information about the stack traces recorded by Valgrind.
This can be used to analyse one possible cause of Valgrind high
memory usage for some programs.
At work, a big set of regression tests crashed out of memory under Valgrind.
Two main causes for out of memory were identified:
1. big memory usage for stacktrace (exe contexts) recording by Valgrind
2. big number of partially initialised bytes.
This patch adds a gdbsrv monitor command that output (very) detailed
information about all the recorded exe context.
This has been used to analyse the problem 1. above,
showing the following identified causes for a (too) big nr of execontexts:
A. When the JIT handles an unknown SP update, even when --track-origins=no,
an execontext is (uselessly) created and recorded
to track the (never used) origin of some uninitialised stack memory.
This creates a whole bunch of 'one IP' execontexts.
B. same problem in handling some system calls (at least the brk system
calls always records an origin, even when --track-origins=yes).
C. The Valgrind unwinder cannot properly unwind some stack traces.
It unwinds a few frames, then go bezerk and stops at a "random" IP.
This then causes the same "logical" stacktrace to be truncated
and records thousands of times with this "differentiating" last IP.
For problem cause 2 above ( a lot of partially initialised bytes),
the idea is to similarly add another gdbsrv commands that will output
statistics about which stack traces are causing a lot of uninitialised bytes.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@13220
This limit is large enough for all practical purposes. It exists
only to sanity check the value specified with --num-callers.
Be frugal in record_ExeContext_wrk and only allocate on the stack
as many frames as needed.
Testcase included.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@12685
support to Memcheck for tracking the origin of uninitialised values,
if you use the --track-origins=yes flag.
This currently causes some Memcheck regression tests to fail, because
they now print an extra line of advisory text in their output. This
will be fixed.
The core-tool interface is slightly changed. The version number for
the interface needs to be incremented.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@7982
a few places have to #include pub_core_stacktrace.h themselves, but
that's ok because explicit is better.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@3976
which I put the wrong log message in for. The log message should have been:
Clean up VG_(core_panic_at)() -- pass it ip/sp/fp rather than a
StackTrace. This allows m_signals.c to not have to know about
m_stacktrace.c, nor does core.h need to import pub_core_stacktrace.h any
more. (Also fixed a bug that was preventing vg_assert2's extra string
from being printed.)
Also moved VG_DEEPEST_BACKTRACE out of core.h into pub_core_execontext.h.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@3832
two halves: stacktrace.c, which deals with getting, traversing and printing
stack traces; and execontext.c, which deals with storing stack traces
permanently in a way that avoids duplicates, and comparing them.
One nice outcome: previously we were often creating ExeContexts, which live
forever, even when they were only needed temporarily. Ie. this was a memory
leak, which has been removed.
As part of this, new headers have been created, carved off core.h and
tool.h. Lots of function names have changed, too.
In Massif, I also changed a lot of "eip" names to "ip" to make them less
x86-specific.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@3429
