This is an approach to anti-defragmentation that does not use a usemap bitmap
but uses a page flag. Rather than using separate free lists in the zone,
it extends the struct free_area which reduces the cache footprint a bit and
eliminates quite a lot of code. In total, it adds 246 lines of code, can be
fully disabled via a compile-time option without any performance penalty and
when enabled, it was able allocate 78% of physical memory in 4MiB contiguous
chunks after a heavy set of benchmarks completed.
This gives best-effort for low fragmentation in all zones. Guaranteed reclaim
for hotplug and high order allocation in a zone still requires a special zone.
Full benchmarks are below the Changelog, but here is the summary
clean anti-defrag
Kernel extract 16 16 (no difference)
Kernel build 736 736 (no difference)
aim9-page_test 132996.67 133802.07 +805.40 (0.61%)
aim9-brk_test 627478.75 628223.93 +745.18 (0.12%)
HugePages under load 79 98 +19 (24%)
HugePages after tests 102 296 +194 (190%)
All the aim9 results measured are within 1% of the standard allocator. Kernel
compiles vary sometimes by about 1-2 seconds or about 0.1%. The HugeTLB pages
under load was 7 simulatanous kernel compiles, each -j1.
When anti-defrag is disabled via the config option, it behaves and
performs just like the standard allocator with no performance impact I can
measure. When it is enabled, there are definite variances but they are small
and on loads like kernel builds, it makes less than 3 seconds of a difference
over 12 minutes.
Diffstat for full set of patches
fs/buffer.c | 3
fs/compat.c | 2
fs/exec.c | 2
fs/inode.c | 2
include/asm-i386/page.h | 3
include/linux/gfp.h | 3
include/linux/highmem.h | 3
include/linux/mmzone.h | 28 ++++-
include/linux/page-flags.h | 8 +
init/Kconfig | 12 ++
mm/memory.c | 6 -
mm/page_alloc.c | 227 ++++++++++++++++++++++++++++++++++++---------
mm/shmem.c | 4
mm/swap_state.c | 3
14 files changed, 246 insertions(+), 60 deletions(-)
Changelog since v20
o Update to 2.6.15-rc1-mm2
o Remove usemap to reduce cache footprint, also does not hurt hotplug add
o Extend free_area not use new free_area lists, reduces footprint and is faster
o Leverage existing support for per-cpu page drain to avoid free but pinned
pages
o Small number of micro-optimisations
Changelog since complex anti-defragmentation v19
o Updated to 2.6.15-rc1-mm2
o Removed the fallback area and balancing code
o Only differentiate between kernel and easy reclaimable allocations
- Removes almost all the code that deals with usemaps
- Made a number of simplifications based on two allocation lists
- Fallback code is drastically simpler
o Do not change behaviour for high-order allocations
o Drop stats patch - unnecessary complications
Changelog since v18
o Resync against 2.6.14-rc5-mm1
o 004_markfree dropped
o Documentation note added on the behavior of free_area.nr_free
Changelog since v17
o Update to 2.6.14-rc4-mm1
o Remove explicit casts where implicit casts were in place
o Change __GFP_USER to __GFP_EASYRCLM, RCLM_USER to RCLM_EASY and PCPU_USER to
PCPU_EASY
o Print a warning and return NULL if both RCLM flags are set in the GFP flags
o Reduce size of fallback_allocs
o Change magic number 64 to FREE_AREA_USEMAP_SIZE
o CodingStyle regressions cleanup
o Move sparsemen setup_usemap() out of header
o Changed fallback_balance to a mechanism that depended on zone->present_pages
to avoid hotplug problems later
o Many superfluous parenthesis removed
Changlog since v16
o Variables using bit operations now are unsigned long. Note that when used
as indices, they are integers and cast to unsigned long when necessary.
This is because aim9 shows regressions when used as unsigned longs
throughout (~10% slowdown)
o 004_showfree added to provide more debugging information
o 008_stats dropped. Even with CONFIG_ALLOCSTATS disabled, it is causing
severe performance regressions. No explanation as to why
o for_each_rclmtype_order moved to header
o More coding style cleanups
Changelog since V14 (V15 not released)
o Update against 2.6.14-rc3
o Resync with Joel's work. All suggestions made on fix-ups to his last
set of patches should also be in here. e.g. __GFP_USER is still __GFP_USER
but is better commented.
o Large amount of CodingStyle, readability cleanups and corrections pointed
out by Dave Hansen.
o Fix CONFIG_NUMA error that corrupted per-cpu lists
o Patches broken out to have one-feature-per-patch rather than
more-code-per-patch
o Fix fallback bug where pages for RCLM_NORCLM end up on random other
free lists.
Changelog since V13
o Patches are now broken out
o Added per-cpu draining of userrclm pages
o Brought the patch more in line with memory hotplug work
o Fine-grained use of the __GFP_USER and __GFP_KERNRCLM flags
o Many coding-style corrections
o Many whitespace-damage corrections
Changelog since V12
o Minor whitespace damage fixed as pointed by Joel Schopp
Changelog since V11
o Mainly a redefiff against 2.6.12-rc5
o Use #defines for indexing into pcpu lists
o Fix rounding error in the size of usemap
Changelog since V10
o All allocation types now use per-cpu caches like the standard allocator
o Removed all the additional buddy allocator statistic code
o Elimated three zone fields that can be lived without
o Simplified some loops
o Removed many unnecessary calculations
Changelog since V9
o Tightened what pools are used for fallbacks, less likely to fragment
o Many micro-optimisations to have the same performance as the standard
allocator. Modified allocator now faster than standard allocator using
gcc 3.3.5
o Add counter for splits/coalescing
Changelog since V8
o rmqueue_bulk() allocates pages in large blocks and breaks it up into the
requested size. Reduces the number of calls to __rmqueue()
o Beancounters are now a configurable option under "Kernel Hacking"
o Broke out some code into inline functions to be more Hotplug-friendly
o Increased the size of reserve for fallbacks from 10% to 12.5%.
Changelog since V7
o Updated to 2.6.11-rc4
o Lots of cleanups, mainly related to beancounters
o Fixed up a miscalculation in the bitmap size as pointed out by Mike Kravetz
(thanks Mike)
o Introduced a 10% reserve for fallbacks. Drastically reduces the number of
kernnorclm allocations that go to the wrong places
o Don't trigger OOM when large allocations are involved
Changelog since V6
o Updated to 2.6.11-rc2
o Minor change to allow prezeroing to be a cleaner looking patch
Changelog since V5
o Fixed up gcc-2.95 errors
o Fixed up whitespace damage
Changelog since V4
o No changes. Applies cleanly against 2.6.11-rc1 and 2.6.11-rc1-bk6. Applies
with offsets to 2.6.11-rc1-mm1
Changelog since V3
o inlined get_pageblock_type() and set_pageblock_type()
o set_pageblock_type() now takes a zone parameter to avoid a call to page_zone()
o When taking from the global pool, do not scan all the low-order lists
Changelog since V2
o Do not to interfere with the "min" decay
o Update the __GFP_BITS_SHIFT properly. Old value broke fsync and probably
anything to do with asynchronous IO
Changelog since V1
o Update patch to 2.6.11-rc1
o Cleaned up bug where memory was wasted on a large bitmap
o Remove code that needed the binary buddy bitmaps
o Update flags to avoid colliding with __GFP_ZERO changes
o Extended fallback_count bean counters to show the fallback count for each
allocation type
o In-code documentation
Version 1
o Initial release against 2.6.9
This patch is designed to reduce fragmentation in the standard buddy allocator
without impairing the performance of the allocator. High fragmentation
in the standard binary buddy allocator means that high-order allocations
can rarely be serviced. This patch works by dividing allocations into two
different types of allocations;
EasyReclaimable - These are userspace pages that are easily reclaimable. This
flag is set when it is known that the pages will be trivially reclaimed
by writing the page out to swap or syncing with backing storage
KernelNonReclaimable - These are pages that are allocated by the kernel that
are not trivially reclaimed. For example, the memory allocated for a
loaded module would be in this category. By default, allocations are
considered to be of this type
Instead of having one global MAX_ORDER-sized array of free lists, there are
two, one for each type of allocation. Once a 2^MAX_ORDER block of pages is
split for a type of allocation, it is added to the free-lists for that type,
in effect reserving it. Hence, over time, pages of the different types can
be clustered together. When a page is allocated, the page-flags are updated
with a value indicating it's type of page so that it is placed on the correct
list on free.
When the preferred freelists are expired, the largest possible block is taken
from the alternative list. Buddies that are split from that large block are
placed on the preferred allocation-type freelists to mitigate fragmentation.
Four benchmark results are included all based on a 2.6.15-rc1-mm2 kernel
compiled with gcc 3.4. These benchmarks were run in the order you see them
*without* rebooting. This means that when the final highorder stress test,
the system is already running with any fragmentation introduced by other
benchmarks.
The first test called bench-kbuild.sh times a kernel build. Time is in seconds
clean anti-defrag
Kernel extract 16 16
Kernel build 736 736
The second is the output of portions of AIM9 for the vanilla
allocator and the modified one;
(Tests run with bench-aim9.sh from VMRegress 0.18)
2.6.15-rc1-mm2-clean
------------------------------------------------------------------------------------------------------------
Test Test Elapsed Iteration Iteration Operation
Number Name Time (sec) Count Rate (loops/sec) Rate (ops/sec)
------------------------------------------------------------------------------------------------------------
1 creat-clo 60.02 958 15.96135 15961.35 File Creations and Closes/second
2 page_test 60.00 4694 78.23333 132996.67 System Allocations & Pages/second
3 brk_test 60.01 2215 36.91051 627478.75 System Memory Allocations/second
4 jmp_test 60.00 264119 4401.98333 4401983.33 Non-local gotos/second
5 signal_test 60.02 5068 84.43852 84438.52 Signal Traps/second
6 exec_test 60.07 761 12.66855 63.34 Program Loads/second
7 fork_test 60.06 1073 17.86547 1786.55 Task Creations/second
8 link_test 60.02 5282 88.00400 5544.25 Link/Unlink Pairs/second
------------------------------------------------------------------------------------------------------------
2.6.14-rc1-mm2-mbuddy-v21
------------------------------------------------------------------------------------------------------------
Test Test Elapsed Iteration Iteration Operation
Number Name Time (sec) Count Rate (loops/sec) Rate (ops/sec)
------------------------------------------------------------------------------------------------------------
1 creat-clo 60.04 957 15.93937 15939.37 File Creations and Closes/second
2 page_test 60.02 4724 78.70710 133802.07 System Allocations & Pages/second
3 brk_test 60.02 2218 36.95435 628223.93 System Memory Allocations/second
4 jmp_test 60.01 264389 4405.74904 4405749.04 Non-local gotos/second
5 signal_test 60.01 5022 83.68605 83686.05 Signal Traps/second
6 exec_test 60.00 756 12.60000 63.00 Program Loads/second
7 fork_test 60.06 1076 17.91542 1791.54 Task Creations/second
8 link_test 60.01 5318 88.61856 5582.97 Link/Unlink Pairs/second
------------------------------------------------------------------------------------------------------------
Difference in performance operations report generated by diff-aim9.sh
Clean mbuddy-v21
---------- ----------
1 creat-clo 15961.35 15939.37 -21.98 -0.14% File Creations and Closes/second
2 page_test 132996.67 133802.07 805.40 0.61% System Allocations & Pages/second
3 brk_test 627478.75 628223.93 745.18 0.12% System Memory Allocations/second
4 jmp_test 4401983.33 4405749.04 3765.71 0.09% Non-local gotos/second
5 signal_test 84438.52 83686.05 -752.47 -0.89% Signal Traps/second
6 exec_test 63.34 63.00 -0.34 -0.54% Program Loads/second
7 fork_test 1786.55 1791.54 4.99 0.28% Task Creations/second
8 link_test 5544.25 5582.97 38.72 0.70% Link/Unlink Pairs/second
The second benchmark tested the CPU cache usage to make sure it was not
getting clobbered. The test was to repeatedly render a large postscript file
10 times and get the average. The result is;
2.6.15-rc1-mm2-clean: Average: 7.42 real, 7.25 user, 0.078 sys
2.6.15-rc1-mm2-mbuddy-v21: Average: 7.406 real, 6.916 user, 0.088 sys
So there are no adverse cache effects. The last test is to show that the
allocator can satisfy more high-order allocations, especially under load,
than the standard allocator. The test performs the following;
1. Start updatedb running in the background
2. Load kernel modules that tries to allocate high-order blocks on demand
3. Clean a kernel tree
4. Make 6 copies of the tree. As each copy finishes, a compile starts at -j1
5. Start compiling the primary tree
6. Sleep 1 minute while the 7 trees are being compiled
7. Use the kernel module to attempt 160 times to allocate a 2^10 block of pages
- note, it only attempts 275 times, no matter how often it succeeds
- An allocation is attempted every 1/10th of a second
- Performance will get badly shot as it forces considerable amounts of
pageout
8. At rest, dd a file from /dev/zero that is the size of physical memory, cat
it and delete it again to flush as much of buffer cache as possible. Then
try and allocate as many pages as possible
The result of the allocations under load;
2.6.15-rc1-mm2 Clean
rder: 10
Allocation type: HighMem
Attempted allocations: 275
Success allocs: 79
Failed allocs: 196
DMA zone allocs: 1
Normal zone allocs: 4
HighMem zone allocs: 74
% Success: 28
2.6.15-rc1-mm2 MBuddy V21
Order: 10
Allocation type: HighMem
Attempted allocations: 275
Success allocs: 98
Failed allocs: 177
DMA zone allocs: 1
Normal zone allocs: 2
HighMem zone allocs: 95
% Success: 35
There is not a large difference there but a lot of it depends on what decisions
LRU takes and under load, your chances are not good. An indicator of how
well fragmentation is addressed is testing again after the load decreases.
After the tests completed, the standard allocator was able to allocate 102
order-10 pages and the modified allocator allocated 296.
The results show that the modified allocator has comparable speed, has no
adverse cache effects but is far less fragmented and in a better position
to satisfy high-order allocations.
--
Mel Gorman
Part-time Phd Student Java Applications Developer
University of Limerick IBM Dublin Software Lab
This patch adds a flag __GFP_EASYRCLM. Allocations using the __GFP_EASYRCLM
flag are expected to be easily reclaimed by syncing with backing storage (be
it a file or swap) or cleaning the buffers and discarding.
