gmail.com
>>> <mailto:aravind.ht at gmail.com>> wrote:
>>>
>>> Sorry about that, not sure how it got missed.
>>> Here it is.
>>>
>>>
>>> # This needs a two node set up (1. local current node 2. remote node
>>> )
>>> # relayd runs on the current node where traces are captured from
>>> the remote node
>>> # remote node runs test applications which generate traces.
>>> # the launch_script_RN is executed on the current node and uses
>>> ssh to execute commands on the remote node. So this part may not
>>> work in every case and may prompt for a password.
>>> # if experiencing problems with ssh , kindly check
>>>
>>> http://serverfault.com/questions/241588/how-to-automate-ssh-login-with-password
>>>
>>> # ==================== To Run =============================
>>> launch_script_RN.py self_profile -c /tmp/configFile.txt
>>>
>>>
>>>
>>> # configFile.txt is the file which has configuration params that
>>> launchScript
>>> # needs to configure lttng sessions. Below is an explanation of
>>> the different options.
>>> # =================== configFile.txt
>>> =============================
>>>
>>> [section1]
>>> # final out put file path
>>> OutputFile = /tmp/Final_report.txt
>>> # the remote node hostname on which test applications run and the
>>> test sessions will be created; this should be something that could
>>> be used with ssh. Traces will be transported from this node to the
>>> lttng_relayd running on the current node.
>>> Node = MY_REMOTE_NODE
>>> # Sub buffer size to start this with
>>> SubBufSize = 16k
>>> # Sub buffer count
>>> SubBufCount = 4
>>> # per uid buffer
>>> BufferScheme = --buffers-uid
>>> # yes
>>> EnableTracing = yes
>>> # Bit rate of the test applications. Comman seperated example "1,
>>> 3, 3, 50" sayss 4 test applications producing 1, 3, 3, and 50 Kb/s
>>> traces.
>>> # So with the below, we just start with 1 test application
>>> producing 10 kb/s
>>> TestApps = 10
>>> # session life time in seonds
>>> TestTime = 10
>>> # Max number of successive sessions to configure. if n then n-1
>>> sessions are run, ex MaxRun = 2 will run 1 session.
>>> MaxRun = 100
>>>
>>>
>>> # ==================== Place the following files under
>>> ===============
>>>
>>> # /tmp on the remote node
>>> clean_RemNode_apps.sh
>>> report_lttng_script.sh
>>>
>>> # rest of the scripts under /usr/sbin on the current local node on
>>> which lttng_realyd runs
>>> # Define a trace point MY_TRACE to take a single string arg with
>>> LOG_TEST_APP_PROFILING as the provider, compile test lttng_testApp
>>> and place it under /usr/sbin of the remote host
>>>
>>> # in launch_script_RN.py change currentNodeIP to the IP address on
>>> which relayd is receiving, default ports are used.
>>>
>>> # lttng_relayd is started as
>>> /usr/bin/lttng-relayd -o /var/log/lttng-traces -d
>>>
>>> # lttng_sessiond is started as
>>> /usr/bin/lttng-sessiond --consumerd32-path
>>> /usr/lib/lttng/libexec/lttng-consumerd --consumerd32-libdir
>>> /usr/lib/ --consumerd64-path
>>> /usr/lib64/lttng/libexec/lttng-consumerd --consumerd64-libdir
>>> /usr/lib64/ -b --no-kernel
>>>
>>>
>>>
>>>
>>> Regards,
>>> Aravind.
>>>
>>> On Wed, Dec 9, 2015 at 12:57 AM, Jonathan Rajotte Julien
>>> <Jonathan.rajotte-julien at efficios.com
>>> <mailto:Jonathan.rajotte-julien at efficios.com>> wrote:
>>>
>>> Hi Aravind,
>>>
>>> There is no README in the archive you sent.
>>>
>>> Cheers
>>>
>>> On 2015-12-08 07:51 AM, Aravind HT wrote:
>>>
>>> Hi,
>>>
>>> I am trying to upgrade in parallel, but this issue may
>>> still be present after I upgrade or may be temporarily
>>> masked. So I need to find the root cause for this and then
>>> see if its available on the latest before committing to
>>> upgrade.
>>>
>>> There is another issue i'm hitting, the lttng list command
>>> hangs after lttng destroy session when running the profiling.
>>>
>>> I found that consumerd 64 goes into an infinite loop
>>> waiting to flush metadata in
>>> lttng_ustconsumer_recv_metadata() :: while
>>> (consumer_metadata_cache_flushed(channel, offset + len,
>>> timer)) .
>>> In consumer_metadata_cache,
>>> channel->metadata_stream->endpoint_status is
>>> CONSUMER_ENDPOINT_ACTIVE,
>>> metadata_stream->ust_metadata_pushed is 0 with offset
>>> having some value. This call always returns a 1 from the
>>> last else{} block resulting in an infinite loop. Upon
>>> searching the forum, I found the same issue being reported
>>> here :
>>>
>>> https://www.mail-archive.com/lttng-dev@lists.lttng.org/msg07982.html
>>>
>>> Regards,
>>> Aravind.
>>>
>>>
>>> On Tue, Dec 8, 2015 at 12:43 AM, Jonathan Rajotte
>>> <jonathan.r.julien at gmail.com
>>> <mailto:jonathan.r.julien at gmail.com>> wrote:
>>>
>>> Hi Aravind,
>>>
>>> Did you have the chance to upgrade to 2.6.1.If so
>>> where you able
>>> to reproduce?
>>>
>>> Cheers
>>>
>>> On Mon, Dec 7, 2015 at 1:07 PM, Aravind HT
>>> <aravind.ht at gmail.com <mailto:aravind.ht at gmail.com>>
>>> wrote:
>>>
>>> Hi,
>>>
>>> I have attached the complete profiling scripts
>>> here, its a bit
>>> shabby, im new to python.
>>>
>>> There is a README which has the details on how to
>>> execute it.
>>> Im using a Yocto 1.6 on x86_64 platforms on both
>>> the nodes.
>>>
>>>
>>> Running this script when there are other sessions
>>> running
>>> seems to reproduce this problem easily.
>>> Please try it and let me know if you have any issues
>>> reproducing the problem.
>>>
>>> Regards,
>>> Aravind.
>>>
>>> On Sat, Dec 5, 2015 at 5:23 PM, Jérémie Galarneau
>>> <jeremie.galarneau at efficios.com
>>> <mailto:jeremie.galarneau at efficios.com>
>>> <mailto:jeremie.galarneau at efficios.com
>>> <mailto:jeremie.galarneau at efficios.com>>> wrote:
>>>
>>> On Fri, Dec 4, 2015 at 11:06 PM, Aravind HT
>>> <aravind.ht at gmail.com
>>> <mailto:aravind.ht at gmail.com> <mailto:aravind.ht at gmail.com
>>>
>>> <mailto:aravind.ht at gmail.com>>> wrote:
>>> > I am using 2.6.0 .I will try to share the
>>> code that I'm
>>> using here in some
>>> > time. If there are any specific fixes that
>>> are relevant
>>> to this issue, see
>>> > if you can provide a link to them. I would
>>> ideally like
>>> to try them out
>>> > before trying a full upgrade to the latest
>>> versions.
>>>
>>> Hi,
>>>
>>> Can you provide more information on the
>>> system? Which
>>> distribution,
>>> architecture, kernel version?
>>>
>>> The verbose sessiond logs might help pinpoint any
>>> unexpected behaviour
>>> here (are all applications registering as
>>> expected?).
>>>
>>> Jérémie
>>>
>>> >
>>> > On Fri, Dec 4, 2015 at 6:11 PM, Jérémie
>>> Galarneau
>>> > <jeremie.galarneau at efficios.com
>>> <mailto:jeremie.galarneau at efficios.com>
>>> <mailto:jeremie.galarneau at efficios.com
>>> <mailto:jeremie.galarneau at efficios.com>>> wrote:
>>> >>
>>> >> Hi Aravind,
>>> >>
>>> >> Can't say I have looked at everything you
>>> sent yet, but
>>> as a
>>> >> preemptive question, which version are we
>>> talking about
>>> here? 2.6.0 or
>>> >> 2.6.1? 2.6.1 contains a lot of relay daemon
>>> fixes.
>>> >>
>>> >> Thanks,
>>> >> Jérémie
>>> >>
>>> >> On Thu, Dec 3, 2015 at 7:01 AM, Aravind HT
>>> <aravind.ht at gmail.com
>>> <mailto:aravind.ht at gmail.com>> wrote:
>>> >> > Hi,
>>> >> >
>>> >> > I am trying to obtain the performance
>>> characteristics
>>> of lttng with the
>>> >> > use
>>> >> > of test applications. Traces are being
>>> produced on a
>>> local node and
>>> >> > delivered to relayd that is running on a
>>> separate
>>> node for storage.
>>> >> >
>>> >> > An lttng session with the test
>>> applications producing
>>> an initial bit
>>> >> > rate of
>>> >> > 10 kb/s is started and run for about 30
>>> seconds. The
>>> starting sub-buffer
>>> >> > size is kept at 128 kb and sub-buf count
>>> at 4. The
>>> session is then
>>> >> > stopped
>>> >> > and destroyed and traces are analyzed to
>>> see if there
>>> are any drops.
>>> >> > This is
>>> >> > being done in a loop with every
>>> subsequent session
>>> having an increment
>>> >> > of 2
>>> >> > kb/s as long as there are no drops. If
>>> there are
>>> drops, I increase the
>>> >> > buffer size by a factor of x2 without
>>> incrementing
>>> the bit rate.
