The NUMA API change that introduced kmalloc_node was accepted for
2.6.12-rc3. Now it is possible to do slab allocations on a node to
localize memory structures. This API was used by the pageset localization
patch and the block layer localization patch now in mm. The existing
kmalloc_node is slow since it simply searches through all pages of the slab
to find a page that is on the node requested. The two patches do a one
time allocation of slab structures at initialization and therefore the
speed of kmalloc node does not matter.
This patch allows kmalloc_node to be as fast as kmalloc by introducing node
specific page lists for partial, free and full slabs. Slab allocation
improves in a NUMA system so that we are seeing a performance gain in AIM7
of about 5% with this patch alone.
More NUMA localizations are possible if kmalloc_node operates in an fast
way like kmalloc.
Test run on a 32p systems with 32G Ram.
w/o patch
Tasks jobs/min jti jobs/min/task real cpu
1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005
100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005
200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005
300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005
400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005
500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005
600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005
700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005
800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005
900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005
1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005
w/patch
Tasks jobs/min jti jobs/min/task real cpu
1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005
100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005
200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005
300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005
400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005
500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005
600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005
700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005
800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005
900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005
1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005
These are measurement taken directly after boot and show a greater
improvement than 5%. However, the performance improvements become less
over time if the AIM7 runs are repeated and settle down at around 5%.
Links to earlier discussions:
http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2
Changelog V4-V5:
- alloc_arraycache and alloc_aliencache take node parameter instead of cpu
- fix initialization so that nodes without cpus are properly handled.
- simplify code in kmem_cache_init
- patch against Andrews temp mm3 release
- Add Shai to credits
- fallback to __cache_alloc from __cache_alloc_node if the node's cache
is not available yet.
Changelog V3-V4:
- Patch against 2.6.12-rc5-mm1
- Cleanup patch integrated
- More and better use of for_each_node and for_each_cpu
- GCC 2.95 fix (do not use [] use [0])
- Correct determination of INDEX_AC
- Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes.
- Remove list3_data and list3_data_ptr macros for better readability
Changelog V2-V3:
- Made to patch against 2.6.12-rc4-mm1
- Revised bootstrap mechanism so that larger size kmem_list3 structs can be
supported. Do a generic solution so that the right slab can be found
for the internal structs.
- use for_each_online_node
Changelog V1-V2:
- Batching for freeing of wrong-node objects (alien caches)
- Locking changes and NUMA #ifdefs as requested by Manfred
Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com>
Signed-off-by: Shai Fultheim <Shai@Scalex86.org>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Manfred Spraul <manfred@colorfullife.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch modifies tmpfs to call the inode_init_security LSM hook to set
up the incore inode security state for new inodes before the inode becomes
accessible via the dcache.
As there is no underlying storage of security xattrs in this case, it is
not necessary for the hook to return the (name, value, len) triple to the
tmpfs code, so this patch also modifies the SELinux hook function to
correctly handle the case where the (name, value, len) pointers are NULL.
The hook call is needed in tmpfs in order to support proper security
labeling of tmpfs inodes (e.g. for udev with tmpfs /dev in Fedora). With
this change in place, we should then be able to remove the
security_inode_post_create/mkdir/... hooks safely.
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
Cc: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Update the file systems in fs/ implementing a delete_inode() callback to
call truncate_inode_pages(). One implementation note: In developing this
patch I put the calls to truncate_inode_pages() at the very top of those
filesystems delete_inode() callbacks in order to retain the previous
behavior. I'm guessing that some of those could probably be optimized.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Acked-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Run PCI driver initialization on local node
Instead of adding messy kmalloc_node()s everywhere run the
PCI driver probe on the node local to the device.
This would not have helped for IDE, but should for
other more clean drivers that do more initialization in probe().
It won't help for drivers that do most of the work
on first open (like many network drivers)
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
This patch introduces a kzalloc wrapper and converts kernel/ to use it. It
saves a little program text.
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Now the real motivation for this cpuset mem_exclusive patch series seems
trivial.