Signed-off-by: Mel Gorman <[email protected]>
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/fs/buffer.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/buffer.c
--- linux-2.6.15-rc1-mm2-clean/fs/buffer.c 2005-11-21 19:44:32.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/buffer.c 2005-11-22 16:49:23.000000000 +0000
@@ -1113,7 +1113,8 @@ grow_dev_page(struct block_device *bdev,
struct page *page;
struct buffer_head *bh;
- page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
+ page = find_or_create_page(inode->i_mapping, index,
+ GFP_NOFS|__GFP_EASYRCLM);
if (!page)
return NULL;
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/fs/compat.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/compat.c
--- linux-2.6.15-rc1-mm2-clean/fs/compat.c 2005-11-21 19:44:32.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/compat.c 2005-11-22 16:49:23.000000000 +0000
@@ -1359,7 +1359,7 @@ static int compat_copy_strings(int argc,
page = bprm->page[i];
new = 0;
if (!page) {
- page = alloc_page(GFP_HIGHUSER);
+ page = alloc_page(GFP_HIGHUSER|__GFP_EASYRCLM);
bprm->page[i] = page;
if (!page) {
ret = -ENOMEM;
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/fs/exec.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/exec.c
--- linux-2.6.15-rc1-mm2-clean/fs/exec.c 2005-11-21 19:44:32.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/exec.c 2005-11-22 16:49:23.000000000 +0000
@@ -238,7 +238,7 @@ static int copy_strings(int argc, char _
page = bprm->page[i];
new = 0;
if (!page) {
- page = alloc_page(GFP_HIGHUSER);
+ page = alloc_page(GFP_HIGHUSER|__GFP_EASYRCLM);
bprm->page[i] = page;
if (!page) {
ret = -ENOMEM;
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/fs/inode.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/inode.c
--- linux-2.6.15-rc1-mm2-clean/fs/inode.c 2005-11-21 19:44:32.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/fs/inode.c 2005-11-22 16:49:23.000000000 +0000
@@ -146,7 +146,7 @@ static struct inode *alloc_inode(struct
mapping->a_ops = &empty_aops;
mapping->host = inode;
mapping->flags = 0;
- mapping_set_gfp_mask(mapping, GFP_HIGHUSER);
+ mapping_set_gfp_mask(mapping, GFP_HIGHUSER|__GFP_EASYRCLM);
mapping->assoc_mapping = NULL;
mapping->backing_dev_info = &default_backing_dev_info;
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/include/asm-i386/page.h linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/asm-i386/page.h
--- linux-2.6.15-rc1-mm2-clean/include/asm-i386/page.h 2005-10-28 01:02:08.000000000 +0100
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/asm-i386/page.h 2005-11-22 16:49:23.000000000 +0000
@@ -36,7 +36,8 @@
#define clear_user_page(page, vaddr, pg) clear_page(page)
#define copy_user_page(to, from, vaddr, pg) copy_page(to, from)
-#define alloc_zeroed_user_highpage(vma, vaddr) alloc_page_vma(GFP_HIGHUSER | __GFP_ZERO, vma, vaddr)
+#define alloc_zeroed_user_highpage(vma, vaddr) \
+ alloc_page_vma(GFP_HIGHUSER | __GFP_ZERO | __GFP_EASYRCLM, vma, vaddr)
#define __HAVE_ARCH_ALLOC_ZEROED_USER_HIGHPAGE
/*
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/include/linux/gfp.h linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/gfp.h
--- linux-2.6.15-rc1-mm2-clean/include/linux/gfp.h 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/gfp.h 2005-11-22 16:49:23.000000000 +0000
@@ -47,6 +47,7 @@ struct vm_area_struct;
#define __GFP_ZERO ((__force gfp_t)0x8000u)/* Return zeroed page on success */
#define __GFP_NOMEMALLOC ((__force gfp_t)0x10000u) /* Don't use emergency reserves */
#define __GFP_HARDWALL ((__force gfp_t)0x20000u) /* Enforce hardwall cpuset memory allocs */
+#define __GFP_EASYRCLM ((__force gfp_t)0x40000u) /* Easily reclaimed page */
#define __GFP_BITS_SHIFT 20 /* Room for 20 __GFP_FOO bits */
#define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))
@@ -55,7 +56,7 @@ struct vm_area_struct;
#define GFP_LEVEL_MASK (__GFP_WAIT|__GFP_HIGH|__GFP_IO|__GFP_FS| \
__GFP_COLD|__GFP_NOWARN|__GFP_REPEAT| \
__GFP_NOFAIL|__GFP_NORETRY|__GFP_NO_GROW|__GFP_COMP| \
- __GFP_NOMEMALLOC|__GFP_HARDWALL)
+ __GFP_NOMEMALLOC|__GFP_HARDWALL|__GFP_EASYRCLM)
/* GFP_ATOMIC means both !wait (__GFP_WAIT not set) and use emergency pool */
#define GFP_ATOMIC (__GFP_HIGH)
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/include/linux/highmem.h linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/highmem.h
--- linux-2.6.15-rc1-mm2-clean/include/linux/highmem.h 2005-10-28 01:02:08.000000000 +0100
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/highmem.h 2005-11-22 16:49:23.000000000 +0000
@@ -47,7 +47,8 @@ static inline void clear_user_highpage(s
static inline struct page *
alloc_zeroed_user_highpage(struct vm_area_struct *vma, unsigned long vaddr)
{
- struct page *page = alloc_page_vma(GFP_HIGHUSER, vma, vaddr);
+ struct page *page = alloc_page_vma(GFP_HIGHUSER|__GFP_EASYRCLM,
+ vma, vaddr);
if (page)
clear_user_highpage(page, vaddr);
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/mm/memory.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/memory.c
--- linux-2.6.15-rc1-mm2-clean/mm/memory.c 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/memory.c 2005-11-22 16:49:23.000000000 +0000
@@ -1378,7 +1378,8 @@ gotten:
if (!new_page)
goto oom;
} else {
- new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
+ new_page = alloc_page_vma(GFP_HIGHUSER|__GFP_EASYRCLM,
+ vma, address);
if (!new_page)
goto oom;
copy_user_highpage(new_page, src_page, address);
@@ -1985,7 +1986,8 @@ retry:
if (unlikely(anon_vma_prepare(vma)))
goto oom;
- page = alloc_page_vma(GFP_HIGHUSER, vma, address);
+ page = alloc_page_vma(GFP_HIGHUSER|__GFP_EASYRCLM,
+ vma, address);
if (!page)
goto oom;
copy_user_highpage(page, new_page, address);
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/mm/shmem.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/shmem.c
--- linux-2.6.15-rc1-mm2-clean/mm/shmem.c 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/shmem.c 2005-11-22 16:49:23.000000000 +0000
@@ -921,7 +921,7 @@ shmem_alloc_page(gfp_t gfp, struct shmem
pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
pvma.vm_pgoff = idx;
pvma.vm_end = PAGE_SIZE;
- page = alloc_page_vma(gfp | __GFP_ZERO, &pvma, 0);
+ page = alloc_page_vma(gfp | __GFP_ZERO | __GFP_EASYRCLM, &pvma, 0);
mpol_free(pvma.vm_policy);
return page;
}
@@ -936,7 +936,7 @@ shmem_swapin(struct shmem_inode_info *in
static inline struct page *
shmem_alloc_page(gfp_t gfp,struct shmem_inode_info *info, unsigned long idx)
{
- return alloc_page(gfp | __GFP_ZERO);
+ return alloc_page(gfp | __GFP_ZERO | __GFP_EASYRCLM);
}
#endif
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-clean/mm/swap_state.c linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/swap_state.c
--- linux-2.6.15-rc1-mm2-clean/mm/swap_state.c 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/swap_state.c 2005-11-22 16:49:23.000000000 +0000
@@ -341,7 +341,8 @@ struct page *read_swap_cache_async(swp_e
* Get a new page to read into from swap.
