Currently hugepage migration is available only for soft offlining, but
it's also useful for some other users of page migration (clearly because
users of hugepage can enjoy the benefit of mempolicy and memory hotplug.)
So this patchset tries to extend such users to support hugepage migration.
The target of this patchset is to enable hugepage migration for NUMA
related system calls (migrate_pages(2), move_pages(2), and mbind(2)), and
memory hotplug.
This patchset does not add hugepage migration for memory compaction,
because users of memory compaction mainly expect to construct thp by
arranging raw pages, and there's little or no need to compact hugepages.
CMA, another user of page migration, can have benefit from hugepage
migration, but is not enabled to support it for now (just because of lack
of testing and expertise in CMA.)
Hugepage migration of non pmd-based hugepage (for example 1GB hugepage in
x86_64, or hugepages in architectures like ia64) is not enabled for now
(again, because of lack of testing.)
As for how these are achived, I extended the API (migrate_pages()) to
handle hugepage (with patch 1 and 2) and adjusted code of each caller to
check and collect movable hugepages (with patch 3-7). Remaining 2 patches
are kind of miscellaneous ones to avoid unexpected behavior. Patch 8 is
about making sure that we only migrate pmd-based hugepages. And patch 9
is about choosing appropriate zone for hugepage allocation.
My test is mainly functional one, simply kicking hugepage migration via
each entry point and confirm that migration is done correctly. Test code
is available here:
git://github.com/Naoya-Horiguchi/test_hugepage_migration_extension.git
And I always run libhugetlbfs test when changing hugetlbfs's code. With
this patchset, no regression was found in the test.
This patch (of 9):
Before enabling each user of page migration to support hugepage,
this patch enables the list of pages for migration to link not only
LRU pages, but also hugepages. As a result, putback_movable_pages()
and migrate_pages() can handle both of LRU pages and hugepages.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Andi Kleen <ak@linux.intel.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If we fail with a reserved page, just calling put_page() is not
sufficient, because put_page() invoke free_huge_page() at last step and it
doesn't know whether a page comes from a reserved pool or not. So it
doesn't do anything related to reserved count. This makes reserve count
lower than how we need, because reserve count already decrease in
dequeue_huge_page_vma(). This patch fix this situation.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Davidlohr Bueso <davidlohr@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We don't need to grab a page_table_lock when we try to release a page.
So, defer to grab a page_table_lock.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Reviewed-by: Davidlohr Bueso <davidlohr@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
is_vma_resv_set(vma, HPAGE_RESV_OWNER) implys that this mapping is for
private. So we don't need to check whether this mapping is for shared or
not.
This patch is just for clean-up.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Reviewed-by: Davidlohr Bueso <davidlohr@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If we alloc hugepage with avoid_reserve, we don't dequeue reserved one.
So, we should check subpool counter when avoid_reserve. This patch
implement it.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Davidlohr Bueso <davidlohr@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
'reservations' is so long name as a variable and we use 'resv_map' to
represent 'struct resv_map' in other place. To reduce confusion and
unreadability, change it.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Reviewed-by: Davidlohr Bueso <davidlohr@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Don't use the reserve pool when soft offlining a hugepage. Check we have
free pages outside the reserve pool before we dequeue the huge page.
Otherwise, we can steal other's reserve page.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Reviewed-by: Davidlohr Bueso <davidlohr@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
lock_device_hotplug() serializes hotplug & online/offline operations. The
lock is held in common sysfs online/offline interfaces and ACPI hotplug
code paths.
And here are the code paths:
- CPU & Mem online/offline via sysfs online
store_online()->lock_device_hotplug()
- Mem online via sysfs state:
store_mem_state()->lock_device_hotplug()
- ACPI CPU & Mem hot-add:
acpi_scan_bus_device_check()->lock_device_hotplug()
- ACPI CPU & Mem hot-delete:
acpi_scan_hot_remove()->lock_device_hotplug()
try_offline_node() off-lines a node if all memory sections and cpus are
removed on the node. It is called from acpi_processor_remove() and
acpi_memory_remove_memory()->remove_memory() paths, both of which are in
the ACPI hotplug code.
try_offline_node() calls stop_machine() to stop all cpus while checking
all cpu status with the assumption that the caller is not protected from
CPU hotplug or CPU online/offline operations. However, the caller is
always serialized with lock_device_hotplug(). Also, the code needs to be
properly serialized with a lock, not by stopping all cpus at a random
place with stop_machine().
