Gilles Filippini a écrit le 04/10/2015 22:43 : > Samuel Thibault a écrit le 04/10/2015 21:19 : >> Hello, >> >> Gilles Filippini, le Sun 04 Oct 2015 19:48:51 +0200, a écrit : >>> I think the cause of this bus error is the upgrade of glibc from >>> 2.19-19 to 2.19-22 because building with libc0.3 from glibc 2.19-19 is >>> successful. >>> >>> Using bisection on a porterbox I've found out that the guilty commit is >>> r6489 [3]. >> >> Thanks for the investigation! >> >> That change indeed changes the implementation of raise(). Actually it's >> even a fix, but apparently there are side effects, I'll have a look. > > I've had a closer look at H5detect.c and from what I understand it > relies heavily on C macros such as H5_HAVE_SIGSETJMP and > H5_HAVE_SIGLONGJMP. And it happens that these macros aren't defined > anywhere when using the autotools build system, while they are taken > care of by the cmake alternative build system. > > Setting these macros by hand, H5detect doesn't fail on Bus error anymore. > > Thus I think part of the problem lies in the hdf5 build configuration. Actually no. It appears that H5_ prefixed macro are correctly defined into src/H5pubconf.h. This file is created from src/H5config.h by a configure snippet which prefixes each macro with H5_. H5_HAVE_SIGSETJMP isn't defined because: configure:25814: checking for sigsetjmp configure:25814: /usr/bin/gcc -o conftest -g -O2 -fstack-protector-strong -Wformat -Werror=format-security -D_FILE_OFFSET_BITS=64 -D_FORTIFY_SOURCE=2 -Wl,-z,relro conftest.c -lpthread -lz -ldl -lm >&5 /tmp/ccEVZwR6.o: In function `main': /home/pini/hdf5-1.8.15-patch1+docs/debian/build/conftest.c:160: undefined reference to `sigsetjmp' collect2: error: ld returned 1 exit status It's weird that sigsetjmp isn't correctly detected but it's not the side effect we are looking for, because successful build logs have "checking for sigsetjmp... no" as well Can't go further right now because exodar.debian.net tells me "Computer bought the farm". I hope I'm not the cause of this system failure with my repeated signal handling tests :/ Samuel Thibault a écrit le 05/10/2015 00:20 :> I'm building the newest hdf5 against libc0.3 2.19-19+b1 so that the > binNMUs can take place, but I'll continue investigating. Thanks! _g.
Attachment:
signature.asc
Description: OpenPGP digital signature