*/
if (!new_page) {
- new_page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
+ new_page = alloc_page_vma(GFP_HIGHUSER|__GFP_EASYRCLM,
+ vma, addr);
if (!new_page)
break; /* Out of memory */
}
Per-cpu pages can accidentally cause fragmentation because they are free, but
pinned pages in an otherwise contiguous block. When this patch is applied,
the per-cpu caches are drained after the direct-reclaim is entered if the
requested order is greater than 3. It simply reuses the code used by suspend
and hotplug and only is triggered when anti-defragmentation is enabled.
Signed-off-by: Mel Gorman <[email protected]>
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-004_configurable/mm/page_alloc.c linux-2.6.15-rc1-mm2-005_drainpercpu/mm/page_alloc.c
--- linux-2.6.15-rc1-mm2-004_configurable/mm/page_alloc.c 2005-11-22 16:53:03.000000000 +0000
+++ linux-2.6.15-rc1-mm2-005_drainpercpu/mm/page_alloc.c 2005-11-22 16:53:45.000000000 +0000
@@ -689,7 +689,9 @@ void drain_remote_pages(void)
}
#endif
-#if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
+#if defined(CONFIG_PM) || \
+ defined(CONFIG_HOTPLUG_CPU) || \
+ defined(CONFIG_PAGEALLOC_ANTIDEFRAG)
static void __drain_pages(unsigned int cpu)
{
struct zone *zone;
@@ -716,10 +718,9 @@ static void __drain_pages(unsigned int c
}
}
}
-#endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
+#endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU || CONFIG_PAGEALLOC_ANTIDEFRAG */
#ifdef CONFIG_PM
-
void mark_free_pages(struct zone *zone)
{
unsigned long zone_pfn, flags;
@@ -746,7 +747,9 @@ void mark_free_pages(struct zone *zone)
}
spin_unlock_irqrestore(&zone->lock, flags);
}
+#endif /* CONFIG_PM */
+#if defined(CONFIG_PM) || defined(CONFIG_PAGEALLOC_ANTIDEFRAG)
/*
* Spill all of this CPU's per-cpu pages back into the buddy allocator.
*/
@@ -758,7 +761,28 @@ void drain_local_pages(void)
__drain_pages(smp_processor_id());
local_irq_restore(flags);
}
-#endif /* CONFIG_PM */
+
+void smp_drain_local_pages(void *arg)
+{
+ drain_local_pages();
+}
+
+/*
+ * Spill all the per-cpu pages from all CPUs back into the buddy allocator
+ */
+void drain_all_local_pages(void)
+{
+ unsigned long flags;
+
+ local_irq_save(flags);
+ __drain_pages(smp_processor_id());
+ local_irq_restore(flags);
+
+ smp_call_function(smp_drain_local_pages, NULL, 0, 1);
+}
+#else
+void drain_all_local_pages(void) {}
+#endif /* CONFIG_PAGEALLOC_ANTIDEFRAG */
void zone_statistics(struct zonelist *zonelist, struct zone *z)
{
@@ -1109,6 +1133,9 @@ rebalance:
did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
+ if (order > 3)
+ drain_all_local_pages();
+
p->reclaim_state = NULL;
p->flags &= ~PF_MEMALLOC;
This patch adds the core of the anti-fragmentation strategy. It works by
grouping related allocation types together. The idea is that large groups of
pages that may be reclaimed are placed near each other. The zone->free_area
list is broken into RCLM_TYPES number of lists.
Signed-off-by: Mel Gorman <[email protected]>
Signed-off-by: Joel Schopp <[email protected]>
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/mmzone.h linux-2.6.15-rc1-mm2-002_fragcore/include/linux/mmzone.h
--- linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/mmzone.h 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-002_fragcore/include/linux/mmzone.h 2005-11-22 16:50:09.000000000 +0000
@@ -22,8 +22,16 @@
#define MAX_ORDER CONFIG_FORCE_MAX_ZONEORDER
#endif
+#define RCLM_NORCLM 0
+#define RCLM_EASY 1
+#define RCLM_TYPES 2
+
+#define for_each_rclmtype_order(type, order) \
+ for (order = 0; order < MAX_ORDER; order++) \
+ for (type = 0; type < RCLM_TYPES; type++)
+
struct free_area {
- struct list_head free_list;
+ struct list_head free_list[RCLM_TYPES];
unsigned long nr_free;
};
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/page-flags.h linux-2.6.15-rc1-mm2-002_fragcore/include/linux/page-flags.h
--- linux-2.6.15-rc1-mm2-001_antidefrag_flags/include/linux/page-flags.h 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-002_fragcore/include/linux/page-flags.h 2005-11-22 16:50:09.000000000 +0000
@@ -76,6 +76,7 @@
#define PG_reclaim 17 /* To be reclaimed asap */
#define PG_nosave_free 18 /* Free, should not be written */
#define PG_uncached 19 /* Page has been mapped as uncached */
+#define PG_easyrclm 20 /* Page is in an easy reclaim block */
/*
* Global page accounting. One instance per CPU. Only unsigned longs are
@@ -304,6 +305,12 @@ extern void __mod_page_state(unsigned lo
#define SetPageUncached(page) set_bit(PG_uncached, &(page)->flags)
#define ClearPageUncached(page) clear_bit(PG_uncached, &(page)->flags)
+#define PageEasyRclm(page) test_bit(PG_easyrclm, &(page)->flags)
+#define SetPageEasyRclm(page) set_bit(PG_easyrclm, &(page)->flags)
+#define ClearPageEasyRclm(page) clear_bit(PG_easyrclm, &(page)->flags)
+#define __SetPageEasyRclm(page) __set_bit(PG_easyrclm, &(page)->flags)
+#define __ClearPageEasyRclm(page) __clear_bit(PG_easyrclm, &(page)->flags)
+
struct page; /* forward declaration */
int test_clear_page_dirty(struct page *page);
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/page_alloc.c linux-2.6.15-rc1-mm2-002_fragcore/mm/page_alloc.c
--- linux-2.6.15-rc1-mm2-001_antidefrag_flags/mm/page_alloc.c 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-002_fragcore/mm/page_alloc.c 2005-11-22 16:50:09.000000000 +0000
@@ -68,6 +68,16 @@ int sysctl_lowmem_reserve_ratio[MAX_NR_Z
EXPORT_SYMBOL(totalram_pages);
+static inline int get_pageblock_type(struct page *page)
+{
+ return (PageEasyRclm(page) != 0);
+}
+
+static inline int gfpflags_to_alloctype(unsigned long gfp_flags)
+{
+ return ((gfp_flags & __GFP_EASYRCLM) != 0);
+}
+
/*
* Used by page_zone() to look up the address of the struct zone whose
* id is encoded in the upper bits of page->flags
@@ -314,11 +324,13 @@ static inline void __free_pages_bulk (st
{
unsigned long page_idx;
int order_size = 1 << order;
+ int alloctype = get_pageblock_type(page);
if (unlikely(order))
destroy_compound_page(page, order);
page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
+ __SetPageEasyRclm(page);
BUG_ON(page_idx & (order_size - 1));
BUG_ON(bad_range(zone, page));
@@ -326,7 +338,6 @@ static inline void __free_pages_bulk (st
zone->free_pages += order_size;
while (order < MAX_ORDER-1) {
unsigned long combined_idx;
- struct free_area *area;
struct page *buddy;
combined_idx = __find_combined_index(page_idx, order);
