Maximal readahead size is limited now by two values:
1) by global 2Mb constant (MAX_READAHEAD in max_sane_readahead())
2) by configurable per-device value* (bdi->ra_pages)
There are devices, which require custom readahead limit.
For instance, for RAIDs it's calculated as number of devices
multiplied by chunk size times 2.
Readahead size can never be larger than bdi->ra_pages * 2 value
(POSIX_FADV_SEQUNTIAL doubles readahead size).
If so, why do we need two limits?
I suggest to completely remove this max_sane_readahead() stuff and
use per-device readahead limit everywhere.
Also, using right readahead size for RAID disks can significantly
increase i/o performance:
before:
dd if=/dev/md2 of=/dev/null bs=100M count=100
100+0 records in
100+0 records out
10485760000 bytes (10 GB) copied, 12.9741 s, 808 MB/s
after:
$ dd if=/dev/md2 of=/dev/null bs=100M count=100
100+0 records in
100+0 records out
10485760000 bytes (10 GB) copied, 8.91317 s, 1.2 GB/s
(It's an 8-disks RAID5 storage).
This patch doesn't change sys_readahead and madvise(MADV_WILLNEED)
behavior introduced by 6d2be915e5 ("mm/readahead.c: fix readahead
failure for memoryless NUMA nodes and limit readahead pages").
Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: onstantin Khlebnikov <khlebnikov@yandex-team.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 6afdb859b7 ("mm: do not ignore mapping_gfp_mask in page cache
allocation paths") has caught some users of hardcoded GFP_KERNEL used in
the page cache allocation paths. This, however, wasn't complete and
there were others which went unnoticed.
Dave Chinner has reported the following deadlock for xfs on loop device:
: With the recent merge of the loop device changes, I'm now seeing
: XFS deadlock on my single CPU, 1GB RAM VM running xfs/073.
:
: The deadlocked is as follows:
:
: kloopd1: loop_queue_read_work
: xfs_file_iter_read
: lock XFS inode XFS_IOLOCK_SHARED (on image file)
: page cache read (GFP_KERNEL)
: radix tree alloc
: memory reclaim
: reclaim XFS inodes
: log force to unpin inodes
: <wait for log IO completion>
:
: xfs-cil/loop1: <does log force IO work>
: xlog_cil_push
: xlog_write
: <loop issuing log writes>
: xlog_state_get_iclog_space()
: <blocks due to all log buffers under write io>
: <waits for IO completion>
:
: kloopd1: loop_queue_write_work
: xfs_file_write_iter
: lock XFS inode XFS_IOLOCK_EXCL (on image file)
: <wait for inode to be unlocked>
:
: i.e. the kloopd, with it's split read and write work queues, has
: introduced a dependency through memory reclaim. i.e. that writes
: need to be able to progress for reads make progress.
:
: The problem, fundamentally, is that mpage_readpages() does a
: GFP_KERNEL allocation, rather than paying attention to the inode's
: mapping gfp mask, which is set to GFP_NOFS.
:
: The didn't used to happen, because the loop device used to issue
: reads through the splice path and that does:
:
: error = add_to_page_cache_lru(page, mapping, index,
: GFP_KERNEL & mapping_gfp_mask(mapping));
This has changed by commit aa4d86163e ("block: loop: switch to VFS
ITER_BVEC").
This patch changes mpage_readpage{s} to follow gfp mask set for the
mapping. There are, however, other places which are doing basically the
same.
lustre:ll_dir_filler is doing GFP_KERNEL from the function which
apparently uses GFP_NOFS for other allocations so let's make this
consistent.
cifs:readpages_get_pages is called from cifs_readpages and
__cifs_readpages_from_fscache called from the same path obeys mapping
gfp.
ramfs_nommu_expand_for_mapping is hardcoding GFP_KERNEL as well
regardless it uses mapping_gfp_mask for the page allocation.
ext4_mpage_readpages is the called from the page cache allocation path
same as read_pages and read_cache_pages
As I've noticed in my previous post I cannot say I would be happy about
sprinkling mapping_gfp_mask all over the place and it sounds like we
should drop gfp_mask argument altogether and use it internally in
__add_to_page_cache_locked that would require all the filesystems to use
mapping gfp consistently which I am not sure is the case here. From a
quick glance it seems that some file system use it all the time while
others are selective.
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reported-by: Dave Chinner <david@fromorbit.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Ming Lei <ming.lei@canonical.com>
Cc: Andreas Dilger <andreas.dilger@intel.com>
Cc: Oleg Drokin <oleg.drokin@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In several places, bdi_congested() and its wrappers are used to
determine whether more IOs should be issued. With cgroup writeback
support, this question can't be answered solely based on the bdi
(backing_dev_info). It's dependent on whether the filesystem and bdi
support cgroup writeback and the blkcg the inode is associated with.
This patch implements inode_congested() and its wrappers which take
@inode and determines the congestion state considering cgroup
writeback. The new functions replace bdi_*congested() calls in places
where the query is about specific inode and task.
There are several filesystem users which also fit this criteria but
they should be updated when each filesystem implements cgroup
writeback support.
v2: Now that a given inode is associated with only one wb, congestion
state can be determined independent from the asking task. Drop
@task. Spotted by Vivek. Also, converted to take @inode instead
of @mapping and renamed to inode_congested().
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jan Kara <jack@suse.cz>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that we got rid of the bdi abuse on character devices we can always use
sb->s_bdi to get at the backing_dev_info for a file, except for the block
device special case. Export inode_to_bdi and replace uses of
mapping->backing_dev_info with it to prepare for the removal of
mapping->backing_dev_info.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit f9acc8c7b3 ("readahead: sanify file_ra_state names") left
ra_submit with a single function call.
Move ra_submit to internal.h and inline it to save some stack. Thanks
to Andrew Morton for commenting different versions.
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently max_sane_readahead() returns zero on the cpu whose NUMA node
has no local memory which leads to readahead failure. Fix this
readahead failure by returning minimum of (requested pages, 512). Users
running applications on a memory-less cpu which needs readahead such as
streaming application see considerable boost in the performance.
Result:
fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless
CPU with 1GB testfile (12 iterations) yielded around 46.66% improvement.
fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB
testfile 32GB* 4G RAM numa machine (12 iterations) showed no impact on
the normal NUMA cases w/ patch.
Kernel Avg Stddev
base 7.4975 3.92%
patched 7.4174 3.26%
[Andrew: making return value PAGE_SIZE independent]
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Acked-by: Jan Kara <jack@suse.cz>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 63d0f0a3c7 ("mm/readahead.c:do_readhead(): don't check for
->readpage") unintentionally made do_readahead return 0 for all valid
files regardless of whether readahead was supported, rather than the
expected -EINVAL. This gets forwarded on to userspace, and results in
sys_readahead appearing to succeed in cases that don't make sense (e.g.
when called on pipes or sockets). This issue is detected by the LTP
readahead01 testcase.
As the exact return value of force_page_cache_readahead is currently
never used, we can simplify it to return only 0 or -EINVAL (when
readpage or readpages is missing). With that in place we can simply
forward on the return value of force_page_cache_readahead in
do_readahead.
This patch performs said change, restoring the expected semantics.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The kernel's readahead algorithm sometimes interprets random read
accesses as sequential and triggers unnecessary data prefecthing from
storage device (impacting random read average latency).
In order to identify sequential cache read misses, the readahead
algorithm intends to check whether offset - previous offset == 1
(trivial sequential reads) or offset - previous offset == 0 (sequential
reads not aligned on page boundary):
if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
The current offset is stored in the "offset" variable of type "pgoff_t"
(unsigned long), while previous offset is stored in "ra->prev_pos" of
type "loff_t" (long long). Therefore, operands of the if statement are
implicitly converted to type long long. Consequently, when previous
offset > current offset (which happens on random pattern), the if
condition is true and access is wrongly interpeted as sequential. An
unnecessary data prefetching is triggered, impacting the average random
read latency.
Storing the previous offset value in a "pgoff_t" variable (unsigned
long) fixes the sequential read detection logic.
Signed-off-by: Damien Ramonda <damien.ramonda@intel.com>
Reviewed-by: Fengguang Wu <fengguang.wu@intel.com>
Acked-by: Pierre Tardy <pierre.tardy@intel.com>
Acked-by: David Cohen <david.a.cohen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The callee force_page_cache_readahead() already does this and unlike
do_readahead(), force_page_cache_readahead() remembers to check for
->readpages() as well.
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>
Currently there is no way to truncate partial page where the end
truncate point is not at the end of the page. This is because it was not
needed and the functionality was enough for file system truncate
operation to work properly. However more file systems now support punch
hole feature and it can benefit from mm supporting truncating page just
up to the certain point.
Specifically, with this functionality truncate_inode_pages_range() can
be changed so it supports truncating partial page at the end of the
range (currently it will BUG_ON() if 'end' is not at the end of the
page).
This commit changes the invalidatepage() address space operation
prototype to accept range to be invalidated and update all the instances
for it.
We also change the block_invalidatepage() in the same way and actually
make a use of the new length argument implementing range invalidation.
Actual file system implementations will follow except the file systems
where the changes are really simple and should not change the behaviour
in any way .Implementation for truncate_page_range() which will be able
to accept page unaligned ranges will follow as well.
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Hugh Dickins <hughd@google.com>
... and convert a bunch of SYSCALL_DEFINE ones to SYSCALL_DEFINE<n>,
killing the boilerplate crap around them.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The files changed within are only using the EXPORT_SYMBOL
macro variants. They are not using core modular infrastructure
and hence don't need module.h but only the export.h header.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Code has been converted over to the new explicit on-stack plugging,
and delay users have been converted to use the new API for that.
So lets kill off the old plugging along with aops->sync_page().
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
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>
Roman Gushchin
Michal Hocko
Tejun Heo
Christoph Hellwig
Fabian Frederick
Raghavendra K T
Johannes Weiner
Mark Rutland
Damien Ramonda
Andrew Morton
Fengguang Wu
Lukas Czerner
Al Viro
Cong Wang
Paul Gortmaker
Jens Axboe
Huang Shijie