On Wed, Mar 05, 2008 at 03:48:05PM +0200, Nikodemus Siivola wrote: > On 3/5/08, Aurelien Jarno <[EMAIL PROTECTED]> wrote: > > Nikodemus Siivola a écrit : > > > > > On 3/5/08, Debian Bug Tracking System <[EMAIL PROTECTED]> wrote: > > > > > >> tag 469058 + patch > > >> Bug#469058: sbcl doesn't reset direction flag upon exit > > >> There were no tags set. > > >> Tags added: patch > > > > > > Thanks for the patch, but... while I agree that it is good to change > > > SBCL to reset the direction flag every time it is diddled, instead of > > > just before calling C, I don't think SBCL is actually at fault here. > > > > > > 1. SBCL does actually reset DF before any call to foreign (GCC > > generated) code. > > > See line 236 in src/compiler/x86/c-call.lisp, and line 125 in > > > src/runtime/x86-assem.S. > > > > > > (It is possible I'm missing out a call-path here, but even so, read > > on and > > > see if my fears are unfounded or not.) > > > > > > 2. If the problem was due to a foreign call, it should be deterministic. > > > > > > 3. If the problem was due to _returning_ to main(), it should be > > deterministic. > > > > > > Looks correct. > > > > > > > What I suspect is actually going on (especially considering your > > > statement that compiling signals/ with 4.2 avoided the issue) is that > > > a signal handler is entered while DF is set. > > > > > > What I am sure is that sigemptyset() from the glibc is called with the > > direction flag set, and that should not happen. > > Right. > > I'm about to merge a patch to SBCL based on yours, which moves all DF > resets to immediate vicinity of STDs for easier auditing, and removed > the then-unnecessary CLD instructions from foreign call sequences. > This will fix them symptoms, and be good for SBCL, but I think the > underlying problem is still there in signal handling. :/ > > > > If this is the case, then clearing it right after each REP loop where > > > SBCL uses it just makes seeing the bug much more unlikely -- but not > > > impossible in the presence of async signals. > > > > > > Seems correct, though I have made half a dozen of build here, without > > any problem. > > That is not too suprising: the are normally no asynch signals > delivered during the build, but SIGSEGV is a regular occurance (it is > used by the GC), so SIGSEGV handlers may have been seeing the DF set. > > What _is_ strange is that this appears to have been random. (At least > all the reporters seemed to characterize it as semirandom behaviour.) > Multiple builds from the same source with the same host compiler > should have essentially identical GC characteristics. > > > > If so, this may also explain some _very_ hard to reproduce faults we > > > have seen over the years: using a pre 4.3-GCC compiled libc, a signal > > > at an in opportune moment in the middle of a REP loop could clear DF! > > > Yikes! > > > > > > I'm not sure what is The Right Thing here, though. Should SBCL (and > > > _any_ program that ever sets DF!) save, clear, and restore DF in its > > > signal handlers? Should libc/kernel do that? Should signals be blocked > > > > > > I currently have no idea about that. > > I'll see if I can cook up a small test-case using async signals. (One > that doesn't need SBCL so that it can be passed to upstream libc / > kernel people if necessary without too much friction.) >
The small code below exhibits the problem. It was there already with gcc-4.2, but in that case, gcc generates a cld or std instruction before any instruction that uses the direction flag. #include <stdint.h> #include <stdlib.h> #include <stdio.h> #include <signal.h> void handler(int signal) { uint64_t rflags; asm volatile("pushfq ; popq %0" : "=g" (rflags)); if (rflags & (1 << 10)) printf("DF = 1\n"); else printf("DF = 0\n"); } int main() { signal(SIGUSR1, handler); while(1) { asm volatile("std\r\n"); } } -- .''`. Aurelien Jarno | GPG: 1024D/F1BCDB73 : :' : Debian developer | Electrical Engineer `. `' [EMAIL PROTECTED] | [EMAIL PROTECTED] `- people.debian.org/~aurel32 | www.aurel32.net -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]