This patch removes the use of stop_machine() in try_offline_node() and
adds comments to try_offline_node() and remove_memory() that
lock_device_hotplug() is required.
Signed-off-by: Toshi Kani <toshi.kani@hp.com>
Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
add_memory() and remove_memory() can only handle a memory range aligned
with section. There are problems when an unaligned range is added and
then deleted as follows:
- add_memory() with an unaligned range succeeds, but __add_pages()
called from add_memory() adds a whole section of pages even though
a given memory range is less than the section size.
- remove_memory() to the added unaligned range hits BUG_ON() in
__remove_pages().
This patch changes add_memory() and remove_memory() to check if a given
memory range is aligned with section at the beginning. As the result,
add_memory() fails with -EINVAL when a given range is unaligned, and does
not add such memory range. This prevents remove_memory() to be called
with an unaligned range as well. Note that remove_memory() has to use
BUG_ON() since this function cannot fail.
[akpm@linux-foundation.org: avoid printk warnings]
Signed-off-by: Toshi Kani <toshi.kani@hp.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Explicitly mention/recommend using the libhugetlbfs test cases when
changing related kernel code. Developers that are unaware of the project
can easily miss this and introduce potential regressions that may or may
not be caught by community review.
Also do some cleanups that make the document visually easier to view at a
first glance.
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This helps performance on moderately dense random reads on SSD.
Transaction-Per-Second numbers provided by Taobao:
QPS case
-------------------------------------------------------
7536 disable context readahead totally
w/ patch: 7129 slower size rampup and start RA on the 3rd read
6717 slower size rampup
w/o patch: 5581 unmodified context readahead
Before, readahead will be started whenever reading page N+1 when it happen
to read N recently. After patch, we'll only start readahead when *three*
random reads happen to access pages N, N+1, N+2. The probability of this
happening is extremely low for pure random reads, unless they are very
dense, which actually deserves some readahead.
Also start with a smaller readahead window. The impact to interleaved
sequential reads should be small, because for a long run stream, the the
small readahead window rampup phase is negletable.
The context readahead actually benefits clustered random reads on HDD
whose seek cost is pretty high. However as SSD is increasingly used for
random read workloads it's better for the context readahead to concentrate
on interleaved sequential reads.
Another SSD rand read test from Miao
# file size: 2GB
# read IO amount: 625MB
sysbench --test=fileio \
--max-requests=10000 \
--num-threads=1 \
--file-num=1 \
--file-block-size=64K \
--file-test-mode=rndrd \
--file-fsync-freq=0 \
--file-fsync-end=off run
shows the performance of btrfs grows up from 69MB/s to 121MB/s, ext4 from
104MB/s to 121MB/s.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Tested-by: Tao Ma <tm@tao.ma>
Tested-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use "zone_is_initialized()" instead of "if (zone->wait_table)".
Simplify the code, no functional change.
Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use "zone_is_empty()" instead of "if (zone->spanned_pages)".
Simplify the code, no functional change.
Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In struct gen_pool_chunk, end_addr means the end address of memory chunk
(inclusive), but in the implementation it is treated as address + size of
memory chunk (exclusive), so it points to the address plus one instead of
correct ending address.
The ending address of memory chunk plus one will cause overflow on the
memory chunk including the last address of memory map, e.g. when starting
address is 0xFFF00000 and size is 0x100000 on 32bit machine, ending
address will be 0x100000000.
Use correct ending address like starting address + size - 1.
[akpm@linux-foundation.org: add comment to struct gen_pool_chunk:end_addr]
Signed-off-by: Joonyoung Shim <jy0922.shim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I think we can remove "BUG_ON(start_pfn >= end_pfn)" in __offline_pages(),
because in memory_block_action() "nr_pages = PAGES_PER_SECTION * sections_per_block"
is always greater than 0.
memory_block_action()
offline_pages()
__offline_pages()
BUG_ON(start_pfn >= end_pfn)
In v2.6.32, If info->length==0, this way may hit this BUG_ON().
acpi_memory_disable_device()
remove_memory(info->start_addr, info->length)
offline_pages()
A later Fujitsu patch renamed this function and the BUG_ON() is
unnecessary.
Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Reviewed-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Toshi Kani <toshi.kani@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Our intention in here is to find last_bit within the region to flush.
There is well-defined function, find_last_bit() for this purpose and its
performance may be slightly better than current implementation. So change
it.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
vbq in vmap_block isn't used. So remove it.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Disabling interrupts repeatedly can be avoided in the inner loop if we use
a this_cpu operation.
Signed-off-by: Christoph Lameter <cl@linux.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
CC: Tejun Heo <tj@kernel.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Both functions that update global counters use the same mechanism.
Create a function that contains the common code.
Signed-off-by: Christoph Lameter <cl@linux.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
CC: Tejun Heo <tj@kernel.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The main idea behind this patchset is to reduce the vmstat update overhead
by avoiding interrupt enable/disable and the use of per cpu atomics.
This patch (of 3):
It is better to have a separate folding function because
refresh_cpu_vm_stats() also does other things like expire pages in the
page allocator caches.
If we have a separate function then refresh_cpu_vm_stats() is only called
from the local cpu which allows additional optimizations.
The folding function is only called when a cpu is being downed and
therefore no other processor will be accessing the counters. Also
simplifies synchronization.
[akpm@linux-foundation.org: fix UP build]
Signed-off-by: Christoph Lameter <cl@linux.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
CC: Tejun Heo <tj@kernel.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
PageSwapCache() is always false when !CONFIG_SWAP, so compiler
properly discard related code. Therefore, we don't need #ifdef explicitly.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
pgtable related functions are mostly in pgtable-generic.c.
So move remaining functions from memory.c to pgtable-generic.c.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We rarely allocate a page with ALLOC_NO_WATERMARKS and it is used in slow
path. For helping compiler optimization, add unlikely macro to
ALLOC_NO_WATERMARKS checking.
This patch doesn't have any effect now, because gcc already optimize this
properly. But we cannot assume that gcc always does right and nobody
re-evaluate if gcc do proper optimization with their change, for example,
it is not optimized properly on v3.10. So adding compiler hint here is
reasonable.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If a vma with VM_NORESERVE allocate a new page for page cache, we should
check whether this area is reserved or not. If this address is already
reserved by other process(in case of chg == 0), we should decrement
reserve count, because this allocated page will go into page cache and
currently, there is no way to know that this page comes from reserved pool
or not when releasing inode. This may introduce over-counting problem to
reserved count. With following example code, you can easily reproduce
this situation.
Assume 2MB, nr_hugepages = 100
size = 20 * MB;
flag = MAP_SHARED;
p = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, fd, 0);
if (p == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
return -1;
}
flag = MAP_SHARED | MAP_NORESERVE;
q = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, fd, 0);
if (q == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
}
q[0] = 'c';
After finish the program, run 'cat /proc/meminfo'. You can see below
result.
HugePages_Free: 100
HugePages_Rsvd: 1
To fix this, we should check our mapping type and tracked region. If our
mapping is VM_NORESERVE, VM_MAYSHARE and chg is 0, this imply that current
allocated page will go into page cache which is already reserved region
when mapping is created. In this case, we should decrease reserve count.
As implementing above, this patch solve the problem.
[akpm@linux-foundation.org: fix spelling in comment]
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now, Checking condition of decrement_hugepage_resv_vma() and
vma_has_reserves() is same, so we can clean-up this function with
vma_has_reserves(). Additionally, decrement_hugepage_resv_vma() has only
one call site, so we can remove function and embed it into
dequeue_huge_page_vma() directly. This patch implement it.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If we map the region with MAP_NORESERVE and MAP_SHARED, we can skip to
check reserve counting and eventually we cannot be ensured to allocate a
huge page in fault time. With following example code, you can easily find
this situation.
