Remove the dubious attempt to prefer 'compute' over 'read'. Not only is it
wrong given commit c337869d (md: do not compute parity unless it is on a failed
drive), but it can trigger a BUG_ON in handle_parity_checks5().
Cc: <stable@kernel.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
We shouldn't acknowledge that a stripe has been expanded (When
reshaping a raid5 by adding a device) until the moved data has
actually been written out. However we are currently
acknowledging (by calling md_done_sync) when the POST_XOR
is complete and before the write.
So track in s.locked whether there are pending writes, and don't
call md_done_sync yet if there are.
Note: we all set R5_LOCKED on devices which are are about to
read from. This probably isn't technically necessary, but is
usually done when writing a block, and justifies the use of
s.locked here.
This bug can lead to a crash if an array is stopped while an reshape
is in progress.
Cc: <stable@kernel.org>
Signed-off-by: Neil Brown <neilb@suse.de>
If, while assembling an array, we find a device which is not fully
in-sync with the array, it is important to set the "fullsync" flags.
This is an exact analog to the setting of this flag in hot_add_disk
methods.
Currently, only v1.x metadata supports having devices in an array
which are not fully in-sync (it keep track of how in sync they are).
The 'fullsync' flag only makes a difference when a write-intent bitmap
is being used. In this case it tells recovery to ignore the bitmap
and recovery all blocks.
This fix is already in place for raid1, but not raid5/6 or raid10.
So without this fix, a raid1 ir raid4/5/6 array with version 1.x
metadata and a write intent bitmaps, that is stopped in the middle
of a recovery, will appear to complete the recovery instantly
after it is reassembled, but the recovery will not be correct.
If you might have an array like that, issueing
echo repair > /sys/block/mdXX/md/sync_action
will make sure recovery completes properly.
Cc: <stable@kernel.org>
Signed-off-by: Neil Brown <neilb@suse.de>
If a block is computed (rather than read) then a check/repair operation
may be lead to believe that the data on disk is correct, when infact it
isn't. So only compute blocks for failed devices.
This issue has been around since at least 2.6.12, but has become harder to
hit in recent kernels since most reads bypass the cache.
echo repair > /sys/block/mdN/md/sync_action will set the parity blocks to the
correct state.
Cc: <stable@kernel.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
During the initial array synchronization process there is a window between
when a prexor operation is scheduled to a specific stripe and when it
completes for a sync_request to be scheduled to the same stripe. When
this happens the prexor completes and the stripe is unconditionally marked
"insync", effectively canceling the sync_request for the stripe. Prior to
2.6.23 this was not a problem because the prexor operation was done under
sh->lock. The effect in older kernels being that the prexor would still
erroneously mark the stripe "insync", but sync_request would be held off
and re-mark the stripe as "!in_sync".
Change the write completion logic to not mark the stripe "in_sync" if a
prexor was performed. The effect of the change is to sometimes not set
STRIPE_INSYNC. The worst this can do is cause the resync to stall waiting
for STRIPE_INSYNC to be set. If this were happening, then STRIPE_SYNCING
would be set and handle_issuing_new_read_requests would cause all
available blocks to eventually be read, at which point prexor would never
be used on that stripe any more and STRIPE_INSYNC would eventually be set.
echo repair > /sys/block/mdN/md/sync_action will correct arrays that may
have lost this race.
Cc: <stable@kernel.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When we get any IO error during a recovery (rebuilding a spare), we abort
the recovery and restart it.
For RAID6 (and multi-drive RAID1) it may not be best to restart at the
beginning: when multiple failures can be tolerated, the recovery may be
able to continue and re-doing all that has already been done doesn't make
sense.
We already have the infrastructure to record where a recovery is up to
and restart from there, but it is not being used properly.
This is because:
- We sometimes abort with MD_RECOVERY_ERR rather than just MD_RECOVERY_INTR,
which causes the recovery not be be checkpointed.
- We remove spares and then re-added them which loses important state
information.
