Valgrind-developer notes, re the MacOSX port ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ JRS 22 Mar 09: re these comments in m_libc* and m_debuglog: /* IMPORTANT: on Darwin it is essential to use the _nocancel versions of syscalls rather than the vanilla version, if a _nocancel version is available. See docs/internals/Darwin-notes.txt for the reason why. */ when Valgrind does (for its own purposes, not for the client) read/write/open/close etc syscalls, it really is critical to use the _nocancel versions of syscalls rather than the vanilla versions. This holds throughout the entire code base: whenever V does a syscall for its own purposes, we must use the _nocancel version if it exists. This is of course most prevalent in m_libc* since all of our own-purpose (non-client) syscalls should get routed through there. Why? Because on Darwin, pthread cancellation is done within the kernel (unlike on Linux, iiuc). And read/write/open/close and a whole bunch of other syscalls to do with stream I/O are cancellation points. So what can happen is, client informs the kernel that a given thread is to be cancelled. Then at the next (eg) VG_(printf) call by that thread, which leads to a sys_write, the write syscall gets hit by the cancellation request, and is duly nuked by the kernel. Of course from the outside it looks as if the thread had mysteriously disappeared off the radar for no reason. In short, we need to use _nocancel versions in order to ensure that cancellation requests only take effect at the places where the client does a syscall, and not the places where Valgrind does syscalls. How observed: using the standard pipe-based implementation in coregrind/m_scheduler/sema.c, none/tests/pth_cancel1 would hang (compared to succeeding using native Darwin semaphores). And if the "pause()" call in said test is turned into a spin ("while (1) ;") then the entire Valgrind run mysteriously disappears, rather than spinning using native Darwin semaphores. Because the pipe-based semaphore intensively uses sys_read/sys_write, it is not surprising that it inadvertantly was eating up cancellation requests directed to client threads. With abovementioned change in force the pipe-based semaphore appears to work correctly. Valgrind-developer notes, things removed from the original MacOSX port ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ There was a broken debugstub implementation. It was removed over several commits: r9477, which removed most of it, and r9711, r9759, and r10012, which cleaned up remaining bits. There was machinery to read function names from Dwarf3 debug info. But we already read function names from the symbol tables, so this was duplicated functionality. Furthermore, a Darwin-specific hack was required in storage.c to choose between symbol table names vs. Dwarf3 names. So this machinery was removed in r10155. Valgrind-developer notes, todos re the MacOSX port ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ * m_syswrap/syscall-amd64-darwin.S - correct signal mask is not applied during syscall - restart-labels are completely bogus * m_syswrap/syswrap-darwin.c: - PRE(sys_posix_spawn) completely ignores signal issues, and also ignores the file_actions argument * env var handling w/ exec on Darwin: is there something odd? Compare "valgrind env" on Darwin and Linux. On the former there are settings VALGRIND_LIB and VALGRIND_LIB_INNER, but not for the former. There's a suspicious-looking "#if defined(VGO_darwin)" in VG_(env_remove_valgrind_env_stuff). Maybe related? * Cleanups: sort wrappers in syswrap-darwin.c and priv_syswrap-darwin.h alphabetically. Also, some aren't properly implemented -- check and print warnings * Cleanups: m_scheduler/sema.c: use pipe implementation (but this apparently causes none/tests/pth_cancel1 to hang. I have no idea why, despite quite some investigation). * Cleanups: m_debugstub: move to attic * syswrap-darwin.c: sys_{f,}chmod_extended: handling of ARG5 is way wrong * Cleanups (Linux,AIX5): bogus launcher-path mangling logic in PRE(sys_execve) * Cleanups (ALL PLATFORMS): m_signals.c: are the _MY_SIGRETURN assembly stubs actually necessary for anything? I don't know. * Cleanups: check that changes to VG_(stat) and VG_(stat64) have not broken 64-bit statting on 32-bit Linux * Cleanups: #if !HAVE_PROC in m_main (to do with /proc/<pid>/cmdline -------- m_main doesn't read symbols for the valgrind exe itself, which is annoying. On minimal investigation it seems that the executable isn't even listed by aspacem. This is very strange and not in accordance with the Linux or AIX ports. m_main: relatedly, Darwin version does not collect/give out initial debuginfo handles; hence ptrcheck won't work m_main: Darwin port relies on blocking out big sections of address space with mmap at startup. We know from history that this is a bad idea. (It's also really slow on 64-bit builds, taking 3--4 seconds.) Also, startup is not done on the interim startup stack -- why not? VG_(di_notify_mmap): Linux version is also used for Darwin, and contains some ifdeffery. Clean up. PRE(sys_fork), #ifdeffery syswrap-generic.c: VG_(init_preopened_fds) is #ifdefd for Darwin scheduler.c: #ifdeffery in VG_(get_thread_out_of_syscall) look at notes in coregrind/Makefile.am re Mach RPC interface definitions. See if we can get rid of any more stuff now that m_debugstub is gone.