This patch keeps a task in or under one mem_exclusive cpuset from provoking an
oom kill of a task under a non-overlapping mem_exclusive cpuset. Since only
interrupt and GFP_ATOMIC allocations are allowed to escape mem_exclusive
containment, there is little to gain from oom killing a task under a
non-overlapping mem_exclusive cpuset, as almost all kernel and user memory
allocation must come from disjoint memory nodes.
This patch enables configuring a system so that a runaway job under one
mem_exclusive cpuset cannot cause the killing of a job in another such cpuset
that might be using very high compute and memory resources for a prolonged
time.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch makes use of the previously underutilized cpuset flag
'mem_exclusive' to provide what amounts to another layer of memory placement
resolution. With this patch, there are now the following four layers of
memory placement available:
1) The whole system (interrupt and GFP_ATOMIC allocations can use this),
2) The nearest enclosing mem_exclusive cpuset (GFP_KERNEL allocations can use),
3) The current tasks cpuset (GFP_USER allocations constrained to here), and
4) Specific node placement, using mbind and set_mempolicy.
These nest - each layer is a subset (same or within) of the previous.
Layer (2) above is new, with this patch. The call used to check whether a
zone (its node, actually) is in a cpuset (in its mems_allowed, actually) is
extended to take a gfp_mask argument, and its logic is extended, in the case
that __GFP_HARDWALL is not set in the flag bits, to look up the cpuset
hierarchy for the nearest enclosing mem_exclusive cpuset, to determine if
placement is allowed. The definition of GFP_USER, which used to be identical
to GFP_KERNEL, is changed to also set the __GFP_HARDWALL bit, in the previous
cpuset_gfp_hardwall_flag patch.
GFP_ATOMIC and GFP_KERNEL allocations will stay within the current tasks
cpuset, so long as any node therein is not too tight on memory, but will
escape to the larger layer, if need be.
The intended use is to allow something like a batch manager to handle several
jobs, each job in its own cpuset, but using common kernel memory for caches
and such. Swapper and oom_kill activity is also constrained to Layer (2). A
task in or below one mem_exclusive cpuset should not cause swapping on nodes
in another non-overlapping mem_exclusive cpuset, nor provoke oom_killing of a
task in another such cpuset. Heavy use of kernel memory for i/o caching and
such by one job should not impact the memory available to jobs in other
non-overlapping mem_exclusive cpusets.
This patch enables providing hardwall, inescapable cpusets for memory
allocations of each job, while sharing kernel memory allocations between
several jobs, in an enclosing mem_exclusive cpuset.
Like Dinakar's patch earlier to enable administering sched domains using the
cpu_exclusive flag, this patch also provides a useful meaning to a cpuset flag
that had previously done nothing much useful other than restrict what cpuset
configurations were allowed.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch series extends the use of the cpuset attribute 'mem_exclusive'
to support cpuset configurations that:
1) allow GFP_KERNEL allocations to come from a potentially larger
set of memory nodes than GFP_USER allocations, and
2) can constrain the oom killer to tasks running in cpusets in
a specified subtree of the cpuset hierarchy.
Here's an example usage scenario. For a few hours or more, a large NUMA
system at a University is to be divided in two halves, with a bunch of student
jobs running in half the system under some form of batch manager, and with a
big research project running in the other half. Each of the student jobs is
placed in a small cpuset, but should share the classic Unix time share
facilities, such as buffered pages of files in /bin and /usr/lib. The big
research project wants no interference whatsoever from the student jobs, and
has highly tuned, unusual memory and i/o patterns that intend to make full use
of all the main memory on the nodes available to it.
In this example, we have two big sibling cpusets, one of which is further
divided into a more dynamic set of child cpusets.
We want kernel memory allocations constrained by the two big cpusets, and user
allocations constrained by the smaller child cpusets where present. And we
require that the oom killer not operate across the two halves of this system,
or else the first time a student job runs amuck, the big research project will
likely be first inline to get shot.
Tweaking /proc/<pid>/oom_adj is not ideal -- if the big research project
really does run amuck allocating memory, it should be shot, not some other
task outside the research projects mem_exclusive cpuset.
I propose to extend the use of the 'mem_exclusive' flag of cpusets to manage
such scenarios. Let memory allocations for user space (GFP_USER) be
constrained by a tasks current cpuset, but memory allocations for kernel space
(GFP_KERNEL) by constrained by the nearest mem_exclusive ancestor of the
current cpuset, even though kernel space allocations will still _prefer_ to
remain within the current tasks cpuset, if memory is easily available.
Let the oom killer be constrained to consider only tasks that are in
overlapping mem_exclusive cpusets (it won't help much to kill a task that
normally cannot allocate memory on any of the same nodes as the ones on which
the current task can allocate.)
The current constraints imposed on setting mem_exclusive are unchanged. A
cpuset may only be mem_exclusive if its parent is also mem_exclusive, and a
mem_exclusive cpuset may not overlap any of its siblings memory nodes.
This patch was presented on linux-mm in early July 2005, though did not
generate much feedback at that time. It has been built for a variety of
arch's using cross tools, and built, booted and tested for function on SN2
(ia64).
There are 4 patches in this set:
1) Some minor cleanup, and some improvements to the code layout
of one routine to make subsequent patches cleaner.
2) Add another GFP flag - __GFP_HARDWALL. It marks memory
requests for USER space, which are tightly confined by the
current tasks cpuset.
3) Now memory requests (such as KERNEL) that not marked HARDWALL can
if short on memory, look in the potentially larger pool of memory
defined by the nearest mem_exclusive ancestor cpuset of the current
tasks cpuset.
4) Finally, modify the oom killer to skip any task whose mem_exclusive
cpuset doesn't overlap ours.
Patch (1), the one time I looked on an SN2 (ia64) build, actually saved 32
bytes of kernel text space. Patch (2) has no affect on the size of kernel
text space (it just adds a preprocessor flag). Patches (3) and (4) added
about 600 bytes each of kernel text space, mostly in kernel/cpuset.c, which
matters only if CONFIG_CPUSET is enabled.
This patch:
This patch applies a few comment and code cleanups to mm/oom_kill.c prior to
applying a few small patches to improve cpuset management of memory placement.
The comment changed in oom_kill.c was seriously misleading. The code layout
change in select_bad_process() makes room for adding another condition on
which a process can be spared the oom killer (see the subsequent
cpuset_nodes_overlap patch for this addition).
Also a couple typos and spellos that bugged me, while I was here.
This patch should have no material affect.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Mark variables which are usually accessed for reads with __readmostly.
Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
We don't reset the cache hit count until after readahead does a successful
readahead. This seems to leave a corner case open where we miss in cache,
but don't restart the readhead right away.
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Move some more frequently read variables that showed up during some of our
performance tests as sometimes ending up in hot cachelines to the
read_mostly section.
Fix: Move the __read_mostly from before hpet_usec_quotient to follow the
variable like the other uses of __read_mostly.
Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Christoph Lameter <christoph@scalex86.org>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch modifies the VFS setxattr, getxattr, and listxattr code to fall
back to the security module for security xattrs if the filesystem does not
support xattrs natively. This allows security modules to export the incore
inode security label information to userspace even if the filesystem does
not provide xattr storage, and eliminates the need to individually patch
various pseudo filesystem types to provide such access. The patch removes
the existing xattr code from devpts and tmpfs as it is then no longer
needed.
The patch restructures the code flow slightly to reduce duplication between
the normal path and the fallback path, but this should only have one
user-visible side effect - a program may get -EACCES rather than
-EOPNOTSUPP if policy denied access but the filesystem didn't support the
operation anyway. Note that the post_setxattr hook call is not needed in
the fallback case, as the inode_setsecurity hook call handles the incore
inode security state update directly. In contrast, we do call fsnotify in
both cases.
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
Acked-by: James Morris <jmorris@namei.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add page_state info to the per-node meminfo file in sysfs. This is mostly
just for informational purposes.
