Re: System freeze w/display strangeness on amd64/current

[Gary Duzan <gary%duzan.org@localhost>]

In Message <814af848282deef0f1b462c9515a728f.squirrel%www.duzan.org@localhost>,
   gary@duzan.orgwrote:

=>=> On Mon, 2 Jan 2012 09:13:37 -0800
=>=> Chuck Silvers <chuq%chuq.com@localhost> wrote:
=>=>
=>=>> On Tue, Dec 27, 2011 at 07:52:29AM -0500, Gary Duzan wrote:
=>=>> > In Message <20111225114643.GD11619%mail.duskware.de@localhost>,
=>=>> >    Martin Husemann <martin%duskware.de@localhost>wrote:
=>=>> >
=>=>> > =>On Fri, Dec 23, 2011 at 08:20:51PM -0500, Gary Duzan wrote:
=>=>> > =>>    Any thoughts for diagnosing the problem? Loading older
=>=>> > versions =>> would be possible, if a bit messy. dmesg included
=>=>> > below. =>
=>=>> > =>Just a very tiny straw, but you could try to comment out the
=>=>> > =>
=>=>> > =>#define __HAVE_DIRECT_MAP 1
=>=>> > =>#define __HAVE_MM_MD_DIRECT_MAPPED_IO
=>=>> > =>#define __HAVE_MM_MD_DIRECT_MAPPED_PHYS
=>=>> > =>
=>=>> > =>block in /usr/src/sys/arch/amd64/include/types.h and build a new
=>=>> > kernel.
=>=>> >
=>=>> >    Thanks for this. I #if 0'd out those lines, and things seem much
=>=>> > more stable now. I ran the full test suite and have been browsing
=>=>> > using Firefox without incident for a while. From the commit note,
=>=>> > I guess the video memory is being grabbed by a large page allocation
=>=>> > improperly?
=>=>>
=>=>>
=>=>> if the physical address of the video memory is less than the highest
=>=>> physical address of normal memory, then it's being mapped by the
=>=>> direct-map, yes.  that shouldn't cause a problem since nothing should
=>=>> be referencing non-RAM addresses via the direct map, but there may be
=>=>> something subtle going on.  openbsd and freebsd also have a direct
=>=>> map that includes the gaps between RAM regions, and they apparently
=>=>> don't have these problems, so something else that we're doing
=>=>> differently must be involved.
=>=>>
=>=>> could you enable "options DEBUG_MEMLOAD" and "options PCI_CONFIG_DUMP"
=>=>> and send me the boot output?  you may also need to increase
=>=>> MSGBUFSIZE.
=>=>
=>=> Just curious: is there a PR filed for this issue, and/or is it known to
=>=> be fixed?
=>
=>   The last time I tried to recreate this problem I ran into other
=>problems before I could trigger this one. It did seem harder to
=>trigger, at least. I have a fresh build at home that I will try to
=>break sometime this weekend. I'm not aware of a PR.

   With yesterday's -current I wasn't able to recreate the problem
by running tests, but less than a minute after starting browsing
using Firefox the system froze with the screen showing vertical
yellow stripes, which is exactly what I was seeing before. I had
a PS/2 keyboard attached this time, and it was also non-responsive.
With the DIRECT_MAP block (plus UCODE) #if'd out the worst I've had
so far is a Firefox core.

                                        Gary Duzan