Re: glibc 2.19 - asyn-signal safe TLS and ASan.

[Andrew Hunter <ahh at google dot com>]

FYI -- you could easily do the same thing with calls to
signal_safe_memalign from libc as a stopgap.  (Well, we'd need to
export the symbols from libc, like I wanted to in the first place.)

On Thu, Jan 23, 2014 at 7:35 AM, Kostya Serebryany <kcc@google.com> wrote:
> [plain text now]
>
> Thanks again for reaching us.
> The new TLS implementation *is* a problem for us as it turns out.
> At least for LeakSanitizer (lsan) it will cause a false positive leak report.
> Admittedly, the current lsan's implementation has an ugly hack around TLS,
> which was the major reason for filing
> https://sourceware.org/bugzilla/show_bug.cgi?id=16291
> In short, we treat __libc_memalign called from elf/dl-tls.c in a special way
> which allows us to include the dynamic TLS into the leak detector's
> memory root set.
> More:
> http://llvm.org/viewvc/llvm-project/compiler-rt/trunk/lib/lsan/lsan_common_linux.cc?view=diff&r1=199899&r2=199900&pathrev=199900
>
> I suggest we continue the discussion in
> https://sourceware.org/bugzilla/show_bug.cgi?id=16291
> unless you prefer otherwise. Let me post more details there.
>
> I also suspect that the new TLS implementation may cause us trouble in
> MemorySanitizer,
> but I haven't tried to verify that yet.
>
>
> --kcc
>
>
>
> On Tue, Jan 21, 2014 at 4:27 AM, Carlos O'Donell <carlos@redhat.com> wrote:
>> Konstantin,
>>
>> I've forwarded your response to libc-alpha which I assume rejected
>> your multi-part plain/html email.
>>
>> I've also corrected the small mistake that the next release is 2.19
>> not 2.20. Sorry.
>>
>> +address-sanitizer@googlegroups.com
>>
>> Hi Carlos,
>>
>> Thanks for the heads up!
>> I don't expect any impact on ASan from this change.
>> We'd still test ASan with the new glibc to make sure.
>>
>> --kcc
>>
>>
>>
>> On Sat, Jan 11, 2014 at 7:39 AM, Carlos O'Donell <carlos@redhat.com> wrote:
>>
>>     Hello Konstantin!,
>>
>>     You're getting this email because you're the only ASan expert I know,
>>     and I was at your talk at LFCS2013 ;-)
>>
>>     We have a problem and we'd like your input if you have time.
>>
>>     The GNU C Library version 2.20 (coming out at the end of the month)
>>     plans to stop using malloc for TLS allocations. The reason for this
>>     is that malloc is async-signal unsafe, and TLS accessed in signal
>>     handlers may need to allocate storage at the time of access. This
>>     is particularly true of signal handlers provided by dlopened shared
>>     libraries. There is no way to interpose yourself here because the
>>     non-malloc signal-safe allocator being used is internal to glibc.
>>
>>     What kind of impact do you see this having on ASan?
>>
>>     Do you see any way we can mitigate this impact?
>>
>>     Cheers,
>>     Carlos.
>>