sfrench/cifs-2.6.git
13 years agoBe more agressive about stealing when MIGRATE_RECLAIMABLE allocations fallback
Mel Gorman [Tue, 16 Oct 2007 08:25:55 +0000 (01:25 -0700)]
Be more agressive about stealing when MIGRATE_RECLAIMABLE allocations fallback

MIGRATE_RECLAIMABLE allocations tend to be very bursty in nature like when
updatedb starts.  It is likely this will occur in situations where MAX_ORDER
blocks of pages are not free.  This means that updatedb can scatter
MIGRATE_RECLAIMABLE pages throughout the address space.  This patch is more
agressive about stealing blocks of pages for MIGRATE_RECLAIMABLE.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoBias the placement of kernel pages at lower PFNs
Mel Gorman [Tue, 16 Oct 2007 08:25:54 +0000 (01:25 -0700)]
Bias the placement of kernel pages at lower PFNs

This patch chooses blocks with lower PFNs when placing kernel allocations.
This is particularly important during fallback in low memory situations to
stop unmovable pages being placed throughout the entire address space.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoDo not group pages by mobility type on low memory systems
Mel Gorman [Tue, 16 Oct 2007 08:25:54 +0000 (01:25 -0700)]
Do not group pages by mobility type on low memory systems

Grouping pages by mobility can only successfully operate when there are more
MAX_ORDER_NR_PAGES areas than mobility types.  When there are insufficient
areas, fallbacks cannot be avoided.  This has noticeable performance impacts
on machines with small amounts of memory in comparison to MAX_ORDER_NR_PAGES.
For example, on IA64 with a configuration including huge pages spans 1GiB with
MAX_ORDER_NR_PAGES so would need at least 4GiB of RAM before grouping pages by
mobility would be useful.  In comparison, an x86 would need 16MB.

This patch checks the size of vm_total_pages in build_all_zonelists(). If
there are not enough areas,  mobility is effectivly disabled by considering
all allocations as the same type (UNMOVABLE).  This is achived via a
__read_mostly flag.

With this patch, performance is comparable to disabling grouping pages
by mobility at compile-time on a test machine with insufficient memory.
With this patch, it is reasonable to get rid of grouping pages by mobility
a compile-time option.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoGroup high-order atomic allocations
Mel Gorman [Tue, 16 Oct 2007 08:25:53 +0000 (01:25 -0700)]
Group high-order atomic allocations

In rare cases, the kernel needs to allocate a high-order block of pages
without sleeping.  For example, this is the case with e1000 cards configured
to use jumbo frames.  Migrating or reclaiming pages in this situation is not
an option.

This patch groups these allocations together as much as possible by adding a
new MIGRATE_TYPE.  The MIGRATE_HIGHATOMIC type are exactly what they sound
like.  Care is taken that pages of other migrate types do not use the same
blocks as high-order atomic allocations.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoGroup short-lived and reclaimable kernel allocations
Mel Gorman [Tue, 16 Oct 2007 08:25:52 +0000 (01:25 -0700)]
Group short-lived and reclaimable kernel allocations

This patch marks a number of allocations that are either short-lived such as
network buffers or are reclaimable such as inode allocations.  When something
like updatedb is called, long-lived and unmovable kernel allocations tend to
be spread throughout the address space which increases fragmentation.

This patch groups these allocations together as much as possible by adding a
new MIGRATE_TYPE.  The MIGRATE_RECLAIMABLE type is for allocations that can be
reclaimed on demand, but not moved.  i.e.  they can be migrated by deleting
them and re-reading the information from elsewhere.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMove free pages between lists on steal
Mel Gorman [Tue, 16 Oct 2007 08:25:51 +0000 (01:25 -0700)]
Move free pages between lists on steal

When a fallback occurs, there will be free pages for one allocation type
stored on the list for another.  When a large steal occurs, this patch will
move all the free pages within one list to the other.

[y-goto@jp.fujitsu.com: fix BUG_ON check at move_freepages()]
[apw@shadowen.org: Move to using pfn_valid_within()]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Bjorn Helgaas <bjorn.helgaas@hp.com>
Signed-off-by: Andy Whitcroft <andyw@uk.ibm.com>
Cc: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoDrain per-cpu lists when high-order allocations fail
Mel Gorman [Tue, 16 Oct 2007 08:25:50 +0000 (01:25 -0700)]
Drain per-cpu lists when high-order allocations fail

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 0.  It simply reuses the code used by suspend
and hotplug.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoAdd a configure option to group pages by mobility
Mel Gorman [Tue, 16 Oct 2007 08:25:50 +0000 (01:25 -0700)]
Add a configure option to group pages by mobility

The grouping mechanism has some memory overhead and a more complex allocation
path.  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 page groupings strategy interacts with the standard
buddy allocator.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Joel Schopp <jschopp@austin.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoChoose pages from the per-cpu list based on migration type
Mel Gorman [Tue, 16 Oct 2007 08:25:49 +0000 (01:25 -0700)]
Choose pages from the per-cpu list based on migration type

The freelists for each migrate type can slowly become polluted due to the
per-cpu list.  Consider what happens when the following happens

1. A 2^(MAX_ORDER-1) list is reserved for __GFP_MOVABLE 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 migratable region. This
patch prevents this leak occuring by storing the MIGRATE_ type of the page in
page->private. On allocate, a page will only be returned of the desired type,
else more pages will be allocated. This may temporarily allow a per-cpu list
to go over the pcp->high limit but it'll be corrected on the next free. Care
is taken to preserve the hotness of pages recently freed.

The additional code is not measurably slower for the workloads we've tested.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoSplit the free lists for movable and unmovable allocations
Mel Gorman [Tue, 16 Oct 2007 08:25:48 +0000 (01:25 -0700)]
Split the free lists for movable and unmovable allocations

This patch adds the core of the fragmentation reduction strategy.  It works by
grouping pages together based on their ability to migrate or be reclaimed.
Basically, it works by breaking the list in zone->free_area list into
MIGRATE_TYPES number of lists.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoAdd a bitmap that is used to track flags affecting a block of pages
Mel Gorman [Tue, 16 Oct 2007 08:25:47 +0000 (01:25 -0700)]
Add a bitmap that is used to track flags affecting a block of pages

Here is the latest revision of the anti-fragmentation patches.  Of particular
note in this version is special treatment of high-order atomic allocations.
Care is taken to group them together and avoid grouping pages of other types
near them.  Artifical tests imply that it works.  I'm trying to get the
hardware together that would allow setting up of a "real" test.  If anyone
already has a setup and test that can trigger the atomic-allocation problem,
I'd appreciate a test of these patches and a report.  The second major change
is that these patches will apply cleanly with patches that implement
anti-fragmentation through zones.

kernbench shows effectively no performance difference varying between -0.2%
and +2% on a variety of test machines.  Success rates for huge page allocation
are dramatically increased.  For example, on a ppc64 machine, the vanilla
kernel was only able to allocate 1% of memory as a hugepage and this was due
to a single hugepage reserved as min_free_kbytes.  With these patches applied,
17% was allocatable as superpages.  With reclaim-related fixes from Andy
Whitcroft, it was 40% and further reclaim-related improvements should increase
this further.

