Commit 022333427a ("dm: optimize dm_mq_queue_rq to _not_ use kthread if
using pure blk-mq") mistakenly removed free_rq_clone()'s clone->q check
before testing clone->q->mq_ops. It was an oversight to discontinue
that check for 1 of the 2 use-cases for free_rq_clone():
1) free_rq_clone() called when an unmapped original request is requeued
2) free_rq_clone() called in the request-based IO completion path
The clone->q check made sense for case #1 but not for #2. However, we
cannot just reinstate the check as it'd mask a serious bug in the IO
completion case #2 -- no in-flight request should have an uninitialized
request_queue (basic block layer refcounting _should_ ensure this).
The NULL pointer seen for case #1 is detailed here:
https://www.redhat.com/archives/dm-devel/2015-April/msg00160.html
Fix this free_rq_clone() NULL pointer by simply checking if the
mapped_device's type is DM_TYPE_MQ_REQUEST_BASED (clone's queue is
blk-mq) rather than checking clone->q->mq_ops. This avoids the need to
dereference clone->q, but a WARN_ON_ONCE is added to let us know if an
uninitialized clone request is being completed.
Reported-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Commit bfebd1cdb4 ("dm: add full blk-mq support to request-based DM")
didn't properly account for the need to short-circuit re-initializing
DM's blk-mq request_queue if it was already initialized.
Otherwise, reloading a blk-mq request-based DM table (either manually
or via multipathd) resulted in errors, see:
https://www.redhat.com/archives/dm-devel/2015-April/msg00132.html
Fix is to only initialize the request_queue on the initial table load
(when the mapped_device type is assigned).
This is better than having dm_init_request_based_blk_mq_queue() return
early if the queue was already initialized because it elevates the
constraint to a more meaningful location in DM core. As such the
pre-existing early return in dm_init_request_based_queue() can now be
removed.
Fixes: bfebd1cdb4 ("dm: add full blk-mq support to request-based DM")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Pull md updates from Neil Brown:
"More updates that usual this time. A few have performance impacts
which hould mostly be positive, but RAID5 (in particular) can be very
work-load ensitive... We'll have to wait and see.
Highlights:
- "experimental" code for managing md/raid1 across a cluster using
DLM. Code is not ready for general use and triggers a WARNING if
used. However it is looking good and mostly done and having in
mainline will help co-ordinate development.
- RAID5/6 can now batch multiple (4K wide) stripe_heads so as to
handle a full (chunk wide) stripe as a single unit.
- RAID6 can now perform read-modify-write cycles which should help
performance on larger arrays: 6 or more devices.
- RAID5/6 stripe cache now grows and shrinks dynamically. The value
set is used as a minimum.
- Resync is now allowed to go a little faster than the 'mininum' when
there is competing IO. How much faster depends on the speed of the
devices, so the effective minimum should scale with device speed to
some extent"
* tag 'md/4.1' of git://neil.brown.name/md: (58 commits)
md/raid5: don't do chunk aligned read on degraded array.
md/raid5: allow the stripe_cache to grow and shrink.
md/raid5: change ->inactive_blocked to a bit-flag.
md/raid5: move max_nr_stripes management into grow_one_stripe and drop_one_stripe
md/raid5: pass gfp_t arg to grow_one_stripe()
md/raid5: introduce configuration option rmw_level
md/raid5: activate raid6 rmw feature
md/raid6 algorithms: xor_syndrome() for SSE2
md/raid6 algorithms: xor_syndrome() for generic int
md/raid6 algorithms: improve test program
md/raid6 algorithms: delta syndrome functions
raid5: handle expansion/resync case with stripe batching
raid5: handle io error of batch list
RAID5: batch adjacent full stripe write
raid5: track overwrite disk count
raid5: add a new flag to track if a stripe can be batched
raid5: use flex_array for scribble data
md raid0: access mddev->queue (request queue member) conditionally because it is not set when accessed from dm-raid
md: allow resync to go faster when there is competing IO.
md: remove 'go_faster' option from ->sync_request()
...
When array is degraded, read data landed on failed drives will result in
reading rest of data in a stripe. So a single sequential read would
result in same data being read twice.
This patch is to avoid chunk aligned read for degraded array. The
downside is to involve stripe cache which means associated CPU overhead
and extra memory copy.
Test Results:
Following test are done on a enterprise storage node with Seagate 6T SAS
drives and Xeon E5-2648L CPU (10 cores, 1.9Ghz), 10 disks MD RAID6 8+2,
chunk size 128 KiB.
I use FIO, using direct-io with various bs size, enough queue depth,
tested sequential and 100% random read against 3 array config:
1) optimal, as baseline;
2) degraded;
3) degraded with this patch.
Kernel version is 4.0-rc3.
Each individual test I only did once so there might be some variations,
but we just focus on big trend.
Sequential Read:
bs=(KiB) optimal(MiB/s) degraded(MiB/s) degraded-with-patch (MiB/s)
1024 1608 656 995
512 1624 710 956
256 1635 728 980
128 1636 771 983
64 1612 1119 1000
32 1580 1420 1004
16 1368 688 986
8 768 647 953
4 411 413 850
Random Read:
bs=(KiB) optimal(IOPS) degraded(IOPS) degraded-with-patch (IOPS)
1024 163 160 156
512 274 273 272
256 426 428 424
128 576 592 591
64 726 724 726
32 849 848 837
16 900 970 971
8 927 940 929
4 948 940 955
Some notes:
* In sequential + optimal, as bs size getting smaller, the FIO thread
become CPU bound.
* In sequential + degraded, there's big increase when bs is 64K and
32K, I don't have explanation.
* In sequential + degraded-with-patch, the MD thread mostly become CPU
bound.
If you want to we can discuss specific data point in those data. But in
general it seems with this patch, we have more predictable and in most
cases significant better sequential read performance when array is
degraded, and almost no noticeable impact on random read.
Performance is a complicated thing, the patch works well for this
particular configuration, but may not be universal. For example I
imagine testing on all SSD array may have very different result. But I
personally think in most cases IO bandwidth is more scarce resource than
CPU.
Signed-off-by: Eric Mei <eric.mei@seagate.com>
Signed-off-by: NeilBrown <neilb@suse.de>
The default setting of 256 stripe_heads is probably
much too small for many configurations. So it is best to make it
auto-configure.
Shrinking the cache under memory pressure is easy. The only
interesting part here is that we put a fairly high cost
('seeks') on shrinking the cache as the cost is greater than
just having to read more data, it reduces parallelism.
Growing the cache on demand needs to be done carefully. If we allow
fast growth, that can upset memory balance as lots of dirty memory can
quickly turn into lots of memory queued in the stripe_cache.
It is important for the raid5 block device to appear congested to
allow write-throttling to work.
So we only add stripes slowly. We set a flag when an allocation
fails because all stripes are in use, allocate at a convenient
time when that flag is set, and don't allow it to be set again
until at least one stripe_head has been released for re-use.
This means that a spurt of requests will only cause one stripe_head
to be allocated, but a steady stream of requests will slowly
increase the cache size - until memory pressure puts it back again.
It could take hours to reach a steady state.
The value written to, and displayed in, stripe_cache_size is
used as a minimum. The cache can grow above this and shrink back
down to it. The actual size is not directly visible, though it can
be deduced to some extent by watching stripe_cache_active.
Signed-off-by: NeilBrown <neilb@suse.de>
This allows us to easily add more (atomic) flags. Signed-off-by: NeilBrown <neilb@suse.de>
Rather than adjusting max_nr_stripes whenever {grow,drop}_one_stripe()
succeeds, do it inside the functions.
Also choose the correct hash to handle next inside the functions.
This removes duplication and will help with future new uses of
{grow,drop}_one_stripe.
This also fixes a minor bug where the "md/raid:%md: allocate XXkB"
message always said "0kB".
Signed-off-by: NeilBrown <neilb@suse.de>
This is needed for future improvement to stripe cache management. Signed-off-by: NeilBrown <neilb@suse.de>
Depending on the available coding we allow optimized rmw logic for write operations. To support easier testing this patch allows manual control of the rmw/rcw descision through the interface /sys/block/mdX/md/rmw_level. The configuration can handle three levels of control. rmw_level=0: Disable rmw for all RAID types. Hardware assisted P/Q calculation has no implementation path yet to factor in/out chunks of a syndrome. Enforcing this level can be benefical for slow CPUs with hardware syndrome support and fast SSDs. rmw_level=1: Estimate rmw IOs and rcw IOs. Execute rmw only if we will save IOs. This equals the "old" unpatched behaviour and will be the default. rmw_level=2: Execute rmw even if calculated IOs for rmw and rcw are equal. We might have higher CPU consumption because of calculating the parity twice but it can be benefical otherwise. E.g. RAID4 with fast dedicated parity disk/SSD. The option is implemented just to be forward-looking and will ONLY work with this patch! Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
Glue it altogehter. The raid6 rmw path should work the same as the
already existing raid5 logic. So emulate the prexor handling/flags
and split functions as needed.
1) Enable xor_syndrome() in the async layer.
2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome
at the start of a rmw run as we did it before for the single parity.
3) Take care of rmw run in ops_run_reconstruct6(). Again process only
the changed pages to get syndrome back into sync.
4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw
run. The lower layers will calculate start & end pages from that and
call the xor_syndrome() correspondingly.
5) Adapt the several places where we ignored Q handling up to now.
Performance numbers for a single E5630 system with a mix of 10 7200k
desktop/server disks. 300 seconds random write with 8 threads onto a
3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4)
bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0
skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0
4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s
8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s
16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s
32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s
64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s
128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s
256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s
512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NeilBrown <neilb@suse.de>