[lttng-dev] RFC: More flexible support for UST event collection in containers

[lttng-dev] RFC: More flexible support for UST event collection in containers

[Loïc Gelle]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1499 bytes --]

Hi all,

I am writing a proof of concept for LTTng (ust + tools) for better  
support of application tracing in containers. My point is that we  
should be able to collect all UST events emitted from within  
containers using a single sessiond / consumerd, at least if you want  
to. The main problem to that approach was that the session deamon was  
unable to handle PID clashes when two apps with the same vpid tried to  
register. I fix this by patching the UST client for it to declare both  
its PID and its PID namespace inode number during registration (and I  
also propagated the change to lttng-tools for app registration in the  
hashtable). My patched versions of tools and UST are here, in the  
branch "containers":

https://github.com/loicgelle/lttng-ust
https://github.com/loicgelle/lttng-tools

One problem that I have now is that LTTng-tools currently uses  
/proc/[pid]/cmdline to fetch the name of each application when listing  
the available tracepoints. This is a problem since we have no (simple)  
way to know the PID of the registered apps in the context of the PID  
namespace of sessiond, so this step basically fails. I thought that a  
fix for this would be to ask the applications themselves to declare  
their name using their dedicated socket. Do you think that would be  
possible and relevant?

Also, if you like the idea that I'm pushing, there is more work ahead  
to make it a feature (for example propagating the changes to pid  
tracking).

Thanks,
Loïc.

[lttng-dev] RFC: More flexible support for UST event collection in containers

[Loïc Gelle]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1845 bytes --]

Hi all,

Any reactions to my previous message?

Thanks,
Loïc.

Loïc Gelle <loic.gelle at polymtl.ca> a écrit :

> Hi all,
>
> I am writing a proof of concept for LTTng (ust + tools) for better  
> support of application tracing in containers. My point is that we  
> should be able to collect all UST events emitted from within  
> containers using a single sessiond / consumerd, at least if you want  
> to. The main problem to that approach was that the session deamon  
> was unable to handle PID clashes when two apps with the same vpid  
> tried to register. I fix this by patching the UST client for it to  
> declare both its PID and its PID namespace inode number during  
> registration (and I also propagated the change to lttng-tools for  
> app registration in the hashtable). My patched versions of tools and  
> UST are here, in the branch "containers":
>
> https://github.com/loicgelle/lttng-ust
> https://github.com/loicgelle/lttng-tools
>
> One problem that I have now is that LTTng-tools currently uses  
> /proc/[pid]/cmdline to fetch the name of each application when  
> listing the available tracepoints. This is a problem since we have  
> no (simple) way to know the PID of the registered apps in the  
> context of the PID namespace of sessiond, so this step basically  
> fails. I thought that a fix for this would be to ask the  
> applications themselves to declare their name using their dedicated  
> socket. Do you think that would be possible and relevant?
>
> Also, if you like the idea that I'm pushing, there is more work  
> ahead to make it a feature (for example propagating the changes to  
> pid tracking).
>
> Thanks,
> Loïc.
>
> _______________________________________________
> lttng-dev mailing list
> lttng-dev at lists.lttng.org
> https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev