From: David Miller <davem@davemloft.net>
Date: Fri, 02 Jun 2017 14:39:06 -0400 (EDT)

From: "Darrick J. Wong" <redacted>
Date: Fri, 2 Jun 2017 11:08:08 -0700

quoted

ext4/jbd2's crc32c implementations will also need a fix like this for
{ext4,jbd2}_chksum.  Note that both of these modules call the crypto api
directly to avoid a static dependence on libcrc32c; this was done to
reduce kernel footprint for applications that don't need it.  (ext2,
ext3, and ext4 before the metadata_csum feature existed).

Good catch.  I even looked at that code the other day so should
have spotted it as well.

I'll integrate fixes for those into my patch when I get a chance,
thanks!

Actually, ext4 doesn't trigger the problem because the on-stack object
used in ext4 is a fixed size at compile time.  Which is technically an
ill-advised assumption to make.  Even the generic libcrc32c.c doesn't
assume that the context area is 4 bytes for crc32c.

Anyways, back to the main point, the gcc bug only triggers when
alloca() like constructs are used.

That's why I scanned for SHASH_DESC_ON_STACK() to see exactly where
the workaround is necessary.