17 Mar, 2006
3 commits
-
Lee Revell reported 28ms latency when process with lots of swapped memory
exits.2.6.15 introduced a latency regression when unmapping: in accounting the
zap_work latency breaker, pte_none counted 1, pte_present PAGE_SIZE, but a
swap entry counted nothing at all. We think of pages present as the slow
case, but Lee's trace shows that free_swap_and_cache's radix tree lookup
can make a lot of work - and we could have been doing it many thousands of
times without a latency break.Move the zap_work update up to account swap entries like pages present.
This does account non-linear pte_file entries, and unmap_mapping_range
skipping over swap entries, by the same amount even though they're quick:
but neither of those cases deserves complicating the code (and they're
treated no worse than they were in 2.6.14).Signed-off-by: Hugh Dickins
Acked-by: Nick Piggin
Acked-by: Ingo Molnar
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
We can call try_to_release_page() with PagePrivate off and a valid
page->mapping This may cause all sorts of trouble for the filesystem
*_releasepage() handlers. XFS bombs out in that case.Lock the page before checking for page private.
Signed-off-by: Christoph Lameter
Cc: Nick Piggin
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Currently the migration of anonymous pages will silently fail if no swap is
setup. This patch makes page migration functions check for available swap
and fail with -ENODEV if no swap space is available.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
15 Mar, 2006
2 commits
-
It seems that setting scheduling policy and priorities is also the kind of
thing that might be performed in apps that also use the NUMA API, so it
would seem consistent to use CAP_SYS_NICE for NUMA also.So use CAP_SYS_NICE for controlling migration permissions.
Signed-off-by: Christoph Lameter
Cc: Michael Kerrisk
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
page migration currently simply retries a couple of times if try_to_unmap()
fails without inspecting the return code.However, SWAP_FAIL indicates that the page is in a vma that has the
VM_LOCKED flag set (if ignore_refs ==1). We can check for that return code
and avoid retrying the migration.migrate_page_remove_references() now needs to return a reason why the
failure occured. So switch migrate_page_remove_references to use -Exx
style error messages.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
10 Mar, 2006
4 commits
-
The cache reaper currently tries to free all alien caches and all remote
per cpu pages in each pass of cache_reap. For a machines with large number
of nodes (such as Altix) this may lead to sporadic delays of around ~10ms.
Interrupts are disabled while reclaiming creating unacceptable delays.This patch changes that behavior by adding a per cpu reap_node variable.
Instead of attempting to free all caches, we free only one alien cache and
the per cpu pages from one remote node. That reduces the time spend in
cache_reap. However, doing so will lengthen the time it takes to
completely drain all remote per cpu pagesets and all alien caches. The
time needed will grow with the number of nodes in the system. All caches
are drained when they overflow their respective capacity. So the drawback
here is only that a bit of memory may be wasted for awhile longer.Details:
1. Rename drain_remote_pages to drain_node_pages to allow the specification
of the node to drain of pcp pages.2. Add additional functions init_reap_node, next_reap_node for NUMA
that manage a per cpu reap_node counter.3. Add a reap_alien function that reaps only from the current reap_node.
For us this seems to be a critical issue. Holdoffs of an average of ~7ms
cause some HPC benchmarks to slow down significantly. F.e. NAS parallel
slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor
compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with
the additional interrupt holdoff reductions.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
When pages are onlined, not only zone->present_pages but also
pgdat->node_present_pages should be refreshed.This parameter is used to show information at
/sys/device/system/node/nodeX/meminfo via si_meminfo_node().So, it shows strange value for MemUsed which is calculated
(node_present_pages - all zones free pages).Signed-off-by: Yasunori Goto
Cc: Dave Hansen
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
If the process has already set PF_MALLOC and is already using
current->reclaim_state then do not try to reclaim memory from the zone.
This is set by kswapd and/or synchrononous global reclaim which will not
take it lightly if we zap the reclaim_state.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Remove two early-development BUG_ONs from page_add_file_rmap.
