> I thought the cache-aligned might make sense, since usually we'd expect the
> failure to be at a cache-line level, but our copy_tofrom_user does accurate
> accounting

That's one of the wrinkles in the current x86 memcpy_mcsafe(). It starts by
checking alignment of the source address, and moves a byte at a time until
it is 8-byte aligned. We do this because current x86 implementations do not
gracefully handle an unaligned read that spans from a good cache-line into
a poisoned one.

This is different from copy_tofrom_user which aligns the destination for speed
reasons (unaligned reads have a lower penalty than unaligned writes).

Linux-nvdimm mailing list

Reply via email to