Changelog Since V28
o Group high-order atomic allocations together
o It is no longer required to set min_free_kbytes to 10% of memory. A value
  of 16384 in most cases will be sufficient
o Now applied with zone-based anti-fragmentation
o Fix incorrect VM_BUG_ON within buffered_rmqueue()
o Reorder the stack so later patches do not back out work from earlier patches
o Fix bug were journal pages were being treated as movable
o Bias placement of non-movable pages to lower PFNs
o More agressive clustering of reclaimable pages in reactions to workloads
  like updatedb that flood the size of inode caches

Changelog Since V27

o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving
  the mistaken impression that it was the 100% solution for high order
  allocations. Instead, it greatly increases the chances high-order
  allocations will succeed and lays the foundation for defragmentation and
  memory hot-remove to work properly
o Redefine page groupings based on ability to migrate or reclaim instead of
  basing on reclaimability alone
o Get rid of spurious inits
o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is
  searched for a page of the appropriate type
o Added more explanation commentary
o Fix up bug in pageblock code where bitmap was used before being initalised

Changelog Since V26
o Fix double init of lists in setup_pageset

Changelog Since V25
o Fix loop order of for_each_rclmtype_order so that order of loop matches args
o gfpflags_to_rclmtype uses gfp_t instead of unsigned long
o Rename get_pageblock_type() to get_page_rclmtype()
o Fix alignment problem in move_freepages()
o Add mechanism for assigning flags to blocks of pages instead of page->flags
o On fallback, do not examine the preferred list of free pages a second time

The purpose of these patches is to reduce external fragmentation by grouping
pages of related types together.  When pages are migrated (or reclaimed under
memory pressure), large contiguous pages will be freed.

This patch works by categorising allocations by their ability to migrate;

Movable - The pages may be moved with the page migration mechanism. These are
generally userspace pages.

Reclaimable - These are allocations for some kernel caches that are
reclaimable or allocations that are known to be very short-lived.

Unmovable - 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

HighAtomic - These are high-order allocations belonging to callers that
cannot sleep or perform any IO. In practice, this is restricted to
jumbo frame allocation for network receive. It is assumed that the
allocations are short-lived

Instead of having one MAX_ORDER-sized array of free lists in struct free_area,
there is one for each type of reclaimability.  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 the preferred freelists are expired, the largest possible block is taken
from an alternative list.  Buddies that are split from that large block are
placed on the preferred allocation-type freelists to mitigate fragmentation.

This implementation gives best-effort for low fragmentation in all zones.
Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 <<
(MAX_ORDER-1)) pages in most cases.  This would be 16384 on x86 and x86_64 for
example.

Our tests show that about 60-70% of physical memory can be allocated on a
desktop after a few days uptime.  In benchmarks and stress tests, we are
finding that 80% of memory is available as contiguous blocks at the end of the
test.  To compare, a standard kernel was getting < 1% of memory as large pages
on a desktop and about 8-12% of memory as large pages at the end of stress
tests.

Following this email are 12 patches that implement thie page grouping feature.
 The first patch introduces a mechanism for storing flags related to a whole
block of pages.  Then allocations are split between movable and all other
allocations.  Following that are patches to deal with per-cpu pages and make
the mechanism configurable.  The next patch moves free pages between lists
when partially allocated blocks are used for pages of another migrate type.
The second last patch groups reclaimable kernel allocations such as inode
caches together.  The final patch related to groupings keeps high-order atomic
allocations.

The last two patches are more concerned with control of fragmentation.  The
second last patch biases placement of non-movable allocations towards the
start of memory.  This is with a view of supporting memory hot-remove of DIMMs
with higher PFNs in the future.  The biasing could be enforced a lot heavier
but it would cost.  The last patch agressively clusters reclaimable pages like
inode caches together.

The fragmentation reduction strategy needs to track if pages within a block
can be moved or reclaimed so that pages are freed to the appropriate list.
This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of
pages.

In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and
allocated during initialisation.  SPARSEMEM statically allocates the bitmap in
a struct mem_section so that bitmaps do not have to be resized during memory
hotadd.  This wastes a small amount of memory per unused section (usually
sizeof(unsigned long)) but the complexity of dynamically allocating the memory
is quite high.

Additional credit to Andy Whitcroft who reviewed up an earlier implementation
of the mechanism an suggested how to make it a *lot* cleaner.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoflush icache before set_pte() on ia64: flush icache at set_pte
KAMEZAWA Hiroyuki [Tue, 16 Oct 2007 08:25:44 +0000 (01:25 -0700)]
flush icache before set_pte() on ia64: flush icache at set_pte

Current ia64 kernel flushes icache by lazy_mmu_prot_update() *after*
set_pte().  This is too late.  This patch removes lazy_mmu_prot_update and
add modfied set_pte() for flushing if necessary.

This patch flush icache of a page when
new pte has exec bit.
&& new pte has present bit
&& new pte is user's page.
&& (old *ptep is not present
            || new pte's pfn is not same to old *ptep's ptn)
&& new pte's page has no Pg_arch_1 bit.
   Pg_arch_1 is set when a page is cache consistent.

I think this condition checks are much easier to understand than considering
"Where sync_icache_dcache() should be inserted ?".

pte_user() for ia64 was removed by http://lkml.org/lkml/2007/6/12/67 as
clean-up. So, I added it again.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoflush cache before installing new page at migraton
KAMEZAWA Hiroyuki [Tue, 16 Oct 2007 08:25:43 +0000 (01:25 -0700)]
flush cache before installing new page at migraton

In migration, a new page should be cache flushed before set_pte() in some
archs which have virtually-tagged cache.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomake swappiness safer to use
Andrea Arcangeli [Tue, 16 Oct 2007 08:25:42 +0000 (01:25 -0700)]
make swappiness safer to use

Swappiness isn't a safe sysctl.  Setting it to 0 for example can hang a
system.  That's a corner case but even setting it to 10 or lower can waste
enormous amounts of cpu without making much progress.  We've customers who
wants to use swappiness but they can't because of the current
implementation (if you change it so the system stops swapping it really
stops swapping and nothing works sane anymore if you really had to swap
something to make progress).

This patch from Kurt Garloff makes swappiness safer to use (no more huge
cpu usage or hangs with low swappiness values).

I think the prev_priority can also be nuked since it wastes 4 bytes per
zone (that would be an incremental patch but I wait the nr_scan_[in]active
to be nuked first for similar reasons).  Clearly somebody at some point
noticed how broken that thing was and they had to add min(priority,
prev_priority) to give it some reliability, but they didn't go the last
mile to nuke prev_priority too.  Calculating distress only in function of
not-racy priority is correct and sure more than enough without having to
add randomness into the equation.

Patch is tested on older kernels but it compiles and it's quite simple
so...

Overall I'm not very satisified by the swappiness tweak, since it doesn't
rally do anything with the dirty pagecache that may be inactive.  We need
another kind of tweak that controls the inactive scan and tunes the
can_writepage feature (not yet in mainline despite having submitted it a
few times), not only the active one.  That new tweak will tell the kernel
how hard to scan the inactive list for pure clean pagecache (something the
mainline kernel isn't capable of yet).  We already have that feature
working in all our enterprise kernels with the default reasonable tune, or
they can't even run a readonly backup with tar without triggering huge
write I/O.  I think it should be available also in mainline later.

Cc: Nick Piggin <npiggin@suse.de>
Signed-off-by: Kurt Garloff <garloff@suse.de>
Signed-off-by: Andrea Arcangeli <andrea@suse.de>
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoCategorize GFP flags
Christoph Lameter [Tue, 16 Oct 2007 08:25:41 +0000 (01:25 -0700)]
Categorize GFP flags

The function of GFP_LEVEL_MASK seems to be unclear.  In order to clear up
the mystery we get rid of it and replace GFP_LEVEL_MASK with 3 sets of GFP
flags:

GFP_RECLAIM_MASK Flags used to control page allocator reclaim behavior.

GFP_CONSTRAINT_MASK Flags used to limit where allocations can occur.

GFP_SLAB_BUG_MASK Flags that the slab allocator BUG()s on.

These replace the uses of GFP_LEVEL mask in the slab allocators and in
vmalloc.c.

The use of the flags not included in these sets may occur as a result of a
slab allocation standing in for a page allocation when constructing scatter
gather lists.  Extraneous flags are cleared and not passed through to the
page allocator.  __GFP_MOVABLE/RECLAIMABLE, __GFP_COLD and __GFP_COMP will
now be ignored if passed to a slab allocator.

Change the allocation of allocator meta data in SLAB and vmalloc to not
pass through flags listed in GFP_CONSTRAINT_MASK.  SLAB already removes the
__GFP_THISNODE flag for such allocations.  Generalize that to also cover
vmalloc.  The use of GFP_CONSTRAINT_MASK also includes __GFP_HARDWALL.