@@ -337,15 +348,14 @@ static inline void __free_pages_bulk (st
if (!page_is_buddy(buddy, order))
break; /* Move the buddy up one level. */
list_del(&buddy->lru);
- area = zone->free_area + order;
- area->nr_free--;
+ zone->free_area[order].nr_free--;
rmv_page_order(buddy);
page = page + (combined_idx - page_idx);
page_idx = combined_idx;
order++;
}
set_page_order(page, order);
- list_add(&page->lru, &zone->free_area[order].free_list);
+ list_add(&page->lru, &zone->free_area[order].free_list[alloctype]);
zone->free_area[order].nr_free++;
}
@@ -450,7 +460,8 @@ void __free_pages_ok(struct page *page,
*/
static inline struct page *
expand(struct zone *zone, struct page *page,
- int low, int high, struct free_area *area)
+ int low, int high, struct free_area *area,
+ int alloctype)
{
unsigned long size = 1 << high;
@@ -459,7 +470,7 @@ expand(struct zone *zone, struct page *p
high--;
size >>= 1;
BUG_ON(bad_range(zone, &page[size]));
- list_add(&page[size].lru, &area->free_list);
+ list_add(&page[size].lru, &area->free_list[alloctype]);
area->nr_free++;
set_page_order(&page[size], high);
}
@@ -520,30 +531,79 @@ static int prep_new_page(struct page *pa
return 0;
}
+/* Remove an element from the buddy allocator from the fallback list */
+static struct page *__rmqueue_fallback(struct zone *zone, int order,
+ int alloctype)
+{
+ struct free_area * area;
+ int current_order;
+ struct page *page;
+
+ /* Find the largest possible block of pages in the other list */
+ alloctype = !alloctype;
+ for (current_order = MAX_ORDER-1; current_order >= order;
+ --current_order) {
+ area = &(zone->free_area[current_order]);
+ if (list_empty(&area->free_list[alloctype]))
+ continue;
+
+ page = list_entry(area->free_list[alloctype].next,
+ struct page, lru);
+ area->nr_free--;
+
+ /*
+ * If breaking a large block of pages, place the buddies
+ * on the preferred allocation list
+ */
+ if (unlikely(current_order >= MAX_ORDER / 2))
+ alloctype = !alloctype;
+
+ list_del(&page->lru);
+ rmv_page_order(page);
+ zone->free_pages -= 1UL << order;
+ return expand(zone, page, order, current_order, area,
+ alloctype);
+
+ }
+
+ return NULL;
+}
+
/*
* Do the hard work of removing an element from the buddy allocator.
* Call me with the zone->lock already held.
*/
-static struct page *__rmqueue(struct zone *zone, unsigned int order)
+static struct page *__rmqueue(struct zone *zone, unsigned int order,
+ int alloctype)
{
struct free_area * area;
unsigned int current_order;
struct page *page;
+ /* Find a page of the appropriate size in the preferred list */
for (current_order = order; current_order < MAX_ORDER; ++current_order) {
- area = zone->free_area + current_order;
- if (list_empty(&area->free_list))
+ area = &(zone->free_area[current_order]);
+ if (list_empty(&area->free_list[alloctype]))
continue;
- page = list_entry(area->free_list.next, struct page, lru);
+ page = list_entry(area->free_list[alloctype].next,
+ struct page, lru);
list_del(&page->lru);
rmv_page_order(page);
area->nr_free--;
zone->free_pages -= 1UL << order;
- return expand(zone, page, order, current_order, area);
+ page = expand(zone, page, order, current_order, area,
+ alloctype);
+ goto got_page;
}
- return NULL;
+ page = __rmqueue_fallback(zone, order, alloctype);
+
+got_page:
+ if (unlikely(alloctype == RCLM_NORCLM))
+ __ClearPageEasyRclm(page);
+
+ return page;
}
/*
@@ -552,7 +612,8 @@ static struct page *__rmqueue(struct zon
* Returns the number of new pages which were placed at *list.
*/
static int rmqueue_bulk(struct zone *zone, unsigned int order,
- unsigned long count, struct list_head *list)
+ unsigned long count, struct list_head *list,
+ int alloctype)
{
unsigned long flags;
int i;
@@ -561,7 +622,7 @@ static int rmqueue_bulk(struct zone *zon
spin_lock_irqsave(&zone->lock, flags);
for (i = 0; i < count; ++i) {
- page = __rmqueue(zone, order);
+ page = __rmqueue(zone, order, alloctype);
if (page == NULL)
break;
allocated++;
@@ -627,7 +688,7 @@ static void __drain_pages(unsigned int c
void mark_free_pages(struct zone *zone)
{
unsigned long zone_pfn, flags;
- int order;
+ int order, t;
struct list_head *curr;
if (!zone->spanned_pages)
@@ -638,13 +699,15 @@ void mark_free_pages(struct zone *zone)
ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
for (order = MAX_ORDER - 1; order >= 0; --order)
- list_for_each(curr, &zone->free_area[order].free_list) {
+ for_each_rclmtype_order(t, order) {
+ list_for_each(curr, &zone->free_area[order].free_list[t]) {
unsigned long start_pfn, i;
start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
for (i=0; i < (1<<order); i++)
SetPageNosaveFree(pfn_to_page(start_pfn+i));
+ }
}
spin_unlock_irqrestore(&zone->lock, flags);
}
@@ -748,6 +811,7 @@ buffered_rmqueue(struct zone *zone, int
unsigned long flags;
struct page *page;
int cold = !!(gfp_flags & __GFP_COLD);
+ int alloctype = gfpflags_to_alloctype(gfp_flags);
again:
if (order == 0) {
@@ -758,7 +822,8 @@ again:
local_irq_save(flags);
if (pcp->count <= pcp->low)
pcp->count += rmqueue_bulk(zone, 0,
- pcp->batch, &pcp->list);
+ pcp->batch, &pcp->list,
+ alloctype);
if (pcp->count) {
page = list_entry(pcp->list.next, struct page, lru);
list_del(&page->lru);
@@ -768,7 +833,7 @@ again:
put_cpu();
} else {
spin_lock_irqsave(&zone->lock, flags);
- page = __rmqueue(zone, order);
+ page = __rmqueue(zone, order, alloctype);
spin_unlock_irqrestore(&zone->lock, flags);
}
@@ -1791,7 +1856,8 @@ void zone_init_free_lists(struct pglist_
{
int order;
for (order = 0; order < MAX_ORDER ; order++) {
- INIT_LIST_HEAD(&zone->free_area[order].free_list);
+ INIT_LIST_HEAD(&zone->free_area[order].free_list[RCLM_NORCLM]);
+ INIT_LIST_HEAD(&zone->free_area[order].free_list[RCLM_EASY]);
zone->free_area[order].nr_free = 0;
}
}
The freelists for each allocation type can slowly become corrupted due to
the per-cpu list. Consider what happens when the following happens
1. A 2^(MAX_ORDER-1) list is reserved for __GFP_EASYRCLM pages
2. An order-0 page is allocated from the newly reserved block
3. The page is freed and placed on the per-cpu list
4. alloc_page() is called with GFP_KERNEL as the gfp_mask
5. The per-cpu list is used to satisfy the allocation
This results in a kernel page is in the middle of a RCLM_EASY region. This
means that over long periods of the time, the anti-fragmentation scheme
slowly degrades to the standard allocator.
