llvm dev - Jun 2018 - XRay feature – pid reporting

I would be happy to help. Could you send me the example patch? Where would
I submit my patch to be reviewed?

Thanks,
Henry

On Mon, Jun 11, 2018 at 8:31 PM, Dean Michael Berris <dean.berris at
gmail.com>
wrote:
>
>
> > On 12 Jun 2018, at 07:49, Henry Zhu via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> >
> > Hello,
> >
> >
> > I’ve noticed that XRay does not report the PID in the trace log. Could
> this be added for multi-process support? (Maybe as a command line
argument).
> >
>
> We can certainly add the PID in the trace header. This will be a format
> change which will require a version increment, but we’ve done that before
> so it should be doable.
>
> Do you have time to implement this feature? It will span LLVM and
> compiler-rt, but it should be straight-forward to do so.
>
> Let me know and I’ll get you an example patch that did this for some of
> the changes we’ve done recently in this regard. Keith can also help as he’s
> worked on updating the FDR format in the past as well.
>
> Cheers
>
> -- Dean
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180625/43e981b7/attachment.html>

Hi Henry,
> On 26 Jun 2018, at 07:07, Henry Zhu <henryzhu at seas.upenn.edu>
wrote:
> 
> I would be happy to help. Could you send me the example patch? Where would
I submit my patch to be reviewed?
> 
Thanks!

First, please see https://llvm.org/docs/Contributing.html#how-to-submit-a-patch
which details the process of contributing to the LLVM project.

There are two patches that show how we changed the version number for Basic mode
(formerly known as “naive” mode):

- https://reviews.llvm.org/D38550 (change in the tooling to support newer
versions of the files)
- https://reviews.llvm.org/D38551 (change in the runtime to write out new kinds
of records; in this case we started writing out tail exits)

We have the chance now to change this in a single patch through the git monorepo
process detailed in
http://llvm.org/docs/GettingStarted.html#for-developers-to-work-with-a-git-monorepo
— but that’s only if you’d like to do that in a single patch.

Cheers

> On 27 Jun 2018, at 01:46, Henry Zhu <henryzhu at seas.upenn.edu>
wrote:
> 
> I'm not too clear on what you mean by changing the version number. Does
that mean that I should change the Header.Version = 3? Therefore, I should add
XRayPidPayload to xray_records.h and install a handler for fetching the pid like
what was done in those commits?
> 
This is the preferred way of doing it (updating the log version number). We have
a set of metadata records in FDR mode to support writing down the thread ID at
the start of the FDR mode block as a trade-off to reduce the log size. We’ll
need a similar change in FDR mode to support writing the process ID at the start
of the block as well.

Basic mode doesn’t make this trade-off which means you need to update the basic
mode file header and function records to always include the process ID.
> Originally, I thought of modifying XRayRecord to contain pid field since
there is an unused buffer at the end that could hold the pid value.
> 
The XRayRecord type in LLVM is only used by the tooling, so this definitely
could work. In compiler-rt though we need to be able to write them down in all
modes that we support — this means writing it down in FDR, Basic, and Profiling
mode.

For Basic mode, adding the PID in every record should be OK but we still need to
update the version number and add the PID in the file header as well.

Cheers

-- Dean

> On 28 Jun 2018, at 07:28, Henry Zhu <henryzhu at seas.upenn.edu>
wrote:
> 
> Thanks, that cleared up my confusion about version numbers.
> 
> I've implemented the handlers and done up to 6) below. I'm unsure
of how to test what I have added.
> 
> 1) Update the Header.Version = 3 (from 2)
> 2) Add a new XRayRecord for pid (XRayPidRecord) in xray_basic_logging.h
> 3) Implement InMemoryRawLogWithPid similar to how InMemoryRawLogWithArg is
implemented
> 4) Implement __xray_set_handler_pid
> 5) call __xray_set_handler_pid, passing in basicLoggingHandlePid in the
initialization function for basic mode
> 6) Add an assembly stub for the [od handler
> 
Huh, I’m sorry for not being clear here — I suspect you don’t need steps 3 to 6.

You may just need to add a metadata record at the beginning of the block, and
getting the PID and TID directly (instead of the cached versions). This way FDR
mode will have the PID record and the TID records at the beginning of the block.

