On Mon, 2009-04-20 at 12:13 +0200, Thomas Hellstrom wrote: > Jerome Glisse wrote: > > On Sun, 2009-04-19 at 18:21 +0200, Thomas Hellström wrote: > > > >> Jerome Glisse wrote: > >> > >>> On Sat, 2009-04-18 at 21:06 +0200, Thomas Hellström wrote: > >>> > >>> > >>>> Jerome Glisse wrote: > >>>> > >>>> > >>>>> Hi Thomas > >>>>> > >>>>> I am getting massive error on x86_64, things like : > >>>>> BUG: Bad page map in process gnome-session pte:1f1d1d1d01000000 > >>>>> pmd:321a6067 > >>>>> keep filling the log until very bad things happen. > >>>>> Do you have any idea what might cause that in ttm ? > >>>>> My assumption is that ttm vm code is guilty their. > >>>>> Note that on x86 exact same code seem to run fine. > >>>>> All this with 2.6.29 final. > >>>>> > >>>>> Cheers, > >>>>> Jerome Glisse > >>>>> > >>>>> > >>>>> > >>>>> > >>>> Hi, Jerome! > >>>> > >>>> The TTM code may well be guilty here. I haven't tested x86-64 for a > >>>> while, but I can probably give it a try on openChrome next week. > >>>> > >>>> /Thomas > >>>> > >>>> > >>> Okay so i really narrowed it down to asking for WC memory, so my guess > >>> is that either my CPU (AMD Athlon(tm) Dual Core Processor 4450e) have > >>> PAT issue either TTM PAT/WC code is wrong somehow. I haven't time yet > >>> to go deeper with this but i think it worked on an Intel Core2 CPU > >>> with x86-64. Also it seems other people doesn't have the issue with > >>> WC on x86-64. > >>> > >>> PS: Sorry for all the noise, the bug didn't always showed up quickly so > >>> i had false feeling. I am yet unsure it's fully fixed but so far all my > >>> test case which triggered it seems to work fine. > >>> > >>> Cheers, > >>> Jerome > >>> > >>> > >>> > >> Jerome, > >> I think I missed what the fix was? > >> > >> Anyway, for PAT-aware kernels the drm git TTM code is always assuming > >> PAT is enabled and working. > >> For production kernels, the function > >> > >> pgprot_ttm_x86_wc > >> > >> should be replaced by an exported version of x86 pgprot_writecombine. > >> > >> FWIW x86-64 seems to work fine here on a 2.6.27 kernel Athlon-64 > >> single-core on openChrome. > >> > >> /Thomas > >> > > > > > > Fix is to not ask for WC memory, i will look at this after i cleanup > > all my hack to track down this. > > > > Cheers, > > Jerome > > > > > OK. > > If there was an erratum causing PAT not to be enabled on your processor, > then definitely that > may have cause strange inconsistencies. > > Thanks, > Thomas.
I think ttm_tt caching stuff does follow kernel policies outlined in Documentation/x86/pat.txt well at least from understanding of code i have right now through (call chains being sometimes hard to fully follow). As also have another issue it seems that calling set_memory_(uc|wc) while suspending lockup the cpu or at least doesn't return, is this somethings i should expect ? Cheers, Jerome ------------------------------------------------------------------------------ Stay on top of everything new and different, both inside and around Java (TM) technology - register by April 22, and save $200 on the JavaOne (SM) conference, June 2-5, 2009, San Francisco. 300 plus technical and hands-on sessions. Register today. Use priority code J9JMT32. http://p.sf.net/sfu/p -- _______________________________________________ Dri-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/dri-devel