The impact of allocator metadata placement on access latency to the
cachelines of the object itself is minimal since metadata is only
referenced on alloc and free.  The attempt is still made to place the meta
data optimally but we consistently allow fallback both in SLAB and vmalloc
(SLUB does not need to allocate metadata like that).

Allocator metadata may serve multiple in kernel users and thus should not
be subject to the limitations arising from a single allocation context.

[akpm@linux-foundation.org: fix fallback_alloc()]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoFix panic of cpu online with memory less node
Yasunori Goto [Tue, 16 Oct 2007 08:25:40 +0000 (01:25 -0700)]
Fix panic of cpu online with memory less node

When a cpu is onlined on memory-less-node box, kernel panics due to touch
NULL pointer of pgdat->kswapd.  Current kswapd runs only nodes which have
memory.  So, calling of set_cpus_allowed() is not necessary for memory-less
node.

This is fix for it.

Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomemoryless nodes: fixup uses of node_online_map in generic code
Lee Schermerhorn [Tue, 16 Oct 2007 08:25:39 +0000 (01:25 -0700)]
memoryless nodes: fixup uses of node_online_map in generic code

Here's a cut at fixing up uses of the online node map in generic code.

mm/shmem.c:shmem_parse_mpol()

Ensure nodelist is subset of nodes with memory.
Use node_states[N_HIGH_MEMORY] as default for missing
nodelist for interleave policy.

mm/shmem.c:shmem_fill_super()

initialize policy_nodes to node_states[N_HIGH_MEMORY]

mm/page-writeback.c:highmem_dirtyable_memory()

sum over nodes with memory

mm/page_alloc.c:zlc_setup()

allowednodes - use nodes with memory.

mm/page_alloc.c:default_zonelist_order()

average over nodes with memory.

mm/page_alloc.c:find_next_best_node()

skip nodes w/o memory.
N_HIGH_MEMORY state mask may not be initialized at this time,
unless we want to depend on early_calculate_totalpages() [see
below].  Will ZONE_MOVABLE ever be configurable?

mm/page_alloc.c:find_zone_movable_pfns_for_nodes()

spread kernelcore over nodes with memory.

This required calling early_calculate_totalpages()
unconditionally, and populating N_HIGH_MEMORY node
state therein from nodes in the early_node_map[].
If we can depend on this, we can eliminate the
population of N_HIGH_MEMORY mask from __build_all_zonelists()
and use the N_HIGH_MEMORY mask in find_next_best_node().

mm/mempolicy.c:mpol_check_policy()

Ensure nodes specified for policy are subset of
nodes with memory.

[akpm@linux-foundation.org: fix warnings]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Christoph Lameter <clameter@sgi.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Use N_HIGH_MEMORY for cpusets
Christoph Lameter [Tue, 16 Oct 2007 08:25:38 +0000 (01:25 -0700)]
Memoryless nodes: Use N_HIGH_MEMORY for cpusets

cpusets try to ensure that any node added to a cpuset's mems_allowed is
on-line and contains memory.  The assumption was that online nodes contained
memory.  Thus, it is possible to add memoryless nodes to a cpuset and then add
tasks to this cpuset.  This results in continuous series of oom-kill and
apparent system hang.

Change cpusets to use node_states[N_HIGH_MEMORY] [a.k.a.  node_memory_map] in
place of node_online_map when vetting memories.  Return error if admin
attempts to write a non-empty mems_allowed node mask containing only
memoryless-nodes.

Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Fix GFP_THISNODE behavior
Christoph Lameter [Tue, 16 Oct 2007 08:25:37 +0000 (01:25 -0700)]
Memoryless nodes: Fix GFP_THISNODE behavior

GFP_THISNODE checks that the zone selected is within the pgdat (node) of the
first zone of a nodelist.  That only works if the node has memory.  A
memoryless node will have its first node on another pgdat (node).

GFP_THISNODE currently will return simply memory on the first pgdat.  Thus it
is returning memory on other nodes.  GFP_THISNODE should fail if there is no
local memory on a node.

Add a new set of zonelists for each node that only contain the nodes that
belong to the zones itself so that no fallback is possible.

Then modify gfp_type to pickup the right zone based on the presence of
__GFP_THISNODE.

Drop the existing GFP_THISNODE checks from the page_allocators hot path.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: drop one memoryless node boot warning
Christoph Lameter [Tue, 16 Oct 2007 08:25:37 +0000 (01:25 -0700)]
Memoryless nodes: drop one memoryless node boot warning

get_pfn_range_for_nid() is called multiple times for each node at boot time.
Each time, it will warn about nodes with no memory, resulting in boot messages
like:

        Node 0 active with no memory
        Node 0 active with no memory
        Node 0 active with no memory
        Node 0 active with no memory
        Node 0 active with no memory
        Node 0 active with no memory
        On node 0 totalpages: 0
        Node 0 active with no memory
        Node 0 active with no memory
          DMA zone: 0 pages used for memmap
        Node 0 active with no memory
        Node 0 active with no memory
          Normal zone: 0 pages used for memmap
        Node 0 active with no memory
        Node 0 active with no memory
          Movable zone: 0 pages used for memmap

and so on for each memoryless node.

We already have the "On node N totalpages: ..." and other related messages, so
drop the "Node N active with no memory" warnings.

Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Add N_CPU node state
Christoph Lameter [Tue, 16 Oct 2007 08:25:36 +0000 (01:25 -0700)]
Memoryless nodes: Add N_CPU node state

We need the check for a node with cpu in zone reclaim.  Zone reclaim will not
allow remote zone reclaim if a node has a cpu.

[Lee.Schermerhorn@hp.com: Move setup of N_CPU node state mask]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Update memory policy and page migration
Christoph Lameter [Tue, 16 Oct 2007 08:25:35 +0000 (01:25 -0700)]
Memoryless nodes: Update memory policy and page migration

Online nodes now may have no memory.  The checks and initialization must
therefore be changed to no longer use the online functions.

This will correctly initialize the interleave on bootup to only target nodes
with memory and will make sys_move_pages return an error when a page is to be
moved to a memoryless node.  Similarly we will get an error if MPOL_BIND and
MPOL_INTERLEAVE is used on a memoryless node.

These are somewhat new semantics.  So far one could specify memoryless nodes
and we would maybe do the right thing and just ignore the node (or we'd do
something strange like with MPOL_INTERLEAVE).  If we want to allow the
specification of memoryless nodes via memory policies then we need to keep
checking for online nodes.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Allow profiling data to fall back to other nodes
Christoph Lameter [Tue, 16 Oct 2007 08:25:34 +0000 (01:25 -0700)]
Memoryless nodes: Allow profiling data to fall back to other nodes

Processors on memoryless nodes must be able to fall back to remote nodes in
order to get a profiling buffer.  This may lead to excessive NUMA traffic but
I think we should allow this rather than failing.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Uncached allocator updates
Christoph Lameter [Tue, 16 Oct 2007 08:25:33 +0000 (01:25 -0700)]
Memoryless nodes: Uncached allocator updates

The checks for node_online in the uncached allocator are made to make sure
that memory is available on these nodes.  Thus switch all the checks to use
N_HIGH_MEMORY and to N_ONLINE.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Jes Sorensen <jes@sgi.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: SLUB support
Christoph Lameter [Tue, 16 Oct 2007 08:25:33 +0000 (01:25 -0700)]
Memoryless nodes: SLUB support

Simply switch all for_each_online_node to for_each_node_state(NORMAL_MEMORY).
That way SLUB only operates on nodes with regular memory.  Any allocation
attempt on a memoryless node or a node with just highmem will fall whereupon
SLUB will fetch memory from a nearby node (depending on how memory policies
and cpuset describe fallback).