The pfn_valid test (originally useful for checking that nobody passed an
artificial struct page) comes too late, since we already have the struct
page.The PageAnon test (useful when anon was first distinguished from file rmap)
prevents ->nopage implementations from reusing ->mapping, which would
otherwise be available.Signed-off-by: Hugh Dickins
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
09 Mar, 2006
4 commits
-
kmem_cache_init() incorrectly assumes that the cache_cache object will fit
in an order 0 allocation. On very large systems, this is not true. Change
the code to try larger order allocations if order 0 fails.Signed-off-by: Jack Steiner
Cc: Manfred Spraul
Cc: Pekka Enberg
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Implement percpu_counter_sum(). This is a more accurate but slower version of
percpu_counter_read_positive().We need this for Alex's speedup-ext3_statfs patch and for the nr_file
accounting fix. Otherwise these things would be too inaccurate on large CPU
counts.Cc: Ravikiran G Thirumalai
Cc: Alex Tomas
Cc: "David S. Miller"
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Fix the mm/mempolicy.c build for !CONFIG_HUGETLB_PAGE.
Cc: Christoph Lameter
Cc: Martin Bligh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Instead of having a hard-to-read and confusing conditional in the
caller, just make the slab order calculation handle this special case,
since it's simple and obvious there.Signed-off-by: Linus Torvalds
07 Mar, 2006
3 commits
-
Change the format of numa_maps to be more compact and contain additional
information that is useful for managing and troubleshooting memory on a
NUMA system. Numa_maps can now also support huge pages.Fixes:
1. More compact format. Only display fields if they contain additional
information.2. Always display information for all vmas. The old numa_maps did not display
vma with no mapped entries. This was a bit confusing because page
migration removes ptes for file backed vmas. After page migration
a part of the vmas vanished.3. Rename maxref to maxmap. This is the maximum mapcount of all the pages
in a vma and may be used as an indicator as to how many processes
may be using a certain vma.4. Include the ability to scan over huge page vmas.
New items shown:
dirty
Number of pages in a vma that have either the dirty bit set in the
page_struct or in the pte.file=
The file backing the pages if anystack
Stack areaheap
Heap areahuge
Huge page area. The number of pages shows is the number of huge
pages not the regular sized pages.swapcache
Number of pages with swap references. Must be >0 in order to
be shown.active
Number of active pages. Only displayed if different from the number
of pages mapped.writeback
Number of pages under writeback. Only displayed if >0.Sample ouput of a process using huge pages:
00000000 default
2000000000000000 default file=/lib/ld-2.3.90.so mapped=13 mapmax=30 N0=13
2000000000044000 default file=/lib/ld-2.3.90.so anon=2 dirty=2 swapcache=2 N2=2
2000000000064000 default file=/lib/librt-2.3.90.so mapped=2 active=1 N1=1 N3=1
2000000000074000 default file=/lib/librt-2.3.90.so
2000000000080000 default file=/lib/librt-2.3.90.so anon=1 swapcache=1 N2=1
2000000000084000 default
2000000000088000 default file=/lib/libc-2.3.90.so mapped=52 mapmax=32 active=48 N0=52
20000000002bc000 default file=/lib/libc-2.3.90.so
20000000002c8000 default file=/lib/libc-2.3.90.so anon=3 dirty=2 swapcache=3 active=2 N1=1 N2=2
20000000002d4000 default anon=1 swapcache=1 N1=1
20000000002d8000 default file=/lib/libpthread-2.3.90.so mapped=8 mapmax=3 active=7 N2=2 N3=6
20000000002fc000 default file=/lib/libpthread-2.3.90.so
2000000000308000 default file=/lib/libpthread-2.3.90.so anon=1 dirty=1 swapcache=1 N1=1
200000000030c000 default anon=1 dirty=1 swapcache=1 N1=1
2000000000320000 default anon=1 dirty=1 N1=1
200000000071c000 default
2000000000720000 default anon=2 dirty=2 swapcache=1 N1=1 N2=1
2000000000f1c000 default
2000000000f20000 default anon=2 dirty=2 swapcache=1 active=1 N2=1 N3=1
200000000171c000 default
2000000001720000 default anon=1 dirty=1 swapcache=1 N1=1
2000000001b20000 default
2000000001b38000 default file=/lib/libgcc_s.so.1 mapped=2 N1=2
2000000001b48000 default file=/lib/libgcc_s.so.1
2000000001b54000 default file=/lib/libgcc_s.so.1 anon=1 dirty=1 active=0 N1=1
2000000001b58000 default file=/lib/libunwind.so.7.0.0 mapped=2 active=1 N1=2
2000000001b74000 default file=/lib/libunwind.so.7.0.0
2000000001b80000 default file=/lib/libunwind.so.7.0.0
2000000001b84000 default
4000000000000000 default file=/media/huge/test9 mapped=1 N1=1
6000000000000000 default file=/media/huge/test9 anon=1 dirty=1 active=0 N1=1
6000000000004000 default heap
607fffff7fffc000 default anon=1 dirty=1 swapcache=1 N2=1
607fffffff06c000 default stack anon=1 dirty=1 active=0 N1=1
8000000060000000 default file=/mnt/huge/test0 huge dirty=3 N1=3
8000000090000000 default file=/mnt/huge/test1 huge dirty=3 N0=1 N2=2
80000000c0000000 default file=/mnt/huge/test2 huge dirty=3 N1=1 N3=2Signed-off-by: Christoph Lameter
Cc: Andi Kleen
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
If we triggered the 'offslab_limit' test, we would return with
cachep->gfporder incremented once too many times.This clarifies the logic somewhat, and fixes that bug.
