Robert Watson wrote:
> On Fri, 15 Nov 2002, John Baldwin wrote:
> > It only happens with P4's.  I haven't seen it locally on a p4 test
> > machine at work that I have built test releases on.  Also, it would be
> > nice to see if just adding one of the options fixed the problems.  As
> > for NOTES, those options should not be enabled in NOTES as they would
> > defeat the purpose of LINT since they disable code.
> Does this apply generally to all P4's, or just a subset?  If all, it may
> be we want to add a P4-workaround to GENERIC so that P4's work better ouf
> of the box.  If it's a select few, I wonder if there's some way to test
> for the problem early in the boot...
> One of the recurring themes here has (a) been P4 processors, and (b) been
> a fear that because of timing changes introduced by the DISABLE_FOO flags,
> the real bug is still there, but less visible in the tests people are
> running.

The amount of RAM will also affect it.  It can also happen on P3's
and AMD K6's.  It is a CPU bug related to the use of 4M pages.

Bosko understands the problem (I have explained it to him under
non-disclosure), and he has a patch which avoids it without really
disclosing the problem, which I'm OK with.  Using the patch cranks
the amount of base memory required for a minimal FreeBSD up to 16M,
and loads the kernel at 4M, instead of 1M.  This avoids the problem
on purpose that the older FreeBSD locore.s used to avoid by accident.

The alternative is to take up to a 15% performance hit by not using
4M and global pages, or to revert the locore.s code so that it does
not tickle the hardware bug.

-- Terry

To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-current" in the body of the message

Reply via email to