[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 Roger Light changed: What|Removed |Added Attachment #118745|0 |1 is obsolete|| Attachment #118746|0 |1 is obsolete|| Attachment #118747|0 |1 is obsolete|| Attachment #118748|0 |1 is obsolete|| --- Comment #35 from Roger Light --- Created attachment 142878 --> https://bugs.kde.org/attachment.cgi?id=142878=edit Patch against 3.18.1 This is the old patches combined and refreshed for 3.18.1. It also includes the failgrind-helper utility, which makes it easy to run failgrind against an executable repeatedly until it completes with all memory allocations/syscalls in the program run succeeding, or the executable segfaults - at which point you can see which failures caused the segfault. It is imperfect, but a big improvement over manual testing. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #34 from Roger Light --- Created attachment 118748 --> https://bugs.kde.org/attachment.cgi?id=118748=edit Patch 4/4, function tracing code from callgrind This adds a cut down version of the function tracing code from callgrind to failgrind, which allows the syscall failure code to work, as well as the --alloc-toggle-fn option, which is likely to be a very common way this tool is used. The tests work under this patch as well. This patch is for demonstrating the tool operation, but I don't think it is a sensible patch for inclusion due to the huge amount of duplication with callgrind. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #33 from Roger Light --- Created attachment 118747 --> https://bugs.kde.org/attachment.cgi?id=118747=edit Patch 3/4, support for syscall failures This adds the Failgrind code that handles syscall failures, in terms of the command line options, keeping track of already seen syscall call stacks, the gdb monitor and client requests and similar, but does not include any instrumentation code, so it doesn't work on its own. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #32 from Roger Light --- Created attachment 118746 --> https://bugs.kde.org/attachment.cgi?id=118746=edit Patch 2/4, add VG_(force_syscall_error) This adds VG_(force_syscall_error) to coregrind, which allows tools to force an error value when a syscall fails. This is necessary for the next patch. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 Roger Light changed: What|Removed |Added Attachment #115711|0 |1 is obsolete|| Attachment #115713|0 |1 is obsolete|| Attachment #115744|0 |1 is obsolete|| --- Comment #31 from Roger Light --- Created attachment 118745 --> https://bugs.kde.org/attachment.cgi?id=118745=edit Patch 1/4, heap memory allocation failures only This patch adds the failgrind tool, with support for failing heap allocations with a variety of control. There is no instrumentation and hence no command line capability of controlling when the failures are enabled. The gdb interface and client requests are available and can be used to control when failures are enabled, but I doubt either of those are used as widely as just command line options. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 Tom Hughes changed: What|Removed |Added CC||t...@compton.nu --- Comment #30 from Tom Hughes --- Nick beat you to it and applied something very similar a few hours ago in https://sourceware.org/git/gitweb.cgi?p=valgrind.git;a=commitdiff;h=8b689c66d9b7f12852048db0e842560761bbf904. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 roger.li...@cedalo.com changed: What|Removed |Added Attachment #115712|0 |1 is obsolete|| --- Comment #29 from roger.li...@cedalo.com --- Created attachment 115744 --> https://bugs.kde.org/attachment.cgi?id=115744=edit Updated apply_ExeContext after comments. Thanks for your comments. The epoch parameter was there because I hadn't seen that an ExeContext contains its own epoch. You're also right that it could just call apply_StackTrace. I've updated the patch to do that. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #28 from Philippe Waroquiers --- (In reply to roger.light from comment #25) > Created attachment 115712 [details] > Patch that implements VG_(apply_ExeContext)(). 2 comments: * As ExeContext maintains internally its epoch, I do not understand why there is a parameter 'ep' in VG_(apply_ExeContext). There should be no reason to call VG_(apply_ExeContext) with another ep than the ExeContext ep, and so VG_(apply_ExeContext) can retrieve it itself. * The code of VG_(apply_ExeContext) looks very similar to VG_(apply_stacktrace). It looks like VG_(apply_ExeContext) can just make a call to VG_(apply_stacktrace). Thanks Philippe -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #27 from roger.li...