The check for super sized slabs where we can no longer move the free
pointer behind the object for debugging purposes etc is accessing a
field that is not setup yet. We must use objsize here since the size of
the slab has not been determined yet.
The effect of this is that a global slab shrink via "slabinfo -s" will
show errors about offsets being wrong if booted with slub_debug.
Potentially there are other troubles with huge slabs under slub_debug
because the calculated free pointer offset is truncated.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The object size calculation is wrong if !CONFIG_SLUB_DEBUG because the
#ifdef CONFIG_SLUB_DEBUG is now switching off the size adjustments for
DESTROY_BY_RCU and ctor.
Signed-off-by: Christoph Lameter <[email protected]>
Acked-by: Hugh Dickins <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The slab manipulation functions should not be triggered by slabs that
are unresovable in the subset of slabs selected on the command line.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
grow_dev_page() simply passes GFP_NOFS to find_or_create_page. This means
the allocation of radix tree nodes is done with GFP_NOFS and the allocation
of a new page is done using GFP_NOFS.
The mapping has a flags field that contains the necessary allocation flags
for the page cache allocation. These need to be consulted in order to get
DMA and HIGHMEM allocations etc right. And yes a blockdev could be
allowing Highmem allocations if its a ramdisk.
Cc: Hugh Dickins <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Currently we have a maze of configuration variables that determine the
maximum slab size. Worst of all it seems to vary between SLAB and SLUB.
So define a common maximum size for kmalloc. For conveniences sake we use
the maximum size ever supported which is 32 MB. We limit the maximum size
to a lower limit if MAX_ORDER does not allow such large allocations.
For many architectures this patch will have the effect of adding large
kmalloc sizes. x86_64 adds 5 new kmalloc sizes. So a small amount of
memory will be needed for these caches (contemporary SLAB has dynamically
sizeable node and cpu structure so the waste is less than in the past)
Most architectures will then be able to allocate object with sizes up to
MAX_ORDER. We have had repeated breakage (in fact whenever we doubled the
number of supported processors) on IA64 because one or the other struct
grew beyond what the slab allocators supported. This will avoid future
issues and f.e. avoid fixes for 2k and 4k cpu support.
CONFIG_LARGE_ALLOCS is no longer necessary so drop it.
It fixes sparc64 with SLAB.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: "David S. Miller" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Consolidate functionality into the #ifdef section.
Extract tracing into one subroutine.
Move object debug processing into the #ifdef section so that the
code in __slab_alloc and __slab_free becomes minimal.
Reduce number of functions we need to provide stubs for in the !SLUB_DEBUG case.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The atomicity when handling flags in SLUB is not necessary since both flags
used by SLUB are not updated in a racy way. Flag updates are either done
during slab creation or destruction or under slab_lock. Some of these flags
do not have the non atomic variants that we need. So define our own.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Align the output of % with K/M/G of sizes.
Check for empty NUMA information to avoid segfault on !NUMA.
-r should work directly not only if we match a single slab
without additional options.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
slub warns on this, and we're working on making kmalloc(0) return NULL.
Let's make slab warn as well so our testers detect such callers more
rapidly.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Use inline functions to access the per cpu bit. Intoduce the notion of
"freezing" a slab to make things more understandable.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Two definitions remained in slab.h that are particular to the SLAB allocator.
Move to slab_def.h
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
No arch sets ARCH_USES_SLAB_PAGE_STRUCT anymore.
Remove the experimental dependency as well since we want to have it as
a real alternative to SLAB.
It all comes down to killing a single line from init/Kconfig.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There is no user of destructors left. There is no reason why we should keep
checking for destructors calls in the slab allocators.
The RFC for this patch was discussed at
http://marc.info/?l=linux-kernel&m=117882364330705&w=2
Destructors were mainly used for list management which required them to take a
spinlock. Taking a spinlock in a destructor is a bit risky since the slab
allocators may run the destructors anytime they decide a slab is no longer
needed.
Patch drops destructor support. Any attempt to use a destructor will BUG().
Acked-by: Pekka Enberg <[email protected]>
Acked-by: Paul Mundt <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We call alloc_page where we should be calling __page_cache_alloc.
__page_cache_alloc performs cpuset memory spreading. alloc_page does not.
There is no reason that pages allocated via find_or_create should be
exempt.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Sorry I screwed up the comparison. It is only an error if we attempt
to allocate a slab larger than the maximum allowed size.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Take MAX_ORDER into consideration when determining KMALLOC_SHIFT_HIGH.
Otherwise we may run into a situation where we attempt to create general
slabs larger than MAX_ORDER.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: "David S. Miller" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
SLUB cannot run on i386 at this point because i386 uses the page->private and
page->index field of slab pages for the pgd cache.
Make SLUB run on i386 by replacing the pgd slab cache with a quicklist.
Limit the changes as much as possible. Leave the improvised linked list in place
etc etc. This has been working here for a couple of weeks now.
Acked-by: William Lee Irwin III <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
IA64 is the origin of the quicklist implementation. So cut out the pieces
that are now in core code and modify the functions called.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Tony Luck <[email protected]>
VM statistics updates do not matter if the kernel is in idle powersaving
mode. So allow the timer to be deferred.
It would be better though if we could switch the timer between deferrable
and nondeferrable based on differentials present. The timer would start
out nondeferrable and if we find that there were no updates in the last
statistics interval then we would switch the timer to deferrable. If the
timer later finds again that there are differentials then go to
nondeferrable again.
And yet another way would be to run the timer shortly before going to idle?
The solution here means that the VM counters may be slightly off during
idle since differentials may be still pending while the timer is deferred.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This was in SLUB in order to head off trouble while the nr_cpu_ids
functionality was not merged. Its merged now so no need to still have this.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Otherwise people get asked about SLUB_DEBUG even if they have another
slab allocator enabled.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Avoid atomic overhead in slab_alloc and slab_free
SLUB needs to use the slab_lock for the per cpu slabs to synchronize with
potential kfree operations. This patch avoids that need by moving all free
objects onto a lockless_freelist. The regular freelist continues to exist
and will be used to free objects. So while we consume the
lockless_freelist the regular freelist may build up objects.
If we are out of objects on the lockless_freelist then we may check the
regular freelist. If it has objects then we move those over to the
lockless_freelist and do this again. There is a significant savings in
terms of atomic operations that have to be performed.
We can even free directly to the lockless_freelist if we know that we are
running on the same processor. So this speeds up short lived objects.