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Slab support
Christoph Lameter [Tue, 16 Oct 2007 08:25:32 +0000 (01:25 -0700)]
Memoryless nodes: Slab support

Slab should not allocate control structures for nodes without memory.  This
may seem to work right now but its unreliable since not all allocations can
fall back due to the use of GFP_THISNODE.

Switching a few for_each_online_node's to N_NORMAL_MEMORY will allow us to
only allocate for nodes that have regular memory.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: No need for kswapd
Christoph Lameter [Tue, 16 Oct 2007 08:25:31 +0000 (01:25 -0700)]
Memoryless nodes: No need for kswapd

A node without memory does not need a kswapd.  So use the memory map instead
of the online map when starting kswapd.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: OOM: use N_HIGH_MEMORY map instead of constructing one on the fly
Christoph Lameter [Tue, 16 Oct 2007 08:25:30 +0000 (01:25 -0700)]
Memoryless nodes: OOM: use N_HIGH_MEMORY map instead of constructing one on the fly

constrained_alloc() builds its own memory map for nodes with memory.  We have
that available in N_HIGH_MEMORY now.  So simplify the code.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Fix interleave behavior for memoryless nodes
Christoph Lameter [Tue, 16 Oct 2007 08:25:30 +0000 (01:25 -0700)]
Memoryless nodes: Fix interleave behavior for memoryless nodes

MPOL_INTERLEAVE currently simply loops over all nodes.  Allocations on
memoryless nodes will be redirected to nodes with memory.  This results in an
imbalance because the neighboring nodes to memoryless nodes will get
significantly more interleave hits that the rest of the nodes on the system.

We can avoid this imbalance by clearing the nodes in the interleave node set
that have no memory.  If we use the node map of the memory nodes instead of
the online nodes then we have only the nodes we want.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: introduce mask of nodes with memory
Christoph Lameter [Tue, 16 Oct 2007 08:25:29 +0000 (01:25 -0700)]
Memoryless nodes: introduce mask of nodes with memory

It is necessary to know if nodes have memory since we have recently begun to
add support for memoryless nodes.  For that purpose we introduce a two new
node states: N_HIGH_MEMORY and N_NORMAL_MEMORY.

A node has its bit in N_HIGH_MEMORY set if it has any memory regardless of the
type of mmemory.  If a node has memory then it has at least one zone defined
in its pgdat structure that is located in the pgdat itself.

A node has its bit in N_NORMAL_MEMORY set if it has a lower zone than
ZONE_HIGHMEM.  This means it is possible to allocate memory that is not
subject to kmap.

N_HIGH_MEMORY and N_NORMAL_MEMORY can then be used in various places to insure
that we do the right thing when we encounter a memoryless node.

[akpm@linux-foundation.org: build fix]
[Lee.Schermerhorn@hp.com: update N_HIGH_MEMORY node state for memory hotadd]
[y-goto@jp.fujitsu.com: Fix memory hotplug + sparsemem build]
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMemoryless nodes: Generic management of nodemasks for various purposes
Christoph Lameter [Tue, 16 Oct 2007 08:25:27 +0000 (01:25 -0700)]
Memoryless nodes: Generic management of nodemasks for various purposes

Why do we need to support memoryless nodes?

KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:

> For fujitsu, problem is called "empty" node.
>
> When ACPI's SRAT table includes "possible nodes", ia64 bootstrap(acpi_numa_init)
> creates nodes, which includes no memory, no cpu.
>
> I tried to remove empty-node in past, but that was denied.
> It was because we can hot-add cpu to the empty node.
> (node-hotplug triggered by cpu is not implemented now. and it will be ugly.)
>
>
> For HP, (Lee can comment on this later), they have memory-less-node.
> As far as I hear, HP's machine can have following configration.
>
> (example)
> Node0: CPU0   memory AAA MB
> Node1: CPU1   memory AAA MB
> Node2: CPU2   memory AAA MB
> Node3: CPU3   memory AAA MB
> Node4: Memory XXX GB
>
> AAA is very small value (below 16MB)  and will be omitted by ia64 bootstrap.
> After boot, only Node 4 has valid memory (but have no cpu.)
>
> Maybe this is memory-interleave by firmware config.

Christoph Lameter <clameter@sgi.com> wrote:

> Future SGI platforms (actually also current one can have but nothing like
> that is deployed to my knowledge) have nodes with only cpus. Current SGI
> platforms have nodes with just I/O that we so far cannot manage in the
> core. So the arch code maps them to the nearest memory node.

Lee Schermerhorn <Lee.Schermerhorn@hp.com> wrote:

> For the HP platforms, we can configure each cell with from 0% to 100%
> "cell local memory".  When we configure with <100% CLM, the "missing
> percentages" are interleaved by hardware on a cache-line granularity to
> improve bandwidth at the expense of latency for numa-challenged
> applications [and OSes, but not our problem ;-)].  When we boot Linux on
> such a config, all of the real nodes have no memory--it all resides in a
> single interleaved pseudo-node.
>
> When we boot Linux on a 100% CLM configuration [== NUMA], we still have
> the interleaved pseudo-node.  It contains a few hundred MB stolen from
> the real nodes to contain the DMA zone.  [Interleaved memory resides at
> phys addr 0].  The memoryless-nodes patches, along with the zoneorder
> patches, support this config as well.
>
> Also, when we boot a NUMA config with the "mem=" command line,
> specifying less memory than actually exists, Linux takes the excluded
> memory "off the top" rather than distributing it across the nodes.  This
> can result in memoryless nodes, as well.
>

This patch:

Preparation for memoryless node patches.

Provide a generic way to keep nodemasks describing various characteristics of
NUMA nodes.

Remove the node_online_map and the node_possible map and realize the same
functionality using two nodes stats: N_POSSIBLE and N_ONLINE.

[Lee.Schermerhorn@hp.com: Initialize N_*_MEMORY and N_CPU masks for non-NUMA config]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: remove some AOP_TRUNCATED_PAGE
Nick Piggin [Tue, 16 Oct 2007 08:25:26 +0000 (01:25 -0700)]
fs: remove some AOP_TRUNCATED_PAGE

prepare/commit_write no longer returns AOP_TRUNCATED_PAGE since OCFS2 and
GFS2 were converted to the new aops, so we can make some simplifications
for that.

[michal.k.k.piotrowski@gmail.com: fix warning]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Michael Halcrow <mhalcrow@us.ibm.com>
Cc: Mark Fasheh <mark.fasheh@oracle.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Michal Piotrowski <michal.k.k.piotrowski@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: restore nobh
Nick Piggin [Tue, 16 Oct 2007 08:25:25 +0000 (01:25 -0700)]
fs: restore nobh

Implement nobh in new aops.  This is a bit tricky.  FWIW, nobh_truncate is
now implemented in a way that does not create blocks in sparse regions,
which is a silly thing for it to have been doing (isn't it?)

ext2 survives fsx and fsstress. jfs is converted as well... ext3
should be easy to do (but not done yet).

[akpm@linux-foundation.org: coding-style fixes]
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoocfs2: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:24 +0000 (01:25 -0700)]
ocfs2: convert to new aops

Plug ocfs2 into the ->write_begin and ->write_end aops.