@cedalo.com --- I've just added some updated patches. The two small ones add some coregrind features that are needed for this tool - implementing VG_(apply_ExeContext)() and adding a way for a tool to force a syscall to fail in the pre_syscall callback. The big one is the patch for the tool. This now has a lot of control over how the allocation failures occur. It can be turned on/off at the start of the run, set to toggle on entry and exit from named functions, limited to a certain number of failures per run, automatically accepted based on size of allocation requested, automatically accepted based on matching function name, automatically accepted based on a dice roll, or to some degree controlled with gdb or client requests. It also implements syscall failures, which is turned off by default. If you turn it on, then all syscalls (bar those that are marked as always succeeding) will fail with errno set to EINVAL. You can change to have all syscalls fail with a different errno, and/or set individual syscalls to fail with specific errors. You can set all syscalls with a matching name to be automatically accepted. You can set all syscalls that aren't given a specific error to be automatically accepted. You can accept based on the number of failures in a run, or based on a dice roll, or to some degree controlled with gdb or client requests. There are tests for most features, but they haven't been tried on different Linux releases or distributions (Kubuntu 18.04 only) nor on Mac or Solaris. The documentation is pretty complete with examples for most situations. The bits that are missing - I've taken a lot of code from callgrind to support the function enter/leave feature, but I may have taken too much out. The bit that needs looking at is function name wildcard support which I initially removed but now see it was included in callgrind to allow C++ names to be dealt with correctly. That needs putting back in, but I could do with a break from this for a few days at least and was hoping for some feedback. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 roger.li...@cedalo.com changed: What|Removed |Added Attachment #114630|0 |1 is obsolete|| Attachment #114817|0 |1 is obsolete|| --- Comment #26 from roger.li...@cedalo.com --- Created attachment 115713 --> https://bugs.kde.org/attachment.cgi?id=115713=edit Tool patch October 2018 -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #25 from roger.li...@cedalo.com --- Created attachment 115712 --> https://bugs.kde.org/attachment.cgi?id=115712=edit Patch that implements VG_(apply_ExeContext)(). -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #24 from roger.li...@cedalo.com --- Created attachment 115711 --> https://bugs.kde.org/attachment.cgi?id=115711=edit Patch that allows tools to set a syscall error. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 --- Comment #23 from roger.li...@cedalo.com --- Created attachment 114817 --> https://bugs.kde.org/attachment.cgi?id=114817=edit Patch with function enter/leave detection code from callgrind. This patch is the same as 114630, but with code borrowed from callgrind for tracking the callstack for detecting function entering/leaving. I've minimised the code from callgrind as much as I think is sensible, but it may be possible to remove more with some thought. It is still 3800 lines of text from callgrind compared to 1100 for the rest of failgrind. There's effectively a lot of duplicate code here, so I'll be interested to hear your comments. -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 roger.li...@cedalo.com changed: What|Removed |Added Attachment #114553|0 |1 is obsolete|| --- Comment #22 from roger.li...@cedalo.com --- Created attachment 114630 --> https://bugs.kde.org/attachment.cgi?id=114630=edit Updated patch - tests and other improvements This new patch includes a set of tests covering the majority of the command line options and client requests. The documentation is improved as well. I've added more client requests for more effective use with a test suite - if there were 0 allocations rejected you know the test covered all options, for example. This has the toggle function removed. I still believe it is a very important feature for the tool, but I haven't got it working yet without ~4000 lines of code from callgrind. The patch is otherwise "complete". -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 396290] [PATCH] Possible tool - failgrind
https://bugs.kde.org/show_bug.cgi?id=396290 roger.li...@cedalo.com changed: What|Removed |Added Summary|[PATCH] Possible tool - |[PATCH] Possible tool - |allocfail |failgrind -- You are receiving this mail because: You are watching all bug changes.
