[lttng-dev] RFC: Fix crash in dlerror()

[amonakov@ispras.ru (Alexander Monakov)]

On Thu, 13 Feb 2014, Stefan Seefeld wrote:
> Our compilation unit defines a bunch of functions with external linkage,
> which access cur_alloc. And since gcc has no way to rule out that the
> call to dlsym() will not cause any of these functions to be called, it
> mustn't make any assumptions about whether or not the first
> initialization of cur_alloc is redundant or not, and thus shouldn't
> elide it.

Stefan, LTTng developers,

The problem here is that glibc declares dlsym() with __attribute__((leaf))
(see the definition of __THROW in /usr/include/sys/cdefs.h and the difference
from __THROWNL).  Presence of the attribute allows the compiler to assume that
no functions from the current compilation unit will be called from dlsym, and
thus there's no need to write back potentially escaping data.

glibc used to have __THROW annotations on dlopen() as well, and they changed
it to __THROWNL (removing the "leaf" attribute) after it was pointed out that
dlopen will call constructors:

https://sourceware.org/ml/libc-alpha/2013-08/msg00465.html

Unfortunately when I pointed out that dlsym is not really "leaf" as well, my
argument was dismissed:

https://sourceware.org/ml/libc-alpha/2013-09/msg00012.html

Please consider filing a glibc bug or otherwise reopening that discussion.

Hope that helps,
Alexander