They may be allocated and freed without taking the slab_lock. This is
particular good for netperf.
In order to maximize the effect of the new faster hotpath we extract the
hottest performance pieces into inlined functions. These are then inlined
into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and
freeing no longer requires a subroutine call within SLUB.
[I am not sure that it is worth doing this because it changes the easy to
read structure of slub just to reduce atomic ops. However, there is
someone out there with a benchmark on 4 way and 8 way processor systems
that seems to show a 5% regression vs. Slab. Seems that the regression is
due to increased atomic operations use vs. SLAB in SLUB). I wonder if
this is applicable or discernable at all in a real workload?]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Currently the slab allocators contain callbacks into the page allocator to
perform the draining of pagesets on remote nodes. This requires SLUB to have
a whole subsystem in order to be compatible with SLAB. Moving node draining
out of the slab allocators avoids a section of code in SLUB.
Move the node draining so that is is done when the vm statistics are updated.
At that point we are already touching all the cachelines with the pagesets of
a processor.
Add a expire counter there. If we have to update per zone or global vm
statistics then assume that the pageset will require subsequent draining.
The expire counter will be decremented on each vm stats update pass until it
reaches zero. Then we will drain one batch from the pageset. The draining
will cause vm counter updates which will then cause another expiration until
the pcp is empty. So we will drain a batch every 3 seconds.
Note that remote node draining is a somewhat esoteric feature that is required
on large NUMA systems because otherwise significant portions of system memory
can become trapped in pcp queues. The number of pcp is determined by the
number of processors and nodes in a system. A system with 4 processors and 2
nodes has 8 pcps which is okay. But a system with 1024 processors and 512
nodes has 512k pcps with a high potential for large amount of memory being
caught in them.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Make it configurable. Code in mm makes the vm statistics intervals
independent from the cache reaper use that opportunity to make it
configurable.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
vmstat is currently using the cache reaper to periodically bring the
statistics up to date. The cache reaper does only exists in SLUB as a way to
provide compatibility with SLAB. This patch removes the vmstat calls from the
slab allocators and provides its own handling.
The advantage is also that we can use a different frequency for the updates.
Refreshing vm stats is a pretty fast job so we can run this every second and
stagger this by only one tick. This will lead to some overlap in large
systems. F.e a system running at 250 HZ with 1024 processors will have 4 vm
updates occurring at once.
However, the vm stats update only accesses per node information. It is only
necessary to stagger the vm statistics updates per processor in each node. Vm
counter updates occurring on distant nodes will not cause cacheline
contention.
We could implement an alternate approach that runs the first processor on each
node at the second and then each of the other processor on a node on a
subsequent tick. That may be useful to keep a large amount of the second free
of timer activity. Maybe the timer folks will have some feedback on this one?
[[email protected]: add missing break]
Cc: Arjan van de Ven <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Jiri Slaby <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Shutdown the cache_reaper if the cpu is brought down and set the
cache_reap.func to NULL. Otherwise hotplug shuts down the reaper for good.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This is done in order to be able to run SLUB which expects no modifications
to its page structs.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: David Howells <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Fix some of the spelling issues. Fix sentences. Discourage SLOB use
since SLUB can pack objects denser.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Matt Mackall <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
In some cases SLUB is creating uselessly slabs that are larger than
slub_max_order. Also the layout of some of the slabs was not satisfactory.
Go to an iterarive approach.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
CONFIG_SLUB_DEBUG can be used to switch off the debugging and sysfs components
of SLUB. Thus SLUB will be able to replace SLOB. SLUB can arrange objects in
a denser way than SLOB and the code size should be minimal without debugging
and sysfs support.
Note that CONFIG_SLUB_DEBUG is materially different from CONFIG_SLAB_DEBUG.
CONFIG_SLAB_DEBUG is used to enable slab debugging in SLAB. SLUB enables
debugging via a boot parameter. SLUB debug code should always be present.
CONFIG_SLUB_DEBUG can be modified in the embedded config section.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Move the tracking definitions and the check_valid_pointer() function away from
the debugging related functions.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Trace in both slab_alloc and slab_free has a lot of common code. Use a single
function for both.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This replaces the PageError() checking. DebugSlab is clearer and allows for
future changes to the page bit used. We also need it to support
CONFIG_SLUB_DEBUG.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Move the resiliency check into the SYSFS section after validate_slab that is
used by the resiliency check. This will avoid a forward declaration.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Scanning of objects happens in a number of functions. Consolidate that code.
DECLARE_BITMAP instead of coding the declaration for bitmaps.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Update comments throughout SLUB to reflect the new developments. Fix up
various awkward sentences.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Its only purpose was to bring some sort of symmetry to sysfs usage when
dealing with bootstrapping per cpu flushing. Since we do not time out slabs
anymore we have no need to run finish_bootstrap even without sysfs. Fold it
back into slab_sysfs_init and drop the initcall for the !SYFS case.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We really do not need all this gaga there.
ksize gives us all the information we need to figure out if the object can
cope with the new size.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We needlessly duplicate code. Also make check_valid_pointer inline.
Signed-off-by: Christoph LAemter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
-e Show empty slabs
-d Modification of slab debug options at runtime
-o Operations. Display of ctor / dtor etc.
-r Report: Display all available information about a slabcache.
Cleanup tracking display and make it work right.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
If no redzoning is selected then we do not need padding before the next
object.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
SLUB currently assumes that the cacheline size is static. However, i386 f.e.
supports dynamic cache line size determination.
Use cache_line_size() instead of L1_CACHE_BYTES in the allocator.
That also explains the purpose of SLAB_HWCACHE_ALIGN. So we will need to keep
that one around to allow dynamic aligning of objects depending on boot
determination of the cache line size.
[[email protected]: need to define it before we use it]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Address spaces contain an allocation flag that specifies restriction on the
zone for pages placed in the mapping. I.e. some device may require pages
to be allocated from a DMA zone. Block devices may not be able to use
pages from HIGHMEM.
Memory policies and the common use of page migration works only on the
highest zone. If the address space does not allow allocation from the
highest zone then the pages in the address space are not migratable simply
because we can only allocate memory for a specified node if we allow
allocation for the highest zone on each node.
Acked-by: Hugh Dickins <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There is no user remaining and I have never seen any use of that flag.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
SLAB_CTOR atomic is never used which is no surprise since I cannot imagine
that one would want to do something serious in a constructor or destructor.
