On 06/23/2017 01:48 AM, Jan Stancek wrote: > > > ----- Original Message ----- >> On Wed, May 24, 2017 at 08:46:57AM -0400, Jan Stancek wrote: >>> >>> >>> ----- Original Message ----- >>>> Hi, >>>> >>>> I'm seeing rare crashes during NFS cthon with krb5 auth. After >>>> some digging I arrived at potential problem with sha1-avx2. >>> >>> Adding more sha1_avx2 experts to CC. >>> >>>> >>>> Problem appears to be that sha1_transform_avx2() reads beyond >>>> number of blocks you pass, if it is an odd number. It appears >>>> to try read one block more. This creates a problem if it falls >>>> beyond a page and there's nothing there. >>> >>> As noted in my reply, worst case appears to be read ahead >>> of up to 3 SHA1 blocks beyond end of data: >>> http://marc.info/?l=linux-crypto-vger&m=149373371023377 >>> >>> +----------+---------+---------+---------+ >>> | 2*SHA1_BLOCK_SIZE | 2*SHA1_BLOCK_SIZE | >>> +----------+---------+---------+---------+ >>> ^ page boundary >>> ^ data end >>> >>> It is still reproducible with 4.12-rc2. >> >> Can someone from Intel please look into this? Otherwise we'll have >> to disable sha-avx2. > > So I take it my workaround patch [1] is not acceptable in > short-term as well? > > [1] http://marc.info/?l=linux-crypto-vger&m=149373371023377 > > Regards, > Jan > Megha, Can you take a look at this issue? Thanks. Tim