The lack of this information was brought up recently during a discussion
regarding pagecache clearing, and I put this patch together to test out one
of the suggestions.
It seems like interesting info to have, so I'm submitting the patch.
Signed-off-by: Martin Hicks <mort@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Proposed by and based on a patch from Eric Dumazet <dada1@cosmosbay.com>:
This patch removes unnecessary critical section in ksize() function, as
cli/sti are rather expensive on modern CPUS.
It additionally adds a docbook entry for ksize() and further simplifies the
code.
Signed-Off-By: Manfred Spraul <manfred@colorfullife.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Mostobjects returned by __cache_alloc() will be written by the caller,
(but not all callers want to write all the object, but just at the
begining) prefetchw() tells the modern CPU to think about the future
writes, ie start some memory transactions in advance.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add a new accessor for PTEs, which passes the full hint from the mmu_gather
struct; this allows architectures with hardware pagetables to optimize away
atomic PTE operations when destroying an address space. Removing the
locked operation should allow better pipelining of memory access in this
loop. I measured an average savings of 30-35 cycles per zap_pte_range on
the first 500 destructions on Pentium-M, but I believe the optimization
would win more on older processors which still assert the bus lock on xchg
for an exclusive cacheline.
Update: I made some new measurements, and this saves exactly 26 cycles over
ptep_get_and_clear on Pentium M. On P4, with a PAE kernel, this saves 180
cycles per ptep_get_and_clear, for a whopping 92160 cycles savings for a
full address space destruction.
pte_clear_full is not yet used, but is provided for future optimizations
(in particular, when running inside of a hypervisor that queues page table
updates, the full hint allows us to avoid queueing unnecessary page table
update for an address space in the process of being destroyed.
This is not a huge win, but it does help a bit, and sets the stage for
further hypervisor optimization of the mm layer on all architectures.
Signed-off-by: Zachary Amsden <zach@vmware.com>
Cc: Christoph Lameter <christoph@lameter.com>
Cc: <linux-mm@kvack.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This is used only in slab.c and each architecture gets to define whcih
underlying type is to be used.
Seems a bit silly - move it to slab.c and use the same type for all
architectures: unsigned int.
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial Post (Wed, 17 Aug 2005)
This patch moves the
if (! pte_none(*pte))
hugetlb_clean_stale_pgtable(pte);
logic into huge_pte_alloc() so all of its callers can be immune to the bug
described by Kenneth Chen at http://lkml.org/lkml/2004/6/16/246
> It turns out there is a bug in hugetlb_prefault(): with 3 level page table,
> huge_pte_alloc() might return a pmd that points to a PTE page. It happens
> if the virtual address for hugetlb mmap is recycled from previously used
> normal page mmap. free_pgtables() might not scrub the pmd entry on
> munmap and hugetlb_prefault skips on any pmd presence regardless what type
> it is.
Unless I am missing something, it seems more correct to place the check inside
huge_pte_alloc() to prevent a the same bug wherever a huge pte is allocated.
It also allows checking for this condition when lazily faulting huge pages
later in the series.
Signed-off-by: Adam Litke <agl@us.ibm.com>
Cc: <linux-mm@kvack.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Version 6 of the ARM architecture introduces the concept of 16MB pages
(supersections) and 36-bit (40-bit actually, but nobody uses this) physical
addresses. 36-bit addressed memory and I/O and ARMv6 can only be mapped
using supersections and the requirement on these is that both virtual and
physical addresses be 16MB aligned. In trying to add support for ioremap()
of 36-bit I/O, we run into the issue that get_vm_area() allows for a
maximum of 512K alignment via the IOREMAP_MAX_ORDER constant. To work
around this, we can:
- Allocate a larger VM area than needed (size + (1ul << IOREMAP_MAX_ORDER))
and then align the pointer ourselves, but this ends up with 512K of
wasted VM per ioremap().
- Provide a new __get_vm_area_aligned() API and make __get_vm_area() sit
on top of this. I did this and it works but I don't like the idea
adding another VM API just for this one case.