Signed-off-by: Linus Torvalds
-
We want to use the "struct slab" size, not the size of the pointer to
same. As it is, we'd not print out the last entry pointers in the
slab (where is ~10, depending on whether it's a 32-bit or 64-bit
kernel).Gaah, that slab code was written by somebody who likes unreadable crud.
Signed-off-by: Linus Torvalds
03 Mar, 2006
2 commits
-
numa_maps should not scan over huge vmas in order not to cause problems for
non IA64 platforms that may have pte entries pointing to huge pages in a
variety of ways in their page tables. Add a simple check to ignore vmas
containing huge pages.Signed-off-by: Christoph Lameter
Cc: Hugh Dickins
Cc: Andi Kleen
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
I seem to have lost this read_unlock().
While we're there, let's turn that interruptible sleep unto uninterruptible,
so we don't get a busywait if signal_pending(). (Again. We seem to have a
habit of doing this).Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
01 Mar, 2006
4 commits
-
Under some circumstances `points' can get printed before it's initialised.
Spotted by Carlos Martin .Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Fix oprofile linkage. Pointed out by "Luke Yang" .
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
remove_from_swap() currently attempts to use page_lock_anon_vma to obtain
an anon_vma lock. That is not working since the page may have been
remapped via swap ptes in order to move the page.However, do_migrate_pages() obtain the mmap_sem lock and therefore there is
a guarantee that the anonymous vma will not vanish from under us. There is
therefore no need to use page_lock_anon_vma.Signed-off-by: Christoph Lameter
Acked-by: Hugh Dickins
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Currently sys_migrate_pages only moves pages belonging to a process. This
is okay when invoked from a regular user. But if invoked from root it
should move all pages as documented in the migrate_pages manpage.Signed-off-by: Christoph Lameter
Cc: Andi Kleen
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
25 Feb, 2006
2 commits
-
- PF_SWAPWRITE needs to be set for RECLAIM_SWAP to be able to write
out pages to swap. Currently RECLAIM_SWAP may not do that.- remove setting nr_reclaimed pages after slab reclaim since the slab shrinking
code does not use that and the nr_reclaimed pages is just right for the
intended follow up action.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
migrate_pages_to() allocates a list of new pages on the intended target
node or with the intended policy and then uses the list of new pages as
targets for the migration of a list of pages out of place.When the pages are allocated it is not clear which of the out of place
pages will be moved to the new pages. So we cannot specify an address as
needed by alloc_page_vma(). This causes problem for MPOL_INTERLEAVE which
will currently allocate the pages on the first node of the set. If mbind
is used with vma that has the policy of MPOL_INTERLEAVE then the
interleaving of pages may be destroyed.This patch fixes that by generating a fake address for each alloc_page_vma
which will result is a distribution of pages as prescribed by
MPOL_INTERLEAVE.Lee also noted that the sequence of nodes for the new pages seems to be
inverted. So we also invert the way the lists of pages for migration are
build.Signed-off-by: Christoph Lameter
Signed-off-by: Lee Schermerhorn
Looks-ok-to: Andi Kleen
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
22 Feb, 2006
1 commit
-
I've been dissatisfied with the mpol_nodelist mount option which was
added to tmpfs earlier in -rc. Replace it by mpol=policy:nodelist.And it was broken: a nodelist is a comma-separated list of numbers and
ranges; the mount options are a comma-separated list of token=values.
