raid10 splits requests in two different ways for two different
reasons.
First, bio_split() is used to ensure the bio fits with a chunk.
Second, multiple r10bio structures are allocated to represent the
different sections that need to go to different devices, to avoid
known bad blocks.
This can be simplified to just use bio_split() once, and not to use
multiple r10bios.
We delay the split until we know a maximum bio size that can
be handled with a single r10bio, and then split the bio and queue
the remainder for later handling.
As with raid1, we allocate a new bio_set to help with the splitting.
It is not correct to use fs_bio_set in a device driver.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
If a device is marked FailFast, and it is not the only
device we can read from, we mark the bio as MD_FAILFAST.
If this does fail-fast, we don't try read repair but just
allow failure.
If it was the last device, it doesn't get marked Faulty so
the retry happens on the same device - this time without
FAILFAST. A subsequent failure will not retry but will just
pass up the error.
During resync we may use FAILFAST requests, and on a failure
we will simply use the other device(s).
During recovery we will only use FAILFAST in the unusual
case were there are multiple places to read from - i.e. if
there are > 2 devices. If we get a failure we will fail the
device and complete the resync/recovery with remaining
devices.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
RAID10 random read performance is lower than expected due to excessive spinlock
utilisation which is required mostly for rebuild/resync. Simplify allow_barrier
as it's in IO path and encounters a lot of unnecessary congestion.
As lower_barrier just takes a lock in order to decrement a counter, convert
counter (nr_pending) into atomic variable and remove the spin lock. There is
also a congestion for wake_up (it uses lock internally) so call it only when
it's really needed. As wake_up is not called constantly anymore, ensure process
waiting to raise a barrier is notified when there are no more waiting IOs.
Signed-off-by: Tomasz Majchrzak <tomasz.majchrzak@intel.com>
Signed-off-by: Shaohua Li <shli@fb.com>
When a write to one of the legs of a RAID10 fails, the failure is
recorded in the metadata of the other legs so that after a restart
the data on the failed drive wont be trusted even if that drive seems
to be working again (maybe a cable was unplugged).
Currently there is no interlock between the write request completing
and the metadata update. So it is possible that the write will
complete, the app will confirm success in some way, and then the
machine will crash before the metadata update completes.
This is an extremely small hole for a racy to fit in, but it is
theoretically possible and so should be closed.
So:
- set MD_CHANGE_PENDING when requesting a metadata update for a
failed device, so we can know with certainty when it completes
- queue requests that experienced an error on a new queue which
is only processed after the metadata update completes
- call raid_end_bio_io() on bios in that queue when the time comes.
Signed-off-by: NeilBrown <neilb@suse.com>
There is currently no locking around calls to the 'congested'
bdi function. If called at an awkward time while an array is
being converted from one level (or personality) to another, there
is a tiny chance of running code in an unreferenced module etc.
So add a 'congested' function to the md_personality operations
structure, and call it with appropriate locking from a central
'mddev_congested'.
When the array personality is changing the array will be 'suspended'
so no IO is processed.
If mddev_congested detects this, it simply reports that the
array is congested, which is a safe guess.
As mddev_suspend calls synchronize_rcu(), mddev_congested can
avoid races by included the whole call inside an rcu_read_lock()
region.
This require that the congested functions for all subordinate devices
can be run under rcu_lock. Fortunately this is the case.
Signed-off-by: NeilBrown <neilb@suse.de>
The MD RAID10 'far' and 'offset' algorithms make copies of entire stripe
widths - copying them to a different location on the same devices after
shifting the stripe. An example layout of each follows below:
"far" algorithm
dev1 dev2 dev3 dev4 dev5 dev6
==== ==== ==== ==== ==== ====
A B C D E F
G H I J K L
...
F A B C D E --> Copy of stripe0, but shifted by 1
L G H I J K
...
"offset" algorithm
dev1 dev2 dev3 dev4 dev5 dev6
==== ==== ==== ==== ==== ====
A B C D E F
F A B C D E --> Copy of stripe0, but shifted by 1
G H I J K L
L G H I J K
...
Redundancy for these algorithms is gained by shifting the copied stripes
one device to the right. This patch proposes that array be divided into
sets of adjacent devices and when the stripe copies are shifted, they wrap
on set boundaries rather than the array size boundary. That is, for the
purposes of shifting, the copies are confined to their sets within the
array. The sets are 'near_copies * far_copies' in size.