This patch divides the per-cpu lists into RCLM_TYPES number of lists.
Signed-off-by: Mel Gorman <[email protected]>
Signed-off-by: Joel Schopp <[email protected]>
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-002_fragcore/include/linux/mmzone.h linux-2.6.15-rc1-mm2-003_percpu/include/linux/mmzone.h
--- linux-2.6.15-rc1-mm2-002_fragcore/include/linux/mmzone.h 2005-11-22 16:50:09.000000000 +0000
+++ linux-2.6.15-rc1-mm2-003_percpu/include/linux/mmzone.h 2005-11-22 16:52:10.000000000 +0000
@@ -26,6 +26,8 @@
#define RCLM_EASY 1
#define RCLM_TYPES 2
+#define for_each_rclmtype(type) \
+ for (type = 0; type < RCLM_TYPES; type++)
#define for_each_rclmtype_order(type, order) \
for (order = 0; order < MAX_ORDER; order++) \
for (type = 0; type < RCLM_TYPES; type++)
@@ -53,11 +55,11 @@ struct zone_padding {
#endif
struct per_cpu_pages {
- int count; /* number of pages in the list */
+ int count[RCLM_TYPES]; /* Number of pages on the lists */
int low; /* low watermark, refill needed */
int high; /* high watermark, emptying needed */
int batch; /* chunk size for buddy add/remove */
- struct list_head list; /* the list of pages */
+ struct list_head list[RCLM_TYPES]; /* the lists of pages */
};
struct per_cpu_pageset {
@@ -72,6 +74,11 @@ struct per_cpu_pageset {
#endif
} ____cacheline_aligned_in_smp;
+static inline int pcp_count(struct per_cpu_pages *pcp)
+{
+ return pcp->count[RCLM_NORCLM] + pcp->count[RCLM_EASY];
+}
+
#ifdef CONFIG_NUMA
#define zone_pcp(__z, __cpu) ((__z)->pageset[(__cpu)])
#else
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-002_fragcore/mm/page_alloc.c linux-2.6.15-rc1-mm2-003_percpu/mm/page_alloc.c
--- linux-2.6.15-rc1-mm2-002_fragcore/mm/page_alloc.c 2005-11-22 16:50:09.000000000 +0000
+++ linux-2.6.15-rc1-mm2-003_percpu/mm/page_alloc.c 2005-11-22 16:52:10.000000000 +0000
@@ -637,7 +637,7 @@ static int rmqueue_bulk(struct zone *zon
void drain_remote_pages(void)
{
struct zone *zone;
- int i;
+ int i, pindex;
unsigned long flags;
local_irq_save(flags);
@@ -653,9 +653,16 @@ void drain_remote_pages(void)
struct per_cpu_pages *pcp;
pcp = &pset->pcp[i];
- if (pcp->count)
- pcp->count -= free_pages_bulk(zone, pcp->count,
- &pcp->list, 0);
+ for_each_rclmtype(pindex) {
+ if (!pcp->count[pindex])
+ continue;
+
+ /* Try remove all pages from the pcpu list */
+ pcp->count[pindex] -=
+ free_pages_bulk(zone,
+ pcp->count[pindex],
+ &pcp->list[pindex], 0);
+ }
}
}
local_irq_restore(flags);
@@ -666,7 +673,7 @@ void drain_remote_pages(void)
static void __drain_pages(unsigned int cpu)
{
struct zone *zone;
- int i;
+ int i, pindex;
for_each_zone(zone) {
struct per_cpu_pageset *pset;
@@ -676,8 +683,16 @@ static void __drain_pages(unsigned int c
struct per_cpu_pages *pcp;
pcp = &pset->pcp[i];
- pcp->count -= free_pages_bulk(zone, pcp->count,
- &pcp->list, 0);
+ for_each_rclmtype(pindex) {
+ if (!pcp->count[pindex])
+ continue;
+
+ /* Try remove all pages from the pcpu list */
+ pcp->count[pindex] -=
+ free_pages_bulk(zone,
+ pcp->count[pindex],
+ &pcp->list[pindex], 0);
+ }
}
}
}
@@ -758,6 +773,7 @@ static void FASTCALL(free_hot_cold_page(
static void fastcall free_hot_cold_page(struct page *page, int cold)
{
struct zone *zone = page_zone(page);
+ int pindex = get_pageblock_type(page);
struct per_cpu_pages *pcp;
unsigned long flags;
@@ -773,10 +789,11 @@ static void fastcall free_hot_cold_page(
pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
local_irq_save(flags);
- list_add(&page->lru, &pcp->list);
- pcp->count++;
- if (pcp->count >= pcp->high)
- pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
+ list_add(&page->lru, &pcp->list[pindex]);
+ pcp->count[pindex]++;
+ if (pcp->count[pindex] >= pcp->high)
+ pcp->count[pindex] -= free_pages_bulk(zone, pcp->batch,
+ &pcp->list[pindex], 0);
local_irq_restore(flags);
put_cpu();
}
@@ -820,14 +837,16 @@ again:
page = NULL;
pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
local_irq_save(flags);
- if (pcp->count <= pcp->low)
- pcp->count += rmqueue_bulk(zone, 0,
- pcp->batch, &pcp->list,
+ if (pcp->count[alloctype] <= pcp->low)
+ pcp->count[alloctype] += rmqueue_bulk(zone, 0,
+ pcp->batch,
+ &pcp->list[alloctype],
alloctype);
- if (pcp->count) {
- page = list_entry(pcp->list.next, struct page, lru);
+ if (pcp->count[alloctype]) {
+ page = list_entry(pcp->list[alloctype].next,
+ struct page, lru);
list_del(&page->lru);
- pcp->count--;
+ pcp->count[alloctype]--;
}
local_irq_restore(flags);
put_cpu();
@@ -1478,7 +1497,7 @@ void show_free_areas(void)
pageset->pcp[temperature].low,
pageset->pcp[temperature].high,
pageset->pcp[temperature].batch,
- pageset->pcp[temperature].count);
+ pcp_count(&pageset->pcp[temperature]));
}
}
@@ -1920,18 +1939,23 @@ inline void setup_pageset(struct per_cpu
memset(p, 0, sizeof(*p));
pcp = &p->pcp[0]; /* hot */
- pcp->count = 0;
+ pcp->count[RCLM_NORCLM] = 0;
+ pcp->count[RCLM_EASY] = 0;
pcp->low = 0;
pcp->high = 6 * batch;
pcp->batch = max(1UL, 1 * batch);
- INIT_LIST_HEAD(&pcp->list);
+ INIT_LIST_HEAD(&pcp->list[RCLM_NORCLM]);
+ INIT_LIST_HEAD(&pcp->list[RCLM_EASY]);
pcp = &p->pcp[1]; /* cold*/
- pcp->count = 0;
+
+ pcp->count[RCLM_NORCLM] = 0;
+ pcp->count[RCLM_EASY] = 0;
pcp->low = 0;
pcp->high = 2 * batch;
pcp->batch = max(1UL, batch/2);
- INIT_LIST_HEAD(&pcp->list);
+ INIT_LIST_HEAD(&pcp->list[RCLM_NORCLM]);
+ INIT_LIST_HEAD(&pcp->list[RCLM_EASY]);
}
#ifdef CONFIG_NUMA
@@ -2328,7 +2352,7 @@ static int zoneinfo_show(struct seq_file
"\n high: %i"
"\n batch: %i",
i, j,
- pageset->pcp[j].count,
+ pcp_count(&pageset->pcp[j]),
pageset->pcp[j].low,
pageset->pcp[j].high,
pageset->pcp[j].batch);
The anti-defragmentation strategy has memory overhead. This patch allows
the strategy to be disabled for small memory systems or if it is known the
workload is suffering because of the strategy. It also acts to show where
the anti-defrag strategy interacts with the standard buddy allocator.