- My preferred solution which is to allow the architecture to override
the IOREMAP_MAX_ORDER constant with it's own version.
Signed-off-by: Deepak Saxena <dsaxena@plexity.net>
Cc: Russell King <rmk@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
If !vma->vm-ops we already BUG above, so retesting it is useless. The
compiler cannot optimize this because BUG is a macro and is not thus marked
noreturn; that should possibly be fixed.
Signed-off-by: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Either shmem_getpage returns a failure, or it found a page, or it was told
it couldn't do any I/O. So it's useless to check nonblock in the else
branch. We could add a BUG() there but I preferred to comment the
offending function.
This was taken out from one Ingo Molnar's old patch I'm resurrecting.
Signed-off-by: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Fix a small spelling mistake. subtile->subtle
Signed-off-by: Martin Hicks <mort@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Better late than never, I've at last reviewed the madvise vma merging
going into 2.6.13. Remove a pointless check and fix two little bugs -
a simple test (with /proc/<pid>/maps hacked to show ReadHints) showed
both mismerges in practice: though being madvise, neither was disastrous.
1. Correct placement of the success label in madvise_behavior: as in
mprotect_fixup and mlock_fixup, it is necessary to update vm_flags
when vma_merge succeeds (to handle the exceptional Case 8 noted in
the comments above vma_merge itself).
2. Correct initial value of prev when starting part way into a vma: as
in sys_mprotect and do_mlock, it needs to be set to vma in this case
(vma_merge handles only that minimum of cases shown in its comments).
3. If find_vma_prev sets prev, then the vma it returns is prev->vm_next,
so it's pointless to make that same assignment again in sys_madvise.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Christoph Lameter and Marcelo Tosatti asked to get rid of the
atomic_inc_and_test() to cleanup the atomic ops in the zone reclaim code.
Signed-off-by: Martin Hicks <mort@sgi.com>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add a capability check to sys_set_zone_reclaim(). This syscall is not
something that should be available to a user.
Signed-off-by: Martin Hicks <mort@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This bitop does not need to be atomic because it is performed when there will
be no references to the page (ie. the page is being freed).
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
filemap_xip's nopage routine maps the ZERO_PAGE into readonly mappings, if it
has no data page to map there: then if the hole in the file is later filled,
__xip_unmap uses an rmap technique to replace the ZERO_PAGEs mapped for that
offset by the newly allocated file page, so that established mappings will see
the newly written data.
However, on MIPS (alone) there's not one but as many as eight ZERO_PAGEs,
chosen for coloring by user virtual address; and if mremap has meanwhile been
used to move a mapping containing a ZERO_PAGE, it will generally not match the
ZERO_PAGE(address) __xip_unmap is looking for.
To maintain XIP's established mappings correctly on MIPS, we need Nick's fix
to mremap's move_one_page (originally presented as an optimization), to
replace the ZERO_PAGE appropriate to the old address by the ZERO_PAGE
appropriate to the new address.
(But when I first saw this, I was thinking the ZERO_PAGEs themselves would get
corrupted, very bad. Now I think it's the other way round, that the
established mappings will fail to see the newly written data: incorrect, but
not corrupting everything else. Whether filemap_xip's technique is generally
safe, I'd hesitate to say in a hurry: it's interesting, but we've never tried
to do that in tmpfs.)