Assume 2MB, nr_hugepages = 100
fd = hugetlbfs_unlinked_fd();
if (fd < 0)
return 1;
size = 200 * MB;
flag = MAP_SHARED;
p = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, fd, 0);
if (p == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
return -1;
}
size = 2 * MB;
flag = MAP_ANONYMOUS | MAP_SHARED | MAP_HUGETLB | MAP_NORESERVE;
p = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, -1, 0);
if (p == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
}
p[0] = '0';
sleep(10);
During executing sleep(10), run 'cat /proc/meminfo' on another process.
HugePages_Free: 99
HugePages_Rsvd: 100
Number of free should be higher or equal than number of reserve, but this
aren't. This represent that non reserved shared mapping steal a reserved
page. Non reserved shared mapping should not eat into reserve space.
If we consider VM_NORESERVE in vma_has_reserve() and return 0 which mean
that we don't have reserved pages, then we check that we have enough free
pages in dequeue_huge_page_vma(). This prevent to steal a reserved page.
With this change, above test generate a SIGBUG which is correct, because
all free pages are reserved and non reserved shared mapping can't get a
free page.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently, we use a page with mapped count 1 in page cache for cow
optimization. If we find this condition, we don't allocate a new page and
copy contents. Instead, we map this page directly. This may introduce a
problem that writting to private mapping overwrite hugetlb file directly.
You can find this situation with following code.
size = 20 * MB;
flag = MAP_SHARED;
p = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, fd, 0);
if (p == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
return -1;
}
p[0] = 's';
fprintf(stdout, "BEFORE STEAL PRIVATE WRITE: %c\n", p[0]);
munmap(p, size);
flag = MAP_PRIVATE;
p = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, fd, 0);
if (p == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
}
p[0] = 'c';
munmap(p, size);
flag = MAP_SHARED;
p = mmap(NULL, size, PROT_READ|PROT_WRITE, flag, fd, 0);
if (p == MAP_FAILED) {
fprintf(stderr, "mmap() failed: %s\n", strerror(errno));
return -1;
}
fprintf(stdout, "AFTER STEAL PRIVATE WRITE: %c\n", p[0]);
munmap(p, size);
We can see that "AFTER STEAL PRIVATE WRITE: c", not "AFTER STEAL PRIVATE
WRITE: s". If we turn off this optimization to a page in page cache, the
problem is disappeared.
So, I change the trigger condition of optimization. If this page is not
AnonPage, we don't do optimization. This makes this optimization turning
off for a page cache.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Current node iteration code have a minor problem which do one more node
rotation if we can't succeed to allocate. For example, if we start to
allocate at node 0, we stop to iterate at node 0. Then we start to
allocate at node 1 for next allocation.
I introduce new macros "for_each_node_mask_to_[alloc|free]" and fix and
clean-up node iteration code to alloc or free. This makes code more
understandable.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In this time we are holding a hugetlb_lock, so hstate values can't be
changed. If we don't have any usable free huge page in this time, we
don't need to proceed with the processing. So move this code up.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit 75f7ad8e04. It was the result of a problem
observed with a 3.2 kernel and merged in 3.9, while the issue had been
resolved upstream in 3.3 (commit ab8fabd46f: "mm: exclude reserved
pages from dirtyable memory").
The "reserved pages" are a superset of min_free_kbytes, thus this change
is redundant and confusing. Revert it.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Paul Szabo <psz@maths.usyd.edu.au>
Cc: Rik van Riel <riel@redhat.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Each zone that holds userspace pages of one workload must be aged at a
speed proportional to the zone size. Otherwise, the time an individual
page gets to stay in memory depends on the zone it happened to be
allocated in. Asymmetry in the zone aging creates rather unpredictable
aging behavior and results in the wrong pages being reclaimed, activated
etc.
But exactly this happens right now because of the way the page allocator
and kswapd interact. The page allocator uses per-node lists of all zones
in the system, ordered by preference, when allocating a new page. When
the first iteration does not yield any results, kswapd is woken up and the
allocator retries. Due to the way kswapd reclaims zones below the high
watermark while a zone can be allocated from when it is above the low
watermark, the allocator may keep kswapd running while kswapd reclaim
ensures that the page allocator can keep allocating from the first zone in
the zonelist for extended periods of time. Meanwhile the other zones
rarely see new allocations and thus get aged much slower in comparison.
