Opteron bugs

Matthew Dillon dillon at apollo.backplane.com
Wed Jun 23 12:49:50 PDT 2004


:Since amd64 is a future target, it may be worthwhile to note these 
:hardware bugs.
:
:from
:http://www.theregister.com/2004/06/22/amd_opteron_glitches/
:
:'AMD has pledged to fix three bugs in its Opteron chips that, if 
:unremedied, could cause host systems to crash in certain circumstances.
:
:In a June 2004 Opteron and Athlon 64 Revision Guide (PDF), the chip 
:maker lists three processor glitches - known in the trade as 'errata' - 
:that could mean the "system may hang".'

    The first one might allow a user to intentionally crash a machine,
    but no normal code will ever generate those circumstances.   

    The second one is s non-starter.  Kernel developers already know all
    about livelocking cpus on mutex spin operations.  No such situations
    exist in DragonFly.

    The rep/movs issue is more problematic but, again, it is not a
    sequence that the compiler normally generates.  The compiler will
    use rep/movs to copy structures, but it does so in the forward
    direction and the bug only occurs with reverse direction copies.

    The FreeBSD kernel bcopy code uses a std/rep/movs sequence but I actually
    ripped that out of DragonFly over a month ago when I redid DFly's
    optimized bcopy/bzero code.

    There are lots of errata (In the low hundreds) for both Intel and
    AMD cpus, but 99.5% of them are related to circumstances that cannot
    occur under normal system operation.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>





More information about the Kernel mailing list