Re: 2.6.29 git master and PAT problems

["Pallipadi, Venkatesh" <venkatesh.pallipadi@intel.com>]

On Wed, Apr 01, 2009 at 03:23:49AM -0700, Arkadiusz Miskiewicz wrote:
> On Wednesday 01 of April 2009, Pallipadi, Venkatesh wrote:
> > On Tue, Mar 31, 2009 at 12:44:32AM -0700, Arkadiusz Miskiewicz wrote:
> > > On Tuesday 31 of March 2009, Pallipadi, Venkatesh wrote:
> > > > On Mon, Mar 30, 2009 at 04:25:11PM -0700, Arkadiusz Miskiewicz wrote:
> > > > > On Tuesday 31 of March 2009, Arkadiusz Miskiewicz wrote:
> > > > > > On Monday 30 of March 2009, Pallipadi, Venkatesh wrote:
> > > > > >
> > > > > > More info follows. Now I've switched to
> > > > > > e1c502482853f84606928f5a2f2eb6da1993cda1 which contains latest drm
> > > > > > fixes and now I get much lower numbers of PAT errors but still.
> > > > >
> > > > > Also when I switch t400 into discrete mode (radeon hd 3400 instead
> > > > > of integrated intel GM45) I get such errors (probably unrelated
> > > > > to these seen when using intel):
> > > > >
> > > > > [  419.187657] X:10550 conflicting memory types cfff0000-d0000000
> > > > > uncached<->uncached-minus [  419.187670] reserve_memtype failed
> > > > > 0xcfff0000-0xd0000000, track uncached, req write-back [  419.553914]
> > > > > X:10550 conflicting memory types cfff0000-d0000000
> > > > > uncached<->uncached-minus [  419.553923] reserve_memtype failed
> > > > > 0xcfff0000-0xd0000000, track uncached, req write-back [  419.813592]
> > > > > X:10550 conflicting memory types cfff0000-d0000000
> > > > > uncached<->uncached-minus [  419.813601] reserve_memtype failed
> > > > > 0xcfff0000-0xd0000000, track uncached, req write-back [  420.100102]
> > > > > X:10550 conflicting memory types cfff0000-d0000000
> > > > > uncached<->uncached-minus [  420.100111] reserve_memtype failed
> > > > > 0xcfff0000-0xd0000000, track uncached, req write-back
> > > >
> > > > Yes. This is a different problem than the freeing invalid type one. Are
> > > > these errors also with latest git kernel? Can you try the patch below
> > > > (which is a part of a bigger cleanup patch I have lined up).
> > >
> > > It's a latest git kernel as of today morning
> > > (latest commit is 15f7176eb1cccec0a332541285ee752b935c1c85)
> > > + your patch. Problem persists:
> > >
> > > [   74.696353] [drm] Setting GART location based on new memory map
> > > [   74.711520] [drm] Loading RV620 CP Microcode
> > > [   74.711792] [drm] Loading RV620 PFP Microcode
> > > [   74.726719] [drm] Resetting GPU
> > > [   74.726776] [drm] writeback test succeeded in 1 usecs
> > > [   75.256034] X:5366 conflicting memory types d0000000-e0000000
> > > uncached-minus<->write-combining [   75.256043] reserve_memtype failed
> > > 0xd0000000-0xe0000000, track uncached-minus, req write-back [
> > > 75.849951] X:5366 conflicting memory types d0000000-e0000000
> > > uncached-minus<->write-combining [   75.849960] reserve_memtype failed
> > > 0xd0000000-0xe0000000, track uncached-minus, req write-back [
> > > 76.054374] X:5366 conflicting memory types d0000000-e0000000
> > > uncached-minus<->write-combining [   76.054377] reserve_memtype failed
> > > 0xd0000000-0xe0000000, track uncached-minus, req write-back [
> > > 76.074481] X:5378 freeing invalid memtype d0000000-e0000000
> > > [   76.176881] X:5366 conflicting memory types d0000000-e0000000
> > > uncached-minus<->write-combining [   76.176885] reserve_memtype failed
> > > 0xd0000000-0xe0000000, track uncached-minus, req write-back [
> > > 76.207734] X:5380 freeing invalid memtype d0000000-e0000000
> >
> > OK. We now have a theory on what is going wrong here.
> >
> > The problem seems to be pci mmap uses vm_page_prot flag to remember the
> > memtype for this region. Looks like that memtype is somehow getting cleared
> > in this case. We still don't know where it is getting cleared. But, with
> > debug patch below we can be sure that it is indeed getting cleared, which
> > is causing problems on fork() as child wont know the memtype that parent
> > got.
> >
> > Can you please try the below debug patch over upstream git and check
> > whether you indeed hit the warnon.
> 
> warns triggered:
> 

OK. Below patch should resolve this problem.