The result is that the occasional page placed in lower zones gets
relatively more time in memory, even gets promoted to the active list
after its peers have long been evicted. Meanwhile, the bulk of the
working set may be thrashing on the preferred zone even though there may
be significant amounts of memory available in the lower zones.
Even the most basic test -- repeatedly reading a file slightly bigger than
memory -- shows how broken the zone aging is. In this scenario, no single
page should be able stay in memory long enough to get referenced twice and
activated, but activation happens in spades:
$ grep active_file /proc/zoneinfo
nr_inactive_file 0
nr_active_file 0
nr_inactive_file 0
nr_active_file 8
nr_inactive_file 1582
nr_active_file 11994
$ cat data data data data >/dev/null
$ grep active_file /proc/zoneinfo
nr_inactive_file 0
nr_active_file 70
nr_inactive_file 258753
nr_active_file 443214
nr_inactive_file 149793
nr_active_file 12021
Fix this with a very simple round robin allocator. Each zone is allowed a
batch of allocations that is proportional to the zone's size, after which
it is treated as full. The batch counters are reset when all zones have
been tried and the allocator enters the slowpath and kicks off kswapd
reclaim. Allocation and reclaim is now fairly spread out to all
available/allowable zones:
$ grep active_file /proc/zoneinfo
nr_inactive_file 0
nr_active_file 0
nr_inactive_file 174
nr_active_file 4865
nr_inactive_file 53
nr_active_file 860
$ cat data data data data >/dev/null
$ grep active_file /proc/zoneinfo
nr_inactive_file 0
nr_active_file 0
nr_inactive_file 666622
nr_active_file 4988
nr_inactive_file 190969
nr_active_file 937
When zone_reclaim_mode is enabled, allocations will now spread out to all
zones on the local node, not just the first preferred zone (which on a 4G
node might be a tiny Normal zone).
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Paul Bolle <paul.bollee@gmail.com>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Tested-by: Kevin Hilman <khilman@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Allocations that do not have to respect the watermarks are rare
high-priority events. Reorder the code such that per-zone dirty limits
and future checks important only to regular page allocations are ignored
in these extraordinary situations.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Paul Bolle <paul.bollee@gmail.com>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The way the page allocator interacts with kswapd creates aging imbalances,
where the amount of time a userspace page gets in memory under reclaim
pressure is dependent on which zone, which node the allocator took the
page frame from.
#1 fixes missed kswapd wakeups on NUMA systems, which lead to some
nodes falling behind for a full reclaim cycle relative to the other
nodes in the system
#3 fixes an interaction where kswapd and a continuous stream of page
allocations keep the preferred zone of a task between the high and
low watermark (allocations succeed + kswapd does not go to sleep)
indefinitely, completely underutilizing the lower zones and
thrashing on the preferred zone
These patches are the aging fairness part of the thrash-detection based
file LRU balancing. Andrea recommended to submit them separately as they
are bugfixes in their own right.
The following test ran a foreground workload (memcachetest) with
background IO of various sizes on a 4 node 8G system (similar results were
observed with single-node 4G systems):
parallelio
BAS FAIRALLO
BASE FAIRALLOC
Ops memcachetest-0M 5170.00 ( 0.00%) 5283.00 ( 2.19%)
Ops memcachetest-791M 4740.00 ( 0.00%) 5293.00 ( 11.67%)
Ops memcachetest-2639M 2551.00 ( 0.00%) 4950.00 ( 94.04%)
Ops memcachetest-4487M 2606.00 ( 0.00%) 3922.00 ( 50.50%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-791M 55.00 ( 0.00%) 18.00 ( 67.27%)
Ops io-duration-2639M 235.00 ( 0.00%) 103.00 ( 56.17%)
Ops io-duration-4487M 278.00 ( 0.00%) 173.00 ( 37.77%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-791M 245184.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-2639M 468069.00 ( 0.00%) 108778.00 ( 76.76%)
Ops swaptotal-4487M 452529.00 ( 0.00%) 76623.00 ( 83.07%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-791M 108297.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2639M 169537.00 ( 0.00%) 50031.00 ( 70.49%)
Ops swapin-4487M 167435.00 ( 0.00%) 34178.00 ( 79.59%)
Ops minorfaults-0M 1518666.00 ( 0.00%) 1503993.00 ( 0.97%)
Ops minorfaults-791M 1676963.00 ( 0.00%) 1520115.00 ( 9.35%)
Ops minorfaults-2639M 1606035.00 ( 0.00%) 1799717.00 (-12.06%)
Ops minorfaults-4487M 1612118.00 ( 0.00%) 1583825.00 ( 1.76%)
Ops majorfaults-0M 6.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-791M 13836.00 ( 0.00%) 10.00 ( 99.93%)
Ops majorfaults-2639M 22307.00 ( 0.00%) 6490.00 ( 70.91%)
Ops majorfaults-4487M 21631.00 ( 0.00%) 4380.00 ( 79.75%)
BAS FAIRALLO
BASE FAIRALLOC
User 287.78 460.97
System 2151.67 3142.51
Elapsed 9737.00 8879.34
BAS FAIRALLO
BASE FAIRALLOC
Minor Faults 53721925 57188551
Major Faults 392195 15157
Swap Ins 2994854 112770
Swap Outs 4907092 134982
Direct pages scanned 0 41824
Kswapd pages scanned 32975063 8128269
Kswapd pages reclaimed 6323069 7093495
Direct pages reclaimed 0 41824
Kswapd efficiency 19% 87%
Kswapd velocity 3386.573 915.414
Direct efficiency 100% 100%
Direct velocity 0.000 4.710
Percentage direct scans 0% 0%
Zone normal velocity 2011.338 550.661
Zone dma32 velocity 1365.623 369.221
Zone dma velocity 9.612 0.242
Page writes by reclaim 18732404.000 614807.000
Page writes file 13825312 479825
Page writes anon 4907092 134982
Page reclaim immediate 85490 5647
Sector Reads 12080532 483244
Sector Writes 88740508 65438876
Page rescued immediate 0 0
Slabs scanned 82560 12160
Direct inode steals 0 0
Kswapd inode steals 24401 40013
Kswapd skipped wait 0 0
THP fault alloc 6 8
THP collapse alloc 5481 5812
THP splits 75 22
THP fault fallback 0 0
THP collapse fail 0 0
Compaction stalls 0 54
Compaction success 0 45
Compaction failures 0 9
Page migrate success 881492 82278
Page migrate failure 0 0
Compaction pages isolated 0 60334
Compaction migrate scanned 0 53505
Compaction free scanned 0 1537605
Compaction cost 914 86
NUMA PTE updates 46738231 41988419
NUMA hint faults 31175564 24213387
NUMA hint local faults 10427393 6411593
NUMA pages migrated 881492 55344
AutoNUMA cost 156221 121361
The overall runtime was reduced, throughput for both the foreground
workload as well as the background IO improved, major faults, swapping and
reclaim activity shrunk significantly, reclaim efficiency more than
quadrupled.
This patch:
When the page allocator fails to get a page from all zones in its given
zonelist, it wakes up the per-node kswapds for all zones that are at their
low watermark.
However, with a system under load the free pages in a zone can fluctuate
enough that the allocation fails but the kswapd wakeup is also skipped
while the zone is still really close to the low watermark.
When one node misses a wakeup like this, it won't be aged before all the
other node's zones are down to their low watermarks again. And skipping a
full aging cycle is an obvious fairness problem.
Kswapd runs until the high watermarks are restored, so it should also be
woken when the high watermarks are not met. This ages nodes more equally
and creates a safety margin for the page counter fluctuation.
By using zone_balanced(), it will now check, in addition to the watermark,
if compaction requires more order-0 pages to create a higher order page.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Paul Bolle <paul.bollee@gmail.com>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In collapse_huge_page() there is a race window between releasing the
mmap_sem read lock and taking the mmap_sem write lock, so find_vma() may
return NULL. So check the return value to avoid NULL pointer dereference.
collapse_huge_page
khugepaged_alloc_page
up_read(&mm->mmap_sem)
down_write(&mm->mmap_sem)
vma = find_vma(mm, address)
Signed-off-by: Libin <huawei.libin@huawei.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Cc: <stable@vger.kernel.org> # v3.0+
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We should not check loop+1 with loop end in loop body. Just duplicate two
lines code to avoid it.
That will help a bit when we have huge amount of pages on system with
16TiB memory.
Signed-off-by: Yinghai Lu <yinghai@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In the current code, the value of fallback_migratetype that is printed
using the mm_page_alloc_extfrag tracepoint, is the value of the
migratetype *after* it has been set to the preferred migratetype (if the
ownership was changed). Obviously that wouldn't have been the original
intent. (We already have a separate 'change_ownership' field to tell
whether the ownership of the pageblock was changed from the
fallback_migratetype to the preferred type.)
The intent of the fallback_migratetype field is to show the migratetype
from which we borrowed pages in order to satisfy the allocation request.
So fix the code to print that value correctly.
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix all errors reported by checkpatch and some small spelling mistakes.
Signed-off-by: Pintu Kumar <pintu.k@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
swap cluster allocation is to get better request merge to improve
performance. But the cluster is shared globally, if multiple tasks are
doing swap, this will cause interleave disk access. While multiple tasks
swap is quite common, for example, each numa node has a kswapd thread
doing swap and multiple threads/processes doing direct page reclaim.
ioscheduler can't help too much here, because tasks don't send swapout IO
down to block layer in the meantime. Block layer does merge some IOs, but
a lot not, depending on how many tasks are doing swapout concurrently. In
practice, I've seen a lot of small size IO in swapout workloads.
We makes the cluster allocation per-cpu here. The interleave disk access
issue goes away. All tasks swapout to their own cluster, so swapout will
become sequential, which can be easily merged to big size IO. If one CPU
can't get its per-cpu cluster (for example, there is no free cluster
anymore in the swap), it will fallback to scan swap_map. The CPU can
still continue swap. We don't need recycle free swap entries of other
CPUs.
In my test (swap to a 2-disk raid0 partition), this improves around 10%
swapout throughput, and request size is increased significantly.
How does this impact swap readahead is uncertain though. On one side,
page reclaim always isolates and swaps several adjancent pages, this will
make page reclaim write the pages sequentially and benefit readahead. On
the other side, several CPU write pages interleave means the pages don't
live _sequentially_ but relatively _near_. In the per-cpu allocation
case, if adjancent pages are written by different cpus, they will live
relatively _far_. So how this impacts swap readahead depends on how many
pages page reclaim isolates and swaps one time. If the number is big,
this patch will benefit swap readahead. Of course, this is about
sequential access pattern. The patch has no impact for random access
pattern, because the new cluster allocation algorithm is just for SSD.
Alternative solution is organizing swap layout to be per-mm instead of
this per-cpu approach. In the per-mm layout, we allocate a disk range for
each mm, so pages of one mm live in swap disk adjacently. per-mm layout
has potential issues of lock contention if multiple reclaimers are swap
pages from one mm. For a sequential workload, per-mm layout is better to
implement swap readahead, because pages from the mm are adjacent in disk.
But per-cpu layout isn't very bad in this workload, as page reclaim always
isolates and swaps several pages one time, such pages will still live in
disk sequentially and readahead can utilize this. For a random workload,
per-mm layout isn't beneficial of request merge, because it's quite
possible pages from different mm are swapout in the meantime and IO can't
be merged in per-mm layout. while with per-cpu layout we can merge
requests from any mm. Considering random workload is more popular in
workloads with swap (and per-cpu approach isn't too bad for sequential
workload too), I'm choosing per-cpu layout.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Kyungmin Park <kmpark@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The previous patch can expose races, according to Hugh:
swapoff was sometimes failing with "Cannot allocate memory", coming from
try_to_unuse()'s -ENOMEM: it needs to allow for swap_duplicate() failing
on a free entry temporarily SWAP_MAP_BAD while being discarded.
We should use ACCESS_ONCE() there, and whenever accessing swap_map
locklessly; but rather than peppering it throughout try_to_unuse(), just
declare *swap_map with volatile.
try_to_unuse() is accustomed to *swap_map going down racily, but not
necessarily to it jumping up from 0 to SWAP_MAP_BAD: we'll be safer to
prevent that transition once SWP_WRITEOK is switched off, when it's a
waste of time to issue discards anyway (swapon can do a whole discard).
Another issue is:
In swapin_readahead(), read_swap_cache_async() can read a bad swap entry,
because we don't check if readahead swap entry is bad. This doesn't break
anything but such swapin page is wasteful and can only be freed at page
reclaim. We should avoid read such swap entry. And in discard, we mark
swap entry SWAP_MAP_BAD and then switch it to normal when discard is
finished. If readahead reads such swap entry, we have the same issue, so
we much check if swap entry is bad too.
Thanks Hugh to inspire swapin_readahead could use bad swap entry.
[include Hugh's patch 'swap: fix swapoff ENOMEMs from discard']
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Kyungmin Park <kmpark@infradead.org>
Cc: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
swap can do cluster discard for SSD, which is good, but there are some
problems here:
1. swap do the discard just before page reclaim gets a swap entry and
writes the disk sectors. This is useless for high end SSD, because an
overwrite to a sector implies a discard to original sector too. A
discard + overwrite == overwrite.
2. the purpose of doing discard is to improve SSD firmware garbage
collection. Idealy we should send discard as early as possible, so
firmware can do something smart. Sending discard just after swap entry
is freed is considered early compared to sending discard before write.
Of course, if workload is already bound to gc speed, sending discard
earlier or later doesn't make
3. block discard is a sync API, which will delay scan_swap_map()
significantly.
4. Write and discard command can be executed parallel in PCIe SSD.
Making swap discard async can make execution more efficiently.
This patch makes swap discard async and moves discard to where swap entry
is freed. Discard and write have no dependence now, so above issues can
be avoided. Idealy we should do discard for any freed sectors, but some
SSD discard is very slow. This patch still does discard for a whole
cluster.
My test does a several round of 'mmap, write, unmap', which will trigger a
lot of swap discard. In a fusionio card, with this patch, the test
runtime is reduced to 18% of the time without it, so around 5.5x faster.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Kyungmin Park <kmpark@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to
20~30% CPU time (when cluster is hard to find, the CPU time can be up to
80%), which becomes a bottleneck. scan_swap_map() scans a byte array to
search a 256 page cluster, which is very slow.
Here I introduced a simple algorithm to search cluster. Since we only
care about 256 pages cluster, we can just use a counter to track if a
cluster is free. Every 256 pages use one int to store the counter. If
the counter of a cluster is 0, the cluster is free. All free clusters
will be added to a list, so searching cluster is very efficient. With
this, scap_swap_map() overhead disappears.
This might help low end SD card swap too. Because if the cluster is
aligned, SD firmware can do flash erase more efficiently.
We only enable the algorithm for SSD. Hard disk swap isn't fast enough
and has downside with the algorithm which might introduce regression (see
below).
The patch slightly changes which cluster is choosen. It always adds free
cluster to list tail. This can help wear leveling for low end SSD too.
And if no cluster found, the scan_swap_map() will do search from the end
of last cluster. So if no cluster found, the scan_swap_map() will do
search from the end of last free cluster, which is random. For SSD, this
isn't a problem at all.
Another downside is the cluster must be aligned to 256 pages, which will
reduce the chance to find a cluster. I would expect this isn't a big
problem for SSD because of the non-seek penality. (And this is the reason
I only enable the algorithm for SSD).
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Kyungmin Park <kmpark@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
set_pageblock_order() may be called when memory hotplug, so need use
'__paginginit' instead of '__init'.
The related warning:
The function __meminit .free_area_init_node() references
a function __init .set_pageblock_order().
If .set_pageblock_order is only used by .free_area_init_node then
annotate .set_pageblock_order with a matching annotation.
Signed-off-by: Chen Gang <gang.chen@asianux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When PAGE_SHIFT > 20, the result of "20 - PAGE_SHIFT" is negative. The
previous calculating here will generate an unexpected result. In
addition, if PAGE_SIZE >= 1MB, The memory size of "numentries" was
already integral multiple of 1MB.
Signed-off-by: Jerry Zhou <uulinux@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