A bunch of custom code is now gone - the iovec iteration stuff during write
and the ocfs2 splice write actor.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: affs convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:24 +0000 (01:25 -0700)]
fs: affs convert to new aops

Cc: Roman Zippel <zippel@linux-m68k.org>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: adfs convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:23 +0000 (01:25 -0700)]
fs: adfs convert to new aops

Acked-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agojfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:22 +0000 (01:25 -0700)]
jfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Dave Kleikamp <shaggy@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agominixfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:21 +0000 (01:25 -0700)]
minixfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Andries Brouwer <Andries.Brouwer@cwi.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agosysv: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:21 +0000 (01:25 -0700)]
sysv: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoudf: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:20 +0000 (01:25 -0700)]
udf: convert to new aops

Convert udf to new aops.  Also seem to have fixed pagecache corruption in
udf_adinicb_commit_write -- page was marked uptodate when it is not.  Also,
fixed the silly setup where prepare_write was doing a kmap to be used in
commit_write: just do kmap_atomic in write_end.  Use libfs helpers to make
this easier.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: <bfennema@falcon.csc.calpoly.edu>
Cc: Jan Kara <jack@ucw.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoufs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:19 +0000 (01:25 -0700)]
ufs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Evgeniy Dushistov <dushistov@mail.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agojffs2: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:18 +0000 (01:25 -0700)]
jffs2: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agohostfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:17 +0000 (01:25 -0700)]
hostfs: convert to new aops

This also gets rid of a lot of useless read_file stuff. And also
optimises the full page write case by marking a !uptodate page uptodate.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeff Dike <jdike@addtoit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofuse: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:17 +0000 (01:25 -0700)]
fuse: convert to new aops

[mszeredi]
 - don't send zero length write requests
 - it is not legal for the filesystem to return with zero written bytes

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agosmbfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:16 +0000 (01:25 -0700)]
smbfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agonfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:16 +0000 (01:25 -0700)]
nfs: convert to new aops

[akpm@linux-foundation.org: fix against git-nfs]
[peterz@infradead.org: fix against git-nfs]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoWith reiserfs no longer using the weird generic_cont_expand, remove it completely.
Nick Piggin [Tue, 16 Oct 2007 08:25:15 +0000 (01:25 -0700)]
With reiserfs no longer using the weird generic_cont_expand, remove it completely.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoreiserfs: use generic_cont_expand_simple
Vladimir Saveliev [Tue, 16 Oct 2007 08:25:14 +0000 (01:25 -0700)]
reiserfs: use generic_cont_expand_simple

This patch makes reiserfs to use AOP_FLAG_CONT_EXPAND
in order to get rid of the special generic_cont_expand routine

Signed-off-by: Vladimir Saveliev <vs@namesys.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoreiserfs: convert to new aops
Vladimir Saveliev [Tue, 16 Oct 2007 08:25:14 +0000 (01:25 -0700)]
reiserfs: convert to new aops

Convert reiserfs to new aops

Signed-off-by: Vladimir Saveliev <vs@namesys.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoreiserfs: use generic write
Vladimir Saveliev [Tue, 16 Oct 2007 08:25:12 +0000 (01:25 -0700)]
reiserfs: use generic write

Make reiserfs to write via generic routines.
Original reiserfs write optimized for big writes is deadlock rone

Signed-off-by: Vladimir Saveliev <vs@namesys.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoqnx4: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:12 +0000 (01:25 -0700)]
qnx4: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Anders Larsen <al@alarsen.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agobfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:11 +0000 (01:25 -0700)]
bfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Tigran Aivazian <tigran@aivazian.fsnet.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agohpfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:10 +0000 (01:25 -0700)]
hpfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: <mikulas@artax.karlin.mff.cuni.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agohfsplus: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:09 +0000 (01:25 -0700)]
hfsplus: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Roman Zippel <zippel@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agohfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:09 +0000 (01:25 -0700)]
hfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Roman Zippel <zippel@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofat: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:08 +0000 (01:25 -0700)]
fat: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: new cont helpers
Nick Piggin [Tue, 16 Oct 2007 08:25:07 +0000 (01:25 -0700)]
fs: new cont helpers

Rework the generic block "cont" routines to handle the new aops.  Supporting
cont_prepare_write would take quite a lot of code to support, so remove it
instead (and we later convert all filesystems to use it).

write_begin gets passed AOP_FLAG_CONT_EXPAND when called from
generic_cont_expand, so filesystems can avoid the old hacks they used.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agogfs2: convert to new aops
Steven Whitehouse [Tue, 16 Oct 2007 08:25:07 +0000 (01:25 -0700)]
gfs2: convert to new aops

Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoxfs: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:06 +0000 (01:25 -0700)]
xfs: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: David Chinner <dgc@sgi.com>
Cc: Timothy Shimmin <tes@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoext4: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:05 +0000 (01:25 -0700)]
ext4: convert to new aops

Convert ext4 to use write_begin()/write_end() methods.

Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Dmitriy Monakhov <dmonakhov@sw.ru>
Cc: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoext3: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:05 +0000 (01:25 -0700)]
ext3: convert to new aops

Various fixes and improvements

Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoext2: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:04 +0000 (01:25 -0700)]
ext2: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoblock_dev: convert to new aops
Nick Piggin [Tue, 16 Oct 2007 08:25:04 +0000 (01:25 -0700)]
block_dev: convert to new aops

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoimplement simple fs aops
Nick Piggin [Tue, 16 Oct 2007 08:25:03 +0000 (01:25 -0700)]
implement simple fs aops

Implement new aops for some of the simpler filesystems.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: restore KERNEL_DS optimisations
Nick Piggin [Tue, 16 Oct 2007 08:25:03 +0000 (01:25 -0700)]
mm: restore KERNEL_DS optimisations

Restore the KERNEL_DS optimisation, especially helpful to the 2copy write
path.

This may be a pretty questionable gain in most cases, especially after the
legacy 2copy write path is removed, but it doesn't cost much.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agodeny partial write for loop dev fd
Dmitry Monakhov [Tue, 16 Oct 2007 08:25:02 +0000 (01:25 -0700)]
deny partial write for loop dev fd

Partial write can be easily supported by LO_CRYPT_NONE mode, but it is not
easy in LO_CRYPT_CRYPTOAPI case, because of its block nature.  I don't know
who still used cryptoapi, but theoretically it is possible.  So let's leave
things as they are.  Loop device doesn't support partial write before
Nick's "write_begin/write_end" patch set, and let's it behave the same way
after.

Signed-off-by: Dmitriy Monakhov <dmonakhov@openvz.org>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: introduce write_begin, write_end, and perform_write aops
Nick Piggin [Tue, 16 Oct 2007 08:25:01 +0000 (01:25 -0700)]
fs: introduce write_begin, write_end, and perform_write aops

These are intended to replace prepare_write and commit_write with more
flexible alternatives that are also able to avoid the buffered write
deadlock problems efficiently (which prepare_write is unable to do).

[mark.fasheh@oracle.com: API design contributions, code review and fixes]
[akpm@linux-foundation.org: various fixes]
[dmonakhov@sw.ru: new aop block_write_begin fix]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Dmitriy Monakhov <dmonakhov@openvz.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: fix data-loss on error
Nick Piggin [Tue, 16 Oct 2007 08:25:00 +0000 (01:25 -0700)]
fs: fix data-loss on error

New buffers against uptodate pages are simply be marked uptodate, while the
buffer_new bit remains set.  This causes error-case code to zero out parts of
those buffers because it thinks they contain stale data: wrong, they are
actually uptodate so this is a data loss situation.

Fix this by actually clearning buffer_new and marking the buffer dirty.  It
makes sense to always clear buffer_new before setting a buffer uptodate.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: buffered write iterator
Nick Piggin [Tue, 16 Oct 2007 08:24:59 +0000 (01:24 -0700)]
mm: buffered write iterator

Add an iterator data structure to operate over an iovec.  Add usercopy
operators needed by generic_file_buffered_write, and convert that function
over.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: fix pagecache write deadlocks
Nick Piggin [Tue, 16 Oct 2007 08:24:59 +0000 (01:24 -0700)]
mm: fix pagecache write deadlocks

Modify the core write() code so that it won't take a pagefault while holding a
lock on the pagecache page. There are a number of different deadlocks possible
if we try to do such a thing:

1.  generic_buffered_write
2.   lock_page
3.    prepare_write
4.     unlock_page+vmtruncate
5.     copy_from_user
6.      mmap_sem(r)
7.       handle_mm_fault
8.        lock_page (filemap_nopage)
9.    commit_write
10.  unlock_page

a. sys_munmap / sys_mlock / others
b.  mmap_sem(w)
c.   make_pages_present
d.    get_user_pages
e.     handle_mm_fault
f.      lock_page (filemap_nopage)

2,8 - recursive deadlock if page is same
2,8;2,8 - ABBA deadlock is page is different
2,6;b,f - ABBA deadlock if page is same

The solution is as follows:
1.  If we find the destination page is uptodate, continue as normal, but use
    atomic usercopies which do not take pagefaults and do not zero the uncopied
    tail of the destination. The destination is already uptodate, so we can
    commit_write the full length even if there was a partial copy: it does not
    matter that the tail was not modified, because if it is dirtied and written
    back to disk it will not cause any problems (uptodate *means* that the
    destination page is as new or newer than the copy on disk).

