[PATCH] wcsmbs: Fix data race in __wcsmbs_clone_conv [BZ #24584]
Andreas Schwab
schwab@suse.de
Tue May 21 07:16:00 GMT 2019
On Mai 20 2019, Florian Weimer <fweimer@redhat.com> wrote:
> * Andreas Schwab:
>
>> On Mai 20 2019, Florian Weimer <fweimer@redhat.com> wrote:
>>
>>> diff --git a/wcsmbs/wcsmbsload.c b/wcsmbs/wcsmbsload.c
>>> index 5494d0a23e..e33a9c1312 100644
>>> --- a/wcsmbs/wcsmbsload.c
>>> +++ b/wcsmbs/wcsmbsload.c
>>> @@ -20,6 +20,7 @@
>>> #include <langinfo.h>
>>> #include <limits.h>
>>> #include <stdlib.h>
>>> +#include <stdio.h>
>>> #include <string.h>
>>>
>>> #include <locale/localeinfo.h>
>>> @@ -223,12 +224,24 @@ __wcsmbs_clone_conv (struct gconv_fcts *copy)
>>> /* Copy the data. */
>>> *copy = *orig;
>>>
>>> - /* Now increment the usage counters.
>>> - Note: This assumes copy->*_nsteps == 1. */
>>> + /* Now increment the usage counters. Note: This assumes
>>> + copy->*_nsteps == 1. The current locale holds a reference, so it
>>> + is still there after acquiring the lock. */
>>> +
>>> + __libc_lock_lock (__gconv_lock);
>>> +
>>> + bool overflow = false;
>>> if (copy->towc->__shlib_handle != NULL)
>>> - ++copy->towc->__counter;
>>> + overflow |= __builtin_add_overflow (copy->towc->__counter, 1,
>>> + ©->towc->__counter);
>>> if (copy->tomb->__shlib_handle != NULL)
>>> - ++copy->tomb->__counter;
>>> + overflow |= __builtin_add_overflow (copy->tomb->__counter, 1,
>>> + ©->tomb->__counter);
>>> + if (overflow)
>>> + __libc_fatal ("\
>>> +Fatal glibc error: gconv module reference counter overflow\n");
>>> +
>>> + __libc_lock_unlock (__gconv_lock);
>>
>> Should the lock be dropped before __libc_fatal?
>
> I think this is purely a matter of style because __libc_fatal does not
> return. Do you have a preference?
I think it would be a nice to avoid leaving internal locks locked when
calling abort, in case a SIGABRT handler does something stupid.
Andreas.
--
Andreas Schwab, SUSE Labs, schwab@suse.de
GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE 1748 E4D4 88E3 0EEA B9D7
"And now for something completely different."
More information about the Libc-alpha
mailing list