tracing/user_events: Tracking broken status and feedback

tracing/user_events: Tracking broken status and feedback

[Beau Belgrave]

To help keep track of what is required to remove the broken status from
user_events I am starting this thread. I would like to use this thread
to provide status on the work that has been done so far and to have a
discussion about when the broken status can be removed.

Feedback threads from 5.18 version of user_events:
1. https://lore.kernel.org/all/2059213643.196683.1648499088753.JavaMail.zimbra@efficios.com/
2. https://lore.kernel.org/all/1651771383.54437.1652370439159.JavaMail.zimbra@efficios.com/

All the feedback has been addressed in the patchsets 1 and 2 (see below).
Here are the details: 

1. Use bits vs bytes in event enabled memory mappings.
Fixed in patchset 1.

2. Pack ABI structures where size is used.
Fixed in patchset 1.

3. Don't trust user strings for string formatting.
Fixed in patchset 1.

4. Move to refcount APIs vs atomic for tracking references.
Fixed in patchset 1.

5. Ensure event_mutex is held during registration.
Fixed in patchset 2.

With these, I believe I have addressed all issues to remove the
"broken status".

Here is the list of additional feedback (and status) that I don't believe
should have a bearing on removing the "broken status":

1. Kernel vs user tracers in ABI.
This is not done, the plan is to build a libtracepoint library that allows
working with both kernel and user tracers in user programs. Steven is
working on this at the moment. I don't believe this is required to remove
the broken status, but it will help prove the ABI by having it.

No patchset yet.

2. Container/namespace isolation of events.
user_events utilizes tracefs for user facing files in the ABI. I've created
an RFC patchset showing how if tracefs offered an isolated directory
structure per-namespace the user_events ABI is unaffected. This is true for
other ABIs that tracefs hosts, if they would like to integrate. I don't
believe this is required to remove the broken status, however, it's useful
to have to see how the ABI is unaffected while we work toward enabling
isolation within tracing.

See patchset 3.

Patchsets:
1. https://lore.kernel.org/all/20220425184631.2068-1-beaub@linux.microsoft.com/
2. https://lore.kernel.org/all/20220328223225.1992-1-beaub@linux.microsoft.com/
3. https://lore.kernel.org/all/20220707215828.2021-1-beaub@linux.microsoft.com/

Thanks,
-Beau

Re: tracing/user_events: Tracking broken status and feedback

[Beau Belgrave]

On Mon, Jul 25, 2022 at 04:19:00PM -0700, Beau Belgrave wrote:
> To help keep track of what is required to remove the broken status from
> user_events I am starting this thread. I would like to use this thread
> to provide status on the work that has been done so far and to have a
> discussion about when the broken status can be removed.
> 
> Feedback threads from 5.18 version of user_events:
> 1. https://lore.kernel.org/all/2059213643.196683.1648499088753.JavaMail.zimbra@efficios.com/
> 2. https://lore.kernel.org/all/1651771383.54437.1652370439159.JavaMail.zimbra@efficios.com/
> 
> All the feedback has been addressed in the patchsets 1 and 2 (see below).
> Here are the details: 
> 
> 1. Use bits vs bytes in event enabled memory mappings.
> Fixed in patchset 1.
> 
> 2. Pack ABI structures where size is used.
> Fixed in patchset 1.
> 
> 3. Don't trust user strings for string formatting.
> Fixed in patchset 1.
> 
> 4. Move to refcount APIs vs atomic for tracking references.
> Fixed in patchset 1.
> 
> 5. Ensure event_mutex is held during registration.
> Fixed in patchset 2.
> 
> With these, I believe I have addressed all issues to remove the
> "broken status".
> 
> Here is the list of additional feedback (and status) that I don't believe
> should have a bearing on removing the "broken status":
> 
> 1. Kernel vs user tracers in ABI.
> This is not done, the plan is to build a libtracepoint library that allows
> working with both kernel and user tracers in user programs. Steven is
> working on this at the moment. I don't believe this is required to remove
> the broken status, but it will help prove the ABI by having it.
> 
> No patchset yet.
> 
> 2. Container/namespace isolation of events.
> user_events utilizes tracefs for user facing files in the ABI. I've created
> an RFC patchset showing how if tracefs offered an isolated directory
> structure per-namespace the user_events ABI is unaffected. This is true for
> other ABIs that tracefs hosts, if they would like to integrate. I don't
> believe this is required to remove the broken status, however, it's useful
> to have to see how the ABI is unaffected while we work toward enabling
> isolation within tracing.
> 
> See patchset 3.
> 
> Patchsets:
> 1. https://lore.kernel.org/all/20220425184631.2068-1-beaub@linux.microsoft.com/
> 2. https://lore.kernel.org/all/20220328223225.1992-1-beaub@linux.microsoft.com/
> 3. https://lore.kernel.org/all/20220707215828.2021-1-beaub@linux.microsoft.com/
> 
> Thanks,
> -Beau

Steven had feedback on patchset 1 here:
https://lore.kernel.org/all/20220726180115.69320865@gandalf.local.home/

I've updated patchset 1 and 3 to address this feedback.

Updated patchsets:
1. https://lore.kernel.org/all/20220728233309.1896-1-beaub@linux.microsoft.com/
3. https://lore.kernel.org/all/20220728235241.2249-1-beaub@linux.microsoft.com/

Thanks,
-Beau