The distinction between MD_RECOVERY_ERR and MD_RECOVERY_INTR really isn't
needed. If there is an error, the relevant drive will be marked as
Faulty, and that is enough to ensure correct handling of the error. So we
first remove MD_RECOVERY_ERR, changing some of the uses of it to
MD_RECOVERY_INTR.
Then we cause the attempt to remove a non-faulty device from an array to
fail (unless recovery is impossible as the array is too degraded). Then
when remove_and_add_spares attempts to remove the devices on which
recovery can continue, it will fail, they will remain in place, and
recovery will continue on them as desired.
Issue: If we are halfway through rebuilding a spare and another drive
fails, and a new spare is immediately available, do we want to:
1/ complete the current rebuild, then go back and rebuild the new spare or
2/ restart the rebuild from the start and rebuild both devices in
parallel.
Both options can be argued for. The code currently takes option 2 as
a/ this requires least code change
b/ this results in a minimally-degraded array in minimal time.
Cc: "Eivind Sarto" <ivan@kasenna.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Last night we had scsi problems and a hardware raid unit was offlined
during heavy i/o. While this happened we got for about 3 minutes a huge
number messages like these
Apr 12 03:36:07 pfs1n14 kernel: [197510.696595] raid5:md7: read error not correctable (sector 2993096568 on sdj2).
I guess the high error rate is responsible for not scheduling other events
- during this time the system was not pingable and in the end also other
devices run into scsi command timeouts causing problems on these unrelated
devices as well.
Signed-off-by: Bernd Schubert <bernd-schubert@gmx.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As setting and clearing queue flags now requires that we hold a spinlock
on the queue, and as blk_queue_stack_limits is called without that lock,
get the lock inside blk_queue_stack_limits.
For blk_queue_stack_limits to be able to find the right lock, each md
personality needs to set q->queue_lock to point to the appropriate lock.
Those personalities which didn't previously use a spin_lock, us
q->__queue_lock. So always initialise that lock when allocated.
With this in place, setting/clearing of the QUEUE_FLAG_PLUGGED bit will no
longer cause warnings as it will be clear that the proper lock is held.
Thanks to Dan Williams for review and fixing the silly bugs.
Signed-off-by: NeilBrown <neilb@suse.de>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Alistair John Strachan <alistair@devzero.co.uk>
Cc: Nick Piggin <npiggin@suse.de>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Jacek Luczak <difrost.kernel@gmail.com>
Cc: Prakash Punnoor <prakash@punnoor.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
commit bd2ab67030 "md: close a livelock window
in handle_parity_checks5" introduced a bug in handling 'repair' operations.
After a repair operation completes we clear the state bits tracking this
operation. However, they are cleared too early and this results in the code
deciding to re-run the parity check operation. Since we have done the repair
in memory the second check does not find a mismatch and thus does not do a
writeback.
Test results:
$ echo repair > /sys/block/md0/md/sync_action
$ cat /sys/block/md0/md/mismatch_cnt
51072
$ echo repair > /sys/block/md0/md/sync_action
$ cat /sys/block/md0/md/mismatch_cnt
0
(also fix incorrect indentation)
Cc: <stable@kernel.org>
Tested-by: George Spelvin <linux@horizon.com>
Acked-by: NeilBrown <neilb@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Allows a userspace metadata handler to take action upon detecting a device
failure.
Based on an original patch by Neil Brown.
Changes:
-added blocked_wait waitqueue to rdev
-don't qualify Blocked with Faulty always let userspace block writes
-added md_wait_for_blocked_rdev to wait for the block device to be clear, if
userspace misses the notification another one is sent every 5 seconds
-set MD_RECOVERY_NEEDED after clearing "blocked"
-kill DoBlock flag, just test mddev->external
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
MD drivers use one printk() call to print 2 log messages and the second line
may be prefixed by a TAB character. It may also output a trailing space
before newline. klogd (I think) turns the TAB character into the 2 characters
'^I' when logging to a file. This looks ugly.
Instead of a leading TAB to indicate continuation, prefix both output lines
with 'raid:' or similar. Also remove any trailing space in the vicinity of
the affected code and consistently end the sentences with a period.
