PG_uncached, CONFIG_EXTENDED_PAGEFLAGS and !NUMA

PG_uncached, CONFIG_EXTENDED_PAGEFLAGS and !NUMA

[H. Peter Anvin]

Hi all,

I am looking at a patchset by Venkatesh Pallipadi which cleans up a lot
of the corner cases in x86 PAT.

http://marc.info/?i=cover.1247162373.git.venkatesh.pallipadi@intel.com

This patchset pages PG_uncached available to other architectures than
IA64 on an opt-in basis.  Unfortunately, it means we run out of page
flags on X86_32+PAE+SPARSEMEM.

Rather than increasing SECTION_SIZE_BITS further, it seems more
reasonable to disable CONFIG_EXTENDED_PAGEFLAGS in this case:

diff --git a/mm/Kconfig b/mm/Kconfig
index c948d4c..fe221c7 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -153,7 +153,7 @@ config MEMORY_HOTREMOVE
 #
 config PAGEFLAGS_EXTENDED
        def_bool y
-       depends on 64BIT || SPARSEMEM_VMEMMAP || !NUMA || !SPARSEMEM
+       depends on 64BIT || SPARSEMEM_VMEMMAP || !SPARSEMEM

 # Heavily threaded applications may benefit from splitting the mm-wide

Dropping the !NUMA requirement here seems reasonable, since we already
have generic code that handles removing the node number from the page
flags when there are too many.

We could make this an x86-specific change, but the above generic change
would be cleaner in terms of Kconfig complexity.  Would people object to
this as a general change?

	-hpa

[tip:x86/pat] mm: remove !NUMA condition from PAGEFLAGS_EXTENDED condition set

[tip-bot for H. Peter Anvin]

Commit-ID:  a269cca9926faf8e44b340b017be0d884203141b
Gitweb:     http://git.kernel.org/tip/a269cca9926faf8e44b340b017be0d884203141b
Author:     H. Peter Anvin <hpa@zytor.com>
AuthorDate: Mon, 31 Aug 2009 11:17:44 -0700
Committer:  H. Peter Anvin <hpa@zytor.com>
CommitDate: Mon, 31 Aug 2009 11:17:44 -0700

mm: remove !NUMA condition from PAGEFLAGS_EXTENDED condition set

CONFIG_PAGEFLAGS_EXTENDED disables a trick to conserve pageflags.
This trick is indended to be enabled when the pressure on page flags
is very high.

The previous condition was:

-       depends on 64BIT || SPARSEMEM_VMEMMAP || !NUMA || !SPARSEMEM

... however, the sparsemem code already has a way to crowd out the
node number from the pageflags, which means that !NUMA actually
doesn't contribute to hard pageflags exhaustion.

This is required for the new PG_uncached flag to not cause pageflags
exhaustion on x86_32 + PAE + SPARSEMEM + !NUMA.

Signed-off-by: H. Peter Anvin <hpa@zytor.com>
LKML-Reference: <4A9828F4.4040905@zytor.com>
Cc: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Cc: Suresh Siddha <suresh.siddha@intel.com>


---
 mm/Kconfig |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/mm/Kconfig b/mm/Kconfig
index c948d4c..fe221c7 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -153,7 +153,7 @@ config MEMORY_HOTREMOVE
 #
 config PAGEFLAGS_EXTENDED
 	def_bool y
-	depends on 64BIT || SPARSEMEM_VMEMMAP || !NUMA || !SPARSEMEM
+	depends on 64BIT || SPARSEMEM_VMEMMAP || !SPARSEMEM
 
 # Heavily threaded applications may benefit from splitting the mm-wide
 # page_table_lock, so that faults on different parts of the user address