1a. The above requires that fault_in_pages_readable correctly returns access
    information, because atomic usercopies cannot distinguish between
    non-present pages in a readable mapping, from lack of a readable mapping.

2.  If we find the destination page is non uptodate, unlock it (this could be
    made slightly more optimal), then allocate a temporary page to copy the
    source data into. Relock the destination page and continue with the copy.
    However, instead of a usercopy (which might take a fault), copy the data
    from the pinned temporary page via the kernel address space.

(also, rename maxlen to seglen, because it was confusing)

This increases the CPU/memory copy cost by almost 50% on the affected
workloads. That will be solved by introducing a new set of pagecache write
aops in a subsequent patch.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: write iovec cleanup
Nick Piggin [Tue, 16 Oct 2007 08:24:58 +0000 (01:24 -0700)]
mm: write iovec cleanup

Hide some of the open-coded nr_segs tests into the iovec helpers.  This is all
to simplify generic_file_buffered_write, because that gets more complex in the
next patch.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: buffered write cleanup
Nick Piggin [Tue, 16 Oct 2007 08:24:57 +0000 (01:24 -0700)]
mm: buffered write cleanup

Quite a bit of code is used in maintaining these "cached pages" that are
probably pretty unlikely to get used. It would require a narrow race where
the page is inserted concurrently while this process is allocating a page
in order to create the spare page. Then a multi-page write into an uncached
part of the file, to make use of it.

Next, the buffered write path (and others) uses its own LRU pagevec when it
should be just using the per-CPU LRU pagevec (which will cut down on both data
and code size cacheline footprint). Also, these private LRU pagevecs are
emptied after just a very short time, in contrast with the per-CPU pagevecs
that are persistent. Net result: 7.3 times fewer lru_lock acquisitions required
to add the pages to pagecache for a bulk write (in 4K chunks).

[this gets rid of some cond_resched() calls in readahead.c and mpage.c due
 to clashes in -mm. What put them there, and why? ]

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: trim more holes
Nick Piggin [Tue, 16 Oct 2007 08:24:56 +0000 (01:24 -0700)]
mm: trim more holes

If prepare_write fails with AOP_TRUNCATED_PAGE, or if commit_write fails, then
we may have failed the write operation despite prepare_write having
instantiated blocks past i_size.  Fix this, and consolidate the trimming into
one place.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: debug write deadlocks
Nick Piggin [Tue, 16 Oct 2007 08:24:56 +0000 (01:24 -0700)]
mm: debug write deadlocks

Allow CONFIG_DEBUG_VM to switch off the prefaulting logic, to simulate the
Makes the race much easier to hit.

This is useful for demonstration and testing purposes, but is removed in a
subsequent patch.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: clean up buffered write code
Andrew Morton [Tue, 16 Oct 2007 08:24:55 +0000 (01:24 -0700)]
mm: clean up buffered write code

Rename some variables and fix some types.

Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoRevert "[PATCH] generic_file_buffered_write(): deadlock on vectored write"
Andrew Morton [Tue, 16 Oct 2007 08:24:54 +0000 (01:24 -0700)]
Revert "[PATCH] generic_file_buffered_write(): deadlock on vectored write"

This reverts commit 6527c2bdf1f833cc18e8f42bd97973d583e4aa83, which
fixed the following bug:

  When prefaulting in the pages in generic_file_buffered_write(), we only
  faulted in the pages for the firts segment of the iovec.  If the second of
  successive segment described a mmapping of the page into which we're
  write()ing, and that page is not up-to-date, the fault handler tries to lock
  the already-locked page (to bring it up to date) and deadlocks.

  An exploit for this bug is in writev-deadlock-demo.c, in
  http://www.zip.com.au/~akpm/linux/patches/stuff/ext3-tools.tar.gz.

  (These demos assume blocksize < PAGE_CACHE_SIZE).

The problem with this fix is that it takes the kernel back to doing a single
prepare_write()/commit_write() per iovec segment.  So in the worst case we'll
run prepare_write+commit_write 1024 times where we previously would have run
it once. The other problem with the fix is that it fix all the locking problems.

<insert numbers obtained via ext3-tools's writev-speed.c here>

And apparently this change killed NFS overwrite performance, because, I
suppose, it talks to the server for each prepare_write+commit_write.

So just back that patch out - we'll be fixing the deadlock by other means.

Nick says: also it only ever actually papered over the bug, because after
faulting in the pages, they might be unmapped or reclaimed.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoRevert "[PATCH] generic_file_buffered_write(): handle zero-length iovec segments"
Andrew Morton [Tue, 16 Oct 2007 08:24:54 +0000 (01:24 -0700)]
Revert "[PATCH] generic_file_buffered_write(): handle zero-length iovec segments"

This reverts commit 81b0c8713385ce1b1b9058e916edcf9561ad76d6, which was
a bugfix against 6527c2bdf1f833cc18e8f42bd97973d583e4aa83 ("[PATCH]
generic_file_buffered_write(): deadlock on vectored write"), which we
also revert.

Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: revert KERNEL_DS buffered write optimisation
Nick Piggin [Tue, 16 Oct 2007 08:24:53 +0000 (01:24 -0700)]
mm: revert KERNEL_DS buffered write optimisation

Revert the patch from Neil Brown to optimise NFSD writev handling.

Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: use pagevec to rotate reclaimable page
Hisashi Hifumi [Tue, 16 Oct 2007 08:24:52 +0000 (01:24 -0700)]
mm: use pagevec to rotate reclaimable page

While running some memory intensive load, system response deteriorated just
after swap-out started.

The cause of this problem is that when a PG_reclaim page is moved to the tail
of the inactive LRU list in rotate_reclaimable_page(), lru_lock spin lock is
acquired every page writeback .  This deteriorates system performance and
makes interrupt hold off time longer when swap-out started.

Following patch solves this problem.  I use pagevec in rotating reclaimable
pages to mitigate LRU spin lock contention and reduce interrupt hold off time.

I did a test that allocating and touching pages in multiple processes, and
pinging to the test machine in flooding mode to measure response under memory
intensive load.

The test result is:

-2.6.23-rc5
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 53222ms
rtt min/avg/max/mdev = 0.074/0.652/172.228/7.176 ms, pipe 11, ipg/ewma
17.746/0.092 ms

-2.6.23-rc5-patched
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 51924ms
rtt min/avg/max/mdev = 0.072/0.108/3.884/0.114 ms, pipe 2, ipg/ewma
17.314/0.091 ms

Max round-trip-time was improved.

The test machine spec is that 4CPU(3.16GHz, Hyper-threading enabled)
8GB memory , 8GB swap.

I did ping test again to observe performance deterioration caused by taking
a ref.

-2.6.23-rc6-with-modifiedpatch
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 53386ms
rtt min/avg/max/mdev = 0.074/0.110/4.716/0.147 ms, pipe 2, ipg/ewma 17.801/0.129 ms

The result for my original patch is as follows.

-2.6.23-rc5-with-originalpatch
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 51924ms
rtt min/avg/max/mdev = 0.072/0.108/3.884/0.114 ms, pipe 2, ipg/ewma 17.314/0.091 ms

The influence to response was small.