The above "far" algorithm example would change to:
"far" algorithm
dev1 dev2 dev3 dev4 dev5 dev6
==== ==== ==== ==== ==== ====
A B C D E F
G H I J K L
...
B A D C F E --> Copy of stripe0, shifted 1, 2-dev sets
H G J I L K Dev sets are 1-2, 3-4, 5-6
...
This has the affect of improving the redundancy of the array. We can
always sustain at least one failure, but sometimes more than one can
be handled. In the first examples, the pairs of devices that CANNOT fail
together are:
(1,2) (2,3) (3,4) (4,5) (5,6) (1, 6) [40% of possible pairs]
In the example where the copies are confined to sets, the pairs of
devices that cannot fail together are:
(1,2) (3,4) (5,6) [20% of possible pairs]
We cannot simply replace the old algorithms, so the 17th bit of the 'layout'
variable is used to indicate whether we use the old or new method of computing
the shift. (This is similar to the way the 16th bit indicates whether the
"far" algorithm or the "offset" algorithm is being used.)
This patch only handles the cases where the number of total raid disks is
a multiple of 'far_copies'. A follow-on patch addresses the condition where
this is not true.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
A 'struct r10bio' has an array of per-copy information at the end.
This array is declared with size [0] and r10bio_pool_alloc allocates
enough extra space to store the per-copy information depending on the
number of copies needed.
So declaring a 'struct r10bio on the stack isn't going to work. It
won't allocate enough space, and memory corruption will ensue.
So in the two places where this is done, declare a sufficiently large
structure and use that instead.
The two call-sites of this bug were introduced in 3.4 and 3.5
so this is suitable for both those kernels. The patch will have to
be modified for 3.4 as it only has one bug.
Cc: stable@vger.kernel.org
Reported-by: Ivan Vasilyev <ivan.vasilyev@gmail.com>
Tested-by: Ivan Vasilyev <ivan.vasilyev@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
md/raid10: Export is_congested test.
In similar fashion to commits
11d8a6e3711ed7242e59
we export the RAID10 congestion checking function so that dm-raid.c can
make use of it and make use of the personality. The 'queue' and 'gendisk'
structures will not be available to the MD code when device-mapper sets
up the device, so we conditionalize access to these fields also.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
MD RAID1/RAID10: Move some macros from .h file to .c file
There are three macros (IO_BLOCKED,IO_MADE_GOOD,BIO_SPECIAL) which are defined
in both raid1.h and raid10.h. They are only used in there respective .c files.
However, if we wish to make RAID10 accessible to the device-mapper RAID
target (dm-raid.c), then we need to move these macros into the .c files where
they are used so that they do not conflict with each other.
The macros from the two files are identical and could be moved into md.h, but
I chose to leave the duplication and have them remain in the personality
files.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
MD RAID10: Rename the structure 'mirror_info' to 'raid10_info'
The same structure name ('mirror_info') is used by raid1. Each of these
structures are defined in there respective header files. If dm-raid is
to support both RAID1 and RAID10, the header files will be included and
the structure names must not collide.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
A 'near' or 'offset' lay RAID10 array can be reshaped to a different
'near' or 'offset' layout, a different chunk size, and a different
number of devices.
However the number of copies cannot change.
Unlike RAID5/6, we do not support having user-space backup data that
is being relocated during a 'critical section'. Rather, the
data_offset of each device must change so that when writing any block
to a new location, it will not over-write any data that is still
'live'.
This means that RAID10 reshape is not supportable on v0.90 metadata.
The different between the old data_offset and the new_offset must be
at least the larger of the chunksize multiplied by offset copies of
each of the old and new layout. (for 'near' mode, offset_copies == 1).
A larger difference of around 64M seems useful for in-place reshapes
as more data can be moved between metadata updates.
Very large differences (e.g. 512M) seem to slow the process down due
to lots of long seeks (on oldish consumer graded devices at least).
Metadata needs to be updated whenever the place we are about to write
to is considered - by the current metadata - to still contain data in
the old layout.
[unbalanced locking fix from Dan Carpenter <dan.carpenter@oracle.com>]
Signed-off-by: NeilBrown <neilb@suse.de>
When RAID10 supports reshape it will need a 'previous' and a 'current'
geometry, so introduce that here.
Use the 'prev' geometry when before the reshape_position, and the
current 'geo' when beyond it. At other times, use both as
appropriate.
For now, both are identical (And reshape_position is never set).
When we use the 'prev' geometry, we must use the old data_offset.
When we use the current (And a reshape is happening) we must use
the new_data_offset.
