Created attachment 104448 [details] Fix ptrace(traceme) In glibc2.24 implementation of function raise() is changed, so this test crashes when it is run under Valgrind. Signal SIGUSR1 should stop the process, but when it is called with ptrace(traceme), expected behaviour is to stop process. In Valgrind there is no valid support for ptrace(traceme) syscall. I propose fix for this problem.
Created attachment 104449 [details] Test case This is example that shows problem on boards that have version of glibc older then glibc2.24. $ gcc -g traceme.c -o traceme $ ./vg-in-place --tool=none ./traceme ==23581== Nulgrind, the minimal Valgrind tool ==23581== Copyright (C) 2002-2015, and GNU GPL'd, by Nicholas Nethercote. ==23581== Using Valgrind-3.13.0.SVN and LibVEX; rerun with -h for copyright info ==23581== Command: /home/aca/test-primeri/traceme ==23581== ==23582== ==23582== Process terminating with default action of signal 10 (SIGUSR1) ==23582== at 0x4C5CEE0: sigprocmask (sigprocmask.c:39) ==23582== by 0x400736: do_child (traceme.c:17) ==23582== by 0x400758: main (traceme.c:26) ==23582== ==23581== valgrind: the 'impossible' happened: main(): signal was supposed to be fatal host stacktrace: ==23582== at 0x380E15D3: show_sched_status_wrk (m_libcassert.c:378) sched status: running_tid=1
Created attachment 104456 [details] Fix ptrace(traceme)
The patch looks good to me. Anyone has any objections?
Please could you point me to a resource which describes interaction between tracer and tracee with respect to PTRACE_TRACEME. In particular, I'd like to know when flag tst->ptrace (as set in linux_POST_traceme()) gets reset back to 0? Are you going to incorporate the provided test case into Valgrind regression test suite? That would be useful!
I don`t have any specific documentation, I was looking implementation in kernel, file kernel/ptrace.c:339. It looks like flag tst->ptrace is reset only when the tracer exits, I am afraid that we can not implement this in Valgrind. This test is already in test suits memcheck/linux/getregset.c, we just add masking signal before calling raise(), which exists in wrapper for raise in glibc-2.24.
Fair enough. Thank you for the explanation.
(In reply to Aleksandra Karadzic from comment #2) > Created attachment 104456 [details] > Fix ptrace(traceme) The patch committed in r16273. Thank you.
Created attachment 104596 [details] Fix for broken build on Solaris and Mac We have some complains from Solaris and Mac developers about this patch, this is our fix for broken build. Ivo can you test this patch on Solaris?
Thank you for the patch. Indeed, it unbreaks the build. "some complaints" is an euphemism here :-) Valgrind build was effectively broken on two thirds of supported OSes :-)
I have committed the build fix as r16275.