On Fri, Feb 14, 2003 at 11:37:20AM -0700, patrick charles wrote:
>On Thursday 13 February 2003 09:03 pm, David Dawes wrote:
>> On Thu, Feb 13, 2003 at 02:11:40PM -0700, patrick charles wrote:
>> >On Wednesday 12 February 2003 10:20 pm, David Dawes wrote:
>> >> On Tue, Feb 11, 2003 at 02:51:04PM -0700, patrick charles wrote:
>> >> >> On Saturday 08 February 2003 05:41 pm, David Dawes wrote:
>> >> >> > On Sat, Feb 08, 2003 at 01:07:25PM -0700, patrick charles wrote:
>> >> >> > >How would I communicate this? Somebody on XFree86 working with or
>> >> >> > > have contact with the appropriate people in kernel/agpgart
>> >> >> > > development?
>> >> >> >
>> >> >> > First of all, how are you "killing" the X server?  I haven't seen
>> >> >> > this behaviour when the X server exits normally, and I've done a
>> >> >> > lot of testing where 32MB is allocated per run on machines with
>> >> >> > only 128MB of physical memory.
>> >> >> >
>> >> >> > There are people here familiar with the kernel agpgart driver.
>> >> >> >
>> >> >> > Note that just because top shows that there's little memory free
>> >> >> > doesn't mean that the agpgart driver isn't freeing it.  Also the
>> >> >> > agpgart driver allocates physical pages, never swap.  I'm not sure
>> >> >> > what the symptoms are when it can't get any free physical pages. 
>> >> >> > On my test system the free memory indicated by top does go up when
>> >> >> > the X server exits, and this is on an otherwise idle system.
>> >> >> >
>> >> >> > So, I'd suggest starting a bare X server (run just 'X') on an
>> >> >> > otherwise idle system, see what top reports, then exit it cleanly
>> >> >> > (<Ctrl><Alt><Backspace>), and see if the free memory amount
>> >> >> > changes. Check the X server log to confirm how much memory was
>> >> >> > allocated via the agpgart mechanism (look for the lines containing
>> >> >> > "Allocated").
>> >> >> >
>> >> >> > If that looks OK, then try the same thing you tried before but with
>> >> >> > a bare X server and an idle system.
>> >> >> >
>> >> >> > David
>> >> >
>> >> >David,
>> >> >
>> >> >I ran some tests as you suggested. I started up a bare X server using
>> >> > the command 'X' on an idle system. I then exited cleanly using
>> >> > ctrl-alt-bak.
>> >> >
>> >> >I recorded the amount of physical RAM free before and after the X
>> >> > start. I repeated this process.
>> >> >
>> >> >After 13 iterations, the machine became very sluggish.
>> >> >
>> >> >After 16 iterations, the machine hung.
>> >> >
>> >> >Still looks like X (or, the agpgart driver?) is not freeing resources.
>> >> >The machine gradually ran out of physical RAM.
>> >>
>> >> I just tried repeating this with what I think should be an even more
>> >> demanding configuration: 845G system with 128MB physical memory, 1MB
>> >> stolen memory (preallocated video memory), X configured to use 32MB
>> >> video memory, so just over 31MB of physical memory needs to be allocated
>> >> at each server start.
>> >>
>> >> After several iterations, I got to a pattern where the free memory
>> >> after the server starts is 2MB, and the free memory when it exits is
>> >> 41MB.  I went as far as 25 iterations without any change in this pattern
>> >> and without any slowdown.
>> >>
>> >> This is with RH 7.3, using the default kernel plus an agpgart driver
>> >> patched for correct 845G support.  The 2.4.20 kernel should already have
>> >> the correct 845G agpgart support.
>> >>
>> >> The source for the agpgart driver I'm using can be found at
>> >> <http://www.xfree86.org/~dawes/intel-85x/agpgart-85x.tar.gz>, in case
>> >> that makes a difference.
>> >>
>> >> David
>> >
>> >Ok.
>> >
>> >To simplify my environment, I did a fresh install of Red Hat 8.0.
>> >
>> >I then installed kernel 2.4.20-2.21 and XFree86-4.2.99.3-20030115,
>> >taken as RPM's from the RH81 'phoebe' beta, required for the i845 support.
>> >
>> >So, I now have a 'clean' setup which doesn't contain any of the pieces
>> > which I previously downloaded/built from various cvs repositories.
>> >
>> >On this machine (which has quite a few services running since it is a
>> > default 8.0 workstation-type install), it only takes 6 restart iterations
>> > of X before the system hangs.
>> >
>> >I (unfortunately) have 4 of these brand new GX60 machines. I see the exact
>> > same behavior on all of them.
>> >
>> >Therefore, I don't think the problem is specific to a particular system.
>> > By using the RH RPM's, also doesn't appear that the problem stems from
>> > something peculiar in my build environment.
>> >
>> >You tried on an i845G and can't reproduce, but you are using RH7.3?
>>
>> Yep, RH7.3 with its default kernel plus the agpgart driver referenced
>> above.  Could you try your setup using that agpgart driver?  That
>> might help narrow down if the problem lies there or elsewhere.  I
>> don't see how the problem could be anywhere other than the kernel
>> or agpgart driver.
>>
>> David
>
>i replaced agpgart.o with one recompiled from agpgart-85x.tar.gz
>
>% uname -a
>Linux localhost.localdomain 2.4.20-2.21 #1 Wed Jan 15 20:31:35 EST 2003 i686 i686 
>i386 GNU/Linux
>
>% ls -l /lib/modules/2.4.20-2.21/kernel/drivers/char/agp
>-rw-r--r--    1 root     root        64920 Feb 14 12:21 agpgart.o
>-rw-r--r--    1 root     root        23738 Jan 15 18:38 agpgart.o.gz.bak
>
>
>I am still seeing the same behavior.

I presume you rebooted after installing the new module, or explicitly
unloaded the old one?

I guess I'll need to try this same kernel.

David
--
David Dawes
Release Engineer/Architect                      The XFree86 Project
www.XFree86.org/~dawes
_______________________________________________
Devel mailing list
[EMAIL PROTECTED]
http://XFree86.Org/mailman/listinfo/devel

Reply via email to