[uclibc-ng-devel] bug: stdio funcs, thread cancellation with futex hold
ysionneau at kalray.eu
Thu Sep 26 17:06:51 CEST 2019
I would like to know if there is consensus on the fact that this program
hanging on a deadlock is a uClibc-ng bug: https://pastebin.com/11qLsTW5
I've asked people to try this on several arch/libc combination, it works
well (no hang) on:
Also, I've run the puts.c test (linked above) on 3 archs with uClibc-ng,
it fails (hangs) on:
* k1c/uClibc-ng (port not completely published yet)
See failure logs + strace -f of what happens when hanging:
My understanding of the issue is that:
puts, and possibly other libc functions, are taking a lock (
) and end up calling write() which is a cancellation point. (
So, if a thread is canceled, is asynchronous mode (which is the default
one), and the cancelation is triggered by the write() inside the puts(),
then the thread will unwind and exit without unlocking the puts lock.
Then, any other thread calling puts() will hang indefinetely (and hang
other threads if it hangs with locks held...).
My understanding of what can be done to fix this issue:
1/ Either make puts a non cancelation point (see man 7 pthreads, puts is
not listed in mandatory cancelation point, only in "may"). For instance
it is not a cancelation point in glibc.
2/ Or keep puts as a cancelation point and fix the puts code so that it
releases the lock upon cancelation (using pthread_cleanup_push/pop for
In case people think this is indeed a bug, here are examples of code
fixes that I have in mind, please don't hesitate to comment or/and
propose something else:
One problem of those fixes is that we need to identify all libc
functions that take a lock and call a cancellable function and apply
such kind of fixes... This is not easy and a bit painful.
More information about the devel