[akpm@linux-foundation.org: fix uninitalised var warning]
[hugh@veritas.com: fix locking]
[randy.dunlap@oracle.com: fix function declaration]
[hugh@veritas.com: fix BUG at include/linux/mm.h:220!]
[hugh@veritas.com: kill redundancy in rotate_reclaimable_page]
[hugh@veritas.com: move_tail_pages into lru_add_drain]
Signed-off-by: Hisashi Hifumi <hifumi.hisashi@oss.ntt.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoMem Policy: add MPOL_F_MEMS_ALLOWED get_mempolicy() flag
Lee Schermerhorn [Tue, 16 Oct 2007 08:24:51 +0000 (01:24 -0700)]
Mem Policy: add MPOL_F_MEMS_ALLOWED get_mempolicy() flag

Allow an application to query the memories allowed by its context.

Updated numa_memory_policy.txt to mention that applications can use this to
obtain allowed memories for constructing valid policies.

TODO:  update out-of-tree libnuma wrapper[s], or maybe add a new
wrapper--e.g.,  numa_get_mems_allowed() ?

Also, update numa syscall man pages.

Tested with memtoy V>=0.13.

Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Christoph Lameter <clameter@sgi.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: prevent kswapd from freeing excessive amounts of lowmem
Rik van Riel [Tue, 16 Oct 2007 08:24:50 +0000 (01:24 -0700)]
mm: prevent kswapd from freeing excessive amounts of lowmem

The current VM can get itself into trouble fairly easily on systems with a
small ZONE_HIGHMEM, which is common on i686 computers with 1GB of memory.

On one side, page_alloc() will allocate down to zone->pages_low, while on
the other side, kswapd() and balance_pgdat() will try to free memory from
every zone, until every zone has more free pages than zone->pages_high.

Highmem can be filled up to zone->pages_low with page tables, ramfs,
vmalloc allocations and other unswappable things quite easily and without
many bad side effects, since we still have a huge ZONE_NORMAL to do future
allocations from.

However, as long as the number of free pages in the highmem zone is below
zone->pages_high, kswapd will continue swapping things out from
ZONE_NORMAL, too!

Sami Farin managed to get his system into a stage where kswapd had freed
about 700MB of low memory and was still "going strong".

The attached patch will make kswapd stop paging out data from zones when
there is more than enough memory free.  We do go above zone->pages_high in
order to keep pressure between zones equal in normal circumstances, but the
patch should prevent the kind of excesses that made Sami's computer totally
unusable.

Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: no need to cast vmalloc() return value in zone_wait_table_init()
Jesper Juhl [Tue, 16 Oct 2007 08:24:49 +0000 (01:24 -0700)]
mm: no need to cast vmalloc() return value in zone_wait_table_init()

vmalloc() returns a void pointer, so there's no need to cast its
return value in mm/page_alloc.c::zone_wait_table_init().

Signed-off-by: Jesper Juhl <jesper.juhl@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofix the max path calculation in radix-tree.c
Jeff Moyer [Tue, 16 Oct 2007 08:24:49 +0000 (01:24 -0700)]
fix the max path calculation in radix-tree.c

A while back, Nick Piggin introduced a patch to reduce the node memory
usage for small files (commit cfd9b7df4abd3257c9e381b0e445817b26a51c0c):

-#define RADIX_TREE_MAP_SHIFT 6
+#define RADIX_TREE_MAP_SHIFT (CONFIG_BASE_SMALL ? 4 : 6)

Unfortunately, he didn't take into account the fact that the
calculation of the maximum path was based on an assumption of having
to round up:

#define RADIX_TREE_MAX_PATH (RADIX_TREE_INDEX_BITS/RADIX_TREE_MAP_SHIFT + 2)

So, if CONFIG_BASE_SMALL is set, you will end up with a
RADIX_TREE_MAX_PATH that is one greater than necessary.  The practical
upshot of this is just a bit of wasted memory (one long in the
height_to_maxindex array, an extra pre-allocated radix tree node per
cpu, and extra stack usage in a couple of functions), but it seems
worth getting right.

It's also worth noting that I never build with CONFIG_BASE_SMALL.
What I did to test this was duplicate the code in a small user-space
program and check the results of the calculations for max path and the
contents of the height_to_maxindex array.

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofs: fix nobh error handling
Nick Piggin [Tue, 16 Oct 2007 08:24:48 +0000 (01:24 -0700)]
fs: fix nobh error handling

nobh mode error handling is not just pretty slack, it's wrong.

One cannot zero out the whole page to ensure new blocks are zeroed, because
it just brings the whole page "uptodate" with zeroes even if that may not
be the correct uptodate data.  Also, other parts of the page may already
contain dirty data which would get lost by zeroing it out.  Thirdly, the
writeback of zeroes to the new blocks will also erase existing blocks.  All
these conditions are pagecache and/or filesystem corruption.

The problem comes about because we didn't keep track of which buffers
actually are new or old.  However it is not enough just to keep only this
state, because at the point we start dirtying parts of the page (new
blocks, with zeroes), the handling of IO errors becomes impossible without
buffers because the page may only be partially uptodate, in which case the
page flags allone cannot capture the state of the parts of the page.

So allocate all buffers for the page upfront, but leave them unattached so
that they don't pick up any other references and can be freed when we're
done.  If the error path is hit, then zero the new buffers as the regular
buffer path does, then attach the buffers to the page so that it can
actually be written out correctly and be subject to the normal IO error
handling paths.

As an upshot, we save 1K of kernel stack on ia64 or powerpc 64K page
systems.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: add end_buffer_read helper function
Dmitry Monakhov [Tue, 16 Oct 2007 08:24:47 +0000 (01:24 -0700)]
mm: add end_buffer_read helper function

Move duplicated code from end_buffer_read_XXX methods to separate helper
function.

Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoSlab allocators: fail if ksize is called with a NULL parameter
Christoph Lameter [Tue, 16 Oct 2007 08:24:46 +0000 (01:24 -0700)]
Slab allocators: fail if ksize is called with a NULL parameter

A NULL pointer means that the object was not allocated.  One cannot
determine the size of an object that has not been allocated.  Currently we
return 0 but we really should BUG() on attempts to determine the size of
something nonexistent.

krealloc() interprets NULL to mean a zero sized object.  Handle that
separately in krealloc().

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Matt Mackall <mpm@selenic.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agocalculation of pgoff in do_linear_fault() uses mixed units
Dean Nelson [Tue, 16 Oct 2007 08:24:45 +0000 (01:24 -0700)]
calculation of pgoff in do_linear_fault() uses mixed units

The calculation of pgoff in do_linear_fault() should use PAGE_SHIFT and not
PAGE_CACHE_SHIFT since vma->vm_pgoff is in units of PAGE_SIZE and not
PAGE_CACHE_SIZE.  At the moment linux/pagemap.h has PAGE_CACHE_SHIFT
defined as PAGE_SHIFT, but should that ever change this calculation would
break.

Signed-off-by: Dean Nelson <dcn@sgi.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years ago{slub, slob}: use unlikely() for kfree(ZERO_OR_NULL_PTR) check
Satyam Sharma [Tue, 16 Oct 2007 08:24:44 +0000 (01:24 -0700)]
{slub, slob}: use unlikely() for kfree(ZERO_OR_NULL_PTR) check

Considering kfree(NULL) would normally occur only in error paths and
kfree(ZERO_SIZE_PTR) is uncommon as well, so let's use unlikely() for the
condition check in SLUB's and SLOB's kfree() to optimize for the common
case.  SLAB has this already.

Signed-off-by: Satyam Sharma <satyam@infradead.org>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomove mm_struct and vm_area_struct
Martin Schwidefsky [Tue, 16 Oct 2007 08:24:43 +0000 (01:24 -0700)]
move mm_struct and vm_area_struct

Move the definitions of struct mm_struct and struct vma_area_struct to
include/mm_types.h.  This allows to define more function in asm/pgtable.h
and friends with inline assemblies instead of macros.  Compile tested on
i386, powerpc, powerpc64, s390-32, s390-64 and x86_64.