Signed-off-by: Nick Andrew <nick@nick-andrew.net>
Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
strict_strtoul handles the open-coded sanity checks in
raid5_store_stripe_cache_size and raid5_store_preread_threshold
Acked-by: NeilBrown <neilb@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Improve write performance by preventing the delayed_list from dumping all its
stripes onto the handle_list in one shot. Delayed stripes are now further
delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when:
* a STRIPE_IO_STARTED stripe is found at the head of 'handle_list'
* 'handle_list' is empty and i/o is being done to satisfy full stripe-width
write requests
* 'bypass_count' is less than 'bypass_threshold'. By default the threshold
is 1, i.e. every other stripe handled is a preread stripe provided the
top two conditions are false.
Benchmark data:
System: 2x Xeon 5150, 4x SATA, mem=1GB
Baseline: 2.6.24-rc7
Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean
Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048
* patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512)
Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS)
* patched: +13%
* patched + preread_bypass_threshold = 0: +37%
Changes since v1:
* reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and
make it configurable. This defaults to fairness and modest performance
gains out of the box.
Changes since v2:
* [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not
necessary, the important change was clearing STRIPE_DELAYED in
add_stripe_bio and this has been moved out to make_request for the hang
fix.
* [neilb@suse.de]: simplify get_priority_stripe
* [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is
sampled empty (+11%)
* [dan.j.williams@intel.com]: decrement the bypass_count at the detection
of stripes being naturally promoted off of hold_list +2%. Note, resetting
bypass_count instead of decrementing on these events yields +4% but that is
probably too aggressive.
Changes since v3:
* cosmetic fixups
Tested-by: James W. Laferriere <babydr@baby-dragons.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
__FUNCTION__ is gcc-specific, use __func__
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If a failure is detected after a parity check operation has been initiated,
but before it completes handle_parity_checks5 will never quiesce operations on
the stripe.
Explicitly handle this case by "canceling" the parity check, i.e. clear the
STRIPE_OP_CHECK flags and queue the stripe on the handle list again to refresh
any non-uptodate blocks.
Kernel versions >= 2.6.23 are susceptible.
Cc: <stable@kernel.org>
Cc: NeilBrown <neilb@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
gcc-3.4.5 on sparc64:
drivers/md/raid5.c: In function `raid5_end_read_request':
drivers/md/raid5.c:1147: warning: long long unsigned int format, long unsigned int arg (arg 4)
drivers/md/raid5.c:1164: warning: long long unsigned int format, long unsigned int arg (arg 3)
drivers/md/raid5.c:1170: warning: long long unsigned int format, long unsigned int arg (arg 3)
sector_t is u64, and we don't know what type the architecture uses to
implement u64 (on some it is unsigned long).
Cc: Neil Brown <neilb@suse.de>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
raid5's 'make_request' function calls generic_make_request on underlying
devices and if we run out of stripe heads, it could end up waiting for one of
those requests to complete. This is bad as recursive calls to
generic_make_request go on a queue and are not even attempted until
make_request completes.
So: don't make any generic_make_request calls in raid5 make_request until all
waiting has been done. We do this by simply setting STRIPE_HANDLE instead of
calling handle_stripe().
If we need more stripe_heads, raid5d will get called to process the pending
stripe_heads which will call generic_make_request from a
This change by itself causes a performance hit. So add a change so that
raid5_activate_delayed is only called at unplug time, never in raid5. This
seems to bring back the performance numbers. Calling it in raid5d was
sometimes too soon...
Neil said:
How about we queue it for 2.6.25-rc1 and then about when -rc2 comes out,
we queue it for 2.6.24.y?
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Tested-by: dean gaudet <dean@arctic.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As this is more in line with common practice in the kernel. Also swap the
args around to be more like list_for_each.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This allows userspace to control resync/reshape progress and synchronise it
with other activities, such as shared access in a SAN, or backing up critical
sections during a tricky reshape.
Writing a number of sectors (which must be a multiple of the chunk size if
such is meaningful) causes a resync to pause when it gets to that point.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently an md array with a write-intent bitmap does not updated that bitmap
to reflect successful partial resync. Rather the entire bitmap is updated
when the resync completes.
