latrace May 2011

latrace@lists.stg.fedorahosted.org

3 participants
5 discussions

Thoughts and quirks
by David Gilbert 02 Jun '11

02 Jun '11

Hi, I thought I'd just summarise some of the latrace quirks that I'd hit over the last few weeks of using it; I'm not sure yet whether all of these are ARM specific (I've not had the chance to try and replicate them all on x86) 1) C++ programs that throw exceptions don't seem to like running under latrace (povray in SPEC 2k6 in particular); it fails with an exception saying that its terminating 'after throwing an instance of 'int''; it looks like it normally does throw exceptions, and it seems to be misreading which exception it is. I found adding: -n_Unwind_RaiseException,__cxa_get_globals,__cxa_throw,__cxa_allocate_exception,__cxxabiv1::__cxa_allocate_exception to latrace seems to help. 2) I'd also had problems with setjmp - I think this was in perl (in the spec wrapper); adding -nlongjmp,siglongjmp,_longjmp,__longjmp_chk,__libc_longjmp,__libc_siglongjmp 3) I don't think the output is opened with O_LARGEFILE, I found that after the 4GB log file got truncated (I now pipe to bzip2) 4) Given (1) and (2) and a few other quirks, I hit the limit on the length of the symbol list (about 200 chars); unfortunately it just fails silently rather than telling you this. 5) For my purposes I modified the code that prints args to print the address and the length of strings passed as arguments; quite useful to see where the strings are coming from. I've got a few other quirks I've not been able to get to the bottom of yet (e.g. one program that I think is stuck in a loop when run under latrace but I've not figured out why); but it's managed to get me probably a few 100GB of logs - thanks! I'd be intereste d to know if anyone has seen (1) or (2) on any other platform, else we need to go digging on ARM and figure out why. Dave

3 6

latrace tests on ARM
by David Gilbert 25 May '11

25 May '11

Hi, I enabled the tests for arm in configure.ac, copied the test/sysdeps/i686 to arm and tried it (I ran test/test.sh from the top level - is that the intended way?) test_char_3 failed: expected pattern='test_char_3$a = 0x01, b = 0xfe, c = 0x07$' received='test_char_3(a = 0x01, b = 0xfe '�', c = 0x07) [/discs/more/daveg/git/latrace/lib-test-args.so] { } test_char_3 = 0xf7 '�' (I modified re_test to print the expected/received). The problem here is that on ARM 'char' is unsigned by default where as on most other platforms it's signed; the code in src/args.c does: if (*((char*) pval) <= ' ') len = snprintf(argbuf, alen, "0x%02x", *((unsigned char*) pval)); else len = snprintf(argbuf, alen, "0x%02x \'%c\'", *((unsigned char*) pval), *((char*) pval)); on ARM the 0xfe is treated as a positive value and hence the test prints the character out. Is there a reason that you didn't use isprint() - perhaps worried about the difference between locales? I'm also getting result mismatches on test_char_6 (getting 0x0e instead of the expected 0x6e). I haven't worked through the math of teh code in test_char_6, but with all those divides and stuff I wouldn't be surprised if this is also a signedness issue. I also had to change the shell in test/test.sh to /bin/bash - the /bin/sh I have seems to be dash which doesn't have functions. Another suggestion is not to separate test_long into CPU specific sysdeps; there are only a few combinations and they'll be shared among lots of CPUs - i686 and ARM are the same for those tests. Dave

2 10

destination strings and arguments
by David Gilbert 25 May '11

25 May '11

Hi, On library functions like strcpy() where there is a destination string, latrace -A ends up trying to print out the destination buffer which may be completely invalid. Maybe it's worth doing the same trick that ltrace does, and just declare the destination parameters as void* Dave

2 5

control file leak
by Akos PASZTORY 25 May '11

25 May '11

Hi, when you Ctrl-C latrace, the control dir/file stays around: $ ./latrace -q /bin/cat controled config: /tmp/lt-config-Bl4RzW/config <Ctrl-C> $ ls -d /tmp/lt-config-* /tmp/lt-config-Bl4RzW/config This happens because the signal terminates latrace itself in lt_run(). It could maybe install a SIGINT handler that sets a flag and wrap the waitpid() into an loop that checks for EINTR plus this flag... -- akos

2 7

[Patch] Build fix for ARM
by David Gilbert 14 May '11

14 May '11

Hi, I had a build problem with latrace on ARM in that it didn't recognise that I was on ARM and thus didn't compile in argument support. I tracked it back to our uname being 'armv7l' rather than 'arm' which it was testing for, so I've got the tiny patch below. (I'm curious actually which ARM system has a uname -m that is 'arm'). (This was against a version from a couple of weeks ago - the latrace git repo seems to be down). Dave ---------------------------------------------------------- diff --git a/configure.ac b/configure.ac index 966f6bb..b3a2ec0 100644 --- a/configure.ac +++ b/configure.ac @@ -74,7 +74,7 @@ AC_TYPE_UINT16_T AC_TYPE_UINT8_T date=`date "+%B %G"` -unamem=`uname -m | sed 's/i.86/i686/'` +unamem=`uname -m | sed -e 's/i.86/i686/' -e 's/^armv.*/arm/'` AC_SUBST(CONFIG_DATE, "$date") AC_SUBST(CONFIG_SYSDEP_DIR, "$unamem")

3 5

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

latrace May 2011