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 <clameter@sgi.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Andi Kleen <ak@suse.de>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: William Lee Irwin III <wli@holomorphy.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
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.
[hugh@veritas.com: fix an oops-causing locking error]
[akpm@linux-foundation.org: various stupid cleanups and small fixes]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Separate out the concept of "queue congestion" from "backing-dev congestion".
Congestion is a backing-dev concept, not a queue concept.
The blk_* congestion functions are retained, as wrappers around the core
backing-dev congestion functions.
This proper layering is needed so that NFS can cleanly use the congestion
functions, and so that CONFIG_BLOCK=n actually links.
Cc: "Thomas Maier" <balagi@justmail.de>
Cc: "Jens Axboe" <jens.axboe@oracle.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: David Howells <dhowells@redhat.com>
Cc: Peter Osterlund <petero2@telia.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Make it possible to disable the block layer. Not all embedded devices require
it, some can make do with just JFFS2, NFS, ramfs, etc - none of which require
the block layer to be present.
This patch does the following:
(*) Introduces CONFIG_BLOCK to disable the block layer, buffering and blockdev
support.
(*) Adds dependencies on CONFIG_BLOCK to any configuration item that controls
an item that uses the block layer. This includes:
(*) Block I/O tracing.
(*) Disk partition code.
(*) All filesystems that are block based, eg: Ext3, ReiserFS, ISOFS.
(*) The SCSI layer. As far as I can tell, even SCSI chardevs use the
block layer to do scheduling. Some drivers that use SCSI facilities -
such as USB storage - end up disabled indirectly from this.
(*) Various block-based device drivers, such as IDE and the old CDROM
drivers.
(*) MTD blockdev handling and FTL.
(*) JFFS - which uses set_bdev_super(), something it could avoid doing by
taking a leaf out of JFFS2's book.
(*) Makes most of the contents of linux/blkdev.h, linux/buffer_head.h and
linux/elevator.h contingent on CONFIG_BLOCK being set. sector_div() is,
however, still used in places, and so is still available.
(*) Also made contingent are the contents of linux/mpage.h, linux/genhd.h and
parts of linux/fs.h.
(*) Makes a number of files in fs/ contingent on CONFIG_BLOCK.
(*) Makes mm/bounce.c (bounce buffering) contingent on CONFIG_BLOCK.
(*) set_page_dirty() doesn't call __set_page_dirty_buffers() if CONFIG_BLOCK
is not enabled.
(*) fs/no-block.c is created to hold out-of-line stubs and things that are
required when CONFIG_BLOCK is not set:
(*) Default blockdev file operations (to give error ENODEV on opening).
(*) Makes some /proc changes:
(*) /proc/devices does not list any blockdevs.
(*) /proc/diskstats and /proc/partitions are contingent on CONFIG_BLOCK.
(*) Makes some compat ioctl handling contingent on CONFIG_BLOCK.
(*) If CONFIG_BLOCK is not defined, makes sys_quotactl() return -ENODEV if
given command other than Q_SYNC or if a special device is specified.
(*) In init/do_mounts.c, no reference is made to the blockdev routines if
CONFIG_BLOCK is not defined. This does not prohibit NFS roots or JFFS2.
(*) The bdflush, ioprio_set and ioprio_get syscalls can now be absent (return
error ENOSYS by way of cond_syscall if so).
(*) The seclvl_bd_claim() and seclvl_bd_release() security calls do nothing if
CONFIG_BLOCK is not set, since they can't then happen.
Signed-Off-By: David Howells <dhowells@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the bounce buffer code from mm/highmem.c to mm/bounce.c so that it can be
more easily disabled when the block layer is disabled.
!!!NOTE!!! There may be a bug in this code: Should init_emergency_pool() be
contingent on CONFIG_HIGHMEM?
Signed-Off-By: David Howells <dhowells@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The allocpercpu functions __alloc_percpu and __free_percpu() are heavily
using the slab allocator. However, they are conceptually slab. This also
simplifies SLOB (at this point slob may be broken in mm. This should fix
it).
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Matt Mackall <mpm@selenic.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
NOTE: ZVC are *not* the lightweight event counters. ZVCs are reliable whereas
event counters do not need to be.
Zone based VM statistics are necessary to be able to determine what the state
of memory in one zone is. In a NUMA system this can be helpful for local
reclaim and other memory optimizations that may be able to shift VM load in
order to get more balanced memory use.
It is also useful to know how the computing load affects the memory
allocations on various zones. This patchset allows the retrieval of that data
from userspace.
The patchset introduces a framework for counters that is a cross between the
existing page_stats --which are simply global counters split per cpu-- and the
approach of deferred incremental updates implemented for nr_pagecache.
Small per cpu 8 bit counters are added to struct zone. If the counter exceeds
certain thresholds then the counters are accumulated in an array of
atomic_long in the zone and in a global array that sums up all zone values.
The small 8 bit counters are next to the per cpu page pointers and so they
will be in high in the cpu cache when pages are allocated and freed.
Access to VM counter information for a zone and for the whole machine is then
possible by simply indexing an array (Thanks to Nick Piggin for pointing out
that approach). The access to the total number of pages of various types does
no longer require the summing up of all per cpu counters.
Benefits of this patchset right now:
- Ability for UP and SMP configuration to determine how memory
is balanced between the DMA, NORMAL and HIGHMEM zones.
- loops over all processors are avoided in writeback and
reclaim paths. We can avoid caching the writeback information
because the needed information is directly accessible.
- Special handling for nr_pagecache removed.
- zone_reclaim_interval vanishes since VM stats can now determine
when it is worth to do local reclaim.
- Fast inline per node page state determination.
- Accurate counters in /sys/devices/system/node/node*/meminfo. Current
counters are counting simply which processor allocated a page somewhere
and guestimate based on that. So the counters were not useful to show
the actual distribution of page use on a specific zone.
- The swap_prefetch patch requires per node statistics in order to
figure out when processors of a node can prefetch. This patch provides
some of the needed numbers.
- Detailed VM counters available in more /proc and /sys status files.
References to earlier discussions:
V1 http://marc.theaimsgroup.com/?l=linux-kernel&m=113511649910826&w=2
V2 http://marc.theaimsgroup.com/?l=linux-kernel&m=114980851924230&w=2
V3 http://marc.theaimsgroup.com/?l=linux-kernel&m=115014697910351&w=2
V4 http://marc.theaimsgroup.com/?l=linux-kernel&m=115024767318740&w=2
Performance tests with AIM7 did not show any regressions. Seems to be a tad
faster even. Tested on ia64/NUMA. Builds fine on i386, SMP / UP. Includes
fixes for s390/arm/uml arch code.
This patch:
Move counter code from page_alloc.c/page-flags.h to vmstat.c/h.
Create vmstat.c/vmstat.h by separating the counter code and the proc
functions.
Move the vm_stat_text array before zoneinfo_show.
[akpm@osdl.org: s390 build fix]
[akpm@osdl.org: HOTPLUG_CPU build fix]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Helper functions for for_each_online_pgdat/for_each_zone look too big to be
inlined. Speed of these helper macro itself is not very important. (inner
loops are tend to do more work than this)
This patch make helper function to be out-of-lined.
inline out-of-line
.text 005c0680 005bf6a0
005c0680 - 005bf6a0 = FE0 = 4Kbytes.
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Centralize the page migration functions in anticipation of additional
tinkering. Creates a new file mm/migrate.c
1. Extract buffer_migrate_page() from fs/buffer.c
2. Extract central migration code from vmscan.c
3. Extract some components from mempolicy.c
4. Export pageout() and remove_from_swap() from vmscan.c
5. Make it possible to configure NUMA systems without page migration
and non-NUMA systems with page migration.
I had to so some #ifdeffing in mempolicy.c that may need a cleanup.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add mm/util.c for functions common between SLAB and SLOB.
Signed-off-by: Matt Mackall <mpm@selenic.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This adds generic memory add/remove and supporting functions for memory
hotplug into a new file as well as a memory hotplug kernel config option.
Individual architecture patches will follow.
For now, disable memory hotplug when swsusp is enabled. There's a lot of
churn there right now. We'll fix it up properly once it calms down.
Signed-off-by: Matt Tolentino <matthew.e.tolentino@intel.com>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
- generic_file* file operations do no longer have a xip/non-xip split
- filemap_xip.c implements a new set of fops that require get_xip_page
aop to work proper. all new fops are exported GPL-only (don't like to
see whatever code use those except GPL modules)
- __xip_unmap now uses page_check_address, which is no longer static
in rmap.c, and defined in linux/rmap.h
- mm/filemap.h is now much more clean, plainly having just Linus'
inline funcs moved here from filemap.c
- fix includes in filemap_xip to make it build cleanly on i386
Signed-off-by: Carsten Otte <cotte@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