This is because there is no guarentee that resync requests will complete in
order, and tracking each request individually is unnecessarily burdensome.
However there is value in regularly updating the bitmap, so add code to
periodically pause while all pending sync requests complete, then update the
bitmap. Doing this only every few seconds (the same as the bitmap update
time) does not notciably affect resync performance.
[snitzer@gmail.com: export bitmap_cond_end_sync]
Signed-off-by: Neil Brown <neilb@suse.de>
Cc: "Mike Snitzer" <snitzer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We currently do not wait for the block from the missing device to be
computed from parity before copying data to the new stripe layout.
The change in the raid6 code is not techincally needed as we don't delay
data block recovery in the same way for raid6 yet. But making the change
now is safer long-term.
This bug exists in 2.6.23 and 2.6.24-rc
Cc: <stable@kernel.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
<debug output from Joel's system>
handling stripe 7629696, state=0x14 cnt=1, pd_idx=2 ops=0:0:0
check 5: state 0x6 toread 0000000000000000 read 0000000000000000 write fffff800ffcffcc0 written 0000000000000000
check 4: state 0x6 toread 0000000000000000 read 0000000000000000 write fffff800fdd4e360 written 0000000000000000
check 3: state 0x1 toread 0000000000000000 read 0000000000000000 write 0000000000000000 written 0000000000000000
check 2: state 0x1 toread 0000000000000000 read 0000000000000000 write 0000000000000000 written 0000000000000000
check 1: state 0x6 toread 0000000000000000 read 0000000000000000 write fffff800ff517e40 written 0000000000000000
check 0: state 0x6 toread 0000000000000000 read 0000000000000000 write fffff800fd4cae60 written 0000000000000000
locked=4 uptodate=2 to_read=0 to_write=4 failed=0 failed_num=0
for sector 7629696, rmw=0 rcw=0
</debug>
These blocks were prepared to be written out, but were never handled in
ops_run_biodrain(), so they remain locked forever. The operations flags
are all clear which means handle_stripe() thinks nothing else needs to be
done.
This state suggests that the STRIPE_OP_PREXOR bit was sampled 'set' when it
should not have been. This patch cleans up cases where the code looks at
sh->ops.pending when it should be looking at the consistent stack-based
snapshot of the operations flags.
Report from Joel:
Resync done. Patch fix this bug.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Tested-by: Joel Bertrand <joel.bertrand@systella.fr>
Cc: <stable@kernel.org>
Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Added blk_unplug interface, allowing all invocations of unplugs to result
in a generated blktrace UNPLUG.
Signed-off-by: Alan D. Brunelle <Alan.Brunelle@hp.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
commit 4ae3f847e4 ("md: raid5: fix
clearing of biofill operations") did not get applied correctly,
presumably due to substantial similarities between handle_stripe5 and
handle_stripe6.
This patch moves the chunk of new code from handle_stripe6 (where it isn't
needed (yet)) to handle_stripe5.
Signed-off-by: Neil Brown <neilb@suse.de>
Cc: "Dan Williams" <dan.j.williams@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ops_complete_biofill() runs outside of spin_lock(&sh->lock) and clears the
'pending' and 'ack' bits. Since the test_and_ack_op() macro only checks
against 'complete' it can get an inconsistent snapshot of pending work.
Move the clearing of these bits to handle_stripe5(), under the lock.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Tested-by: Joel Bertrand <joel.bertrand@systella.fr>
Signed-off-by: Neil Brown <neilb@suse.de>
Cc: Stable <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As bi_end_io is only called once when the reqeust is complete,
the 'size' argument is now redundant. Remove it.
Now there is no need for bio_endio to subtract the size completed
from bi_size. So don't do that either.
While we are at it, change bi_end_io to return void.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
1/ ops_complete_biofill tried to avoid calling handle_stripe since all the
state necessary to return read completions is available. However the
process of determining whether more read requests are pending requires
locking the stripe (to block add_stripe_bio from updating dev->toead).
ops_complete_biofill can run in tasklet context, so rather than upgrading
all the stripe locks from spin_lock to spin_lock_bh this patch just
unconditionally reschedules handle_stripe after completing the read
request.
2/ ops_complete_biofill needlessly qualified processing R5_Wantfill with
dev->toread. The result being that the 'biofill' pending bit is cleared
before handling the pending read-completions on dev->read. R5_Wantfill can
be unconditionally handled because the 'biofill' pending bit prevents new
R5_Wantfill requests from being seen by ops_run_biofill and
ops_complete_biofill.
Found-by: Yuri Tikhonov <yur@emcraft.com>
[neilb@suse.de: simpler fix for bug 1 than moving code]
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The recent changed to raid5 to allow offload of parity calculation etc
introduced some bugs in the code for growing (i.e. adding a disk to) raid5
and raid6. This fixes them
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Some of the code has been gradually transitioned to using the proper
struct request_queue, but there's lots left. So do a full sweet of
the kernel and get rid of this typedef and replace its uses with
the proper type.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Andrew Morton:
[async_memcpy] is very wrong if both ASYNC_TX_KMAP_DST and
ASYNC_TX_KMAP_SRC can ever be set. We'll end up using the same kmap
slot for both src add dest and we get either corrupted data or a BUG.
Evgeniy Polyakov:
Btw, shouldn't it always be kmap_atomic() even if flag is not set.
That pages are usual one returned by alloc_page().
So fix the usage of kmap_atomic and kill the ASYNC_TX_KMAP_DST and
ASYNC_TX_KMAP_SRC flags.
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Evgeniy Polyakov <johnpol@2ka.mipt.ru>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Slab destructors were no longer supported after Christoph's
c59def9f22 change. They've been
BUGs for both slab and slub, and slob never supported them
either.
This rips out support for the dtor pointer from kmem_cache_create()
completely and fixes up every single callsite in the kernel (there were
about 224, not including the slab allocator definitions themselves,
or the documentation references).
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
I/O submission requests were already handled outside of the stripe lock in
handle_stripe. Now that handle_stripe is only tasked with finding work,
this logic belongs in raid5_run_ops.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
When a stripe is being expanded bulk copying takes place to move the data
from the old stripe to the new. Since raid5_run_ops only operates on one
stripe at a time these bulk copies are handled in-line under the stripe
lock. In the dma offload case we poll for the completion of the operation.
After the data has been copied into the new stripe the parity needs to be
recalculated across the new disks. We reuse the existing postxor
functionality to carry out this calculation. By setting STRIPE_OP_POSTXOR
without setting STRIPE_OP_BIODRAIN the completion path in handle stripe
can differentiate expand operations from normal write operations.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
When a read bio is attached to the stripe and the corresponding block is
marked R5_UPTODATE, then a read (biofill) operation is scheduled to copy
the data from the stripe cache to the bio buffer. handle_stripe flags the
blocks to be operated on with the R5_Wantfill flag. If new read requests
arrive while raid5_run_ops is running they will not be handled until
handle_stripe is scheduled to run again.
Changelog:
* cleanup to_read and to_fill accounting
* do not fail reads that have reached the cache
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
Check operations are scheduled when the array is being resynced or an
explicit 'check/repair' command was sent to the array. Previously check
operations would destroy the parity block in the cache such that even if
parity turned out to be correct the parity block would be marked
!R5_UPTODATE at the completion of the check. When the operation can be
carried out by a dma engine the assumption is that it can check parity as a
read-only operation. If raid5_run_ops notices that the check was handled
by hardware it will preserve the R5_UPTODATE status of the parity disk.
When a check operation determines that the parity needs to be repaired we
reuse the existing compute block infrastructure to carry out the operation.
Repair operations imply an immediate write back of the data, so to
differentiate a repair from a normal compute operation the
STRIPE_OP_MOD_REPAIR_PD flag is added.
Changelog:
* remove test_and_set/test_and_clear BUG_ONs, Neil Brown
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
handle_stripe will compute a block when a backing disk has failed, or when
it determines it can save a disk read by computing the block from all the
other up-to-date blocks.
Previously a block would be computed under the lock and subsequent logic in
handle_stripe could use the newly up-to-date block. With the raid5_run_ops
implementation the compute operation is carried out a later time outside
the lock. To preserve the old functionality we take advantage of the
dependency chain feature of async_tx to flag the block as R5_Wantcompute
and then let other parts of handle_stripe operate on the block as if it
were up-to-date. raid5_run_ops guarantees that the block will be ready
before it is used in another operation.
However, this only works in cases where the compute and the dependent
operation are scheduled at the same time. If a previous call to
handle_stripe sets the R5_Wantcompute flag there is no facility to pass the
async_tx dependency chain across successive calls to raid5_run_ops. The
req_compute variable protects against this case.
Changelog:
* remove the req_compute BUG_ON
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
After handle_stripe5 decides whether it wants to perform a
read-modify-write, or a reconstruct write it calls
handle_write_operations5. A read-modify-write operation will perform an
xor subtraction of the blocks marked with the R5_Wantprexor flag, copy the
new data into the stripe (biodrain) and perform a postxor operation across
all up-to-date blocks to generate the new parity. A reconstruct write is run
when all blocks are already up-to-date in the cache so all that is needed
is a biodrain and postxor.
On the completion path STRIPE_OP_PREXOR will be set if the operation was a
read-modify-write. The STRIPE_OP_BIODRAIN flag is used in the completion
path to differentiate write-initiated postxor operations versus
expansion-initiated postxor operations. Completion of a write triggers i/o
to the drives.
Changelog:
* make the 'rcw' parameter to handle_write_operations5 a simple flag, Neil Brown
* remove test_and_set/test_and_clear BUG_ONs, Neil Brown
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
All the handle_stripe operations that are to be transitioned to use
raid5_run_ops need a method to coherently gather work under the stripe-lock
and hand that work off to raid5_run_ops. The 'get_stripe_work' routine
runs under the lock to read all the bits in sh->ops.pending that do not
have the corresponding bit set in sh->ops.ack. This modified 'pending'
bitmap is then passed to raid5_run_ops for processing.
The transition from 'ack' to 'completion' does not need similar protection
as the existing release_stripe infrastructure will guarantee that
handle_stripe will run again after a completion bit is set, and
handle_stripe can tolerate a sh->ops.completed bit being set while the lock
is held.
A call to async_tx_issue_pending_all() is added to raid5d to kick the
offload engines once all pending stripe operations work has been submitted.
This enables batching of the submission and completion of operations.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
When the raid acceleration work was proposed, Neil laid out the following
attack plan:
1/ move the xor and copy operations outside spin_lock(&sh->lock)
2/ find/implement an asynchronous offload api
The raid5_run_ops routine uses the asynchronous offload api (async_tx) and
the stripe_operations member of a stripe_head to carry out xor+copy
operations asynchronously, outside the lock.
To perform operations outside the lock a new set of state flags is needed
to track new requests, in-flight requests, and completed requests. In this
new model handle_stripe is tasked with scanning the stripe_head for work,
updating the stripe_operations structure, and finally dropping the lock and
calling raid5_run_ops for processing. The following flags outline the
requests that handle_stripe can make of raid5_run_ops:
STRIPE_OP_BIOFILL
- copy data into request buffers to satisfy a read request
STRIPE_OP_COMPUTE_BLK
- generate a missing block in the cache from the other blocks
STRIPE_OP_PREXOR
- subtract existing data as part of the read-modify-write process
STRIPE_OP_BIODRAIN
- copy data out of request buffers to satisfy a write request
STRIPE_OP_POSTXOR
- recalculate parity for new data that has entered the cache
STRIPE_OP_CHECK
- verify that the parity is correct
STRIPE_OP_IO
- submit i/o to the member disks (note this was already performed outside
the stripe lock, but it made sense to add it as an operation type
The flow is:
1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending
2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the
operation to the async_tx api
3/ async_tx triggers the completion callback routine to set
sh->ops.complete and release the stripe
4/ handle_stripe runs again to finish the operation and optionally submit
new operations that were previously blocked
Note this patch just defines raid5_run_ops, subsequent commits (one per
major operation type) modify handle_stripe to take advantage of this
routine.
Changelog:
* removed ops_complete_biodrain in favor of ops_complete_postxor and
ops_complete_write.
* removed the raid5_run_ops workqueue
* call bi_end_io for reads in ops_complete_biofill, saves a call to
handle_stripe
* explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown
* fix race between async engines and bi_end_io call for reads, Neil Brown
* remove unnecessary spin_lock from ops_complete_biofill
* remove test_and_set/test_and_clear BUG_ONs, Neil Brown
* remove explicit interrupt handling for channel switching, this feature
was absorbed (i.e. it is now implicit) by the async_tx api
* use return_io in ops_complete_biofill
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
Replaces PRINTK with pr_debug, and kills the RAID5_DEBUG definition in
favor of the global DEBUG definition. To get local debug messages just add
'#define DEBUG' to the top of the file.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
handle_stripe5 and handle_stripe6 have very deep logic paths handling the
various states of a stripe_head. By introducing the 'stripe_head_state'
and 'r6_state' objects, large portions of the logic can be moved to
sub-routines.
'struct stripe_head_state' consumes all of the automatic variables that previously
stood alone in handle_stripe5,6. 'struct r6_state' contains the handle_stripe6
specific variables like p_failed and q_failed.
One of the nice side effects of the 'stripe_head_state' change is that it
allows for further reductions in code duplication between raid5 and raid6.
The following new routines are shared between raid5 and raid6:
handle_completed_write_requests
handle_requests_to_failed_array
handle_stripe_expansion
Changes:
* v2: fixed 'conf->raid_disk-1' for the raid6 'handle_stripe_expansion' path
* v3: removed the unused 'dirty' field from struct stripe_head_state
* v3: coalesced open coded bi_end_io routines into return_io()
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
The async_tx api provides methods for describing a chain of asynchronous
bulk memory transfers/transforms with support for inter-transactional
dependencies. It is implemented as a dmaengine client that smooths over
the details of different hardware offload engine implementations. Code
that is written to the api can optimize for asynchronous operation and the
api will fit the chain of operations to the available offload resources.
I imagine that any piece of ADMA hardware would register with the
'async_*' subsystem, and a call to async_X would be routed as
appropriate, or be run in-line. - Neil Brown
async_tx exploits the capabilities of struct dma_async_tx_descriptor to
provide an api of the following general format:
struct dma_async_tx_descriptor *
async_<operation>(..., struct dma_async_tx_descriptor *depend_tx,
dma_async_tx_callback cb_fn, void *cb_param)
{
struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>);
struct dma_device *device = chan ? chan->device : NULL;
int int_en = cb_fn ? 1 : 0;
struct dma_async_tx_descriptor *tx = device ?
device->device_prep_dma_<operation>(chan, len, int_en) : NULL;
if (tx) { /* run <operation> asynchronously */
...
tx->tx_set_dest(addr, tx, index);
...
tx->tx_set_src(addr, tx, index);
...
async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
} else { /* run <operation> synchronously */
...
<operation>
...
async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param);
}
return tx;
}
async_tx_find_channel() returns a capable channel from its pool. The
channel pool is organized as a per-cpu array of channel pointers. The
async_tx_rebalance() routine is tasked with managing these arrays. In the
uniprocessor case async_tx_rebalance() tries to spread responsibility
evenly over channels of similar capabilities. For example if there are two
copy+xor channels, one will handle copy operations and the other will
handle xor. In the SMP case async_tx_rebalance() attempts to spread the
operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor
channel0 while cpu1 gets copy channel 1 and xor channel 1. When a
dependency is specified async_tx_find_channel defaults to keeping the
operation on the same channel. A xor->copy->xor chain will stay on one
channel if it supports both operation types, otherwise the transaction will
transition between a copy and a xor resource.
Currently the raid5 implementation in the MD raid456 driver has been
converted to the async_tx api. A driver for the offload engines on the
Intel Xscale series of I/O processors, iop-adma, is provided in a later
commit. With the iop-adma driver and async_tx, raid456 is able to offload
copy, xor, and xor-zero-sum operations to hardware engines.
On iop342 tiobench showed higher throughput for sequential writes (20 - 30%
improvement) and sequential reads to a degraded array (40 - 55%
improvement). For the other cases performance was roughly equal, +/- a few
percentage points. On a x86-smp platform the performance of the async_tx
implementation (in synchronous mode) was also +/- a few percentage points
of the original implementation. According to 'top' on iop342 CPU
utilization drops from ~50% to ~15% during a 'resync' while the speed
according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s.
The tiobench command line used for testing was: tiobench --size 2048
--block 4096 --block 131072 --dir /mnt/raid --numruns 5
* iop342 had 1GB of memory available
Details:
* if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making
async_tx_find_channel a static inline routine that always returns NULL
* when a callback is specified for a given transaction an interrupt will
fire at operation completion time and the callback will occur in a
tasklet. if the the channel does not support interrupts then a live
polling wait will be performed
* the api is written as a dmaengine client that requests all available
channels
* In support of dependencies the api implicitly schedules channel-switch
interrupts. The interrupt triggers the cleanup tasklet which causes
pending operations to be scheduled on the next channel
* Xor engines treat an xor destination address differently than a software
xor routine. To the software routine the destination address is an implied
source, whereas engines treat it as a write-only destination. This patch
modifies the xor_blocks routine to take a an explicit destination address
to mirror the hardware.
Changelog:
* fixed a leftover debug print
* don't allow callbacks in async_interrupt_cond
* fixed xor_block changes
* fixed usage of ASYNC_TX_XOR_DROP_DEST
* drop dma mapping methods, suggested by Chris Leech
* printk warning fixups from Andrew Morton
* don't use inline in C files, Adrian Bunk
* select the API when MD is enabled
* BUG_ON xor source counts <= 1
* implicitly handle hardware concerns like channel switching and
interrupts, Neil Brown
* remove the per operation type list, and distribute operation capabilities
evenly amongst the available channels
* simplify async_tx_find_channel to optimize the fast path
* introduce the channel_table_initialized flag to prevent early calls to
the api
* reorganize the code to mimic crypto
* include mm.h as not all archs include it in dma-mapping.h
* make the Kconfig options non-user visible, Adrian Bunk
* move async_tx under crypto since it is meant as 'core' functionality, and
the two may share algorithms in the future
* move large inline functions into c files
* checkpatch.pl fixes
* gpl v2 only correction
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
The async_tx api tries to use a dma engine for an operation, but will fall
back to an optimized software routine otherwise. Xor support is
implemented using the raid5 xor routines. For organizational purposes this
routine is moved to a common area.
The following fixes are also made:
* rename xor_block => xor_blocks, suggested by Adrian Bunk
* ensure that xor.o initializes before md.o in the built-in case
* checkpatch.pl fixes
* mark calibrate_xor_blocks __init, Adrian Bunk
Cc: Adrian Bunk <bunk@stusta.de>
Cc: NeilBrown <neilb@suse.de>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
This reverts commit 5b479c91da.
Quoth Neil Brown:
"It causes an oops when auto-detecting raid arrays, and it doesn't
seem easy to fix.
The array may not be 'open' when do_md_run is called, so
bdev->bd_disk might be NULL, so bd_set_size can oops.
This whole approach of opening an md device before it has been
assembled just seems to get more and more painful. I think I'm going
to have to come up with something clever to provide both backward
comparability with usage expectation, and sane integration into the
rest of the kernel."
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
md currently uses ->media_changed to make sure rescan_partitions
is call on md array after they are assembled.
However that doesn't happen until the array is opened, which is later
than some people would like.
So use blkdev_ioctl to do the rescan immediately that the
array has been assembled.
This means we can remove all the ->change infrastructure as it was only used
to trigger a partition rescan.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
SLUB doesn't like slashes as it wants to use the cache name as the name of a
directory (or symlink) in sysfs.
Signed-off-by: Neil Brown <neilb@suse.de>
Acked-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
... still not sure why we need this ....
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If this mddev and queue got reused for another array that doesn't register a
congested_fn, this function would get called incorretly.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
