We have enough page flags after vmemmap no longer uses section ids.
Reserve 5 of the 6 saved flags for functionality that is currently
under development in the VM.
The new flags are only available if either of these conditions are met:
1. 64 Bit system. (then we have 8 more free of the 32)
2. !NUMA. In that case 2 bits are needed for the zone
id which leaves 30 page flag bits. Of those we use 24. 6 left.
3. !SPARSEMEM. In that case we use 5 bits of the 30 available
for the node. 1 leftover.
4. SPARSEMEM_VMEMMAP. Case 3 applies.
The remaining case is classic sparsemem with NUMA on a 32 bit platform.
In that case we need to use additional bits from the remaining 25 which
does not allow the use of all 5 extended page flags.
We could deal with that case by only allowing sparsemem vmemmap (if
sparsemem is selected as a model) on 32 bit NUMA platforms.
Signed-off-by: Christoph Lameter <[email protected]>
---
include/linux/page-flags.h | 32 +++++++++++++++++++++++++++++---
mm/Kconfig | 11 +++++++++++
2 files changed, 40 insertions(+), 3 deletions(-)
Index: linux-2.6/include/linux/page-flags.h
===================================================================
--- linux-2.6.orig/include/linux/page-flags.h 2008-02-29 19:33:01.000000000 -0800
+++ linux-2.6/include/linux/page-flags.h 2008-02-29 19:37:31.000000000 -0800
@@ -81,11 +81,24 @@ enum pageflags {
PG_reserved,
PG_private, /* If pagecache, has fs-private data */
PG_writeback, /* Page is under writeback */
- PG_compound, /* A compound page */
PG_swapcache, /* Swap page: swp_entry_t in private */
PG_mappedtodisk, /* Has blocks allocated on-disk */
PG_reclaim, /* To be reclaimed asap */
PG_buddy, /* Page is free, on buddy lists */
+#ifdef CONFIG_EXTENDED_PAGEFLAGS
+ /*
+ * Page flags that are only available without sparsemem on 32 bit
+ * (sparsemem vmemmap is ok. Flags are always available on 64 bit.
+ */
+ PG_mlock, /* Page cannot be swapped out */
+ PG_pin, /* Page cannot be moved in memory */
+ PG_tail, /* Tail of a compound page */
+ PG_head, /* Head of a compound page */
+ PG_vcompound, /* Compound page is virtually mapped */
+ PG_filebacked, /* Page is backed by an actual disk (not RAM) */
+#else
+ PG_compound, /* A compound page */
+#endif
#if (BITS_PER_LONG > 32)
/*
@@ -245,8 +258,20 @@ static inline void set_page_writeback(st
test_set_page_writeback(page);
}
-TESTPAGEFLAG(Compound, compound)
-__PAGEFLAG(Head, compound)
+#ifdef CONFIG_EXTENDED_PAGEFLAGS
+__PAGEFLAG(Head, head)
+__PAGEFLAG(Tail, tail)
+__PAGEFLAG(Vcompound, vcompound)
+__PAGEFLAG(Mlock, mlock)
+__PAGEFLAG(Pin, pin)
+__PAGEFLAG(FileBacked, filebacked)
+
+static inline int PageCompound(struct page *page)
+{
+ return (page->flags & ((1 << PG_tail) | (1 << PG_head))) != 0;
+}
+
+#else
/*
* PG_reclaim is used in combination with PG_compound to mark the
@@ -274,5 +299,6 @@ static inline void __ClearPageTail(struc
{
page->flags &= ~PG_head_tail_mask;
}
+#endif
#endif /* PAGE_FLAGS_H */
Index: linux-2.6/mm/Kconfig
===================================================================
--- linux-2.6.orig/mm/Kconfig 2008-02-29 19:13:55.000000000 -0800
+++ linux-2.6/mm/Kconfig 2008-02-29 19:37:31.000000000 -0800
@@ -193,3 +193,14 @@ config NR_QUICK
config VIRT_TO_BUS
def_bool y
depends on !ARCH_NO_VIRT_TO_BUS
+
+#
+# I wish we could get rid of this...... The main problem is the page
+# flag use on 32 bit system with NUMA and SPARSEMEM (no sparsemem_vmemmap).
+# Does not really make any sense to use sparsemem there since 32 bit spaces
+# will typically be backed by contiguous RAM these days. So there is nothing
+# sparse there anymore.
+#
+config EXTENDED_PAGEFLAGS
+ def_bool y
+ depends on 64BIT || !SPARSEMEM || !NUMA || SPARSEMEM_VMEMMAP
--
On Fri, 29 Feb 2008 20:05:44 -0800
Christoph Lameter <[email protected]> wrote:
> We have enough page flags after vmemmap no longer uses section ids.
>
> Reserve 5 of the 6 saved flags for functionality that is currently
> under development in the VM.
>
> The new flags are only available if either of these conditions are met:
>
> 1. 64 Bit system. (then we have 8 more free of the 32)
>
> 2. !NUMA. In that case 2 bits are needed for the zone
> id which leaves 30 page flag bits. Of those we use 24. 6 left.
>
> 3. !SPARSEMEM. In that case we use 5 bits of the 30 available
> for the node. 1 leftover.
>
> 4. SPARSEMEM_VMEMMAP. Case 3 applies.
Nice! I like your patch series.
--
All rights reversed.