Thanks,
Venki

From: Suresh Siddha <suresh.b.siddha@intel.com>
Subject: [PATCH] x86, PAT: Remove duplicate memtype reserve in pci mmap

pci mmap code was doing memtype reserve for a while now. Recently we
added memtype tracking in remap_pfn_range, and pci code indirectly calls
remap_pfn_range. So, we don't need seperate tracking in pci code
anymore. Which means a patch that removes ~50 lines of code :-).

Also, recently we found out that the pci tracking is not working as we expect
it to work in some cases. Specifically, userlevel X mmap of pci, with some
recent version of X, is having a problem with vm_page_prot getting reset.
The pci tracking uses vm_page_prot to pass on the protection type from parent
to child during fork.
a) Parent does a pci mmap
b) We look at PAT and get either UC_MINUS or WC mapping for parent
c) Store that mapping type in vma vm_page_prot for future use
d) This thread does a fork
e) Fork results in mmap_ops ->open for the child process
f) We get the vm_page_prot from vma and reserve that type for the child process

But, between c) and e) above, the vma vm_page_prot is getting reset to zero.
This results in PAT reserve failing at the time of fork as in here.
http://marc.info/?l=linux-kernel&m=123858163103240&w=2

This cleanup makes the above problem go away as we do not depend on
vm_page_prot in our PAT code anymore.

Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
---
 arch/x86/pci/i386.c |   46 ----------------------------------------------
 1 file changed, 46 deletions(-)

Index: linux-2.6/arch/x86/pci/i386.c
===================================================================
--- linux-2.6.orig/arch/x86/pci/i386.c	2009-04-01 10:29:49.000000000 -0700
+++ linux-2.6/arch/x86/pci/i386.c	2009-04-01 10:33:18.000000000 -0700
@@ -258,24 +258,7 @@ void pcibios_set_master(struct pci_dev *
 	pci_write_config_byte(dev, PCI_LATENCY_TIMER, lat);
 }
 
-static void pci_unmap_page_range(struct vm_area_struct *vma)
-{
-	u64 addr = (u64)vma->vm_pgoff << PAGE_SHIFT;
-	free_memtype(addr, addr + vma->vm_end - vma->vm_start);
-}
-
-static void pci_track_mmap_page_range(struct vm_area_struct *vma)
-{
-	u64 addr = (u64)vma->vm_pgoff << PAGE_SHIFT;
-	unsigned long flags = pgprot_val(vma->vm_page_prot)
-						& _PAGE_CACHE_MASK;
-
-	reserve_memtype(addr, addr + vma->vm_end - vma->vm_start, flags, NULL);
-}
-
 static struct vm_operations_struct pci_mmap_ops = {
-	.open  = pci_track_mmap_page_range,
-	.close = pci_unmap_page_range,
 	.access = generic_access_phys,
 };
 
@@ -283,11 +266,6 @@ int pci_mmap_page_range(struct pci_dev *
 			enum pci_mmap_state mmap_state, int write_combine)
 {
 	unsigned long prot;
-	u64 addr = vma->vm_pgoff << PAGE_SHIFT;
-	unsigned long len = vma->vm_end - vma->vm_start;
-	unsigned long flags;
-	unsigned long new_flags;
-	int retval;
 
 	/* I/O space cannot be accessed via normal processor loads and
 	 * stores on this platform.
@@ -308,30 +286,6 @@ int pci_mmap_page_range(struct pci_dev *
 
 	vma->vm_page_prot = __pgprot(prot);
 
-	flags = pgprot_val(vma->vm_page_prot) & _PAGE_CACHE_MASK;
-	retval = reserve_memtype(addr, addr + len, flags, &new_flags);
-	if (retval)
-		return retval;
-
-	if (flags != new_flags) {
-		if (!is_new_memtype_allowed(flags, new_flags)) {
-			free_memtype(addr, addr+len);
-			return -EINVAL;
-		}
-		flags = new_flags;
-		vma->vm_page_prot = __pgprot(
-			(pgprot_val(vma->vm_page_prot) & ~_PAGE_CACHE_MASK) |
-			flags);
-	}
-
-	if (((vma->vm_pgoff < max_low_pfn_mapped) ||
-	     (vma->vm_pgoff >= (1UL<<(32 - PAGE_SHIFT)) &&
-	      vma->vm_pgoff < max_pfn_mapped)) &&
-	    ioremap_change_attr((unsigned long)__va(addr), len, flags)) {
-		free_memtype(addr, addr + len);
-		return -EINVAL;
-	}
-
 	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 			       vma->vm_end - vma->vm_start,
 			       vma->vm_page_prot))