percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Currently, the following three different ways to define percpu arrays
are in use.
1. DEFINE_PER_CPU(elem_type[array_len], array_name);
2. DEFINE_PER_CPU(elem_type, array_name[array_len]);
3. DEFINE_PER_CPU(elem_type, array_name)[array_len];
Unify to #1 which correctly separates the roles of the two parameters
and thus allows more flexibility in the way percpu variables are
defined.
[ Impact: cleanup ]
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: linux-mm@kvack.org
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Impact: cleanup
node_to_cpumask (and the blecherous node_to_cpumask_ptr which
contained a declaration) are replaced now everyone implements
cpumask_of_node.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Quicklists store pages for each CPU as caches. (Each CPU can cache
node_free_pages/16 pages)
It is used for page table cache. exit() will increase the cache size,
while fork() consumes it.
So for example if an apache-style application runs (one parent and many
child model), one CPU process will fork() while another CPU will process
the middleware work and exit().
At that time, the CPU on which the parent runs doesn't have page table
cache at all. Others (on which children runs) have maximum caches.
QList_max = (#ofCPUs - 1) x Free / 16
=> QList_max / (Free + QList_max) = (#ofCPUs - 1) / (16 + #ofCPUs - 1)
So, How much quicklist memory is used in the maximum case?
This is proposional to # of CPUs because the limit of per cpu quicklist
cache doesn't see the number of cpus.
Above calculation mean
Number of CPUs per node 2 4 8 16
============================== ====================
QList_max / (Free + QList_max) 5.8% 16% 30% 48%
Wow! Quicklist can spend about 50% memory at worst case.
My demonstration program is here
--------------------------------------------------------------------------------
#define _GNU_SOURCE
#include <stdio.h>
#include <errno.h>
#include <stdlib.h>
#include <string.h>
#include <sched.h>
#include <unistd.h>
#include <sys/mman.h>
#include <sys/wait.h>
#define BUFFSIZE 512
int max_cpu(void) /* get max number of logical cpus from /proc/cpuinfo */
{
FILE *fd;
char *ret, buffer[BUFFSIZE];
int cpu = 1;
fd = fopen("/proc/cpuinfo", "r");
if (fd == NULL) {
perror("fopen(/proc/cpuinfo)");
exit(EXIT_FAILURE);
}
while (1) {
ret = fgets(buffer, BUFFSIZE, fd);
if (ret == NULL)
break;
if (!strncmp(buffer, "processor", 9))
cpu = atoi(strchr(buffer, ':') + 2);
}
fclose(fd);
return cpu;
}
void cpu_bind(int cpu) /* bind current process to one cpu */
{
cpu_set_t mask;
int ret;
CPU_ZERO(&mask);
CPU_SET(cpu, &mask);
ret = sched_setaffinity(0, sizeof(mask), &mask);
if (ret == -1) {
perror("sched_setaffinity()");
exit(EXIT_FAILURE);
}
sched_yield(); /* not necessary */
}
#define MMAP_SIZE (10 * 1024 * 1024) /* 10 MB */
#define FORK_INTERVAL 1 /* 1 second */
main(int argc, char *argv[])
{
int cpu_max, nextcpu;
long pagesize;
pid_t pid;
/* set max number of logical cpu */
if (argc > 1)
cpu_max = atoi(argv[1]) - 1;
else
cpu_max = max_cpu();
/* get the page size */
pagesize = sysconf(_SC_PAGESIZE);
if (pagesize == -1) {
perror("sysconf(_SC_PAGESIZE)");
exit(EXIT_FAILURE);
}
/* prepare parent process */
cpu_bind(0);
nextcpu = cpu_max;
loop:
/* select destination cpu for child process by round-robin rule */
if (++nextcpu > cpu_max)
nextcpu = 1;
pid = fork();
if (pid == 0) { /* child action */
char *p;
int i;
/* consume page tables */
p = mmap(0, MMAP_SIZE, PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
i = MMAP_SIZE / pagesize;
while (i-- > 0) {
*p = 1;
p += pagesize;
}
/* move to other cpu */
cpu_bind(nextcpu);
/*
printf("a child moved to cpu%d after mmap().\n", nextcpu);
fflush(stdout);
*/
/* back page tables to pgtable_quicklist */
exit(0);
} else if (pid > 0) { /* parent action */
sleep(FORK_INTERVAL);
waitpid(pid, NULL, WNOHANG);
}
goto loop;
}
----------------------------------------
When above program which does task migration runs, my 8GB box spends
800MB of memory for quicklist. This is not memory leak but doesn't seem
good.
% cat /proc/meminfo
MemTotal: 7701568 kB
MemFree: 4724672 kB
(snip)
Quicklists: 844800 kB
because
- My machine spec is
number of numa node: 2
number of cpus: 8 (4CPU x2 node)
total mem: 8GB (4GB x2 node)
free mem: about 5GB
- Then, 4.7GB x 16% ~= 880MB.
So, Quicklist can use 800MB.
So, if following spec machine run that program
CPUs: 64 (8cpu x 8node)
Mem: 1TB (128GB x8node)
Then, quicklist can waste 300GB (= 1TB x 30%). It is too large.
So, I don't like cache policies which is proportional to # of cpus.
My patch changes the number of caches
from:
per-cpu-cache-amount = memory_on_node / 16
to
per-cpu-cache-amount = memory_on_node / 16 / number_of_cpus_on_node.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Keiichiro Tokunaga <tokunaga.keiich@jp.fujitsu.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Tested-by: David Miller <davem@davemloft.net>
Acked-by: Mike Travis <travis@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Quicklists calculates the size of the quicklists based on the number of
free pages. This must be the number of free pages that can be allocated
with GFP_KERNEL. node_page_state() includes the pages in ZONE_HIGHMEM and
ZONE_MOVABLE which may lead the quicklists to become too large causing OOM.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
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>
Tejun Heo
Rusty Russell
KOSAKI Motohiro
Christoph Lameter