Signed-off-by: Mel Gorman <[email protected]>
Signed-off-by: Joel Schopp <[email protected]>
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-003_percpu/include/linux/mmzone.h linux-2.6.15-rc1-mm2-004_configurable/include/linux/mmzone.h
--- linux-2.6.15-rc1-mm2-003_percpu/include/linux/mmzone.h 2005-11-22 16:52:10.000000000 +0000
+++ linux-2.6.15-rc1-mm2-004_configurable/include/linux/mmzone.h 2005-11-22 16:53:03.000000000 +0000
@@ -74,10 +74,17 @@ struct per_cpu_pageset {
#endif
} ____cacheline_aligned_in_smp;
+#ifdef CONFIG_PAGEALLOC_ANTIDEFRAG
static inline int pcp_count(struct per_cpu_pages *pcp)
{
return pcp->count[RCLM_NORCLM] + pcp->count[RCLM_EASY];
}
+#else
+static inline int pcp_count(struct per_cpu_pages *pcp)
+{
+ return pcp->count[RCLM_NORCLM];
+}
+#endif /* CONFIG_PAGEALLOC_ANTIDEFRAG */
#ifdef CONFIG_NUMA
#define zone_pcp(__z, __cpu) ((__z)->pageset[(__cpu)])
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-003_percpu/init/Kconfig linux-2.6.15-rc1-mm2-004_configurable/init/Kconfig
--- linux-2.6.15-rc1-mm2-003_percpu/init/Kconfig 2005-11-21 19:44:33.000000000 +0000
+++ linux-2.6.15-rc1-mm2-004_configurable/init/Kconfig 2005-11-22 16:53:03.000000000 +0000
@@ -396,6 +396,18 @@ config CC_ALIGN_FUNCTIONS
32-byte boundary only if this can be done by skipping 23 bytes or less.
Zero means use compiler's default.
+config PAGEALLOC_ANTIDEFRAG
+ bool "Avoid fragmentation in the page allocator"
+ def_bool n
+ help
+ The standard allocator will fragment memory over time which means that
+ high order allocations will fail even if kswapd is running. If this
+ option is set, the allocator will try and group page types into
+ two groups, kernel and easy reclaimable. The gain is a best effort
+ attempt at lowering fragmentation which a few workloads care about.
+ The loss is a more complex allocactor that performs slower.
+ If unsure, say N
+
config CC_ALIGN_LABELS
int "Label alignment" if EMBEDDED
default 0
diff -rup -X /usr/src/patchset-0.5/bin//dontdiff linux-2.6.15-rc1-mm2-003_percpu/mm/page_alloc.c linux-2.6.15-rc1-mm2-004_configurable/mm/page_alloc.c
--- linux-2.6.15-rc1-mm2-003_percpu/mm/page_alloc.c 2005-11-22 16:52:10.000000000 +0000
+++ linux-2.6.15-rc1-mm2-004_configurable/mm/page_alloc.c 2005-11-22 16:53:03.000000000 +0000
@@ -68,6 +68,7 @@ int sysctl_lowmem_reserve_ratio[MAX_NR_Z
EXPORT_SYMBOL(totalram_pages);
+#ifdef CONFIG_PAGEALLOC_ANTIDEFRAG
static inline int get_pageblock_type(struct page *page)
{
return (PageEasyRclm(page) != 0);
@@ -77,6 +78,17 @@ static inline int gfpflags_to_alloctype(
{
return ((gfp_flags & __GFP_EASYRCLM) != 0);
}
+#else
+static inline int get_pageblock_type(struct page *page)
+{
+ return RCLM_NORCLM;
+}
+
+static inline int gfpflags_to_alloctype(unsigned long gfp_flags)
+{
+ return RCLM_NORCLM;
+}
+#endif /* CONFIG_PAGEALLOC_ANTIDEFRAG */
/*
* Used by page_zone() to look up the address of the struct zone whose
@@ -531,6 +543,7 @@ static int prep_new_page(struct page *pa
return 0;
}
+#ifdef CONFIG_PAGEALLOC_ANTIDEFRAG
/* Remove an element from the buddy allocator from the fallback list */
static struct page *__rmqueue_fallback(struct zone *zone, int order,
int alloctype)
@@ -568,6 +581,13 @@ static struct page *__rmqueue_fallback(s
return NULL;
}
+#else
+static struct page *__rmqueue_fallback(struct zone *zone, unsigned int order,
+ int alloctype)
+{
+ return NULL;
+}
+#endif /* CONFIG_PAGEALLOC_ANTIDEFRAG */
/*
* Do the hard work of removing an element from the buddy allocator.
From: Mel Gorman Sent: Tuesday, November 22, 2005 11:18 AM
>Per-cpu pages can accidentally cause fragmentation because they are
free, >but
>pinned pages in an otherwise contiguous block. When this patch is
applied,
>the per-cpu caches are drained after the direct-reclaim is entered if
the
I don't think this is the right place to drain the pcp. Since direct
reclaim is already done, so it is possible that allocator can service
the request without draining the pcps.
>requested order is greater than 3.
Why this order limit. Most of the previous failures seen (because of my
earlier patches of bigger and more physical contiguous chunks for pcps)
were with order 1 allocation.
>It simply reuses the code used by suspend
>and hotplug and only is triggered when anti-defragmentation is enabled.
>
That code has issues with pre-emptible kernel.
I will be shortly sending the patch to free pages from pcp when higher
order
allocation is not able to get serviced from global list.