For Basic Mode we need to get the TID directly instead of using the cached
version, and also to get the PID directly instead of attempting to cache it.
This would be an update in the handlers.

In Profiling Mode this would be a little tricky, because it may need changes in
more places. I need to think about that I little more.
> 7) Add additional parsing for pid for llvm-xray tool to parse the header
for pid and xray entries for pid
> 
> Q1. For 7), on order to log the pid, one would need to patch the function
to call the pid logger. Should I add an attribute to clang that patches the
function, so that the function calls the pid? Or is there an easier way to test
the functionality of the pid logger?
Please see above.
> Q2. Should the PID always be set in the file header when xray starts?
Yes.
> Q3. How do I run test cases? 
There’s a ‘check-all’ and ‘check-xray’ target when you build
LLVM+Clang+compiler-rt.
> Q4. Would it be possible to always call the pid logger when fork is called?
> 
Unfortunately no. The simpler solution would be to update the handlers to get
the PID alongside the TID, and to always get the TID instead of using a cached
version.

Cheers

-- Dean

I'm still somewhat unclear about what you mean by "metadata record
entry at
the beginning of the block". I understand that I can make a MetadataRecord
that contains the pid/tid since a metadata record contains 16 bytes.
However, I don't understand what do you mean by the "beginning of the
block". Do you mean right after the file header?

My understanding is that at initialization, the generated log file should
contain:
[File header with pid][metadata record containing pid and tid]

Updating handlers to fetch the PID and TID directly can be done. However,
XRayRecord and XRayArgRecord do not have PID fields, Do I replace some of
the padding with the PID field, or should I make another XRayPayload
containing the TID/PID?

For now, what would be the best way to test the new format to make sure the
format has the correct values?

Thanks

On Wed, Jun 27, 2018 at 11:29 PM, Dean Michael Berris <dean.berris at
gmail.com
> wrote:
>
>
> > On 28 Jun 2018, at 07:28, Henry Zhu <henryzhu at seas.upenn.edu>
wrote:
> >
> > Thanks, that cleared up my confusion about version numbers.
> >
> > I've implemented the handlers and done up to 6) below. I'm
unsure of how
> to test what I have added.
> >
> > 1) Update the Header.Version = 3 (from 2)
> > 2) Add a new XRayRecord for pid (XRayPidRecord) in
xray_basic_logging.h
> > 3) Implement InMemoryRawLogWithPid similar to how
InMemoryRawLogWithArg
> is implemented
> > 4) Implement __xray_set_handler_pid
> > 5) call __xray_set_handler_pid, passing in basicLoggingHandlePid in
the
> initialization function for basic mode
> > 6) Add an assembly stub for the [od handler
> >
>
> Huh, I’m sorry for not being clear here — I suspect you don’t need steps 3
> to 6.
>
> You may just need to add a metadata record at the beginning of the block,
> and getting the PID and TID directly (instead of the cached versions). This
> way FDR mode will have the PID record and the TID records at the beginning
> of the block.
>
> For Basic Mode we need to get the TID directly instead of using the cached
> version, and also to get the PID directly instead of attempting to cache
> it. This would be an update in the handlers.
>
> In Profiling Mode this would be a little tricky, because it may need
> changes in more places. I need to think about that I little more.
>
> > 7) Add additional parsing for pid for llvm-xray tool to parse the
header
> for pid and xray entries for pid
> >
> > Q1. For 7), on order to log the pid, one would need to patch the
> function to call the pid logger. Should I add an attribute to clang that
> patches the function, so that the function calls the pid? Or is there an
> easier way to test the functionality of the pid logger?
>
> Please see above.
>
> > Q2. Should the PID always be set in the file header when xray starts?
>
> Yes.
>
> > Q3. How do I run test cases?
>
> There’s a ‘check-all’ and ‘check-xray’ target when you build
> LLVM+Clang+compiler-rt.
>
> > Q4. Would it be possible to always call the pid logger when fork is
> called?
> >
>
> Unfortunately no. The simpler solution would be to update the handlers to
> get the PID alongside the TID, and to always get the TID instead of using a
> cached version.
>
> Cheers
>
> -- Dean
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180628/6d7642ec/attachment.html>