[aurelien@aurel32.net: build fix]
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoradix-tree: use indirect bit
Nick Piggin [Tue, 16 Oct 2007 08:24:42 +0000 (01:24 -0700)]
radix-tree: use indirect bit

Rather than sign direct radix-tree pointers with a special bit, sign the
indirect one that hangs off the root.  This means that, given a lookup_slot
operation, the invalid result will be differentiated from the valid
(previously, valid results could have the bit either set or clear).

This does not affect slot lookups which occur under lock -- they can never
return an invalid result.  Is needed in future for lockless pagecache.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: clarify __add_to_swap_cache locking
Nick Piggin [Tue, 16 Oct 2007 08:24:42 +0000 (01:24 -0700)]
mm: clarify __add_to_swap_cache locking

__add_to_swap_cache unconditionally sets the page locked, which can be a bit
alarming to the unsuspecting reader: in the code paths where the page is
visible to other CPUs, the page should be (and is) already locked.

Instead, just add a check to ensure the page is locked here, and teach the one
path relying on the old behaviour to call SetPageLocked itself.

[hugh@veritas.com: locking fix]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: improve find_lock_page
Nick Piggin [Tue, 16 Oct 2007 08:24:41 +0000 (01:24 -0700)]
mm: improve find_lock_page

find_lock_page does not need to recheck ->index because if the page is in the
right mapping then the index must be the same.  Also, tree_lock does not need
to be retaken after the page is locked in order to test that ->mapping has not
changed, because holding the page lock pins its mapping.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agomm: use lockless radix-tree probe
Nick Piggin [Tue, 16 Oct 2007 08:24:40 +0000 (01:24 -0700)]
mm: use lockless radix-tree probe

Probing pages and radix_tree_tagged are lockless operations with the lockless
radix-tree.  Convert these users to RCU locking rather than using tree_lock.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoremove ZERO_PAGE
Nick Piggin [Tue, 16 Oct 2007 08:24:40 +0000 (01:24 -0700)]
remove ZERO_PAGE

The commit b5810039a54e5babf428e9a1e89fc1940fabff11 contains the note

  A last caveat: the ZERO_PAGE is now refcounted and managed with rmap
  (and thus mapcounted and count towards shared rss).  These writes to
  the struct page could cause excessive cacheline bouncing on big
  systems.  There are a number of ways this could be addressed if it is
  an issue.

And indeed this cacheline bouncing has shown up on large SGI systems.
There was a situation where an Altix system was essentially livelocked
tearing down ZERO_PAGE pagetables when an HPC app aborted during startup.
This situation can be avoided in userspace, but it does highlight the
potential scalability problem with refcounting ZERO_PAGE, and corner
cases where it can really hurt (we don't want the system to livelock!).

There are several broad ways to fix this problem:
1. add back some special casing to avoid refcounting ZERO_PAGE
2. per-node or per-cpu ZERO_PAGES
3. remove the ZERO_PAGE completely

I will argue for 3. The others should also fix the problem, but they
result in more complex code than does 3, with little or no real benefit
that I can see.

Why? Inserting a ZERO_PAGE for anonymous read faults appears to be a
false optimisation: if an application is performance critical, it would
not be doing many read faults of new memory, or at least it could be
expected to write to that memory soon afterwards. If cache or memory use
is critical, it should not be working with a significant number of
ZERO_PAGEs anyway (a more compact representation of zeroes should be
used).

As a sanity check -- mesuring on my desktop system, there are never many
mappings to the ZERO_PAGE (eg. 2 or 3), thus memory usage here should not
increase much without it.

When running a make -j4 kernel compile on my dual core system, there are
about 1,000 mappings to the ZERO_PAGE created per second, but about 1,000
ZERO_PAGE COW faults per second (less than 1 ZERO_PAGE mapping per second
is torn down without being COWed). So removing ZERO_PAGE will save 1,000
page faults per second when running kbuild, while keeping it only saves
less than 1 page clearing operation per second. 1 page clear is cheaper
than a thousand faults, presumably, so there isn't an obvious loss.

Neither the logical argument nor these basic tests give a guarantee of no
regressions. However, this is a reasonable opportunity to try to remove
the ZERO_PAGE from the pagefault path. If it is found to cause regressions,
we can reintroduce it and just avoid refcounting it.

The /dev/zero ZERO_PAGE usage and TLB tricks also get nuked.  I don't see
much use to them except on benchmarks.  All other users of ZERO_PAGE are
converted just to use ZERO_PAGE(0) for simplicity. We can look at
replacing them all and maybe ripping out ZERO_PAGE completely when we are
more satisfied with this solution.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus "snif" Torvalds <torvalds@linux-foundation.org>
13 years agoSLUB: direct pass through of page size or higher kmalloc requests
Christoph Lameter [Tue, 16 Oct 2007 08:24:38 +0000 (01:24 -0700)]
SLUB: direct pass through of page size or higher kmalloc requests

This gets rid of all kmalloc caches larger than page size.  A kmalloc
request larger than PAGE_SIZE > 2 is going to be passed through to the page
allocator.  This works both inline where we will call __get_free_pages
instead of kmem_cache_alloc and in __kmalloc.

kfree is modified to check if the object is in a slab page. If not then
the page is freed via the page allocator instead. Roughly similar to what
SLOB does.

Advantages:
- Reduces memory overhead for kmalloc array
- Large kmalloc operations are faster since they do not
  need to pass through the slab allocator to get to the
  page allocator.
- Performance increase of 10%-20% on alloc and 50% on free for
  PAGE_SIZEd allocations.
  SLUB must call page allocator for each alloc anyways since
  the higher order pages which that allowed avoiding the page alloc calls
  are not available in a reliable way anymore. So we are basically removing
  useless slab allocator overhead.
- Large kmallocs yields page aligned object which is what
  SLAB did. Bad things like using page sized kmalloc allocations to
  stand in for page allocate allocs can be transparently handled and are not
  distinguishable from page allocator uses.
- Checking for too large objects can be removed since
  it is done by the page allocator.

Drawbacks:
- No accounting for large kmalloc slab allocations anymore
- No debugging of large kmalloc slab allocations.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofilemap: convert some unsigned long to pgoff_t
Fengguang Wu [Tue, 16 Oct 2007 08:24:37 +0000 (01:24 -0700)]
filemap: convert some unsigned long to pgoff_t

Convert some 'unsigned long' to pgoff_t.

Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agofilemap: trivial code cleanups
Fengguang Wu [Tue, 16 Oct 2007 08:24:37 +0000 (01:24 -0700)]
filemap: trivial code cleanups

- remove unused local next_index in do_generic_mapping_read()
- remove a redudant page_cache_read() declaration

Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoreadahead: remove the limit max_sectors_kb imposed on max_readahead_kb
Fengguang Wu [Tue, 16 Oct 2007 08:24:36 +0000 (01:24 -0700)]
readahead: remove the limit max_sectors_kb imposed on max_readahead_kb

Remove the size limit max_sectors_kb imposed on max_readahead_kb.

The size restriction is unreasonable.  Especially when max_sectors_kb cannot
grow larger than max_hw_sectors_kb, which can be rather small for some disk
drives.

Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Acked-by: Jens Axboe <jens.axboe@oracle.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoreadahead: remove several readahead macros
Fengguang Wu [Tue, 16 Oct 2007 08:24:36 +0000 (01:24 -0700)]
readahead: remove several readahead macros

Remove VM_MAX_CACHE_HIT, MAX_RA_PAGES and MIN_RA_PAGES.

Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
13 years agoreadahead: remove the local copy of ra in do_generic_mapping_read()
Fengguang Wu [Tue, 16 Oct 2007 08:24:35 +0000 (01:24 -0700)]
readahead: remove the local copy of ra in do_generic_mapping_read()

The local copy of ra in do_generic_mapping_read() can now go away.

It predates readanead(req_size).  In a time when the readahead code was called
on *every* single page.  Hence a local has to be made to reduce the chance of
the readahead state being overwritten by a concurrent reader.  More details
in: Linux: Random File I/O Regressions In 2.6
<http://kerneltrap.org/node/3039>

Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>