Whoops, blindly strsep'ing on commas doesn't work so well: since we've
no numeric tokens, and unlikely to add them, use that to distinguish.Move the mpol= parsing to shmem_parse_mpol under CONFIG_NUMA, reject
all its options as invalid if not NUMA. /proc shows MPOL_PREFERRED
as "prefer", so use that name for the policy instead of "preferred".Enforce that mpol=default has no nodelist; that mpol=prefer has one
node only; that mpol=bind has a nodelist; but let mpol=interleave use
node_online_map if no nodelist given. Describe this in tmpfs.txt.Signed-off-by: Hugh Dickins
Acked-by: Robin Holt
Acked-by: Andi Kleen
Signed-off-by: Linus Torvalds
21 Feb, 2006
5 commits
-
[akpm; it happens that the code was still correct, only inefficient ]
Signed-off-by: Alexey Dobriyan
Cc: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Signed-off-by: Luke Yang
Acked-by: David Howells
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
maxnode is a bit index and can't be directly compared against a byte length
like PAGE_SIZESigned-off-by: Andi Kleen
Cc: Chris Wright
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Some allocations are restricted to a limited set of nodes (due to memory
policies or cpuset constraints). If the page allocator is not able to find
enough memory then that does not mean that overall system memory is low.In particular going postal and more or less randomly shooting at processes
is not likely going to help the situation but may just lead to suicide (the
whole system coming down).It is better to signal to the process that no memory exists given the
constraints that the process (or the configuration of the process) has
placed on the allocation behavior. The process may be killed but then the
sysadmin or developer can investigate the situation. The solution is
similar to what we do when running out of hugepages.This patch adds a check before we kill processes. At that point
performance considerations do not matter much so we just scan the zonelist
and reconstruct a list of nodes. If the list of nodes does not contain all
online nodes then this is a constrained allocation and we should kill the
current process.Signed-off-by: Christoph Lameter
Cc: Nick Piggin
Cc: Andi Kleen
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
In the badness() calculation, there's currently this piece of code:
/*
* Processes which fork a lot of child processes are likely
* a good choice. We add the vmsize of the children if they
* have an own mm. This prevents forking servers to flood the
* machine with an endless amount of children
*/
list_for_each(tsk, &p->children) {
struct task_struct *chld;
chld = list_entry(tsk, struct task_struct, sibling);
if (chld->mm = p->mm && chld->mm)
points += chld->mm->total_vm;
}The intention is clear: If some server (apache) keeps spawning new children
and we run OOM, we want to kill the father rather than picking a child.This -- to some degree -- also helps a bit with getting fork bombs under
control, though I'd consider this a desirable side-effect rather than a
feature.There's one problem with this: No matter how many or few children there are,
if just one of them misbehaves, and all others (including the father) do
everything right, we still always kill the whole family. This hits in real
life; whether it's javascript in konqueror resulting in kdeinit (and thus the
whole KDE session) being hit or just a classical server that spawns children.Sidenote: The killer does kill all direct children as well, not only the
selected father, see oom_kill_process().The idea in attached patch is that we do want to account the memory
consumption of the (direct) children to the father -- however not fully.
This maintains the property that fathers with too many children will still
very likely be picked, whereas a single misbehaving child has the chance to
be picked by the OOM killer.In the patch I account only half (rounded up) of the children's vm_size to
the parent. This means that if one child eats more mem than the rest of
the family, it will be picked, otherwise it's still the father and thus the
whole family that gets selected.This is heuristics -- we could debate whether accounting for a fourth would
be better than for half of it. Or -- if people would consider it worth the
trouble -- make it a sysctl. For now I sticked to accounting for half,
which should IMHO be a significant improvement.The patch does one more thing: As users tend to be irritated by the choice
of killed processes (mainly because the children are killed first, despite
some of them having a very low OOM score), I added some more output: The
selected (father) process will be reported first and it's oom_score printed
to syslog.Description:
Only account for half of children's vm size in oom score calculation
This should still give the parent enough point in case of fork bombs. If
any child however has more than 50% of the vm size of all children
together, it'll get a higher score and be elected.This patch also makes the kernel display the oom_score.