-rohi
On Tue, 22 Nov 2005, Seth, Rohit wrote:
> From: Mel Gorman Sent: Tuesday, November 22, 2005 11:18 AM
>
> >Per-cpu pages can accidentally cause fragmentation because they are
> free, >but
> >pinned pages in an otherwise contiguous block. When this patch is
> applied,
> >the per-cpu caches are drained after the direct-reclaim is entered if
> the
>
> I don't think this is the right place to drain the pcp. Since direct
> reclaim is already done, so it is possible that allocator can service
> the request without draining the pcps.
>
ok, true. A check should be made to see if it's possible yet and if not,
then drain. A more appropriate place might be after this block
if (page)
goto got_pg;
>
> >requested order is greater than 3.
>
> Why this order limit. Most of the previous failures seen (because of my
> earlier patches of bigger and more physical contiguous chunks for pcps)
> were with order 1 allocation.
>
The order 3 is because of this block;
if (!(gfp_mask & __GFP_NORETRY)) {
if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
do_retry = 1;
if (gfp_mask & __GFP_NOFAIL)
do_retry = 1;
}
If it's less than 3, we are retrying anyway and it's something we are
already doing. If it was felt it had a chance of working before, I felt
that draining per-cpu caches was unnecessary.
> >It simply reuses the code used by suspend
> >and hotplug and only is triggered when anti-defragmentation is enabled.
> >
> That code has issues with pre-emptible kernel.
>
ok... why? I thought that we could only be preempted when we were about to
take a spinlock but I have an imperfect understanding of preempt and
things change quickly. The path the drain_all_local_pages() enters
disables the local IRQs before calling __drain_pages() and when
smp_drain_local_pages() is called, the local IRQs are disabled again
before releasing pages. Where can we get preempted?
> I will be shortly sending the patch to free pages from pcp when higher
> order allocation is not able to get serviced from global list.
>
If that works, this part of the patch can be dropped. The intention is to
"drain the per-cpu lists by some mechanism". I am not too particular about
how it happens. Right now, the per-cpu caches make a massive difference on
my 4-way machine at least on whether a large number of contiguous blocks
can be allocated or not.
--
Mel Gorman
Part-time Phd Student Java Applications Developer
University of Limerick IBM Dublin Software Lab
On Wed, 2005-11-23 at 00:17 +0000, Mel Gorman wrote:
> On Tue, 22 Nov 2005, Seth, Rohit wrote:
>
> >
> >
> > >requested order is greater than 3.
> >
> > Why this order limit. Most of the previous failures seen (because of my
> > earlier patches of bigger and more physical contiguous chunks for pcps)
> > were with order 1 allocation.
> >
>
> The order 3 is because of this block;
>
> if (!(gfp_mask & __GFP_NORETRY)) {
> if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
> do_retry = 1;
> if (gfp_mask & __GFP_NOFAIL)
> do_retry = 1;
> }
>
> If it's less than 3, we are retrying anyway and it's something we are
You are retrying (for 0<order<=3) but without draining the pcps (in your
patch).
> > That code has issues with pre-emptible kernel.
> >
>
> ok... why? I thought that we could only be preempted when we were about to
> take a spinlock but I have an imperfect understanding of preempt and
> things change quickly. The path the drain_all_local_pages() enters
> disables the local IRQs before calling __drain_pages() and when
> smp_drain_local_pages() is called, the local IRQs are disabled again
> before releasing pages. Where can we get preempted?
>
Basically the get_cpu(), put_cpu() needs to cover the whole scope of
smp_processor_id usage. (When you enable CONFIG_DEBUG_PREEMPT the
kernel will barf if preempt is enabled while calling smp_processor_id).
If the interrupts are disabled all the way through then you wouldn't be
preempted though. But get/put_cpu is the right mechanism to ensure
smp_processor_id and its derived value is used on same processor.
> > I will be shortly sending the patch to free pages from pcp when higher
> > order allocation is not able to get serviced from global list.
> >
>
> If that works, this part of the patch can be dropped. The intention is to
> "drain the per-cpu lists by some mechanism". I am not too particular about
> how it happens. Right now, the per-cpu caches make a massive difference on
> my 4-way machine at least on whether a large number of contiguous blocks
> can be allocated or not.
>
Please let me know if you see any issues with the patch that I sent out
a bit earlier.
Thanks,
-rohit
+#define __GFP_EASYRCLM ((__force gfp_t)0x40000u) /* Easily reclaimed page */
Acked (this one line) by pj <grin>.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
On Tue, 22 Nov 2005, Rohit Seth wrote:
> On Wed, 2005-11-23 at 00:17 +0000, Mel Gorman wrote:
> > On Tue, 22 Nov 2005, Seth, Rohit wrote:
> >
> > >
> > >
> > > >requested order is greater than 3.
> > >
> > > Why this order limit. Most of the previous failures seen (because of my
> > > earlier patches of bigger and more physical contiguous chunks for pcps)
> > > were with order 1 allocation.
> > >
> >
> > The order 3 is because of this block;
> >
> > if (!(gfp_mask & __GFP_NORETRY)) {
> > if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
> > do_retry = 1;
> > if (gfp_mask & __GFP_NOFAIL)
> > do_retry = 1;
> > }
> >
> > If it's less than 3, we are retrying anyway and it's something we are
>
> You are retrying (for 0<order<=3) but without draining the pcps (in your
> patch).
>
> > > That code has issues with pre-emptible kernel.
> > >
> >
> > ok... why? I thought that we could only be preempted when we were about to
> > take a spinlock but I have an imperfect understanding of preempt and
> > things change quickly. The path the drain_all_local_pages() enters
> > disables the local IRQs before calling __drain_pages() and when
> > smp_drain_local_pages() is called, the local IRQs are disabled again
> > before releasing pages. Where can we get preempted?
> >
>
> Basically the get_cpu(), put_cpu() needs to cover the whole scope of
> smp_processor_id usage. (When you enable CONFIG_DEBUG_PREEMPT the
> kernel will barf if preempt is enabled while calling smp_processor_id).
>
> If the interrupts are disabled all the way through then you wouldn't be
> preempted though. But get/put_cpu is the right mechanism to ensure
> smp_processor_id and its derived value is used on same processor.
>
That can be easily enough fixed.
> > > I will be shortly sending the patch to free pages from pcp when higher
> > > order allocation is not able to get serviced from global list.
> > >
> >
> > If that works, this part of the patch can be dropped. The intention is to
> > "drain the per-cpu lists by some mechanism". I am not too particular about
> > how it happens. Right now, the per-cpu caches make a massive difference on
> > my 4-way machine at least on whether a large number of contiguous blocks
> > can be allocated or not.
> >
>
> Please let me know if you see any issues with the patch that I sent out
> a bit earlier.
>
I don't have access to my test environment for the rest of the week so I
can't actually try them out.
However, reading through the patches, they appear to duplicate a
significant amount of the existing drain_local_pages() functions and they
only drain the pages on the currently running CPU. On a system with a
number of CPUs, you will only be improving your chances slightly.
I think you would get more of what you need with this patch if;
1. Removed the compile time dependency on CONFIG_PM||CONFIG_HOTPLUG
2. Rechecked the usage of smp_processor_id() (although I don't think it's
wrong because it's only called with local IRQs disabled)
3. Draining the CPUs after direct reclaim and the allocation still failing
This patch does everything you need including the draining of remove
per-cpus.
--
Mel Gorman
Part-time Phd Student Java Applications Developer
University of Limerick IBM Dublin Software Lab