In particular given that the slab allocators run with interrupts disabled.
Actions in constructors and destructors are by their nature very limited
and usually do not go beyond initializing variables and list operations.
(The i386 pgd ctor and dtors do take a spinlock in constructor and
destructor..... I think that is the furthest we go at this point.)
There is no flag passed to the destructor so removing SLAB_CTOR_ATOMIC also
establishes a certain symmetry.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I have never seen a use of SLAB_DEBUG_INITIAL. It is only supported by
SLAB.
I think its purpose was to have a callback after an object has been freed
to verify that the state is the constructor state again? The callback is
performed before each freeing of an object.
I would think that it is much easier to check the object state manually
before the free. That also places the check near the code object
manipulation of the object.
Also the SLAB_DEBUG_INITIAL callback is only performed if the kernel was
compiled with SLAB debugging on. If there would be code in a constructor
handling SLAB_DEBUG_INITIAL then it would have to be conditional on
SLAB_DEBUG otherwise it would just be dead code. But there is no such code
in the kernel. I think SLUB_DEBUG_INITIAL is too problematic to make real
use of, difficult to understand and there are easier ways to accomplish the
same effect (i.e. add debug code before kfree).
There is a related flag SLAB_CTOR_VERIFY that is frequently checked to be
clear in fs inode caches. Remove the pointless checks (they would even be
pointless without removeal of SLAB_DEBUG_INITIAL) from the fs constructors.
This is the last slab flag that SLUB did not support. Remove the check for
unimplemented flags from SLUB.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
It is not necessary to tell the slab allocators to align to a cacheline
if an explicit alignment was already specified. It is rather confusing
to specify multiple alignments.
Make sure that the call sites only use one form of alignment.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patch provides a new macro
KMEM_CACHE(<struct>, <flags>)
to simplify slab creation. KMEM_CACHE creates a slab with the name of the
struct, with the size of the struct and with the alignment of the struct.
Additional slab flags may be specified if necessary.
Example
struct test_slab {
int a,b,c;
struct list_head;
} __cacheline_aligned_in_smp;
test_slab_cache = KMEM_CACHE(test_slab, SLAB_PANIC)
will create a new slab named "test_slab" of the size sizeof(struct
test_slab) and aligned to the alignment of test slab. If it fails then we
panic.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patch was recently posted to lkml and acked by Pekka.
The flag SLAB_MUST_HWCACHE_ALIGN is
1. Never checked by SLAB at all.
2. A duplicate of SLAB_HWCACHE_ALIGN for SLUB
3. Fulfills the role of SLAB_HWCACHE_ALIGN for SLOB.
The only remaining use is in sparc64 and ppc64 and their use there
reflects some earlier role that the slab flag once may have had. If
its specified then SLAB_HWCACHE_ALIGN is also specified.
The flag is confusing, inconsistent and has no purpose.
Remove it.
Acked-by: Pekka Enberg <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On x86_64 this cuts allocation overhead for page table pages down to a
fraction (kernel compile / editing load. TSC based measurement of times spend
in each function):
no quicklist
pte_alloc 1569048 4.3s(401ns/2.7us/179.7us)
pmd_alloc 780988 2.1s(337ns/2.7us/86.1us)
pud_alloc 780072 2.2s(424ns/2.8us/300.6us)
pgd_alloc 260022 1s(920ns/4us/263.1us)
quicklist:
pte_alloc 452436 573.4ms(8ns/1.3us/121.1us)
pmd_alloc 196204 174.5ms(7ns/889ns/46.1us)
pud_alloc 195688 172.4ms(7ns/881ns/151.3us)
pgd_alloc 65228 9.8ms(8ns/150ns/6.1us)
pgd allocations are the most complex and there we see the most dramatic
improvement (may be we can cut down the amount of pgds cached somewhat?). But
even the pte allocations still see a doubling of performance.
1. Proven code from the IA64 arch.
The method used here has been fine tuned for years and
is NUMA aware. It is based on the knowledge that accesses
to page table pages are sparse in nature. Taking a page
off the freelists instead of allocating a zeroed pages
allows a reduction of number of cachelines touched
in addition to getting rid of the slab overhead. So
performance improves. This is particularly useful if pgds
contain standard mappings. We can save on the teardown
and setup of such a page if we have some on the quicklists.
This includes avoiding lists operations that are otherwise
necessary on alloc and free to track pgds.
2. Light weight alternative to use slab to manage page size pages
Slab overhead is significant and even page allocator use
is pretty heavy weight. The use of a per cpu quicklist
means that we touch only two cachelines for an allocation.
There is no need to access the page_struct (unless arch code
needs to fiddle around with it). So the fast past just
means bringing in one cacheline at the beginning of the
page. That same cacheline may then be used to store the
page table entry. Or a second cacheline may be used
if the page table entry is not in the first cacheline of
the page. The current code will zero the page which means
touching 32 cachelines (assuming 128 byte). We get down
from 32 to 2 cachelines in the fast path.
3. x86_64 gets lightweight page table page management.
This will allow x86_64 arch code to faster repopulate pgds
and other page table entries. The list operations for pgds
are reduced in the same way as for i386 to the point where
a pgd is allocated from the page allocator and when it is
freed back to the page allocator. A pgd can pass through
the quicklists without having to be reinitialized.
64 Consolidation of code from multiple arches
So far arches have their own implementation of quicklist
management. This patch moves that feature into the core allowing
an easier maintenance and consistent management of quicklists.
Page table pages have the characteristics that they are typically zero or in a
known state when they are freed. This is usually the exactly same state as
needed after allocation. So it makes sense to build a list of freed page
table pages and then consume the pages already in use first. Those pages have
already been initialized correctly (thus no need to zero them) and are likely
already cached in such a way that the MMU can use them most effectively. Page
table pages are used in a sparse way so zeroing them on allocation is not too
useful.
Such an implementation already exits for ia64. Howver, that implementation
did not support constructors and destructors as needed by i386 / x86_64. It
also only supported a single quicklist. The implementation here has
constructor and destructor support as well as the ability for an arch to
specify how many quicklists are needed.
Quicklists are defined by an arch defining CONFIG_QUICKLIST. If more than one
quicklist is necessary then we can define NR_QUICK for additional lists. F.e.
i386 needs two and thus has
config NR_QUICK
int
default 2
If an arch has requested quicklist support then pages can be allocated
from the quicklist (or from the page allocator if the quicklist is
empty) via:
quicklist_alloc(<quicklist-nr>, <gfpflags>, <constructor>)
Page table pages can be freed using:
quicklist_free(<quicklist-nr>, <destructor>, <page>)
Pages must have a definite state after allocation and before
they are freed. If no constructor is specified then pages
will be zeroed on allocation and must be zeroed before they are
freed.
If a constructor is used then the constructor will establish
a definite page state. F.e. the i386 and x86_64 pgd constructors
establish certain mappings.
Constructors and destructors can also be used to track the pages.
i386 and x86_64 use a list of pgds in order to be able to dynamically
update standard mappings.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: "Luck, Tony" <[email protected]>
Cc: William Lee Irwin III <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Add the tool which gets reports about slabs to the VM documentation directory.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Make sure that the check function really only check things and do not perform
activities. Extract the tracing and object seeding out of the two check
functions and place them into slab_alloc and slab_free
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
At kmem_cache_shrink check if we have any empty slabs on the partial
if so then remove them.
Also--as an anti-fragmentation measure--sort the partial slabs so that
the most fully allocated ones come first and the least allocated last.
The next allocations may fill up the nearly full slabs. Having the
least allocated slabs last gives them the maximum chance that their
remaining objects may be freed. Thus we can hopefully minimize the
partial slabs.
I think this is the best one can do in terms antifragmentation
measures. Real defragmentation (meaning moving objects out of slabs with
the least free objects to those that are almost full) can be implemted
by reverse scanning through the list produced here but that would mean
that we need to provide a callback at slab cache creation that allows
the deletion or moving of an object. This will involve slab API
changes, so defer for now.
Cc: Mel Gorman <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patch enables listing the callers who allocated or freed objects in a
cache.
For example to list the allocators for kmalloc-128 do
cat /sys/slab/kmalloc-128/alloc_calls
7 sn_io_slot_fixup+0x40/0x700
7 sn_io_slot_fixup+0x80/0x700
9 sn_bus_fixup+0xe0/0x380
6 param_sysfs_setup+0xf0/0x280
276 percpu_populate+0xf0/0x1a0
19 __register_chrdev_region+0x30/0x360
8 expand_files+0x2e0/0x6e0
1 sys_epoll_create+0x60/0x200
1 __mounts_open+0x140/0x2c0
65 kmem_alloc+0x110/0x280
3 alloc_disk_node+0xe0/0x200
33 as_get_io_context+0x90/0x280
74 kobject_kset_add_dir+0x40/0x140
12 pci_create_bus+0x2a0/0x5c0
1 acpi_ev_create_gpe_block+0x120/0x9e0
41 con_insert_unipair+0x100/0x1c0
1 uart_open+0x1c0/0xba0
1 dma_pool_create+0xe0/0x340
2 neigh_table_init_no_netlink+0x260/0x4c0
6 neigh_parms_alloc+0x30/0x200
1 netlink_kernel_create+0x130/0x320
5 fz_hash_alloc+0x50/0xe0
2 sn_common_hubdev_init+0xd0/0x6e0
28 kernel_param_sysfs_setup+0x30/0x180
72 process_zones+0x70/0x2e0
cat /sys/slab/kmalloc-128/free_calls
558 <not-available>
3 sn_io_slot_fixup+0x600/0x700
84 free_fdtable_rcu+0x120/0x260
2 seq_release+0x40/0x60
6 kmem_free+0x70/0xc0
24 free_as_io_context+0x20/0x200
1 acpi_get_object_info+0x3a0/0x3e0
1 acpi_add_single_object+0xcf0/0x1e40
2 con_release_unimap+0x80/0x140
1 free+0x20/0x40
SLAB_STORE_USER must be enabled for a slab cache by either booting with
"slab_debug" or enabling user tracking specifically for the slab of interest.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We leave a mininum of partial slabs on nodes when we search for
partial slabs on other node. Define a constant for that value.
Then modify slub to keep MIN_PARTIAL slabs around.
This avoids bad situations where a function frees the last object
in a slab (which results in the page being returned to the page
allocator) only to then allocate one again (which requires getting
a page back from the page allocator if the partial list was empty).
Keeping a couple of slabs on the partial list reduces overhead.
Empty slabs are added to the end of the partial list to insure that
partially allocated slabs are consumed first (defragmentation).
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This enables validation of slab. Validation means that all objects are
checked to see if there are redzone violations, if padding has been
overwritten or any pointers have been corrupted. Also checks the consistency
of slab counters.
Validation enables the detection of metadata corruption without the kernel
having to execute code that actually uses (allocs/frees) and object. It
allows one to make sure that the slab metainformation and the guard values
around an object have not been compromised.
A single slabcache can be checked by writing a 1 to the "validate" file.
i.e.
echo 1 >/sys/slab/kmalloc-128/validate
or use the slabinfo tool to check all slabs
slabinfo -v
Error messages will show up in the syslog.
Note that validation can only reach slabs that are on a list. This means that
we are usually restricted to partial slabs and active slabs unless
SLAB_STORE_USER is active which will build a full slab list and allows
validation of slabs that are fully in use. Booting with "slub_debug" set will
enable SLAB_STORE_USER and then full diagnostic are available.
Note that we attempt to push cpu slabs back to the lists when we start the
check. If the cpu slab is reactivated before we get to it (another processor
grabs it before we get to it) then it cannot be checked.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
If slab tracking is on then build a list of full slabs so that we can verify
the integrity of all slabs and are also able to built list of alloc/free
callers.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Object tracking did not work the right way for several call chains. Fix this up
by adding a new parameter to slub_alloc and slub_free that specifies the
caller address explicitly.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The patch adds PageTail(page) and PageHead(page) to check if a page is the
head or the tail of a compound page. This is done by masking the two bits
describing the state of a compound page and then comparing them. So one
comparision and a branch instead of two bit checks and two branches.
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
If we add a new flag so that we can distinguish between the first page and the
tail pages then we can avoid to use page->private in the first page.
page->private == page for the first page, so there is no real information in
there.
Freeing up page->private makes the use of compound pages more transparent.
They become more usable like real pages. Right now we have to be careful f.e.
if we are going beyond PAGE_SIZE allocations in the slab on i386 because we
can then no longer use the private field. This is one of the issues that
cause us not to support debugging for page size slabs in SLAB.
Having page->private available for SLUB would allow more meta information in
the page struct. I can probably avoid the 16 bit ints that I have in there
right now.
Also if page->private is available then a compound page may be equipped with
buffer heads. This may free up the way for filesystems to support larger
blocks than page size.
We add PageTail as an alias of PageReclaim. Compound pages cannot currently
be reclaimed. Because of the alias one needs to check PageCompound first.
The RFC for the this approach was discussed at
http://marc.info/?t=117574302800001&r=1&w=2
[[email protected]: fix hugetlbfs]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Nishanth Aravamudan <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
PowerPC uses the slab allocator to manage the lowest level of the page
table. In high cpu configurations we also use the page struct to split the
page table lock. Disallow the selection of SLUB for that case.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Paul Mackerras <[email protected]>
Acked-by: Benjamin Herrenschmidt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Makes SLUB behave like SLAB in this area to avoid issues....
Throw a stack dump to alert people.
At some point the behavior should be switched back. NULL is no memory as
far as I can tell and if the use asked for 0 bytes then he need to get no
memory.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Structures may contain u64 items on 32 bit platforms that are only able to
address 64 bit items on 64 bit boundaries. Change the mininum alignment of
slabs to conform to those expectations.
ARCH_KMALLOC_MINALIGN must be changed for good since a variety of structure
are mixed in the general slabs.
ARCH_SLAB_MINALIGN is changed because currently there is no consistent
specification of object alignment. We may have that in the future when the
KMEM_CACHE and related macros are used to generate slabs. These pass the
alignment of the structure generated by the compiler to the slab.
With KMEM_CACHE etc we could align structures that do not contain 64
bit values to 32 bit boundaries potentially saving some memory.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This is a new slab allocator which was motivated by the complexity of the
existing code in mm/slab.c. It attempts to address a variety of concerns
with the existing implementation.
A. Management of object queues
A particular concern was the complex management of the numerous object
queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for
each allocating CPU and use objects from a slab directly instead of
queueing them up.
B. Storage overhead of object queues
SLAB Object queues exist per node, per CPU. The alien cache queue even
has a queue array that contain a queue for each processor on each
node. For very large systems the number of queues and the number of
objects that may be caught in those queues grows exponentially. On our
systems with 1k nodes / processors we have several gigabytes just tied up
for storing references to objects for those queues This does not include
the objects that could be on those queues. One fears that the whole
memory of the machine could one day be consumed by those queues.
C. SLAB meta data overhead
SLAB has overhead at the beginning of each slab. This means that data
cannot be naturally aligned at the beginning of a slab block. SLUB keeps
all meta data in the corresponding page_struct. Objects can be naturally
aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte
boundaries and can fit tightly into a 4k page with no bytes left over.
SLAB cannot do this.
D. SLAB has a complex cache reaper
SLUB does not need a cache reaper for UP systems. On SMP systems
the per CPU slab may be pushed back into partial list but that
operation is simple and does not require an iteration over a list
of objects. SLAB expires per CPU, shared and alien object queues
during cache reaping which may cause strange hold offs.
E. SLAB has complex NUMA policy layer support
SLUB pushes NUMA policy handling into the page allocator. This means that
allocation is coarser (SLUB does interleave on a page level) but that
situation was also present before 2.6.13. SLABs application of
policies to individual slab objects allocated in SLAB is
certainly a performance concern due to the frequent references to
memory policies which may lead a sequence of objects to come from
one node after another. SLUB will get a slab full of objects
from one node and then will switch to the next.
F. Reduction of the size of partial slab lists
SLAB has per node partial lists. This means that over time a large
number of partial slabs may accumulate on those lists. These can
only be reused if allocator occur on specific nodes. SLUB has a global
pool of partial slabs and will consume slabs from that pool to
decrease fragmentation.
G. Tunables
SLAB has sophisticated tuning abilities for each slab cache. One can
manipulate the queue sizes in detail. However, filling the queues still
requires the uses of the spin lock to check out slabs. SLUB has a global
parameter (min_slab_order) for tuning. Increasing the minimum slab
order can decrease the locking overhead. The bigger the slab order the
less motions of pages between per CPU and partial lists occur and the
better SLUB will be scaling.
G. Slab merging
We often have slab caches with similar parameters. SLUB detects those
on boot up and merges them into the corresponding general caches. This
leads to more effective memory use. About 50% of all caches can
be eliminated through slab merging. This will also decrease
slab fragmentation because partial allocated slabs can be filled
up again. Slab merging can be switched off by specifying
slub_nomerge on boot up.
Note that merging can expose heretofore unknown bugs in the kernel
because corrupted objects may now be placed differently and corrupt
differing neighboring objects. Enable sanity checks to find those.
H. Diagnostics
The current slab diagnostics are difficult to use and require a
recompilation of the kernel. SLUB contains debugging code that
is always available (but is kept out of the hot code paths).
SLUB diagnostics can be enabled via the "slab_debug" option.
Parameters can be specified to select a single or a group of
slab caches for diagnostics. This means that the system is running
with the usual performance and it is much more likely that
race conditions can be reproduced.
I. Resiliency
If basic sanity checks are on then SLUB is capable of detecting
common error conditions and recover as best as possible to allow the
system to continue.
J. Tracing
Tracing can be enabled via the slab_debug=T,<slabcache> option
during boot. SLUB will then protocol all actions on that slabcache
and dump the object contents on free.
K. On demand DMA cache creation.
Generally DMA caches are not needed. If a kmalloc is used with
__GFP_DMA then just create this single slabcache that is needed.
For systems that have no ZONE_DMA requirement the support is
completely eliminated.
L. Performance increase
Some benchmarks have shown speed improvements on kernbench in the
range of 5-10%. The locking overhead of slub is based on the
underlying base allocation size. If we can reliably allocate
larger order pages then it is possible to increase slub
performance much further. The anti-fragmentation patches may
enable further performance increases.
Tested on:
i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator
SLUB Boot options
slub_nomerge Disable merging of slabs
slub_min_order=x Require a minimum order for slab caches. This
increases the managed chunk size and therefore
reduces meta data and locking overhead.
slub_min_objects=x Mininum objects per slab. Default is 8.
slub_max_order=x Avoid generating slabs larger than order specified.
slub_debug Enable all diagnostics for all caches
slub_debug=<options> Enable selective options for all caches
slub_debug=<o>,<cache> Enable selective options for a certain set of
caches
Available Debug options
F Double Free checking, sanity and resiliency
R Red zoning
P Object / padding poisoning
U Track last free / alloc
T Trace all allocs / frees (only use for individual slabs).
To use SLUB: Apply this patch and then select SLUB as the default slab
allocator.
[[email protected]: fix an oops-causing locking error]
[[email protected]: various stupid cleanups and small fixes]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Hugh Dickins <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
i386 uses kmalloc to allocate the threadinfo structure assuming that the
allocations result in a page sized aligned allocation. That has worked so
far because SLAB exempts page sized slabs from debugging and aligns them in
special ways that goes beyond the restrictions imposed by
KMALLOC_ARCH_MINALIGN valid for other slabs in the kmalloc array.
SLUB also works fine without debugging since page sized allocations neatly
align at page boundaries. However, if debugging is switched on then SLUB
will extend the slab with debug information. The resulting slab is not
longer of page size. It will only be aligned following the requirements
imposed by KMALLOC_ARCH_MINALIGN. As a result the threadinfo structure may
not be page aligned which makes i386 fail to boot with SLUB debug on.
Replace the calls to kmalloc with calls into the page allocator.
An alternate solution may be to create a custom slab cache where the
alignment is set to PAGE_SIZE. That would allow slub debugging to be
applied to the threadinfo structure.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: William Lee Irwin III <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We can use the global ZVC counters to establish the exact size of the LRU
and the free pages. This allows a more accurate determination of the dirty
ratio.
This patch will fix the broken ratio calculations if large amounts of
memory are allocated to huge pags or other consumers that do not put the
pages on to the LRU.
Notes:
- I did not add NR_SLAB_RECLAIMABLE to the calculation of the
dirtyable pages. Those may be reclaimable but they are at this
point not dirtyable. If NR_SLAB_RECLAIMABLE would be considered
then a huge number of reclaimable pages would stop writeback
from occurring.
- This patch used to be in mm as the last one in a series of patches.
It was removed when Linus updated the treatment of highmem because
there was a conflict. I updated the patch to follow Linus' approach.
This patch is neede to fulfill the claims made in the beginning of the
patchset that is now in Linus' tree.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
x86_64 currently simulates a list using the index and private fields of the
page struct. Seems that the code was inherited from i386. But x86_64 does
not use the slab to allocate pgds and pmds etc. So the lru field is not
used by the slab and therefore available.
This patch uses standard list operations on page->lru to realize pgd
tracking.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
NR_FILE_PAGES must be accounted for depending on the zone that the page
belongs to. If we replace the page in the radix tree then we may have to
shift the count to another zone.
Suggested-by: Ethan Solomita <[email protected]>
Eventually-typed-in-by: Christoph Lameter <[email protected]>
Cc: Martin Bligh <[email protected]>
Cc: <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The alien cache is a per cpu per node array allocated for every slab on the
system. Currently we size this array for all nodes that the kernel does
support. For IA64 this is 1024 nodes. So we allocate an array with 1024
objects even if we only boot a system with 4 nodes.
This patch uses "nr_node_ids" to determine the number of possible nodes
supported by a hardware configuration and only allocates an alien cache
sized for possible nodes.
The initialization of nr_node_ids occurred too late relative to the bootstrap
of the slab allocator and so I moved the setup_nr_node_ids() into
free_area_init_nodes().
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We frequently need the maximum number of possible processors in order to
allocate arrays for all processors. So far this was done using
highest_possible_processor_id(). However, we do need the number of
processors not the highest id. Moreover the number was so far dynamically
calculated on each invokation. The number of possible processors does not
change when the system is running. We can therefore calculate that number
once.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Frederik Deweerdt <[email protected]>
Cc: Neil Brown <[email protected]>
Cc: Trond Myklebust <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
highest_possible_node_id() is currently used to calculate the last possible
node idso that the network subsystem can figure out how to size per node
arrays.
I think having the ability to determine the maximum amount of nodes in a
system at runtime is useful but then we should name this entry
correspondingly, it should return the number of node_ids, and the the value
needs to be setup only once on bootup. The node_possible_map does not
change after bootup.
This patch introduces nr_node_ids and replaces the use of
highest_possible_node_id(). nr_node_ids is calculated on bootup when the
page allocators pagesets are initialized.
[[email protected]: fix oops]
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Neil Brown <[email protected]>
Cc: Trond Myklebust <[email protected]>
Signed-off-by: Frederik Deweerdt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
As Andi pointed out: CONFIG_GENERIC_ISA_DMA only disables the ISA DMA
channel management. Other functionality may still expect GFP_DMA to
provide memory below 16M. So we need to make sure that CONFIG_ZONE_DMA is
set independent of CONFIG_GENERIC_ISA_DMA. Undo the modifications to
mm/Kconfig where we made ZONE_DMA dependent on GENERIC_ISA_DMA and set
theses explicitly in each arches Kconfig.
Reviews must occur for each arch in order to determine if ZONE_DMA can be
switched off. It can only be switched off if we know that all devices
supported by a platform are capable of performing DMA transfers to all of
memory (Some arches already support this: uml, avr32, sh sh64, parisc and
IA64/Altix).
In order to switch ZONE_DMA off conditionally, one would have to establish
a scheme by which one can assure that no drivers are enabled that are only
capable of doing I/O to a part of memory, or one needs to provide an
alternate means of performing an allocation from a specific range of memory
(like provided by alloc_pages_range()) and insure that all drivers use that
call. In that case the arches alloc_dma_coherent() may need to be modified
to call alloc_pages_range() instead of relying on GFP_DMA.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
ZONE_DMA less operation for IA64 SGI platform
Disable ZONE_DMA for SGI SN2. All memory is addressable by all devices and we
do not need any special memory pool.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: "Luck, Tony" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Make ZONE_DMA optional in core code.
- ifdef all code for ZONE_DMA and related definitions following the example
for ZONE_DMA32 and ZONE_HIGHMEM.
- Without ZONE_DMA, ZONE_HIGHMEM and ZONE_DMA32 we get to a ZONES_SHIFT of
0.
- Modify the VM statistics to work correctly without a DMA zone.
- Modify slab to not create DMA slabs if there is no ZONE_DMA.
