Created attachment 2071 [details] test program triggering the bug errno access doesn't seem to be threadsafe using linuxthreads / linuxthreads.old. There has been several reports of it on the list, E.G.: http://lists.uclibc.org/pipermail/uclibc/2010-May/044049.html It's easy to reproduce, simply compile and run the attached test program. I'm expecting to see no output (which is what I get if I run it under glibc), but get lots of errno=0 lines instead of the expected EBADF / ENOSPC. If I apply the patch in the mail above, I no longer see any errno=0 lines, but the threads sometimes gets mixed up (E.G. errno=EBADF instead of ENOSPC and the reverse). Tested on ARM9 (EABI) and x86-64 using linuxthreads and linuxthreads.old with the default buildroot config (http://git.buildroot.net/buildroot/tree/toolchain/uClibc/uClibc-0.9.31.config). I also see this on 0.9.30.3, so it isn't a new issue.
yes there seems to be generic problem with errno handing. Lets see what we can do.
I testing it on nptl. So far works fine on arm. However on mips it gives similar errors. 1035654432: (25/25) got 1035654564 instead of 9 1572525344: (20/20) got 1572525476 instead of 28 I will see whats wrong with mips/nptl
(In reply to comment #2) > > I will see whats wrong with mips/nptl and I will see how it goes on sh/nptl
I have fixed it for mips/nptl with commit 8b48f745be1e086d6e486bbb8167e770f3d1fbc5
For this problem I found a fix here: http://github.com/mat-c/uClibc/commits/master_fixes Look for "don't hide __errno_location and __h_errno_location". I'm not sure if the fix is correct but actually it works for 0.9.31 on ARM.
(In reply to comment #5) > For this problem I found a fix here: > > http://github.com/mat-c/uClibc/commits/master_fixes > > Look for "don't hide __errno_location and __h_errno_location". > > I'm not sure if the fix is correct but actually it works for 0.9.31 on ARM. It works as well here, thanks. It does seem correct to me, as we WANT __errno_location / __h_errno_location to be weak and overridable, so the linuxthread version is used instead when using -pthread.
(In reply to comment #6) > (In reply to comment #5) > > For this problem I found a fix here: > > > > http://github.com/mat-c/uClibc/commits/master_fixes > > > > Look for "don't hide __errno_location and __h_errno_location". > > > > I'm not sure if the fix is correct but actually it works for 0.9.31 on ARM. > > It works as well here, thanks. > > It does seem correct to me, as we WANT __errno_location / __h_errno_location to > be weak and overridable, so the linuxthread version is used instead when using > -pthread. the problem is that we should not genetate __GI__* symbols when generating non PIC .a file currently there are refrences to these __GI symbols in libc so when you link statically then the references in libc get resolved to libc's __GI_ errno symbols unhiding these symbols as patch does may not be the right solution. think of shared objects it will have undefined behavior when both libc and libpthread are loaded. it will depend on load order which one is chosen I think there is need to clean up the build for pic and non pic build of libc.a look into this area
(In reply to comment #7) > the problem is that we should not genetate __GI__* symbols when generating non > PIC.a file currently there are refrences to these __GI symbols in libc so when > you link statically then the references in libc get resolved to libc's __GI_ > errno symbols Hmm, isn't this exactly like how we do it for malloc/free? I don't see any references to __GI_malloc or __GI__errno_location when I like statically? > unhiding these symbols as patch does may not be the right solution. think of > shared objects it will have undefined behavior when both libc and libpthread > are loaded. > it will depend on load order which one is chosen Also when the libc version is marked as weak? I haven't seen that issue. Libc gets added very early to the linker, so the libpthread version seems to always override it. > I think there is need to clean up the build for pic and non pic build of libc.a > look into this area So what exactly do you suggest?
(In reply to comment #8) > (In reply to comment #7) > > > the problem is that we should not genetate __GI__* symbols when generating non > > PIC.a file currently there are refrences to these __GI symbols in libc so when > > you link statically then the references in libc get resolved to libc's __GI_ > > errno symbols > > Hmm, isn't this exactly like how we do it for malloc/free? yes but lpthread and lc are also controlled by gcc compiler driver and gcc -pthread will break. Essentially it should not depend on the link order when doing shared linking thats why weak symbols. I don't see any > references to __GI_malloc or __GI__errno_location when I like statically? > yeah because they are not declared with hidden_proto and hidden_def > > unhiding these symbols as patch does may not be the right solution. think of > > shared objects it will have undefined behavior when both libc and libpthread > > are loaded. > > it will depend on load order which one is chosen > > Also when the libc version is marked as weak? I haven't seen that issue. Libc > gets added very early to the linker, so the libpthread version seems to always > override it. > > > I think there is need to clean up the build for pic and non pic build of libc.a > > look into this area > > So what exactly do you suggest? We need to say that when compiling static libc.a then dont use this hidden_proto and hidden_def because it does not matter in that case. but when compiling PIC version then we use it.
(In reply to comment #9) > > So what exactly do you suggest? > > We need to say that when compiling static libc.a then dont use this > hidden_proto and hidden_def because it does not matter in that case. > > but when compiling PIC version then we use it. But that means we're back to square one, right? Today, with a dynamic libc and -pthread we end up using the __errno_location in libc, not in libpthread.
(In reply to comment #10) > (In reply to comment #9) > > > So what exactly do you suggest? > > > > We need to say that when compiling static libc.a then dont use this > > hidden_proto and hidden_def because it does not matter in that case. > > > > but when compiling PIC version then we use it. > > But that means we're back to square one, right? Today, with a dynamic libc and > -pthread we end up using the __errno_location in libc, not in libpthread. the libc one's will be weak symbols and libpthread ones will be strong so when linking libpthread.so will be preferred.
(In reply to comment #11) > > But that means we're back to square one, right? Today, with a dynamic libc and > > -pthread we end up using the __errno_location in libc, not in libpthread. > > the libc one's will be weak symbols and libpthread ones will be strong so when > linking libpthread.so will be preferred. Yes, but that's not what happens as it is (libc version weak + hidden) - The libc version is used within libc because of the hidden. With my patch to remove the hidden it works.
Without this patch, my system would be broken. I am using uClibc-0.9.31 with a PowerPC-based arch (Linux 2.6.30). Without this patch, I would get errno==0 when read() returned -1. The code doesn't handle such a case (and it shouldn't need to!) and so results in the system locking up. I've upped the Priority to P2 Critical. The patch makes the system behave correctly.
This bug affects mpd (it can't populate database), and I suspect that it affects libao+alsalib (it just segfaults when using libao with alsa as output). Tested with uclibc-0.9.31 and uclibc-0.9.32-rc2 via buildroot-2011.02. So, mpd and libao are unusable in mentioned buildroot version.
(In reply to comment #14) > This bug affects mpd (it can't populate database), and I suspect that it > affects libao+alsalib (it just segfaults when using libao with alsa as output). > Tested with uclibc-0.9.31 and uclibc-0.9.32-rc2 via buildroot-2011.02. > > So, mpd and libao are unusable in mentioned buildroot version. Just tested, with mentioned patch, mpd and libao+alsa problems just disappear.
This bug was also causing stunnel to randomly drop SSL connections in the middle of the SSL handshake with the message "SSL_accept: Peer suddenly disconnected" The patch above seems to have fixed the issue.
This works with NPTL.