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Thanks to Bill Irwin for pointing this out.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Microoptimise page_add_anon_rmap. Although these expressions are used only in
the taken branch of the if() statement, the compiler can't reorder them inside
because atomic_inc_and_test is a barrier.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Just be clear that VM_RESERVED pages here are a bug, and the test is not there
because they are expected.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch was recently discussed on linux-mm:
http://marc.theaimsgroup.com/?t=112085728500002&r=1&w=2
I inherited a large code base from Ray for page migration. There was a
small patch in there that I find to be very useful since it allows the
display of the locality of the pages in use by a process. I reworked that
patch and came up with a /proc/<pid>/numa_maps that gives more information
about the vma's of a process. numa_maps is indexes by the start address
found in /proc/<pid>/maps. F.e. with this patch you can see the page use
of the "getty" process:
margin:/proc/12008 # cat maps
00000000-00004000 r--p 00000000 00:00 0
2000000000000000-200000000002c000 r-xp 00000000 08:04 516 /lib/ld-2.3.3.so
2000000000038000-2000000000040000 rw-p 00028000 08:04 516 /lib/ld-2.3.3.so
2000000000040000-2000000000044000 rw-p 2000000000040000 00:00 0
2000000000058000-2000000000260000 r-xp 00000000 08:04 54707842 /lib/tls/libc.so.6.1
2000000000260000-2000000000268000 ---p 00208000 08:04 54707842 /lib/tls/libc.so.6.1
2000000000268000-2000000000274000 rw-p 00200000 08:04 54707842 /lib/tls/libc.so.6.1
2000000000274000-2000000000280000 rw-p 2000000000274000 00:00 0
2000000000280000-20000000002b4000 r--p 00000000 08:04 9126923 /usr/lib/locale/en_US.utf8/LC_CTYPE
2000000000300000-2000000000308000 r--s 00000000 08:04 60071467 /usr/lib/gconv/gconv-modules.cache
2000000000318000-2000000000328000 rw-p 2000000000318000 00:00 0
4000000000000000-4000000000008000 r-xp 00000000 08:04 29576399 /sbin/mingetty
6000000000004000-6000000000008000 rw-p 00004000 08:04 29576399 /sbin/mingetty
6000000000008000-600000000002c000 rw-p 6000000000008000 00:00 0 [heap]
60000fff7fffc000-60000fff80000000 rw-p 60000fff7fffc000 00:00 0
60000ffffff44000-60000ffffff98000 rw-p 60000ffffff44000 00:00 0 [stack]
a000000000000000-a000000000020000 ---p 00000000 00:00 0 [vdso]
cat numa_maps
2000000000000000 default MaxRef=43 Pages=11 Mapped=11 N0=4 N1=3 N2=2 N3=2
2000000000038000 default MaxRef=1 Pages=2 Mapped=2 Anon=2 N0=2
2000000000040000 default MaxRef=1 Pages=1 Mapped=1 Anon=1 N0=1
2000000000058000 default MaxRef=43 Pages=61 Mapped=61 N0=14 N1=15 N2=16 N3=16
2000000000268000 default MaxRef=1 Pages=2 Mapped=2 Anon=2 N0=2
2000000000274000 default MaxRef=1 Pages=3 Mapped=3 Anon=3 N0=3
2000000000280000 default MaxRef=8 Pages=3 Mapped=3 N0=3
2000000000300000 default MaxRef=8 Pages=2 Mapped=2 N0=2
2000000000318000 default MaxRef=1 Pages=1 Mapped=1 Anon=1 N2=1
4000000000000000 default MaxRef=6 Pages=2 Mapped=2 N1=2
6000000000004000 default MaxRef=1 Pages=1 Mapped=1 Anon=1 N0=1
6000000000008000 default MaxRef=1 Pages=1 Mapped=1 Anon=1 N0=1
60000fff7fffc000 default MaxRef=1 Pages=1 Mapped=1 Anon=1 N0=1
60000ffffff44000 default MaxRef=1 Pages=1 Mapped=1 Anon=1 N0=1
getty uses ld.so. The first vma is the code segment which is used by 43
other processes and the pages are evenly distributed over the 4 nodes.
The second vma is the process specific data portion for ld.so. This is
only one page.
The display format is:
<startaddress> Links to information in /proc/<pid>/map
<memory policy> This can be "default" "interleave={}", "prefer=<node>" or "bind={<zones>}"
MaxRef= <maximum reference to a page in this vma>
Pages= <Nr of pages in use>
Mapped= <Nr of pages with mapcount >
Anon= <nr of anonymous pages>
Nx= <Nr of pages on Node x>
The content of the proc-file is self-evident. If this would be tied into
the sparsemem system then the contents of this file would not be too
useful.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Remove the three get_mm_counter(mm, rss) tests from rmap.c: there was a
time when testing rss was important to avoid a particular race between
dup_mmap and the anonmm rmap; but now it's just a rather silly pseudo-
optimization, made even more obscure by the get_mm_counter macro.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Three of the four BUG_ONs in delete_from_swap_cache are immediately
repeated in __delete_from_swap_cache: delete those and add the one. But
perhaps mm/ is altogether overprovisioned with historic BUGs?
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The idea of a swap_device_lock per device, and a swap_list_lock over them all,
is appealing; but in practice almost every holder of swap_device_lock must
already hold swap_list_lock, which defeats the purpose of the split.
The only exceptions have been swap_duplicate, valid_swaphandles and an
untrodden path in try_to_unuse (plus a few places added in this series).
valid_swaphandles doesn't show up high in profiles, but swap_duplicate does
demand attention. However, with the hold time in get_swap_pages so much
reduced, I've not yet found a load and set of swap device priorities to show
even swap_duplicate benefitting from the split. Certainly the split is mere
overhead in the common case of a single swap device.
So, replace swap_list_lock and swap_device_lock by spinlock_t swap_lock
(generally we seem to prefer an _ in the name, and not hide in a macro).
If someone can show a regression in swap_duplicate, then probably we should
add a hashlock for the swap_map entries alone (shorts being anatomic), so as
to help the case of the single swap device too.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The get_swap_page/scan_swap_map latency can be so bad that even those without
preemption configured deserve relief: periodically cond_resched.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
get_swap_page has often shown up on latency traces, doing lengthy scans while
holding two spinlocks. swap_list_lock is already dropped, now scan_swap_map
drop swap_device_lock before scanning the swap_map.
While scanning for an empty cluster, don't worry that racing tasks may
allocate what was free and free what was allocated; but when allocating an
entry, check it's still free after retaking the lock. Avoid dropping the lock
in the expected common path. No barriers beyond the locks, just let the
cookie crumble; highest_bit limit is volatile, but benign.
Guard against swapoff: must check SWP_WRITEOK before allocating, must raise
SWP_SCANNING reference count while in scan_swap_map, swapoff wait for that to
fall - just use schedule_timeout, we don't want to burden scan_swap_map
itself, and it's very unlikely that anyone can really still be in
scan_swap_map once swapoff gets this far.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Rewrite scan_swap_map to allocate in just the same way as before (taking the
next free entry SWAPFILE_CLUSTER-1 times, then restarting at the lowest wholly
empty cluster, falling back to lowest entry if none), but with a view towards
dropping the lock in the next patch.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Rewrite get_swap_page to allocate in just the same sequence as before, but
without holding swap_list_lock across its scan_swap_map. Decrement
nr_swap_pages and update swap_list.next in advance, while still holding
swap_list_lock. Skip full devices by testing highest_bit. Swapoff hold
swap_device_lock as well as swap_list_lock to clear SWP_WRITEOK. Reduces lock
contention when there are parallel swap devices of the same priority.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This makes negligible difference in practice: but swap_list.next should not be
updated to a higher prio in the general helper swap_info_get, but rather in
swap_entry_free; and then only in the case when entry is actually freed.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The swap header's unsigned int last_page determines the range of swap pages,
but swap_info has been using int or unsigned long in some cases: use unsigned
int throughout (except, in several places a local unsigned long is useful to
avoid overflows when adding).
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Jens Axboe <axboe@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The "Adding %dk swap" message shows the number of swap extents, as a guide to
how fragmented the swapfile may be. But a useful further guide is what total
extent they span across (sometimes scarily large).
And there's no need to keep nr_extents in swap_info: it's unused after the
initial message, so save a little space by keeping it on stack.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
There are several comments that swap's extent_list.prev points to the lowest
extent: that's not so, it's extent_list.next which points to it, as you'd
expect. And a couple of loops in add_swap_extent which go all the way through
the list, when they should just add to the other end.