[[email protected]: cleanup]
[[email protected]: build fix]
[[email protected]: Simplify calculation of the number of bits we need for ZONES_SHIFT]
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: "Luck, Tony" <[email protected]>
Cc: Kyle McMartin <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: James Bottomley <[email protected]>
Cc: Paul Mundt <[email protected]>
Signed-off-by: Andy Whitcroft <[email protected]>
Signed-off-by: Jeff Dike <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patch simply defines CONFIG_ZONE_DMA for all arches. We later do special
things with CONFIG_ZONE_DMA after the VM and an arch are prepared to work
without ZONE_DMA.
CONFIG_ZONE_DMA can be defined in two ways depending on how an architecture
handles ISA DMA.
First if CONFIG_GENERIC_ISA_DMA is set by the arch then we know that the arch
needs ZONE_DMA because ISA DMA devices are supported. We can catch this in
mm/Kconfig and do not need to modify arch code.
Second, arches may use ZONE_DMA in an unknown way. We set CONFIG_ZONE_DMA for
all arches that do not set CONFIG_GENERIC_ISA_DMA in order to insure backwards
compatibility. The arches may later undefine ZONE_DMA if their arch code has
been verified to not depend on ZONE_DMA.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: "Luck, Tony" <[email protected]>
Cc: Kyle McMartin <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: James Bottomley <[email protected]>
Cc: Paul Mundt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patchset follows up on the earlier work in Andrew's tree to reduce the
number of zones. The patches allow to go to a minimum of 2 zones. This one
allows also to make ZONE_DMA optional and therefore the number of zones can be
reduced to one.
ZONE_DMA is usually used for ISA DMA devices. There are a number of reasons
why we would not want to have ZONE_DMA
1. Some arches do not need ZONE_DMA at all.
2. With the advent of IOMMUs DMA zones are no longer needed.
The necessity of DMA zones may drastically be reduced
in the future. This patchset allows a compilation of
a kernel without that overhead.
3. Devices that require ISA DMA get rare these days. All
my systems do not have any need for ISA DMA.
4. The presence of an additional zone unecessarily complicates
VM operations because it must be scanned and balancing
logic must operate on its.
5. With only ZONE_NORMAL one can reach the situation where
we have only one zone. This will allow the unrolling of many
loops in the VM and allows the optimization of varous
code paths in the VM.
6. Having only a single zone in a NUMA system results in a
1-1 correspondence between nodes and zones. Various additional
optimizations to critical VM paths become possible.
Many systems today can operate just fine with a single zone. If you look at
what is in ZONE_DMA then one usually sees that nothing uses it. The DMA slabs
are empty (Some arches use ZONE_DMA instead of ZONE_NORMAL, then ZONE_NORMAL
will be empty instead).
On all of my systems (i386, x86_64, ia64) ZONE_DMA is completely empty. Why
constantly look at an empty zone in /proc/zoneinfo and empty slab in
/proc/slabinfo? Non i386 also frequently have no need for ZONE_DMA and zones
stay empty.
The patchset was tested on i386 (UP / SMP), x86_64 (UP, NUMA) and ia64 (NUMA).
The RFC posted earlier (see
http://marc.theaimsgroup.com/?l=linux-kernel&m=115231723513008&w=2) had lots
of #ifdefs in them. An effort has been made to minize the number of #ifdefs
and make this as compact as possible. The job was made much easier by the
ongoing efforts of others to extract common arch specific functionality.
I have been running this for awhile now on my desktop and finally Linux is
using all my available RAM instead of leaving the 16MB in ZONE_DMA untouched:
christoph@pentium940:~$ cat /proc/zoneinfo
Node 0, zone Normal
pages free 4435
min 1448
low 1810
high 2172
active 241786
inactive 210170
scanned 0 (a: 0 i: 0)
spanned 524224
present 524224
nr_anon_pages 61680
nr_mapped 14271
nr_file_pages 390264
nr_slab_reclaimable 27564
nr_slab_unreclaimable 1793
nr_page_table_pages 449
nr_dirty 39
nr_writeback 0
nr_unstable 0
nr_bounce 0
cpu: 0 pcp: 0
count: 156
high: 186
batch: 31
cpu: 0 pcp: 1
count: 9
high: 62
batch: 15
vm stats threshold: 20
cpu: 1 pcp: 0
count: 177
high: 186
batch: 31
cpu: 1 pcp: 1
count: 12
high: 62
batch: 15
vm stats threshold: 20
all_unreclaimable: 0
prev_priority: 12
temp_priority: 12
start_pfn: 0
This patch:
In two places in the VM we use ZONE_DMA to refer to the first zone. If
ZONE_DMA is optional then other zones may be first. So simply replace
ZONE_DMA with zone 0.
This also fixes ZONETABLE_PGSHIFT. If we have only a single zone then
ZONES_PGSHIFT may become 0 because there is no need anymore to encode the zone
number related to a pgdat. However, we still need a zonetable to index all
the zones for each node if this is a NUMA system. Therefore define
ZONETABLE_SHIFT unconditionally as the offset of the ZONE field in page flags.
[[email protected]: fix mismerge]
Acked-by: Christoph Hellwig <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: "Luck, Tony" <[email protected]>
Cc: Kyle McMartin <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: James Bottomley <[email protected]>
Cc: Paul Mundt <[email protected]>
Signed-off-by: Andy Whitcroft <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Values are readily available via ZVC per node and global sums.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Function is unnecessary now. We can use the summing features of the ZVCs to
get the values we need.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
nr_free_pages is now a simple access to a global variable. Make it a macro
instead of a function.
The nr_free_pages now requires vmstat.h to be included. There is one
occurrence in power management where we need to add the include. Directly
refrer to global_page_state() there to clarify why the #include was added.
[[email protected]: arm build fix]
[[email protected]: sparc64 build fix]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The global and per zone counter sums are in arrays of longs. Reorder the ZVCs
so that the most frequently used ZVCs are put into the same cacheline. That
way calculations of the global, node and per zone vm state touches only a
single cacheline. This is mostly important for 64 bit systems were one 128
byte cacheline takes only 8 longs.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This is again simplifies some of the VM counter calculations through the use
of the ZVC consolidated counters.
[[email protected]: build fix]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Michal Piotrowski <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The determination of the dirty ratio to determine writeback behavior is
currently based on the number of total pages on the system.
However, not all pages in the system may be dirtied. Thus the ratio is always
too low and can never reach 100%. The ratio may be particularly skewed if
large hugepage allocations, slab allocations or device driver buffers make
large sections of memory not available anymore. In that case we may get into
a situation in which f.e. the background writeback ratio of 40% cannot be
reached anymore which leads to undesired writeback behavior.
This patchset fixes that issue by determining the ratio based on the actual
pages that may potentially be dirty. These are the pages on the active and
the inactive list plus free pages.
The problem with those counts has so far been that it is expensive to
calculate these because counts from multiple nodes and multiple zones will
have to be summed up. This patchset makes these counters ZVC counters. This
means that a current sum per zone, per node and for the whole system is always
available via global variables and not expensive anymore to calculate.
The patchset results in some other good side effects:
- Removal of the various functions that sum up free, active and inactive
page counts
- Cleanup of the functions that display information via the proc filesystem.
This patch:
The use of a ZVC for nr_inactive and nr_active allows a simplification of some
counter operations. More ZVC functionality is used for sums etc in the
following patches.
[[email protected]: UP build fix]
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Use the pointer passed to cache_reap to determine the work pointer and
consolidate exit paths.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Currently one can specify an arbitrary node mask to mbind that includes
nodes not allowed. If that is done with an interleave policy then we will
go around all the nodes. Those outside of the currently allowed cpuset
will be redirected to the border nodes. Interleave will then create
imbalances at the borders of the cpuset.
This patch restricts the nodes to the currently allowed cpuset.
The RFC for this patch was discussed at
http://marc.theaimsgroup.com/?t=116793842100004&r=1&w=2
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Paul Jackson <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Both process_zones() and drain_node_pages() check for populated zones
before touching pagesets. However, __drain_pages does not do so,
This may result in a NULL pointer dereference for pagesets in unpopulated
zones if a NUMA setup is combined with cpu hotplug.
Initially the unpopulated zone has the pcp pointers pointing to the boot
pagesets. Since the zone is not populated the boot pageset pointers will
not be changed during page allocator and slab bootstrap.
If a cpu is later brought down (first call to __drain_pages()) then the pcp
pointers for cpus in unpopulated zones are set to NULL since __drain_pages
does not first check for an unpopulated zone.
If the cpu is then brought up again then we call process_zones() which will
ignore the unpopulated zone. So the pageset pointers will still be NULL.
If the cpu is then again brought down then __drain_pages will attempt to
drain pages by following the NULL pageset pointer for unpopulated zones.
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The declaration of kmem_ptr_validate in slab.h does not match the
one in slab.c. Remove the fastcall attribute (this is the only use in
slab.c).
Signed-off-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
More cleanups for slab.h
1. Remove tabs from weird locations as suggested by Pekka
2. Drop the check for NUMA and SLAB_DEBUG from the fallback section
as suggested by Pekka.
3. Uses static inline for the fallback defs as also suggested by Pekka.
4. Make kmem_ptr_valid take a const * argument.
5. Separate the NUMA fallback definitions from the kmalloc_track fallback
definitions.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Pekka Enberg <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This is a response to an earlier discussion on linux-mm about splitting
slab.h components per allocator. Patch is against 2.6.19-git11. See
http://marc.theaimsgroup.com/?l=linux-mm&m=116469577431008&w=2
This patch cleans up the slab header definitions. We define the common
functions of slob and slab in slab.h and put the extra definitions needed
for slab's kmalloc implementations in <linux/slab_def.h>. In order to get
a greater set of common functions we add several empty functions to slob.c
and also rename slob's kmalloc to __kmalloc.
Slob does not need any special definitions since we introduce a fallback
case. If there is no need for a slab implementation to provide its own
kmalloc mess^H^H^Hacros then we simply fall back to __kmalloc functions.
That is sufficient for SLOB.
Sort the function in slab.h according to their functionality. First the
functions operating on struct kmem_cache * then the kmalloc related
functions followed by special debug and fallback definitions.
Also redo a lot of comments.
Signed-off-by: Christoph Lameter <[email protected]>?
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Fallback_alloc() does not do the check for GFP_WAIT as done in
cache_grow(). Thus interrupts are disabled when we call kmem_getpages()
which results in the failure.
Duplicate the handling of GFP_WAIT in cache_grow().
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Jay Cliburn <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Trigger softirq less frequently
We trigger the softirq before this patch using offset of sd->interval.
However, if the queue is busy then it is sufficient to schedule the softirq
with sd->interval * busy_factor.
So we modify the calculation of the next time to balance by taking
the interval added to last_balance again. This is only the
right value if the idle/busy situation continues as is.
There are two potential trouble spots:
- If the queue was idle and now gets busy then we call rebalance
early. However, that is not a problem because we will then use
the longer interval for the next period.
- If the queue was busy and becomes idle then we potentially
wait too long before rebalancing. However, when the task
goes idle then idle_balance is called. We add another calculation
of the next balance time based on sd->interval in idle_balance
so that we will rebalance soon.
V2->V3:
- Calculate rebalance time based on current jiffies and not
based on the jiffies at the last time we load balanced.
We no longer rely on staggering and therefore we can
affort to do this now.
V3->V4:
- Use functions to do jiffy comparisons.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Peter Williams <[email protected]>
Cc: Nick Piggin <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: "Siddha, Suresh B" <[email protected]>
Cc: "Chen, Kenneth W" <[email protected]>
Acked-by: Ingo Molnar <[email protected]>
Cc: KAMEZAWA Hiroyuki <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Call rebalance_tick (renamed to run_rebalance_domains) from a newly introduced
softirq.
We calculate the earliest time for each layer of sched domains to be rescanned
(this is the rescan time for idle) and use the earliest of those to schedule
the softirq via a new field "next_balance" added to struct rq.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Peter Williams <[email protected]>
Cc: Nick Piggin <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: "Siddha, Suresh B" <[email protected]>
Cc: "Chen, Kenneth W" <[email protected]>
Acked-by: Ingo Molnar <[email protected]>
Cc: KAMEZAWA Hiroyuki <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Perform the idle state determination in rebalance_tick.
If we separate balancing from sched_tick then we also need to determine the
idle state in rebalance_tick.
V2->V3
Remove useless idlle != 0 check. Checking nr_running seems
to be sufficient. Thanks Suresh.
Signed-off-by: Christoph Lameter <[email protected]>
Cc: Peter Williams <[email protected]>
Cc: Nick Piggin <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: "Siddha, Suresh B" <[email protected]>
Cc: "Chen, Kenneth W" <[email protected]>
Acked-by: Ingo Molnar <[email protected]>
Cc: KAMEZAWA Hiroyuki <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Christoph Lameter