Signed-off-by: Kurt Garloff
Cc: Rik van Riel
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
18 Feb, 2006
4 commits
-
Make sure maxnodes is safe size before calculating nlongs in
get_nodes().Signed-off-by: Chris Wright
Signed-off-by: Linus Torvalds -
Change the find_next_best_node algorithm to correctly skip
over holes in the node online mask. Previously it would not handle
missing nodes correctly and cause crashes at boot.[Written by Linus, tested by AK]
Signed-off-by: Andi Kleen
Signed-off-by: Linus Torvalds -
The memory allocator doesn't like empty zones (which have an
uninitialized freelist), so a x86-64 system with a node fully
in GFP_DMA32 only would crash on mbind.Fix that up by putting all possible zones as fallback into the zonelist
and skipping the empty ones.In fact the code always enough allocated space for all zones,
but only used it for the highest. This change just uses all the
memory that was allocated before.This should work fine for now, but whoever implements node hot removal
needs to fix this somewhere else too (or make sure zone datastructures
by itself never go away, only their memory)Signed-off-by: Andi Kleen
Acked-by: Christoph Lameter
Signed-off-by: Linus Torvalds -
AMD SimNow!'s JIT doesn't like them at all in the guest. For distribution
installation it's easiest if it's a boot time option.Also I moved the variable to a more appropiate place and make
it independent from sysctlAnd marked __read_mostly which it is.
Signed-off-by: Andi Kleen
Signed-off-by: Linus Torvalds
15 Feb, 2006
3 commits
-
Currently, copy-on-write may change the physical address of a page even if the
user requested that the page is pinned in memory (either by mlock or by
get_user_pages). This happens if the process forks meanwhile, and the parent
writes to that page. As a result, the page is orphaned: in case of
get_user_pages, the application will never see any data hardware DMA's into
this page after the COW. In case of mlock'd memory, the parent is not getting
the realtime/security benefits of mlock.In particular, this affects the Infiniband modules which do DMA from and into
user pages all the time.This patch adds madvise options to control whether memory range is inherited
across fork. Useful e.g. for when hardware is doing DMA from/into these
pages. Could also be useful to an application wanting to speed up its forks
by cutting large areas out of consideration.Signed-off-by: Michael S. Tsirkin
Acked-by: Hugh Dickins
Cc: Michael Kerrisk
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Somehow I imagined that calling a NULL destructor would free a compound page
rather than oopsing. No, we must supply a default destructor, __free_pages_ok
using the order noted by prep_compound_page. hugetlb can still replace this
as before with its own free_huge_page pointer.The case that needs this is not common: rarely does put_compound_page's
put_page_testzero bring the count down to 0. But if get_user_pages is applied
to some part of a compound page, without immediate release (e.g. AIO or
Infiniband), then it's possible for its put_page to come after the containing
vma has been unmapped and the driver done its free_pages.That's just the kind of case compound pages are supposed to be guarding
against (but Nick points out, nor did PageReserved handle this right).Signed-off-by: Hugh Dickins
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
If a compound page has its own put_page_testzero destructor (the only current
example is free_huge_page), that is noted in page[1].mapping of the compound
page. But that's rather a poor place to keep it: functions which call
set_page_dirty_lock after get_user_pages (e.g. Infiniband's
__ib_umem_release) ought to be checking first, otherwise set_page_dirty is
liable to crash on what's not the address of a struct address_space.And now I'm about to make that worse: it turns out that every compound page
needs a destructor, so we can no longer rely on hugetlb pages going their own
special way, to avoid further problems of page->mapping reuse. For example,
not many people know that: on 50% of i386 -Os builds, the first tail page of a
compound page purports to be PageAnon (when its destructor has an odd
address), which surprises page_add_file_rmap.Keep the compound page destructor in page[1].lru.next instead. And to free up
the common pairing of mapping and index, also move compound page order from
index to lru.prev. Slab reuses page->lru too: but if we ever need slab to use
compound pages, it can easily stack its use above this.(akpm: decoded version of the above: the tail pages of a compound page now
have ->mapping==NULL, so there's no need for the set_page_dirty[_lock]()
caller to check that they're not compund pages before doing the dirty).Signed-off-by: Hugh Dickins
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
12 Feb, 2006
3 commits
-
This puts the variables and the way to get to reclaim_mapped in one block.
And allows zone_reclaim or other things to skip the determination (maybe
this whole block of code does not belong into refill_inactive_zone()?)Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
shrink_zone() already increments reclaim_in_progress. No need to do it in
balance_pgdat.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
shrink_list() and refill_inactive() check all ptes pointing to a page for
reference bits in order to decide if the page should be put on the active
list. This is not necessary for zone_reclaim since we are only interested
in removing unmapped pages. Skip the checks in both functions.Signed-off-by: Christoph Lameter
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds