summaryrefslogtreecommitdiff
path: root/libc/sysdeps/linux/e1
diff options
context:
space:
mode:
authorKevin Cernekee <cernekee@gmail.com>2012-06-05 15:05:20 -0700
committerBernhard Reutner-Fischer <rep.dot.nop@gmail.com>2012-06-06 09:36:34 +0200
commit2952c70804b48bb5c87eea21df5e401969dc4ec1 (patch)
tree4b76809d374429ba398bb1e266ac02528ebd39f2 /libc/sysdeps/linux/e1
parent3c58d95d918c7e2fda374c37a52f81b34b81e4ca (diff)
MIPS: Use $a0 instead of $v0 for __syscall_error() argument
$a0 is saved across _dl_runtime_resolve(); $v0 is not. Unfortunately, __syscall_error() uses $v0 for its argument, not $a0 as is the MIPS ABI standard. This means that if lazy binding was used for __syscall_error(), the errno value in $v0 could get corrupted. The problem can be easily seen in testcases where syscalls in librt fail; when librt tries to call __syscall_error() in libc, the argument gets lost and errno gets set to a bogus value: # ./tst-mqueue1 ; echo $? mq_receive on O_WRONLY mqd_t did not fail with EBADF: Unknown error 2004684208 1 # ./tst-mqueue2 ; echo $? mq_timedreceive with too small msg_len did not fail with EMSGSIZE: Unknown error 1997360560 1 # ./tst-mqueue4 ; echo $? mq_timedsend did not fail with ETIMEDOUT: Unknown error 2008747440 1 When _dl_runtime_resolve() was taken out of the equation, the same test cases passed: # LD_BIND_NOW=y ./tst-mqueue1 ; echo $? 0 # LD_BIND_NOW=y ./tst-mqueue2 ; echo $? 0 # LD_BIND_NOW=y ./tst-mqueue4 ; echo $? 0 Changing __syscall_error() to look at $a0 instead of $v0 fixed the problem. (Note that there is also a "__syscall_error.c" file which presumably uses the standard C calling conventions, but I do not think it is used on MIPS.) Signed-off-by: Kevin Cernekee <cernekee@gmail.com> Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
Diffstat (limited to 'libc/sysdeps/linux/e1')
0 files changed, 0 insertions, 0 deletions