>>> >> >
>>> >> > I see trace drops happening consistently
>>> with test
>>> apps producing traces
>>> >> > at
>>> >> > less than 40 kb/s, it doesnt seem to help
>>> even if I
>>> started with 1mb x 4
>>> >> > sub-buffers.
>>> >> >
>>> >> > Analysis :
>>> >> >
>>> >> > I have attached the lttng_relayd ,
>>> lttng_consumerd_64
>>> logs and the
>>> >> > entire
>>> >> > trace directory, hope you will be able to
>>> view it.
>>> >> > I have modified lttng_relayd code to dump
>>> the traces
>>> being captured in
>>> >> > the
>>> >> > lttng_relayd logs along with debug info.
>>> >> >
>>> >> > Each test app is producing logs in the
>>> form of :
>>> >> > "TraceApp PID - 31940 THID - 31970
>>> @threadRate - 1032
>>> b/s appRate - 2079
>>> >> > b/s
>>> >> > threadTraceNum - 9 appTraceNum - 18
>>> sleepTime - 192120"
>>> >> >
>>> >> > The test application PID, test
>>> application thread id,
>>> thread bit rate,
>>> >> > test
>>> >> > app bit rate, thread trace number and
>>> application
>>> trace number s are
>>> >> > part of
>>> >> > the trace. So in the above trace, the
>>> thread is
>>> producing at 1 kb/s and
>>> >> > the
>>> >> > whole test app is producing at 2 kb/s.
>>> >> >
>>> >> > If we look at the babeltrace out put, we
>>> see that the
>>> Trace with
>>> >> > TraceApp
>>> >> > PID - 31940 appTraceNum 2 is missing ,
>>> with 1, 3, 4,
>>> 5 and so on being
>>> >> > successfully captured.
>>> >> > I looked at the lttng_relayd logs and
>>> found that
>>> trace of "appTraceNum
>>> >> > 2" is
>>> >> > not delivered/generated by the consumerd
>>> to the
>>> relayd in sequence with
>>> >> > other traces. To rule out that this is
>>> not a test
>>> application problem,
>>> >> > you
>>> >> > can look at line ltttng_relayd log :
>>> 12778 and see
>>> traces from
>>> >> > appTraceNum -
>>> >> > 1 to appTraceNum - 18 including the
>>> appTraceNum 2 are
>>> "re-delivered" by
>>> >> > the
>>> >> > consumerd to the relayd.
>>> >> > Essentially, I see appTraceNum 1 through
>>> appTraceNum
>>> 18 being delivered
>>> >> > twice, once individually where
>>> appTraceNum 2 is
>>> missing and once as a
>>> >> > group
>>> >> > at line 12778 where its present.
>>> >> >
>>> >> >
>>> >> > Request help with
>>> >> > 1. why traces are delivered twice, is it
>>> by design or
>>> a genuine problem
>>> >> > ?
>>> >> > 2. how to avoid traces being dropped even
>>> though
>>> buffers are
>>> >> > sufficiently
>>> >> > large enough ?
>>> >> >
>>> >> >
>>> >> > Regards,
>>> >> > Aravind.
>>> >> >
>>> >> >
>>> _______________________________________________
>>> >> > lttng-dev mailing list
>>> >> > lttng-dev at lists.lttng.org
>>> <mailto:lttng-dev at lists.lttng.org>
>>> >> >
>>> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
>>> >> >
>>> >>
>>> >>
>>> >>
>>> >> --
>>> >> Jérémie Galarneau
>>> >> EfficiOS Inc.
>>> >> http://www.efficios.com
>>> >
>>> >
>>>
>>>
>>>
>>> --
>>> Jérémie Galarneau
>>> EfficiOS Inc.
>>> http://www.efficios.com
>>>
>>>
>>>
>>> _______________________________________________
>>> lttng-dev mailing list
>>> lttng-dev at lists.lttng.org
>>> <mailto:lttng-dev at lists.lttng.org>
>>> <mailto:lttng-dev at lists.lttng.org
>>> <mailto:lttng-dev at lists.lttng.org>>
>>> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
>>>
>>>
>>>
>>>
>>> -- Jonathan Rajotte Julien
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> lttng-dev mailing list
>>> lttng-dev at lists.lttng.org <mailto:lttng-dev at lists.lttng.org>
>>> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
>>>
>>>
>>> -- Jonathan R. Julien
>>> Efficios
>>>
>>>
>>>
>>>
>> --
>> Jonathan R. Julien
>> Efficios
>>
>>
>
> _______________________________________________
> lttng-dev mailing list
> lttng-dev at lists.lttng.org
> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
>
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lttng.org/pipermail/lttng-dev/attachments/20151216/a12df355/attachment-0001.html>
More information about the lttng-dev
mailing list