Fix those up, and let map_swap_page search the list forwards: profiles shows
it to be twice as quick that way - because prefetch works better on how the
structs are typically kmalloc'ed? or because usually more is written to than
read from swap, and swap is allocated ascendingly?
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
sys_swapon's call to destroy_swap_extents on failure is made after the final
swap_list_unlock, which is faintly unsafe: another sys_swapon might already be
setting up that swap_info_struct. Calling it earlier, before taking
swap_list_lock, is safe. sys_swapoff's call to destroy_swap_extents was safe,
but likewise move it earlier, before taking the locks (once try_to_unuse has
completed, nothing can be needing the swap extents).
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
If a regular swapfile lies on a filesystem whose blocksize is less than
PAGE_SIZE, then setup_swap_extents may have to cut the number of usable swap
pages; but sys_swapon's nr_good_pages was not expecting that. Also,
setup_swap_extents takes no account of badpages listed in the swap header: not
worth doing so, but ensure nr_badpages is 0 for a regular swapfile.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This splits up sparse_index_alloc() into two pieces. This is needed
because we'll allocate the memory for the second level in a different place
from where we actually consume it to keep the allocation from happening
underneath a lock
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
With cleanups from Dave Hansen <haveblue@us.ibm.com>
SPARSEMEM_EXTREME makes mem_section a one dimensional array of pointers to
mem_sections. This two level layout scheme is able to achieve smaller
memory requirements for SPARSEMEM with the tradeoff of an additional shift
and load when fetching the memory section. The current SPARSEMEM
implementation is a one dimensional array of mem_sections which is the
default SPARSEMEM configuration. The patch attempts isolates the
implementation details of the physical layout of the sparsemem section
array.
SPARSEMEM_EXTREME requires bootmem to be functioning at the time of
memory_present() calls. This is not always feasible, so architectures
which do not need it may allocate everything statically by using
SPARSEMEM_STATIC.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
A new option for SPARSEMEM is ARCH_SPARSEMEM_EXTREME. Architecture
platforms with a very sparse physical address space would likely want to
select this option. For those architecture platforms that don't select the
option, the code generated is equivalent to SPARSEMEM currently in -mm.
I'll be posting a patch on ia64 ml which uses this new SPARSEMEM feature.
ARCH_SPARSEMEM_EXTREME makes mem_section a one dimensional array of
pointers to mem_sections. This two level layout scheme is able to achieve
smaller memory requirements for SPARSEMEM with the tradeoff of an
additional shift and load when fetching the memory section. The current
SPARSEMEM -mm implementation is a one dimensional array of mem_sections
which is the default SPARSEMEM configuration. The patch attempts isolates
the implementation details of the physical layout of the sparsemem section
array.
ARCH_SPARSEMEM_EXTREME depends on 64BIT and is by default boolean false.
I've boot tested under aim load ia64 configured for ARCH_SPARSEMEM_EXTREME.
I've also boot tested a 4 way Opteron machine with !ARCH_SPARSEMEM_EXTREME
and tested with aim.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Defer copying of ptes until fault time when it is possible to reconstruct
the pte from backing store. Idea from Andi Kleen and Nick Piggin.
Thanks to input from Rik van Riel and Linus and to Hugh for correcting
my blundering.
Ray Fucillo <fucillo@intersystems.com> reports:
"I applied this latest patch to a 2.6.12 kernel and found that it does
resolve the problem. Prior to the patch on this machine, I was
seeing about 23ms spent in fork for ever 100MB of shared memory
segment.
After applying the patch, fork is taking about 1ms regardless of the
shared memory size."
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Christoph Lameter
Stephen Smalley
Mark Fasheh
Andi Kleen
Pekka J Enberg
Paul Jackson
Ravikiran G Thirumalai
Steven Pratt
Christoph Hellwig
Christoph Lameter
Martin Hicks
Manfred Spraul
Eric Dumazet
Zachary Amsden
Kyle Moffett
Adam Litke
Deepak Saxena
Paolo 'Blaisorblade' Giarrusso
Hugh Dickins
Martin Hicks
Nick Piggin
Dave Hansen
Bob Picco