On Wed, Oct 14, 2015 at 11:03 AM, Adhemerval Zanella <adhemerval.zane...@linaro.org> wrote: > > > On 14-10-2015 04:54, Jakub Jelinek wrote: >> On Tue, Oct 13, 2015 at 07:54:33PM +0300, Maxim Ostapenko wrote: >>> On 13/10/15 14:15, Maxim Ostapenko wrote: >>>> This is the raw merge itself. I'm bumping SONAME to libasan.so.3. >>>> >>>> -Maxim >>> >>> I have just noticed that I've misused autoconf stuff (used wrong version). >>> Here a fixed version of the same patch. Sorry for inconvenience. >> >> Is libubsan, libtsan backwards compatible, or do we want to change SONAME >> there too? >> >> The aarch64 changes are terrible, not just because it doesn't yet have >> runtime decision on what VA to use or that it doesn't support 48-bit VA, >> but also that for the 42-bit VA it uses a different shadow offset from >> 39-bit VA. But on the compiler side we have just one... >> Though, only the 39-bit VA is enabled right now by default, so out of the >> box the state is as bad as we had in 5.x - users wanting 42-bit VA or 48-bit >> VA have to patch libsanitizer. > > Yes we are aware with current deficiencies for aarch64 with a 39-bit and > 42-bit vma and the lack of support of 48-bit vma. On LLVM side current > approach is to built compiler support for either 39 or 42 bit (and again > we are aware this is not the ideal approach). This approach was used > mainly for validation and buildbot enablement. > > I am currently working on a way to make the sanitizer (asan and msan) > VMA independent on aarch64 (aiming to current support 39/42 and later > 48-bit vma) and the approach we decide to use is to change the > instrumentation to use a parametrized value to compute the shadow memory > regions (based on VMA address) which will be initialized externally > by the ibsanitizer. TSAN is somewhat easier since instrumentation > does not take in consideration the VMA (only the libsanitizer itself).
Wait. As Jakub correctly pointed out in the other thread, there is no obvious reason why there could not be a single shadow offset value that would work for all 3 possible VMA settings. I suggest figuring this out first. > > The idea is to avoid compiler switches a make is transparent to run > the binary regardless of the VMA in the system. The downside is > instrumentation will require more steps (to lead the parametrized > value to compute shadow memory) and thus slower. > > > >> >> Have you verified libbacktrace sanitization still works properly (that is >> something upstream does not test)? >> >> Do you plan to update the asan tests we have to reflect the changes in >> upstream? >> >> Jakub >>
