ApportImprovements

Differences between revisions 3 and 4
Revision 3 as of 2006-11-27 14:36:39
Size: 3860
Editor: 87
Comment: remove TODO
Revision 4 as of 2006-11-29 11:10:00
Size: 4865
Editor: 87
Comment: specify mono crash handling
Deletions are marked like this. Additions are marked like this.
Line 37: Line 37:
Line 49: Line 50:
To be discussed. There are two cases of crashes here: (1) due to an unhandled exception, and (2) due to a crash-related signal (SIGSEGV/SIGBUS/etc.). In both cases, Mono prints out a backtrace by default, which we want to capture for the apport report.

Thus Mono's backtrace generator needs to capture the backtrace in a global variable `__apport_mono_backtrace`.

==== Unhandled exception ====

After printing the stacktrace to stderr (current default behaviour), Mono's default exception handler additionally calls `/usr/share/apport/mono-hook` ''pid'' and pipes the value of `__apport_mono_backtrace` to its stdin. `mono-hook` will collect the usual generic data from `/proc` and create a proper crash report in `/var/crash/`.

==== Crash due to signal reception ====

Mono's current signal handler already causes the Mono backtrace to be written to stderr. If, as above, it is written to `__apport_mono_backtrace`, then apport's ``report_add_gdb_info()` can fish out the value from the core dump and add it to the report.

Please check the status of this specification in Launchpad before editing it. If it is Approved, contact the Assignee or another knowledgeable person before making changes.

Summary

We want to extend the range of crashes that apport can process, as well as make the process more efficient.

Rationale

In Edgy we do not get any useful information from crashed processes which cannot write into their cwd (like notification-daemon) or are terminated with SIGABRT due to an exception (like interpreted Python/Mono programs). To accelerate processing, the core dump should not be temporarily written to the disk at all.

Use cases

  • notification-daemon, a user-session daemon which does a chmod('/') at program start, crashes. apport is able to get a proper core dump and extract useful information from it.

  • serpentine crashes with an unhandled Python exception. apport picks this up and creates a report which contains the Python backtrace.
  • F-Spot crashes. apport creates a report which contains the Mono backtrace.

Design

  • Have the kernel pipe the core dump to apport instead of writing it to the disk temporarily. This is both faster and also makes core dump creation independent from cwd writability.
  • Add hooks to most common interpreters (Python/Mono) to intercept unhandled exceptions and create an apport report.

Implementation

Kernel

In 2.6.19, Andi Kleen committed a new feature for /proc/sys/kernel/core_pattern: it can start with a pipe ('|') now, in which case the remainder is interpreted as a path. That path is executed, and the core dump is piped to stdin. We will base our solution on this, however, two modifications are still required:

  • 2.6.19 upstream does not support arguments for the called program. In order to avoid having to write and process the core dump, we want to use the %p and %s macros. Thus the kernel should first split core_pattern value at spaces, consider the first field as program path, the rest as arguments, and, after splitting, perform macro substitution (this will work correctly with %e containing spaces). Andi Kleen would welcome to see this fixed upstream, but doesn't want to work on that himself.

  • We do not want to generally enable core dumps, thus we need to leave the default ulimit -c to 0. If core_pattern is a pipe, the kernel should ignore the current ulimit -c; our kernel maintainers consider this safe, since the kernel does not actually write any file in that case. Instead, the called process can decide about an appropriate limit. This should be discussed with upstream and we should aim for their approval, so that eventually apport (and similar crash interception projects) work on a stock upstream kernel across all distributions.

apport

Apport needs to read the core dump from stdin if the core dump path is '-' (this is already implemented in Feisty). The init script needs to set the appropriate core_pattern: |/usr/share/apport/apport %p %s -.

We will also add a new Python package apport (shipped in python-apport deb) which provides a default Python exception handler that creates an apport report (this is already implemented in Feisty).

Python 2.4/2.5

site.py should try to import apport.python_hook, and if that succeeds, call the apport exception handler (python_hook.install()). See https://launchpad.net/bugs/70957 for details and patch.

Mono

There are two cases of crashes here: (1) due to an unhandled exception, and (2) due to a crash-related signal (SIGSEGV/SIGBUS/etc.). In both cases, Mono prints out a backtrace by default, which we want to capture for the apport report.

Thus Mono's backtrace generator needs to capture the backtrace in a global variable __apport_mono_backtrace.

Unhandled exception

After printing the stacktrace to stderr (current default behaviour), Mono's default exception handler additionally calls /usr/share/apport/mono-hook pid and pipes the value of __apport_mono_backtrace to its stdin. mono-hook will collect the usual generic data from /proc and create a proper crash report in /var/crash/.

Crash due to signal reception

Mono's current signal handler already causes the Mono backtrace to be written to stderr. If, as above, it is written to __apport_mono_backtrace, then apport's report_add_gdb_info()` can fish out the value from the core dump and add it to the report.

Data preservation and migration

Not required.


CategorySpec

ApportImprovements (last edited 2008-08-06 16:30:26 by localhost)