Signed-off-by: NeilBrown <neilb@suse.de>
We will shortly be adding reshape support for RAID10 which will
require it having 2 concurrent geometries (before and after).
To make that easier, collect most geometry fields into 'struct geom'
and access them from there. Then we will more easily be able to add
a second set of fields.
Note that 'copies' is not in this struct and so cannot be changed.
There is little need to change this number and doing so is a lot
more difficult as it requires reallocating more things.
So leave it out for now.
Signed-off-by: NeilBrown <neilb@suse.de>
Allow each slot in the RAID10 to have 2 devices, the want_replacement
and the replacement.
Also an r10bio to have 2 bios, and for resync/recovery allocate the
second bio if there are any replacement devices.
Signed-off-by: NeilBrown <neilb@suse.de>
RAID1 and RAID10 handle write requests by queuing them for handling by
a separate thread. This is because when a write-intent-bitmap is
active we might need to update the bitmap first, so it is good to
queue a lot of writes, then do one big bitmap update for them all.
However writeback request devices to appear to be congested after a
while so it can make some guesstimate of throughput. The infinite
queue defeats that (note that RAID5 has already has a finite queue so
it doesn't suffer from this problem).
So impose a limit on the number of pending write requests. By default
it is 1024 which seems to be generally suitable. Make it configurable
via module option just in case someone finds a regression.
Signed-off-by: NeilBrown <neilb@suse.de>
The typedefs are just annoying. 'mdk' probably refers to 'md_k.h'
which used to be an include file that defined this thing.
Signed-off-by: NeilBrown <neilb@suse.de>
When we get a write error (in the data area, not in metadata),
update the badblock log rather than failing the whole device.
As the write may well be many blocks, we trying writing each
block individually and only log the ones which fail.
Signed-off-by: NeilBrown <neilb@suse.de>
If we succeed in writing to a block that was recorded as
being bad, we clear the bad-block record.
This requires some delayed handling as the bad-block-list update has
to happen in process-context.
Signed-off-by: NeilBrown <neilb@suse.de>
This patch just covers the basic read path:
1/ read_balance needs to check for badblocks, and return not only
the chosen slot, but also how many good blocks are available
there.
2/ read submission must be ready to issue multiple reads to
different devices as different bad blocks on different devices
could mean that a single large read cannot be served by any one
device, but can still be served by the array.
This requires keeping count of the number of outstanding requests
per bio. This count is stored in 'bi_phys_segments'
On read error we currently just fail the request if another target
cannot handle the whole request. Next patch refines that a bit.
Signed-off-by: NeilBrown <neilb@suse.de>
When we get a read error during recovery, RAID10 previously
arranged for the recovering device to appear to fail so that
the recovery stops and doesn't restart. This is misleading and wrong.
Instead, make use of the new recovery_disabled handling and mark
the target device and having recovery disabled.
Add appropriate checks in add_disk and remove_disk so that devices
are removed and not re-added when recovery is disabled.
Signed-off-by: NeilBrown <neilb@suse.de>
Most array level changes leave the list of devices largely unchanged,
possibly causing one at the end to become redundant.
However conversions between RAID0 and RAID10 need to renumber
all devices (except 0).
This renumbering is currently being done in the ->run method when the
new personality takes over. However this is too late as the common
code in md.c might already have invalidated some of the devices if
they had a ->raid_disk number that appeared to high.
Moving it into the ->takeover method is too early as the array is
still active at that time and wrong ->raid_disk numbers could cause
confusion.
So add a ->new_raid_disk field to mdk_rdev_s and use it to communicate
the new raid_disk number.
Now the common code knows exactly which devices need to be renumbered,
and which can be invalidated, and can do it all at a convenient time
when the array is suspend.
It can also update some symlinks in sysfs which previously were not be
updated correctly.
Reported-by: Maciej Trela <maciej.trela@intel.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Having a macro just to cast a void* isn't really helpful.
I would must rather see that we are simply de-referencing ->private,
than have to know what the macro does.
So open code the macro everywhere and remove the pointless cast.
Signed-off-by: NeilBrown <neilb@suse.de>
This makes the includes more explicit, and is preparation for moving
md_k.h to drivers/md/md.h
Remove include/raid/md.h as its only remaining use was to #include
other files.
Signed-off-by: NeilBrown <neilb@suse.de>
Move the headers with the local structures for the disciplines and
bitmap.h into drivers/md/ so that they are more easily grepable for
hacking and not far away. md.h is left where it is for now as there
are some uses from the outside.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.de>
NeilBrown