[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: slapd-mtread crash on ppc64{,el} in stretch/sid

hi Ryan,

On Sun, Jul 09, 2017 at 08:56:59AM -0700, Ryan Tandy wrote:
> There seems to be a regression on powerpc64 (both endians) that can corrupt
> the vector-scalar registers (VSRs) in a threaded program.
> I believe the bad commit is this one:
> 4.9.0: https://github.com/torvalds/linux/commit/dc16b553c949e81f37555777dc7bab66d78285a7
> 4.8.6: https://github.com/linux-stable/linux-stable/commit/613036d9e91990f2043130ff8f78fd770432b3de

Nice work!

> My program is not using transactional memory itself, but I don't know what
> libc/libpthread might do internally.

That is interesting. I also found a bug last week related to hardware
transactional memory, and it corrupts vector-scalar register 0 (vs0)
when using hardware transactional memory. In our case, every exception
would cause a wrong reclaim (grab the register values from the
checkpointed area), which will affect later, the recheckpoint (put the
stacked registers into the checkpointed area). This is specific for
hardware transactional memory.

The full description of the bug and a possible patch should be found at:

> 4.8.7-1 is the first Debian kernel where I observe the crash. I rebuilt that
> package with the above commit reverted and it went away.

Anyway, this is what I am going to investigate now:

1) If glibc's pthread method is using hardware transactional memory by
default.  I remember that upstream enabled it once and then disabled by

2) Investigate this commit ID and check for a possible corruption
depending on the result above.

Reply to: